drivers/md/raid5.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * raid5.c : Multiple Devices driver for Linux
   4  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   5  *         Copyright (C) 1999, 2000 Ingo Molnar
   6  *         Copyright (C) 2002, 2003 H. Peter Anvin
   7  *
   8  * RAID-4/5/6 management functions.
   9  * Thanks to Penguin Computing for making the RAID-6 development possible
  10  * by donating a test server!
  11  */
  12
  13 /*
  14  * BITMAP UNPLUGGING:
  15  *
  16  * The sequencing for updating the bitmap reliably is a little
  17  * subtle (and I got it wrong the first time) so it deserves some
  18  * explanation.
  19  *
  20  * We group bitmap updates into batches.  Each batch has a number.
  21  * We may write out several batches at once, but that isn't very important.
  22  * conf->seq_write is the number of the last batch successfully written.
  23  * conf->seq_flush is the number of the last batch that was closed to
  24  *    new additions.
  25  * When we discover that we will need to write to any block in a stripe
  26  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  27  * the number of the batch it will be in. This is seq_flush+1.
  28  * When we are ready to do a write, if that batch hasn't been written yet,
  29  *   we plug the array and queue the stripe for later.
  30  * When an unplug happens, we increment bm_flush, thus closing the current
  31  *   batch.
  32  * When we notice that bm_flush > bm_write, we write out all pending updates
  33  * to the bitmap, and advance bm_write to where bm_flush was.
  34  * This may occasionally write a bit out twice, but is sure never to
  35  * miss any bits.
  36  */
  37
  38 #include <linux/blkdev.h>
  39 #include <linux/kthread.h>
  40 #include <linux/raid/pq.h>
  41 #include <linux/async_tx.h>
  42 #include <linux/module.h>
  43 #include <linux/async.h>
  44 #include <linux/seq_file.h>
  45 #include <linux/cpu.h>
  46 #include <linux/slab.h>
  47 #include <linux/ratelimit.h>
  48 #include <linux/nodemask.h>
  49
  50 #include <trace/events/block.h>
  51 #include <linux/list_sort.h>
  52
  53 #include "md.h"
  54 #include "raid5.h"
  55 #include "raid0.h"
  56 #include "md-bitmap.h"
  57 #include "raid5-log.h"
  58
  59 #define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
  60
  61 #define cpu_to_group(cpu) cpu_to_node(cpu)
  62 #define ANY_GROUP NUMA_NO_NODE
  63
  64 static bool devices_handle_discard_safely = false;
  65 module_param(devices_handle_discard_safely, bool, 0644);
  66 MODULE_PARM_DESC(devices_handle_discard_safely,
  67                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
  68 static struct workqueue_struct *raid5_wq;
  69
  70 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  71 {
  72         int hash = (sect >> RAID5_STRIPE_SHIFT(conf)) & HASH_MASK;
  73         return &conf->stripe_hashtbl[hash];
  74 }
  75
  76 static inline int stripe_hash_locks_hash(struct r5conf *conf, sector_t sect)
  77 {
  78         return (sect >> RAID5_STRIPE_SHIFT(conf)) & STRIPE_HASH_LOCKS_MASK;
  79 }
  80
  81 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
  82 {
  83         spin_lock_irq(conf->hash_locks + hash);
  84         spin_lock(&conf->device_lock);
  85 }
  86
  87 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
  88 {
  89         spin_unlock(&conf->device_lock);
  90         spin_unlock_irq(conf->hash_locks + hash);
  91 }
  92
  93 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
  94 {
  95         int i;
  96         spin_lock_irq(conf->hash_locks);
  97         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
  98                 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
  99         spin_lock(&conf->device_lock);
 100 }
 101
 102 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
 103 {
 104         int i;
 105         spin_unlock(&conf->device_lock);
 106         for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
 107                 spin_unlock(conf->hash_locks + i);
 108         spin_unlock_irq(conf->hash_locks);
 109 }
 110
 111 /* Find first data disk in a raid6 stripe */
 112 static inline int raid6_d0(struct stripe_head *sh)
 113 {
 114         if (sh->ddf_layout)
 115                 /* ddf always start from first device */
 116                 return 0;
 117         /* md starts just after Q block */
 118         if (sh->qd_idx == sh->disks - 1)
 119                 return 0;
 120         else
 121                 return sh->qd_idx + 1;
 122 }
 123 static inline int raid6_next_disk(int disk, int raid_disks)
 124 {
 125         disk++;
 126         return (disk < raid_disks) ? disk : 0;
 127 }
 128
 129 /* When walking through the disks in a raid5, starting at raid6_d0,
 130  * We need to map each disk to a 'slot', where the data disks are slot
 131  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 132  * is raid_disks-1.  This help does that mapping.
 133  */
 134 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 135                              int *count, int syndrome_disks)
 136 {
 137         int slot = *count;
 138
 139         if (sh->ddf_layout)
 140                 (*count)++;
 141         if (idx == sh->pd_idx)
 142                 return syndrome_disks;
 143         if (idx == sh->qd_idx)
 144                 return syndrome_disks + 1;
 145         if (!sh->ddf_layout)
 146                 (*count)++;
 147         return slot;
 148 }
 149
 150 static void print_raid5_conf (struct r5conf *conf);
 151
 152 static int stripe_operations_active(struct stripe_head *sh)
 153 {
 154         return sh->check_state || sh->reconstruct_state ||
 155                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 156                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 157 }
 158
 159 static bool stripe_is_lowprio(struct stripe_head *sh)
 160 {
 161         return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
 162                 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
 163                !test_bit(STRIPE_R5C_CACHING, &sh->state);
 164 }
 165
 166 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
 167 {
 168         struct r5conf *conf = sh->raid_conf;
 169         struct r5worker_group *group;
 170         int thread_cnt;
 171         int i, cpu = sh->cpu;
 172
 173         if (!cpu_online(cpu)) {
 174                 cpu = cpumask_any(cpu_online_mask);
 175                 sh->cpu = cpu;
 176         }
 177
 178         if (list_empty(&sh->lru)) {
 179                 struct r5worker_group *group;
 180                 group = conf->worker_groups + cpu_to_group(cpu);
 181                 if (stripe_is_lowprio(sh))
 182                         list_add_tail(&sh->lru, &group->loprio_list);
 183                 else
 184                         list_add_tail(&sh->lru, &group->handle_list);
 185                 group->stripes_cnt++;
 186                 sh->group = group;
 187         }
 188
 189         if (conf->worker_cnt_per_group == 0) {
 190                 md_wakeup_thread(conf->mddev->thread);
 191                 return;
 192         }
 193
 194         group = conf->worker_groups + cpu_to_group(sh->cpu);
 195
 196         group->workers[0].working = true;
 197         /* at least one worker should run to avoid race */
 198         queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
 199
 200         thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
 201         /* wakeup more workers */
 202         for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
 203                 if (group->workers[i].working == false) {
 204                         group->workers[i].working = true;
 205                         queue_work_on(sh->cpu, raid5_wq,
 206                                       &group->workers[i].work);
 207                         thread_cnt--;
 208                 }
 209         }
 210 }
 211
 212 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
 213                               struct list_head *temp_inactive_list)
 214 {
 215         int i;
 216         int injournal = 0;      /* number of date pages with R5_InJournal */
 217
 218         BUG_ON(!list_empty(&sh->lru));
 219         BUG_ON(atomic_read(&conf->active_stripes)==0);
 220
 221         if (r5c_is_writeback(conf->log))
 222                 for (i = sh->disks; i--; )
 223                         if (test_bit(R5_InJournal, &sh->dev[i].flags))
 224                                 injournal++;
 225         /*
 226          * In the following cases, the stripe cannot be released to cached
 227          * lists. Therefore, we make the stripe write out and set
 228          * STRIPE_HANDLE:
 229          *   1. when quiesce in r5c write back;
 230          *   2. when resync is requested fot the stripe.
 231          */
 232         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
 233             (conf->quiesce && r5c_is_writeback(conf->log) &&
 234              !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
 235                 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
 236                         r5c_make_stripe_write_out(sh);
 237                 set_bit(STRIPE_HANDLE, &sh->state);
 238         }
 239
 240         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 241                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 242                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 243                         list_add_tail(&sh->lru, &conf->delayed_list);
 244                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 245                            sh->bm_seq - conf->seq_write > 0)
 246                         list_add_tail(&sh->lru, &conf->bitmap_list);
 247                 else {
 248                         clear_bit(STRIPE_DELAYED, &sh->state);
 249                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 250                         if (conf->worker_cnt_per_group == 0) {
 251                                 if (stripe_is_lowprio(sh))
 252                                         list_add_tail(&sh->lru,
 253                                                         &conf->loprio_list);
 254                                 else
 255                                         list_add_tail(&sh->lru,
 256                                                         &conf->handle_list);
 257                         } else {
 258                                 raid5_wakeup_stripe_thread(sh);
 259                                 return;
 260                         }
 261                 }
 262                 md_wakeup_thread(conf->mddev->thread);
 263         } else {
 264                 BUG_ON(stripe_operations_active(sh));
 265                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 266                         if (atomic_dec_return(&conf->preread_active_stripes)
 267                             < IO_THRESHOLD)
 268                                 md_wakeup_thread(conf->mddev->thread);
 269                 atomic_dec(&conf->active_stripes);
 270                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 271                         if (!r5c_is_writeback(conf->log))
 272                                 list_add_tail(&sh->lru, temp_inactive_list);
 273                         else {
 274                                 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
 275                                 if (injournal == 0)
 276                                         list_add_tail(&sh->lru, temp_inactive_list);
 277                                 else if (injournal == conf->raid_disks - conf->max_degraded) {
 278                                         /* full stripe */
 279                                         if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
 280                                                 atomic_inc(&conf->r5c_cached_full_stripes);
 281                                         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
 282                                                 atomic_dec(&conf->r5c_cached_partial_stripes);
 283                                         list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
 284                                         r5c_check_cached_full_stripe(conf);
 285                                 } else
 286                                         /*
 287                                          * STRIPE_R5C_PARTIAL_STRIPE is set in
 288                                          * r5c_try_caching_write(). No need to
 289                                          * set it again.
 290                                          */
 291                                         list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
 292                         }
 293                 }
 294         }
 295 }
 296
 297 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
 298                              struct list_head *temp_inactive_list)
 299 {
 300         if (atomic_dec_and_test(&sh->count))
 301                 do_release_stripe(conf, sh, temp_inactive_list);
 302 }
 303
 304 /*
 305  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
 306  *
 307  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
 308  * given time. Adding stripes only takes device lock, while deleting stripes
 309  * only takes hash lock.
 310  */
 311 static void release_inactive_stripe_list(struct r5conf *conf,
 312                                          struct list_head *temp_inactive_list,
 313                                          int hash)
 314 {
 315         int size;
 316         bool do_wakeup = false;
 317         unsigned long flags;
 318
 319         if (hash == NR_STRIPE_HASH_LOCKS) {
 320                 size = NR_STRIPE_HASH_LOCKS;
 321                 hash = NR_STRIPE_HASH_LOCKS - 1;
 322         } else
 323                 size = 1;
 324         while (size) {
 325                 struct list_head *list = &temp_inactive_list[size - 1];
 326
 327                 /*
 328                  * We don't hold any lock here yet, raid5_get_active_stripe() might
 329                  * remove stripes from the list
 330                  */
 331                 if (!list_empty_careful(list)) {
 332                         spin_lock_irqsave(conf->hash_locks + hash, flags);
 333                         if (list_empty(conf->inactive_list + hash) &&
 334                             !list_empty(list))
 335                                 atomic_dec(&conf->empty_inactive_list_nr);
 336                         list_splice_tail_init(list, conf->inactive_list + hash);
 337                         do_wakeup = true;
 338                         spin_unlock_irqrestore(conf->hash_locks + hash, flags);
 339                 }
 340                 size--;
 341                 hash--;
 342         }
 343
 344         if (do_wakeup) {
 345                 wake_up(&conf->wait_for_stripe);
 346                 if (atomic_read(&conf->active_stripes) == 0)
 347                         wake_up(&conf->wait_for_quiescent);
 348                 if (conf->retry_read_aligned)
 349                         md_wakeup_thread(conf->mddev->thread);
 350         }
 351 }
 352
 353 /* should hold conf->device_lock already */
 354 static int release_stripe_list(struct r5conf *conf,
 355                                struct list_head *temp_inactive_list)
 356 {
 357         struct stripe_head *sh, *t;
 358         int count = 0;
 359         struct llist_node *head;
 360
 361         head = llist_del_all(&conf->released_stripes);
 362         head = llist_reverse_order(head);
 363         llist_for_each_entry_safe(sh, t, head, release_list) {
 364                 int hash;
 365
 366                 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
 367                 smp_mb();
 368                 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
 369                 /*
 370                  * Don't worry the bit is set here, because if the bit is set
 371                  * again, the count is always > 1. This is true for
 372                  * STRIPE_ON_UNPLUG_LIST bit too.
 373                  */
 374                 hash = sh->hash_lock_index;
 375                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
 376                 count++;
 377         }
 378
 379         return count;
 380 }
 381
 382 void raid5_release_stripe(struct stripe_head *sh)
 383 {
 384         struct r5conf *conf = sh->raid_conf;
 385         unsigned long flags;
 386         struct list_head list;
 387         int hash;
 388         bool wakeup;
 389
 390         /* Avoid release_list until the last reference.
 391          */
 392         if (atomic_add_unless(&sh->count, -1, 1))
 393                 return;
 394
 395         if (unlikely(!conf->mddev->thread) ||
 396                 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
 397                 goto slow_path;
 398         wakeup = llist_add(&sh->release_list, &conf->released_stripes);
 399         if (wakeup)
 400                 md_wakeup_thread(conf->mddev->thread);
 401         return;
 402 slow_path:
 403         /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
 404         if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
 405                 INIT_LIST_HEAD(&list);
 406                 hash = sh->hash_lock_index;
 407                 do_release_stripe(conf, sh, &list);
 408                 spin_unlock_irqrestore(&conf->device_lock, flags);
 409                 release_inactive_stripe_list(conf, &list, hash);
 410         }
 411 }
 412
 413 static inline void remove_hash(struct stripe_head *sh)
 414 {
 415         pr_debug("remove_hash(), stripe %llu\n",
 416                 (unsigned long long)sh->sector);
 417
 418         hlist_del_init(&sh->hash);
 419 }
 420
 421 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 422 {
 423         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 424
 425         pr_debug("insert_hash(), stripe %llu\n",
 426                 (unsigned long long)sh->sector);
 427
 428         hlist_add_head(&sh->hash, hp);
 429 }
 430
 431 /* find an idle stripe, make sure it is unhashed, and return it. */
 432 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
 433 {
 434         struct stripe_head *sh = NULL;
 435         struct list_head *first;
 436
 437         if (list_empty(conf->inactive_list + hash))
 438                 goto out;
 439         first = (conf->inactive_list + hash)->next;
 440         sh = list_entry(first, struct stripe_head, lru);
 441         list_del_init(first);
 442         remove_hash(sh);
 443         atomic_inc(&conf->active_stripes);
 444         BUG_ON(hash != sh->hash_lock_index);
 445         if (list_empty(conf->inactive_list + hash))
 446                 atomic_inc(&conf->empty_inactive_list_nr);
 447 out:
 448         return sh;
 449 }
 450
 451 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
 452 static void free_stripe_pages(struct stripe_head *sh)
 453 {
 454         int i;
 455         struct page *p;
 456
 457         /* Have not allocate page pool */
 458         if (!sh->pages)
 459                 return;
 460
 461         for (i = 0; i < sh->nr_pages; i++) {
 462                 p = sh->pages[i];
 463                 if (p)
 464                         put_page(p);
 465                 sh->pages[i] = NULL;
 466         }
 467 }
 468
 469 static int alloc_stripe_pages(struct stripe_head *sh, gfp_t gfp)
 470 {
 471         int i;
 472         struct page *p;
 473
 474         for (i = 0; i < sh->nr_pages; i++) {
 475                 /* The page have allocated. */
 476                 if (sh->pages[i])
 477                         continue;
 478
 479                 p = alloc_page(gfp);
 480                 if (!p) {
 481                         free_stripe_pages(sh);
 482                         return -ENOMEM;
 483                 }
 484                 sh->pages[i] = p;
 485         }
 486         return 0;
 487 }
 488
 489 static int
 490 init_stripe_shared_pages(struct stripe_head *sh, struct r5conf *conf, int disks)
 491 {
 492         int nr_pages, cnt;
 493
 494         if (sh->pages)
 495                 return 0;
 496
 497         /* Each of the sh->dev[i] need one conf->stripe_size */
 498         cnt = PAGE_SIZE / conf->stripe_size;
 499         nr_pages = (disks + cnt - 1) / cnt;
 500
 501         sh->pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
 502         if (!sh->pages)
 503                 return -ENOMEM;
 504         sh->nr_pages = nr_pages;
 505         sh->stripes_per_page = cnt;
 506         return 0;
 507 }
 508 #endif
 509
 510 static void shrink_buffers(struct stripe_head *sh)
 511 {
 512         int i;
 513         int num = sh->raid_conf->pool_size;
 514
 515 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 516         for (i = 0; i < num ; i++) {
 517                 struct page *p;
 518
 519                 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
 520                 p = sh->dev[i].page;
 521                 if (!p)
 522                         continue;
 523                 sh->dev[i].page = NULL;
 524                 put_page(p);
 525         }
 526 #else
 527         for (i = 0; i < num; i++)
 528                 sh->dev[i].page = NULL;
 529         free_stripe_pages(sh); /* Free pages */
 530 #endif
 531 }
 532
 533 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
 534 {
 535         int i;
 536         int num = sh->raid_conf->pool_size;
 537
 538 #if PAGE_SIZE == DEFAULT_STRIPE_SIZE
 539         for (i = 0; i < num; i++) {
 540                 struct page *page;
 541
 542                 if (!(page = alloc_page(gfp))) {
 543                         return 1;
 544                 }
 545                 sh->dev[i].page = page;
 546                 sh->dev[i].orig_page = page;
 547                 sh->dev[i].offset = 0;
 548         }
 549 #else
 550         if (alloc_stripe_pages(sh, gfp))
 551                 return -ENOMEM;
 552
 553         for (i = 0; i < num; i++) {
 554                 sh->dev[i].page = raid5_get_dev_page(sh, i);
 555                 sh->dev[i].orig_page = sh->dev[i].page;
 556                 sh->dev[i].offset = raid5_get_page_offset(sh, i);
 557         }
 558 #endif
 559         return 0;
 560 }
 561
 562 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 563                             struct stripe_head *sh);
 564
 565 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 566 {
 567         struct r5conf *conf = sh->raid_conf;
 568         int i, seq;
 569
 570         BUG_ON(atomic_read(&sh->count) != 0);
 571         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 572         BUG_ON(stripe_operations_active(sh));
 573         BUG_ON(sh->batch_head);
 574
 575         pr_debug("init_stripe called, stripe %llu\n",
 576                 (unsigned long long)sector);
 577 retry:
 578         seq = read_seqcount_begin(&conf->gen_lock);
 579         sh->generation = conf->generation - previous;
 580         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 581         sh->sector = sector;
 582         stripe_set_idx(sector, conf, previous, sh);
 583         sh->state = 0;
 584
 585         for (i = sh->disks; i--; ) {
 586                 struct r5dev *dev = &sh->dev[i];
 587
 588                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 589                     test_bit(R5_LOCKED, &dev->flags)) {
 590                         pr_err("sector=%llx i=%d %p %p %p %p %d\n",
 591                                (unsigned long long)sh->sector, i, dev->toread,
 592                                dev->read, dev->towrite, dev->written,
 593                                test_bit(R5_LOCKED, &dev->flags));
 594                         WARN_ON(1);
 595                 }
 596                 dev->flags = 0;
 597                 dev->sector = raid5_compute_blocknr(sh, i, previous);
 598         }
 599         if (read_seqcount_retry(&conf->gen_lock, seq))
 600                 goto retry;
 601         sh->overwrite_disks = 0;
 602         insert_hash(conf, sh);
 603         sh->cpu = smp_processor_id();
 604         set_bit(STRIPE_BATCH_READY, &sh->state);
 605 }
 606
 607 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 608                                          short generation)
 609 {
 610         struct stripe_head *sh;
 611
 612         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 613         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 614                 if (sh->sector == sector && sh->generation == generation)
 615                         return sh;
 616         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 617         return NULL;
 618 }
 619
 620 /*
 621  * Need to check if array has failed when deciding whether to:
 622  *  - start an array
 623  *  - remove non-faulty devices
 624  *  - add a spare
 625  *  - allow a reshape
 626  * This determination is simple when no reshape is happening.
 627  * However if there is a reshape, we need to carefully check
 628  * both the before and after sections.
 629  * This is because some failed devices may only affect one
 630  * of the two sections, and some non-in_sync devices may
 631  * be insync in the section most affected by failed devices.
 632  */
 633 int raid5_calc_degraded(struct r5conf *conf)
 634 {
 635         int degraded, degraded2;
 636         int i;
 637
 638         rcu_read_lock();
 639         degraded = 0;
 640         for (i = 0; i < conf->previous_raid_disks; i++) {
 641                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 642                 if (rdev && test_bit(Faulty, &rdev->flags))
 643                         rdev = rcu_dereference(conf->disks[i].replacement);
 644                 if (!rdev || test_bit(Faulty, &rdev->flags))
 645                         degraded++;
 646                 else if (test_bit(In_sync, &rdev->flags))
 647                         ;
 648                 else
 649                         /* not in-sync or faulty.
 650                          * If the reshape increases the number of devices,
 651                          * this is being recovered by the reshape, so
 652                          * this 'previous' section is not in_sync.
 653                          * If the number of devices is being reduced however,
 654                          * the device can only be part of the array if
 655                          * we are reverting a reshape, so this section will
 656                          * be in-sync.
 657                          */
 658                         if (conf->raid_disks >= conf->previous_raid_disks)
 659                                 degraded++;
 660         }
 661         rcu_read_unlock();
 662         if (conf->raid_disks == conf->previous_raid_disks)
 663                 return degraded;
 664         rcu_read_lock();
 665         degraded2 = 0;
 666         for (i = 0; i < conf->raid_disks; i++) {
 667                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 668                 if (rdev && test_bit(Faulty, &rdev->flags))
 669                         rdev = rcu_dereference(conf->disks[i].replacement);
 670                 if (!rdev || test_bit(Faulty, &rdev->flags))
 671                         degraded2++;
 672                 else if (test_bit(In_sync, &rdev->flags))
 673                         ;
 674                 else
 675                         /* not in-sync or faulty.
 676                          * If reshape increases the number of devices, this
 677                          * section has already been recovered, else it
 678                          * almost certainly hasn't.
 679                          */
 680                         if (conf->raid_disks <= conf->previous_raid_disks)
 681                                 degraded2++;
 682         }
 683         rcu_read_unlock();
 684         if (degraded2 > degraded)
 685                 return degraded2;
 686         return degraded;
 687 }
 688
 689 static int has_failed(struct r5conf *conf)
 690 {
 691         int degraded;
 692
 693         if (conf->mddev->reshape_position == MaxSector)
 694                 return conf->mddev->degraded > conf->max_degraded;
 695
 696         degraded = raid5_calc_degraded(conf);
 697         if (degraded > conf->max_degraded)
 698                 return 1;
 699         return 0;
 700 }
 701
 702 struct stripe_head *
 703 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
 704                         int previous, int noblock, int noquiesce)
 705 {
 706         struct stripe_head *sh;
 707         int hash = stripe_hash_locks_hash(conf, sector);
 708         int inc_empty_inactive_list_flag;
 709
 710         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 711
 712         spin_lock_irq(conf->hash_locks + hash);
 713
 714         do {
 715                 wait_event_lock_irq(conf->wait_for_quiescent,
 716                                     conf->quiesce == 0 || noquiesce,
 717                                     *(conf->hash_locks + hash));
 718                 sh = __find_stripe(conf, sector, conf->generation - previous);
 719                 if (!sh) {
 720                         if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
 721                                 sh = get_free_stripe(conf, hash);
 722                                 if (!sh && !test_bit(R5_DID_ALLOC,
 723                                                      &conf->cache_state))
 724                                         set_bit(R5_ALLOC_MORE,
 725                                                 &conf->cache_state);
 726                         }
 727                         if (noblock && sh == NULL)
 728                                 break;
 729
 730                         r5c_check_stripe_cache_usage(conf);
 731                         if (!sh) {
 732                                 set_bit(R5_INACTIVE_BLOCKED,
 733                                         &conf->cache_state);
 734                                 r5l_wake_reclaim(conf->log, 0);
 735                                 wait_event_lock_irq(
 736                                         conf->wait_for_stripe,
 737                                         !list_empty(conf->inactive_list + hash) &&
 738                                         (atomic_read(&conf->active_stripes)
 739                                          < (conf->max_nr_stripes * 3 / 4)
 740                                          || !test_bit(R5_INACTIVE_BLOCKED,
 741                                                       &conf->cache_state)),
 742                                         *(conf->hash_locks + hash));
 743                                 clear_bit(R5_INACTIVE_BLOCKED,
 744                                           &conf->cache_state);
 745                         } else {
 746                                 init_stripe(sh, sector, previous);
 747                                 atomic_inc(&sh->count);
 748                         }
 749                 } else if (!atomic_inc_not_zero(&sh->count)) {
 750                         spin_lock(&conf->device_lock);
 751                         if (!atomic_read(&sh->count)) {
 752                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 753                                         atomic_inc(&conf->active_stripes);
 754                                 BUG_ON(list_empty(&sh->lru) &&
 755                                        !test_bit(STRIPE_EXPANDING, &sh->state));
 756                                 inc_empty_inactive_list_flag = 0;
 757                                 if (!list_empty(conf->inactive_list + hash))
 758                                         inc_empty_inactive_list_flag = 1;
 759                                 list_del_init(&sh->lru);
 760                                 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 761                                         atomic_inc(&conf->empty_inactive_list_nr);
 762                                 if (sh->group) {
 763                                         sh->group->stripes_cnt--;
 764                                         sh->group = NULL;
 765                                 }
 766                         }
 767                         atomic_inc(&sh->count);
 768                         spin_unlock(&conf->device_lock);
 769                 }
 770         } while (sh == NULL);
 771
 772         spin_unlock_irq(conf->hash_locks + hash);
 773         return sh;
 774 }
 775
 776 static bool is_full_stripe_write(struct stripe_head *sh)
 777 {
 778         BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
 779         return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
 780 }
 781
 782 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 783                 __acquires(&sh1->stripe_lock)
 784                 __acquires(&sh2->stripe_lock)
 785 {
 786         if (sh1 > sh2) {
 787                 spin_lock_irq(&sh2->stripe_lock);
 788                 spin_lock_nested(&sh1->stripe_lock, 1);
 789         } else {
 790                 spin_lock_irq(&sh1->stripe_lock);
 791                 spin_lock_nested(&sh2->stripe_lock, 1);
 792         }
 793 }
 794
 795 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
 796                 __releases(&sh1->stripe_lock)
 797                 __releases(&sh2->stripe_lock)
 798 {
 799         spin_unlock(&sh1->stripe_lock);
 800         spin_unlock_irq(&sh2->stripe_lock);
 801 }
 802
 803 /* Only freshly new full stripe normal write stripe can be added to a batch list */
 804 static bool stripe_can_batch(struct stripe_head *sh)
 805 {
 806         struct r5conf *conf = sh->raid_conf;
 807
 808         if (raid5_has_log(conf) || raid5_has_ppl(conf))
 809                 return false;
 810         return test_bit(STRIPE_BATCH_READY, &sh->state) &&
 811                 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
 812                 is_full_stripe_write(sh);
 813 }
 814
 815 /* we only do back search */
 816 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
 817 {
 818         struct stripe_head *head;
 819         sector_t head_sector, tmp_sec;
 820         int hash;
 821         int dd_idx;
 822         int inc_empty_inactive_list_flag;
 823
 824         /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
 825         tmp_sec = sh->sector;
 826         if (!sector_div(tmp_sec, conf->chunk_sectors))
 827                 return;
 828         head_sector = sh->sector - RAID5_STRIPE_SECTORS(conf);
 829
 830         hash = stripe_hash_locks_hash(conf, head_sector);
 831         spin_lock_irq(conf->hash_locks + hash);
 832         head = __find_stripe(conf, head_sector, conf->generation);
 833         if (head && !atomic_inc_not_zero(&head->count)) {
 834                 spin_lock(&conf->device_lock);
 835                 if (!atomic_read(&head->count)) {
 836                         if (!test_bit(STRIPE_HANDLE, &head->state))
 837                                 atomic_inc(&conf->active_stripes);
 838                         BUG_ON(list_empty(&head->lru) &&
 839                                !test_bit(STRIPE_EXPANDING, &head->state));
 840                         inc_empty_inactive_list_flag = 0;
 841                         if (!list_empty(conf->inactive_list + hash))
 842                                 inc_empty_inactive_list_flag = 1;
 843                         list_del_init(&head->lru);
 844                         if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
 845                                 atomic_inc(&conf->empty_inactive_list_nr);
 846                         if (head->group) {
 847                                 head->group->stripes_cnt--;
 848                                 head->group = NULL;
 849                         }
 850                 }
 851                 atomic_inc(&head->count);
 852                 spin_unlock(&conf->device_lock);
 853         }
 854         spin_unlock_irq(conf->hash_locks + hash);
 855
 856         if (!head)
 857                 return;
 858         if (!stripe_can_batch(head))
 859                 goto out;
 860
 861         lock_two_stripes(head, sh);
 862         /* clear_batch_ready clear the flag */
 863         if (!stripe_can_batch(head) || !stripe_can_batch(sh))
 864                 goto unlock_out;
 865
 866         if (sh->batch_head)
 867                 goto unlock_out;
 868
 869         dd_idx = 0;
 870         while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
 871                 dd_idx++;
 872         if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
 873             bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
 874                 goto unlock_out;
 875
 876         if (head->batch_head) {
 877                 spin_lock(&head->batch_head->batch_lock);
 878                 /* This batch list is already running */
 879                 if (!stripe_can_batch(head)) {
 880                         spin_unlock(&head->batch_head->batch_lock);
 881                         goto unlock_out;
 882                 }
 883                 /*
 884                  * We must assign batch_head of this stripe within the
 885                  * batch_lock, otherwise clear_batch_ready of batch head
 886                  * stripe could clear BATCH_READY bit of this stripe and
 887                  * this stripe->batch_head doesn't get assigned, which
 888                  * could confuse clear_batch_ready for this stripe
 889                  */
 890                 sh->batch_head = head->batch_head;
 891
 892                 /*
 893                  * at this point, head's BATCH_READY could be cleared, but we
 894                  * can still add the stripe to batch list
 895                  */
 896                 list_add(&sh->batch_list, &head->batch_list);
 897                 spin_unlock(&head->batch_head->batch_lock);
 898         } else {
 899                 head->batch_head = head;
 900                 sh->batch_head = head->batch_head;
 901                 spin_lock(&head->batch_lock);
 902                 list_add_tail(&sh->batch_list, &head->batch_list);
 903                 spin_unlock(&head->batch_lock);
 904         }
 905
 906         if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 907                 if (atomic_dec_return(&conf->preread_active_stripes)
 908                     < IO_THRESHOLD)
 909                         md_wakeup_thread(conf->mddev->thread);
 910
 911         if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
 912                 int seq = sh->bm_seq;
 913                 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
 914                     sh->batch_head->bm_seq > seq)
 915                         seq = sh->batch_head->bm_seq;
 916                 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
 917                 sh->batch_head->bm_seq = seq;
 918         }
 919
 920         atomic_inc(&sh->count);
 921 unlock_out:
 922         unlock_two_stripes(head, sh);
 923 out:
 924         raid5_release_stripe(head);
 925 }
 926
 927 /* Determine if 'data_offset' or 'new_data_offset' should be used
 928  * in this stripe_head.
 929  */
 930 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 931 {
 932         sector_t progress = conf->reshape_progress;
 933         /* Need a memory barrier to make sure we see the value
 934          * of conf->generation, or ->data_offset that was set before
 935          * reshape_progress was updated.
 936          */
 937         smp_rmb();
 938         if (progress == MaxSector)
 939                 return 0;
 940         if (sh->generation == conf->generation - 1)
 941                 return 0;
 942         /* We are in a reshape, and this is a new-generation stripe,
 943          * so use new_data_offset.
 944          */
 945         return 1;
 946 }
 947
 948 static void dispatch_bio_list(struct bio_list *tmp)
 949 {
 950         struct bio *bio;
 951
 952         while ((bio = bio_list_pop(tmp)))
 953                 submit_bio_noacct(bio);
 954 }
 955
 956 static int cmp_stripe(void *priv, const struct list_head *a,
 957                       const struct list_head *b)
 958 {
 959         const struct r5pending_data *da = list_entry(a,
 960                                 struct r5pending_data, sibling);
 961         const struct r5pending_data *db = list_entry(b,
 962                                 struct r5pending_data, sibling);
 963         if (da->sector > db->sector)
 964                 return 1;
 965         if (da->sector < db->sector)
 966                 return -1;
 967         return 0;
 968 }
 969
 970 static void dispatch_defer_bios(struct r5conf *conf, int target,
 971                                 struct bio_list *list)
 972 {
 973         struct r5pending_data *data;
 974         struct list_head *first, *next = NULL;
 975         int cnt = 0;
 976
 977         if (conf->pending_data_cnt == 0)
 978                 return;
 979
 980         list_sort(NULL, &conf->pending_list, cmp_stripe);
 981
 982         first = conf->pending_list.next;
 983
 984         /* temporarily move the head */
 985         if (conf->next_pending_data)
 986                 list_move_tail(&conf->pending_list,
 987                                 &conf->next_pending_data->sibling);
 988
 989         while (!list_empty(&conf->pending_list)) {
 990                 data = list_first_entry(&conf->pending_list,
 991                         struct r5pending_data, sibling);
 992                 if (&data->sibling == first)
 993                         first = data->sibling.next;
 994                 next = data->sibling.next;
 995
 996                 bio_list_merge(list, &data->bios);
 997                 list_move(&data->sibling, &conf->free_list);
 998                 cnt++;
 999                 if (cnt >= target)
1000                         break;
1001         }
1002         conf->pending_data_cnt -= cnt;
1003         BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
1004
1005         if (next != &conf->pending_list)
1006                 conf->next_pending_data = list_entry(next,
1007                                 struct r5pending_data, sibling);
1008         else
1009                 conf->next_pending_data = NULL;
1010         /* list isn't empty */
1011         if (first != &conf->pending_list)
1012                 list_move_tail(&conf->pending_list, first);
1013 }
1014
1015 static void flush_deferred_bios(struct r5conf *conf)
1016 {
1017         struct bio_list tmp = BIO_EMPTY_LIST;
1018
1019         if (conf->pending_data_cnt == 0)
1020                 return;
1021
1022         spin_lock(&conf->pending_bios_lock);
1023         dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
1024         BUG_ON(conf->pending_data_cnt != 0);
1025         spin_unlock(&conf->pending_bios_lock);
1026
1027         dispatch_bio_list(&tmp);
1028 }
1029
1030 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
1031                                 struct bio_list *bios)
1032 {
1033         struct bio_list tmp = BIO_EMPTY_LIST;
1034         struct r5pending_data *ent;
1035
1036         spin_lock(&conf->pending_bios_lock);
1037         ent = list_first_entry(&conf->free_list, struct r5pending_data,
1038                                                         sibling);
1039         list_move_tail(&ent->sibling, &conf->pending_list);
1040         ent->sector = sector;
1041         bio_list_init(&ent->bios);
1042         bio_list_merge(&ent->bios, bios);
1043         conf->pending_data_cnt++;
1044         if (conf->pending_data_cnt >= PENDING_IO_MAX)
1045                 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
1046
1047         spin_unlock(&conf->pending_bios_lock);
1048
1049         dispatch_bio_list(&tmp);
1050 }
1051
1052 static void
1053 raid5_end_read_request(struct bio *bi);
1054 static void
1055 raid5_end_write_request(struct bio *bi);
1056
1057 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
1058 {
1059         struct r5conf *conf = sh->raid_conf;
1060         int i, disks = sh->disks;
1061         struct stripe_head *head_sh = sh;
1062         struct bio_list pending_bios = BIO_EMPTY_LIST;
1063         bool should_defer;
1064
1065         might_sleep();
1066
1067         if (log_stripe(sh, s) == 0)
1068                 return;
1069
1070         should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1071
1072         for (i = disks; i--; ) {
1073                 int op, op_flags = 0;
1074                 int replace_only = 0;
1075                 struct bio *bi, *rbi;
1076                 struct md_rdev *rdev, *rrdev = NULL;
1077
1078                 sh = head_sh;
1079                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1080                         op = REQ_OP_WRITE;
1081                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1082                                 op_flags = REQ_FUA;
1083                         if (test_bit(R5_Discard, &sh->dev[i].flags))
1084                                 op = REQ_OP_DISCARD;
1085                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1086                         op = REQ_OP_READ;
1087                 else if (test_and_clear_bit(R5_WantReplace,
1088                                             &sh->dev[i].flags)) {
1089                         op = REQ_OP_WRITE;
1090                         replace_only = 1;
1091                 } else
1092                         continue;
1093                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1094                         op_flags |= REQ_SYNC;
1095
1096 again:
1097                 bi = &sh->dev[i].req;
1098                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1099
1100                 rcu_read_lock();
1101                 rrdev = rcu_dereference(conf->disks[i].replacement);
1102                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1103                 rdev = rcu_dereference(conf->disks[i].rdev);
1104                 if (!rdev) {
1105                         rdev = rrdev;
1106                         rrdev = NULL;
1107                 }
1108                 if (op_is_write(op)) {
1109                         if (replace_only)
1110                                 rdev = NULL;
1111                         if (rdev == rrdev)
1112                                 /* We raced and saw duplicates */
1113                                 rrdev = NULL;
1114                 } else {
1115                         if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1116                                 rdev = rrdev;
1117                         rrdev = NULL;
1118                 }
1119
1120                 if (rdev && test_bit(Faulty, &rdev->flags))
1121                         rdev = NULL;
1122                 if (rdev)
1123                         atomic_inc(&rdev->nr_pending);
1124                 if (rrdev && test_bit(Faulty, &rrdev->flags))
1125                         rrdev = NULL;
1126                 if (rrdev)
1127                         atomic_inc(&rrdev->nr_pending);
1128                 rcu_read_unlock();
1129
1130                 /* We have already checked bad blocks for reads.  Now
1131                  * need to check for writes.  We never accept write errors
1132                  * on the replacement, so we don't to check rrdev.
1133                  */
1134                 while (op_is_write(op) && rdev &&
1135                        test_bit(WriteErrorSeen, &rdev->flags)) {
1136                         sector_t first_bad;
1137                         int bad_sectors;
1138                         int bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
1139                                               &first_bad, &bad_sectors);
1140                         if (!bad)
1141                                 break;
1142
1143                         if (bad < 0) {
1144                                 set_bit(BlockedBadBlocks, &rdev->flags);
1145                                 if (!conf->mddev->external &&
1146                                     conf->mddev->sb_flags) {
1147                                         /* It is very unlikely, but we might
1148                                          * still need to write out the
1149                                          * bad block log - better give it
1150                                          * a chance*/
1151                                         md_check_recovery(conf->mddev);
1152                                 }
1153                                 /*
1154                                  * Because md_wait_for_blocked_rdev
1155                                  * will dec nr_pending, we must
1156                                  * increment it first.
1157                                  */
1158                                 atomic_inc(&rdev->nr_pending);
1159                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
1160                         } else {
1161                                 /* Acknowledged bad block - skip the write */
1162                                 rdev_dec_pending(rdev, conf->mddev);
1163                                 rdev = NULL;
1164                         }
1165                 }
1166
1167                 if (rdev) {
1168                         if (s->syncing || s->expanding || s->expanded
1169                             || s->replacing)
1170                                 md_sync_acct(rdev->bdev, RAID5_STRIPE_SECTORS(conf));
1171
1172                         set_bit(STRIPE_IO_STARTED, &sh->state);
1173
1174                         bio_set_dev(bi, rdev->bdev);
1175                         bio_set_op_attrs(bi, op, op_flags);
1176                         bi->bi_end_io = op_is_write(op)
1177                                 ? raid5_end_write_request
1178                                 : raid5_end_read_request;
1179                         bi->bi_private = sh;
1180
1181                         pr_debug("%s: for %llu schedule op %d on disc %d\n",
1182                                 __func__, (unsigned long long)sh->sector,
1183                                 bi->bi_opf, i);
1184                         atomic_inc(&sh->count);
1185                         if (sh != head_sh)
1186                                 atomic_inc(&head_sh->count);
1187                         if (use_new_offset(conf, sh))
1188                                 bi->bi_iter.bi_sector = (sh->sector
1189                                                  + rdev->new_data_offset);
1190                         else
1191                                 bi->bi_iter.bi_sector = (sh->sector
1192                                                  + rdev->data_offset);
1193                         if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1194                                 bi->bi_opf |= REQ_NOMERGE;
1195
1196                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1197                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1198
1199                         if (!op_is_write(op) &&
1200                             test_bit(R5_InJournal, &sh->dev[i].flags))
1201                                 /*
1202                                  * issuing read for a page in journal, this
1203                                  * must be preparing for prexor in rmw; read
1204                                  * the data into orig_page
1205                                  */
1206                                 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1207                         else
1208                                 sh->dev[i].vec.bv_page = sh->dev[i].page;
1209                         bi->bi_vcnt = 1;
1210                         bi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1211                         bi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1212                         bi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1213                         bi->bi_write_hint = sh->dev[i].write_hint;
1214                         if (!rrdev)
1215                                 sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1216                         /*
1217                          * If this is discard request, set bi_vcnt 0. We don't
1218                          * want to confuse SCSI because SCSI will replace payload
1219                          */
1220                         if (op == REQ_OP_DISCARD)
1221                                 bi->bi_vcnt = 0;
1222                         if (rrdev)
1223                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1224
1225                         if (conf->mddev->gendisk)
1226                                 trace_block_bio_remap(bi,
1227                                                 disk_devt(conf->mddev->gendisk),
1228                                                 sh->dev[i].sector);
1229                         if (should_defer && op_is_write(op))
1230                                 bio_list_add(&pending_bios, bi);
1231                         else
1232                                 submit_bio_noacct(bi);
1233                 }
1234                 if (rrdev) {
1235                         if (s->syncing || s->expanding || s->expanded
1236                             || s->replacing)
1237                                 md_sync_acct(rrdev->bdev, RAID5_STRIPE_SECTORS(conf));
1238
1239                         set_bit(STRIPE_IO_STARTED, &sh->state);
1240
1241                         bio_set_dev(rbi, rrdev->bdev);
1242                         bio_set_op_attrs(rbi, op, op_flags);
1243                         BUG_ON(!op_is_write(op));
1244                         rbi->bi_end_io = raid5_end_write_request;
1245                         rbi->bi_private = sh;
1246
1247                         pr_debug("%s: for %llu schedule op %d on "
1248                                  "replacement disc %d\n",
1249                                 __func__, (unsigned long long)sh->sector,
1250                                 rbi->bi_opf, i);
1251                         atomic_inc(&sh->count);
1252                         if (sh != head_sh)
1253                                 atomic_inc(&head_sh->count);
1254                         if (use_new_offset(conf, sh))
1255                                 rbi->bi_iter.bi_sector = (sh->sector
1256                                                   + rrdev->new_data_offset);
1257                         else
1258                                 rbi->bi_iter.bi_sector = (sh->sector
1259                                                   + rrdev->data_offset);
1260                         if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1261                                 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1262                         sh->dev[i].rvec.bv_page = sh->dev[i].page;
1263                         rbi->bi_vcnt = 1;
1264                         rbi->bi_io_vec[0].bv_len = RAID5_STRIPE_SIZE(conf);
1265                         rbi->bi_io_vec[0].bv_offset = sh->dev[i].offset;
1266                         rbi->bi_iter.bi_size = RAID5_STRIPE_SIZE(conf);
1267                         rbi->bi_write_hint = sh->dev[i].write_hint;
1268                         sh->dev[i].write_hint = RWH_WRITE_LIFE_NOT_SET;
1269                         /*
1270                          * If this is discard request, set bi_vcnt 0. We don't
1271                          * want to confuse SCSI because SCSI will replace payload
1272                          */
1273                         if (op == REQ_OP_DISCARD)
1274                                 rbi->bi_vcnt = 0;
1275                         if (conf->mddev->gendisk)
1276                                 trace_block_bio_remap(rbi,
1277                                                 disk_devt(conf->mddev->gendisk),
1278                                                 sh->dev[i].sector);
1279                         if (should_defer && op_is_write(op))
1280                                 bio_list_add(&pending_bios, rbi);
1281                         else
1282                                 submit_bio_noacct(rbi);
1283                 }
1284                 if (!rdev && !rrdev) {
1285                         if (op_is_write(op))
1286                                 set_bit(STRIPE_DEGRADED, &sh->state);
1287                         pr_debug("skip op %d on disc %d for sector %llu\n",
1288                                 bi->bi_opf, i, (unsigned long long)sh->sector);
1289                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1290                         set_bit(STRIPE_HANDLE, &sh->state);
1291                 }
1292
1293                 if (!head_sh->batch_head)
1294                         continue;
1295                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1296                                       batch_list);
1297                 if (sh != head_sh)
1298                         goto again;
1299         }
1300
1301         if (should_defer && !bio_list_empty(&pending_bios))
1302                 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1303 }
1304
1305 static struct dma_async_tx_descriptor *
1306 async_copy_data(int frombio, struct bio *bio, struct page **page,
1307         unsigned int poff, sector_t sector, struct dma_async_tx_descriptor *tx,
1308         struct stripe_head *sh, int no_skipcopy)
1309 {
1310         struct bio_vec bvl;
1311         struct bvec_iter iter;
1312         struct page *bio_page;
1313         int page_offset;
1314         struct async_submit_ctl submit;
1315         enum async_tx_flags flags = 0;
1316         struct r5conf *conf = sh->raid_conf;
1317
1318         if (bio->bi_iter.bi_sector >= sector)
1319                 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1320         else
1321                 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1322
1323         if (frombio)
1324                 flags |= ASYNC_TX_FENCE;
1325         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1326
1327         bio_for_each_segment(bvl, bio, iter) {
1328                 int len = bvl.bv_len;
1329                 int clen;
1330                 int b_offset = 0;
1331
1332                 if (page_offset < 0) {
1333                         b_offset = -page_offset;
1334                         page_offset += b_offset;
1335                         len -= b_offset;
1336                 }
1337
1338                 if (len > 0 && page_offset + len > RAID5_STRIPE_SIZE(conf))
1339                         clen = RAID5_STRIPE_SIZE(conf) - page_offset;
1340                 else
1341                         clen = len;
1342
1343                 if (clen > 0) {
1344                         b_offset += bvl.bv_offset;
1345                         bio_page = bvl.bv_page;
1346                         if (frombio) {
1347                                 if (conf->skip_copy &&
1348                                     b_offset == 0 && page_offset == 0 &&
1349                                     clen == RAID5_STRIPE_SIZE(conf) &&
1350                                     !no_skipcopy)
1351                                         *page = bio_page;
1352                                 else
1353                                         tx = async_memcpy(*page, bio_page, page_offset + poff,
1354                                                   b_offset, clen, &submit);
1355                         } else
1356                                 tx = async_memcpy(bio_page, *page, b_offset,
1357                                                   page_offset + poff, clen, &submit);
1358                 }
1359                 /* chain the operations */
1360                 submit.depend_tx = tx;
1361
1362                 if (clen < len) /* hit end of page */
1363                         break;
1364                 page_offset +=  len;
1365         }
1366
1367         return tx;
1368 }
1369
1370 static void ops_complete_biofill(void *stripe_head_ref)
1371 {
1372         struct stripe_head *sh = stripe_head_ref;
1373         int i;
1374         struct r5conf *conf = sh->raid_conf;
1375
1376         pr_debug("%s: stripe %llu\n", __func__,
1377                 (unsigned long long)sh->sector);
1378
1379         /* clear completed biofills */
1380         for (i = sh->disks; i--; ) {
1381                 struct r5dev *dev = &sh->dev[i];
1382
1383                 /* acknowledge completion of a biofill operation */
1384                 /* and check if we need to reply to a read request,
1385                  * new R5_Wantfill requests are held off until
1386                  * !STRIPE_BIOFILL_RUN
1387                  */
1388                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1389                         struct bio *rbi, *rbi2;
1390
1391                         BUG_ON(!dev->read);
1392                         rbi = dev->read;
1393                         dev->read = NULL;
1394                         while (rbi && rbi->bi_iter.bi_sector <
1395                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1396                                 rbi2 = r5_next_bio(conf, rbi, dev->sector);
1397                                 bio_endio(rbi);
1398                                 rbi = rbi2;
1399                         }
1400                 }
1401         }
1402         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1403
1404         set_bit(STRIPE_HANDLE, &sh->state);
1405         raid5_release_stripe(sh);
1406 }
1407
1408 static void ops_run_biofill(struct stripe_head *sh)
1409 {
1410         struct dma_async_tx_descriptor *tx = NULL;
1411         struct async_submit_ctl submit;
1412         int i;
1413         struct r5conf *conf = sh->raid_conf;
1414
1415         BUG_ON(sh->batch_head);
1416         pr_debug("%s: stripe %llu\n", __func__,
1417                 (unsigned long long)sh->sector);
1418
1419         for (i = sh->disks; i--; ) {
1420                 struct r5dev *dev = &sh->dev[i];
1421                 if (test_bit(R5_Wantfill, &dev->flags)) {
1422                         struct bio *rbi;
1423                         spin_lock_irq(&sh->stripe_lock);
1424                         dev->read = rbi = dev->toread;
1425                         dev->toread = NULL;
1426                         spin_unlock_irq(&sh->stripe_lock);
1427                         while (rbi && rbi->bi_iter.bi_sector <
1428                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1429                                 tx = async_copy_data(0, rbi, &dev->page,
1430                                                      dev->offset,
1431                                                      dev->sector, tx, sh, 0);
1432                                 rbi = r5_next_bio(conf, rbi, dev->sector);
1433                         }
1434                 }
1435         }
1436
1437         atomic_inc(&sh->count);
1438         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1439         async_trigger_callback(&submit);
1440 }
1441
1442 static void mark_target_uptodate(struct stripe_head *sh, int target)
1443 {
1444         struct r5dev *tgt;
1445
1446         if (target < 0)
1447                 return;
1448
1449         tgt = &sh->dev[target];
1450         set_bit(R5_UPTODATE, &tgt->flags);
1451         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1452         clear_bit(R5_Wantcompute, &tgt->flags);
1453 }
1454
1455 static void ops_complete_compute(void *stripe_head_ref)
1456 {
1457         struct stripe_head *sh = stripe_head_ref;
1458
1459         pr_debug("%s: stripe %llu\n", __func__,
1460                 (unsigned long long)sh->sector);
1461
1462         /* mark the computed target(s) as uptodate */
1463         mark_target_uptodate(sh, sh->ops.target);
1464         mark_target_uptodate(sh, sh->ops.target2);
1465
1466         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1467         if (sh->check_state == check_state_compute_run)
1468                 sh->check_state = check_state_compute_result;
1469         set_bit(STRIPE_HANDLE, &sh->state);
1470         raid5_release_stripe(sh);
1471 }
1472
1473 /* return a pointer to the address conversion region of the scribble buffer */
1474 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1475 {
1476         return percpu->scribble + i * percpu->scribble_obj_size;
1477 }
1478
1479 /* return a pointer to the address conversion region of the scribble buffer */
1480 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1481                                  struct raid5_percpu *percpu, int i)
1482 {
1483         return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1484 }
1485
1486 /*
1487  * Return a pointer to record offset address.
1488  */
1489 static unsigned int *
1490 to_addr_offs(struct stripe_head *sh, struct raid5_percpu *percpu)
1491 {
1492         return (unsigned int *) (to_addr_conv(sh, percpu, 0) + sh->disks + 2);
1493 }
1494
1495 static struct dma_async_tx_descriptor *
1496 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1497 {
1498         int disks = sh->disks;
1499         struct page **xor_srcs = to_addr_page(percpu, 0);
1500         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1501         int target = sh->ops.target;
1502         struct r5dev *tgt = &sh->dev[target];
1503         struct page *xor_dest = tgt->page;
1504         unsigned int off_dest = tgt->offset;
1505         int count = 0;
1506         struct dma_async_tx_descriptor *tx;
1507         struct async_submit_ctl submit;
1508         int i;
1509
1510         BUG_ON(sh->batch_head);
1511
1512         pr_debug("%s: stripe %llu block: %d\n",
1513                 __func__, (unsigned long long)sh->sector, target);
1514         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1515
1516         for (i = disks; i--; ) {
1517                 if (i != target) {
1518                         off_srcs[count] = sh->dev[i].offset;
1519                         xor_srcs[count++] = sh->dev[i].page;
1520                 }
1521         }
1522
1523         atomic_inc(&sh->count);
1524
1525         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1526                           ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1527         if (unlikely(count == 1))
1528                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
1529                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1530         else
1531                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1532                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1533
1534         return tx;
1535 }
1536
1537 /* set_syndrome_sources - populate source buffers for gen_syndrome
1538  * @srcs - (struct page *) array of size sh->disks
1539  * @offs - (unsigned int) array of offset for each page
1540  * @sh - stripe_head to parse
1541  *
1542  * Populates srcs in proper layout order for the stripe and returns the
1543  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1544  * destination buffer is recorded in srcs[count] and the Q destination
1545  * is recorded in srcs[count+1]].
1546  */
1547 static int set_syndrome_sources(struct page **srcs,
1548                                 unsigned int *offs,
1549                                 struct stripe_head *sh,
1550                                 int srctype)
1551 {
1552         int disks = sh->disks;
1553         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1554         int d0_idx = raid6_d0(sh);
1555         int count;
1556         int i;
1557
1558         for (i = 0; i < disks; i++)
1559                 srcs[i] = NULL;
1560
1561         count = 0;
1562         i = d0_idx;
1563         do {
1564                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1565                 struct r5dev *dev = &sh->dev[i];
1566
1567                 if (i == sh->qd_idx || i == sh->pd_idx ||
1568                     (srctype == SYNDROME_SRC_ALL) ||
1569                     (srctype == SYNDROME_SRC_WANT_DRAIN &&
1570                      (test_bit(R5_Wantdrain, &dev->flags) ||
1571                       test_bit(R5_InJournal, &dev->flags))) ||
1572                     (srctype == SYNDROME_SRC_WRITTEN &&
1573                      (dev->written ||
1574                       test_bit(R5_InJournal, &dev->flags)))) {
1575                         if (test_bit(R5_InJournal, &dev->flags))
1576                                 srcs[slot] = sh->dev[i].orig_page;
1577                         else
1578                                 srcs[slot] = sh->dev[i].page;
1579                         /*
1580                          * For R5_InJournal, PAGE_SIZE must be 4KB and will
1581                          * not shared page. In that case, dev[i].offset
1582                          * is 0.
1583                          */
1584                         offs[slot] = sh->dev[i].offset;
1585                 }
1586                 i = raid6_next_disk(i, disks);
1587         } while (i != d0_idx);
1588
1589         return syndrome_disks;
1590 }
1591
1592 static struct dma_async_tx_descriptor *
1593 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1594 {
1595         int disks = sh->disks;
1596         struct page **blocks = to_addr_page(percpu, 0);
1597         unsigned int *offs = to_addr_offs(sh, percpu);
1598         int target;
1599         int qd_idx = sh->qd_idx;
1600         struct dma_async_tx_descriptor *tx;
1601         struct async_submit_ctl submit;
1602         struct r5dev *tgt;
1603         struct page *dest;
1604         unsigned int dest_off;
1605         int i;
1606         int count;
1607
1608         BUG_ON(sh->batch_head);
1609         if (sh->ops.target < 0)
1610                 target = sh->ops.target2;
1611         else if (sh->ops.target2 < 0)
1612                 target = sh->ops.target;
1613         else
1614                 /* we should only have one valid target */
1615                 BUG();
1616         BUG_ON(target < 0);
1617         pr_debug("%s: stripe %llu block: %d\n",
1618                 __func__, (unsigned long long)sh->sector, target);
1619
1620         tgt = &sh->dev[target];
1621         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1622         dest = tgt->page;
1623         dest_off = tgt->offset;
1624
1625         atomic_inc(&sh->count);
1626
1627         if (target == qd_idx) {
1628                 count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1629                 blocks[count] = NULL; /* regenerating p is not necessary */
1630                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1631                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1632                                   ops_complete_compute, sh,
1633                                   to_addr_conv(sh, percpu, 0));
1634                 tx = async_gen_syndrome(blocks, offs, count+2,
1635                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1636         } else {
1637                 /* Compute any data- or p-drive using XOR */
1638                 count = 0;
1639                 for (i = disks; i-- ; ) {
1640                         if (i == target || i == qd_idx)
1641                                 continue;
1642                         offs[count] = sh->dev[i].offset;
1643                         blocks[count++] = sh->dev[i].page;
1644                 }
1645
1646                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1647                                   NULL, ops_complete_compute, sh,
1648                                   to_addr_conv(sh, percpu, 0));
1649                 tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1650                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1651         }
1652
1653         return tx;
1654 }
1655
1656 static struct dma_async_tx_descriptor *
1657 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1658 {
1659         int i, count, disks = sh->disks;
1660         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1661         int d0_idx = raid6_d0(sh);
1662         int faila = -1, failb = -1;
1663         int target = sh->ops.target;
1664         int target2 = sh->ops.target2;
1665         struct r5dev *tgt = &sh->dev[target];
1666         struct r5dev *tgt2 = &sh->dev[target2];
1667         struct dma_async_tx_descriptor *tx;
1668         struct page **blocks = to_addr_page(percpu, 0);
1669         unsigned int *offs = to_addr_offs(sh, percpu);
1670         struct async_submit_ctl submit;
1671
1672         BUG_ON(sh->batch_head);
1673         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1674                  __func__, (unsigned long long)sh->sector, target, target2);
1675         BUG_ON(target < 0 || target2 < 0);
1676         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1677         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1678
1679         /* we need to open-code set_syndrome_sources to handle the
1680          * slot number conversion for 'faila' and 'failb'
1681          */
1682         for (i = 0; i < disks ; i++) {
1683                 offs[i] = 0;
1684                 blocks[i] = NULL;
1685         }
1686         count = 0;
1687         i = d0_idx;
1688         do {
1689                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1690
1691                 offs[slot] = sh->dev[i].offset;
1692                 blocks[slot] = sh->dev[i].page;
1693
1694                 if (i == target)
1695                         faila = slot;
1696                 if (i == target2)
1697                         failb = slot;
1698                 i = raid6_next_disk(i, disks);
1699         } while (i != d0_idx);
1700
1701         BUG_ON(faila == failb);
1702         if (failb < faila)
1703                 swap(faila, failb);
1704         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1705                  __func__, (unsigned long long)sh->sector, faila, failb);
1706
1707         atomic_inc(&sh->count);
1708
1709         if (failb == syndrome_disks+1) {
1710                 /* Q disk is one of the missing disks */
1711                 if (faila == syndrome_disks) {
1712                         /* Missing P+Q, just recompute */
1713                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1714                                           ops_complete_compute, sh,
1715                                           to_addr_conv(sh, percpu, 0));
1716                         return async_gen_syndrome(blocks, offs, syndrome_disks+2,
1717                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1718                                                   &submit);
1719                 } else {
1720                         struct page *dest;
1721                         unsigned int dest_off;
1722                         int data_target;
1723                         int qd_idx = sh->qd_idx;
1724
1725                         /* Missing D+Q: recompute D from P, then recompute Q */
1726                         if (target == qd_idx)
1727                                 data_target = target2;
1728                         else
1729                                 data_target = target;
1730
1731                         count = 0;
1732                         for (i = disks; i-- ; ) {
1733                                 if (i == data_target || i == qd_idx)
1734                                         continue;
1735                                 offs[count] = sh->dev[i].offset;
1736                                 blocks[count++] = sh->dev[i].page;
1737                         }
1738                         dest = sh->dev[data_target].page;
1739                         dest_off = sh->dev[data_target].offset;
1740                         init_async_submit(&submit,
1741                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1742                                           NULL, NULL, NULL,
1743                                           to_addr_conv(sh, percpu, 0));
1744                         tx = async_xor_offs(dest, dest_off, blocks, offs, count,
1745                                        RAID5_STRIPE_SIZE(sh->raid_conf),
1746                                        &submit);
1747
1748                         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_ALL);
1749                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1750                                           ops_complete_compute, sh,
1751                                           to_addr_conv(sh, percpu, 0));
1752                         return async_gen_syndrome(blocks, offs, count+2,
1753                                                   RAID5_STRIPE_SIZE(sh->raid_conf),
1754                                                   &submit);
1755                 }
1756         } else {
1757                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1758                                   ops_complete_compute, sh,
1759                                   to_addr_conv(sh, percpu, 0));
1760                 if (failb == syndrome_disks) {
1761                         /* We're missing D+P. */
1762                         return async_raid6_datap_recov(syndrome_disks+2,
1763                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1764                                                 faila,
1765                                                 blocks, offs, &submit);
1766                 } else {
1767                         /* We're missing D+D. */
1768                         return async_raid6_2data_recov(syndrome_disks+2,
1769                                                 RAID5_STRIPE_SIZE(sh->raid_conf),
1770                                                 faila, failb,
1771                                                 blocks, offs, &submit);
1772                 }
1773         }
1774 }
1775
1776 static void ops_complete_prexor(void *stripe_head_ref)
1777 {
1778         struct stripe_head *sh = stripe_head_ref;
1779
1780         pr_debug("%s: stripe %llu\n", __func__,
1781                 (unsigned long long)sh->sector);
1782
1783         if (r5c_is_writeback(sh->raid_conf->log))
1784                 /*
1785                  * raid5-cache write back uses orig_page during prexor.
1786                  * After prexor, it is time to free orig_page
1787                  */
1788                 r5c_release_extra_page(sh);
1789 }
1790
1791 static struct dma_async_tx_descriptor *
1792 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1793                 struct dma_async_tx_descriptor *tx)
1794 {
1795         int disks = sh->disks;
1796         struct page **xor_srcs = to_addr_page(percpu, 0);
1797         unsigned int *off_srcs = to_addr_offs(sh, percpu);
1798         int count = 0, pd_idx = sh->pd_idx, i;
1799         struct async_submit_ctl submit;
1800
1801         /* existing parity data subtracted */
1802         unsigned int off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
1803         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1804
1805         BUG_ON(sh->batch_head);
1806         pr_debug("%s: stripe %llu\n", __func__,
1807                 (unsigned long long)sh->sector);
1808
1809         for (i = disks; i--; ) {
1810                 struct r5dev *dev = &sh->dev[i];
1811                 /* Only process blocks that are known to be uptodate */
1812                 if (test_bit(R5_InJournal, &dev->flags)) {
1813                         /*
1814                          * For this case, PAGE_SIZE must be equal to 4KB and
1815                          * page offset is zero.
1816                          */
1817                         off_srcs[count] = dev->offset;
1818                         xor_srcs[count++] = dev->orig_page;
1819                 } else if (test_bit(R5_Wantdrain, &dev->flags)) {
1820                         off_srcs[count] = dev->offset;
1821                         xor_srcs[count++] = dev->page;
1822                 }
1823         }
1824
1825         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1826                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1827         tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
1828                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1829
1830         return tx;
1831 }
1832
1833 static struct dma_async_tx_descriptor *
1834 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1835                 struct dma_async_tx_descriptor *tx)
1836 {
1837         struct page **blocks = to_addr_page(percpu, 0);
1838         unsigned int *offs = to_addr_offs(sh, percpu);
1839         int count;
1840         struct async_submit_ctl submit;
1841
1842         pr_debug("%s: stripe %llu\n", __func__,
1843                 (unsigned long long)sh->sector);
1844
1845         count = set_syndrome_sources(blocks, offs, sh, SYNDROME_SRC_WANT_DRAIN);
1846
1847         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1848                           ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1849         tx = async_gen_syndrome(blocks, offs, count+2,
1850                         RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
1851
1852         return tx;
1853 }
1854
1855 static struct dma_async_tx_descriptor *
1856 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1857 {
1858         struct r5conf *conf = sh->raid_conf;
1859         int disks = sh->disks;
1860         int i;
1861         struct stripe_head *head_sh = sh;
1862
1863         pr_debug("%s: stripe %llu\n", __func__,
1864                 (unsigned long long)sh->sector);
1865
1866         for (i = disks; i--; ) {
1867                 struct r5dev *dev;
1868                 struct bio *chosen;
1869
1870                 sh = head_sh;
1871                 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1872                         struct bio *wbi;
1873
1874 again:
1875                         dev = &sh->dev[i];
1876                         /*
1877                          * clear R5_InJournal, so when rewriting a page in
1878                          * journal, it is not skipped by r5l_log_stripe()
1879                          */
1880                         clear_bit(R5_InJournal, &dev->flags);
1881                         spin_lock_irq(&sh->stripe_lock);
1882                         chosen = dev->towrite;
1883                         dev->towrite = NULL;
1884                         sh->overwrite_disks = 0;
1885                         BUG_ON(dev->written);
1886                         wbi = dev->written = chosen;
1887                         spin_unlock_irq(&sh->stripe_lock);
1888                         WARN_ON(dev->page != dev->orig_page);
1889
1890                         while (wbi && wbi->bi_iter.bi_sector <
1891                                 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
1892                                 if (wbi->bi_opf & REQ_FUA)
1893                                         set_bit(R5_WantFUA, &dev->flags);
1894                                 if (wbi->bi_opf & REQ_SYNC)
1895                                         set_bit(R5_SyncIO, &dev->flags);
1896                                 if (bio_op(wbi) == REQ_OP_DISCARD)
1897                                         set_bit(R5_Discard, &dev->flags);
1898                                 else {
1899                                         tx = async_copy_data(1, wbi, &dev->page,
1900                                                              dev->offset,
1901                                                              dev->sector, tx, sh,
1902                                                              r5c_is_writeback(conf->log));
1903                                         if (dev->page != dev->orig_page &&
1904                                             !r5c_is_writeback(conf->log)) {
1905                                                 set_bit(R5_SkipCopy, &dev->flags);
1906                                                 clear_bit(R5_UPTODATE, &dev->flags);
1907                                                 clear_bit(R5_OVERWRITE, &dev->flags);
1908                                         }
1909                                 }
1910                                 wbi = r5_next_bio(conf, wbi, dev->sector);
1911                         }
1912
1913                         if (head_sh->batch_head) {
1914                                 sh = list_first_entry(&sh->batch_list,
1915                                                       struct stripe_head,
1916                                                       batch_list);
1917                                 if (sh == head_sh)
1918                                         continue;
1919                                 goto again;
1920                         }
1921                 }
1922         }
1923
1924         return tx;
1925 }
1926
1927 static void ops_complete_reconstruct(void *stripe_head_ref)
1928 {
1929         struct stripe_head *sh = stripe_head_ref;
1930         int disks = sh->disks;
1931         int pd_idx = sh->pd_idx;
1932         int qd_idx = sh->qd_idx;
1933         int i;
1934         bool fua = false, sync = false, discard = false;
1935
1936         pr_debug("%s: stripe %llu\n", __func__,
1937                 (unsigned long long)sh->sector);
1938
1939         for (i = disks; i--; ) {
1940                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1941                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1942                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1943         }
1944
1945         for (i = disks; i--; ) {
1946                 struct r5dev *dev = &sh->dev[i];
1947
1948                 if (dev->written || i == pd_idx || i == qd_idx) {
1949                         if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1950                                 set_bit(R5_UPTODATE, &dev->flags);
1951                                 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1952                                         set_bit(R5_Expanded, &dev->flags);
1953                         }
1954                         if (fua)
1955                                 set_bit(R5_WantFUA, &dev->flags);
1956                         if (sync)
1957                                 set_bit(R5_SyncIO, &dev->flags);
1958                 }
1959         }
1960
1961         if (sh->reconstruct_state == reconstruct_state_drain_run)
1962                 sh->reconstruct_state = reconstruct_state_drain_result;
1963         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1964                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1965         else {
1966                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1967                 sh->reconstruct_state = reconstruct_state_result;
1968         }
1969
1970         set_bit(STRIPE_HANDLE, &sh->state);
1971         raid5_release_stripe(sh);
1972 }
1973
1974 static void
1975 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1976                      struct dma_async_tx_descriptor *tx)
1977 {
1978         int disks = sh->disks;
1979         struct page **xor_srcs;
1980         unsigned int *off_srcs;
1981         struct async_submit_ctl submit;
1982         int count, pd_idx = sh->pd_idx, i;
1983         struct page *xor_dest;
1984         unsigned int off_dest;
1985         int prexor = 0;
1986         unsigned long flags;
1987         int j = 0;
1988         struct stripe_head *head_sh = sh;
1989         int last_stripe;
1990
1991         pr_debug("%s: stripe %llu\n", __func__,
1992                 (unsigned long long)sh->sector);
1993
1994         for (i = 0; i < sh->disks; i++) {
1995                 if (pd_idx == i)
1996                         continue;
1997                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1998                         break;
1999         }
2000         if (i >= sh->disks) {
2001                 atomic_inc(&sh->count);
2002                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
2003                 ops_complete_reconstruct(sh);
2004                 return;
2005         }
2006 again:
2007         count = 0;
2008         xor_srcs = to_addr_page(percpu, j);
2009         off_srcs = to_addr_offs(sh, percpu);
2010         /* check if prexor is active which means only process blocks
2011          * that are part of a read-modify-write (written)
2012          */
2013         if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2014                 prexor = 1;
2015                 off_dest = off_srcs[count] = sh->dev[pd_idx].offset;
2016                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
2017                 for (i = disks; i--; ) {
2018                         struct r5dev *dev = &sh->dev[i];
2019                         if (head_sh->dev[i].written ||
2020                             test_bit(R5_InJournal, &head_sh->dev[i].flags)) {
2021                                 off_srcs[count] = dev->offset;
2022                                 xor_srcs[count++] = dev->page;
2023                         }
2024                 }
2025         } else {
2026                 xor_dest = sh->dev[pd_idx].page;
2027                 off_dest = sh->dev[pd_idx].offset;
2028                 for (i = disks; i--; ) {
2029                         struct r5dev *dev = &sh->dev[i];
2030                         if (i != pd_idx) {
2031                                 off_srcs[count] = dev->offset;
2032                                 xor_srcs[count++] = dev->page;
2033                         }
2034                 }
2035         }
2036
2037         /* 1/ if we prexor'd then the dest is reused as a source
2038          * 2/ if we did not prexor then we are redoing the parity
2039          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
2040          * for the synchronous xor case
2041          */
2042         last_stripe = !head_sh->batch_head ||
2043                 list_first_entry(&sh->batch_list,
2044                                  struct stripe_head, batch_list) == head_sh;
2045         if (last_stripe) {
2046                 flags = ASYNC_TX_ACK |
2047                         (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
2048
2049                 atomic_inc(&head_sh->count);
2050                 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
2051                                   to_addr_conv(sh, percpu, j));
2052         } else {
2053                 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
2054                 init_async_submit(&submit, flags, tx, NULL, NULL,
2055                                   to_addr_conv(sh, percpu, j));
2056         }
2057
2058         if (unlikely(count == 1))
2059                 tx = async_memcpy(xor_dest, xor_srcs[0], off_dest, off_srcs[0],
2060                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2061         else
2062                 tx = async_xor_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2063                                 RAID5_STRIPE_SIZE(sh->raid_conf), &submit);
2064         if (!last_stripe) {
2065                 j++;
2066                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2067                                       batch_list);
2068                 goto again;
2069         }
2070 }
2071
2072 static void
2073 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
2074                      struct dma_async_tx_descriptor *tx)
2075 {
2076         struct async_submit_ctl submit;
2077         struct page **blocks;
2078         unsigned int *offs;
2079         int count, i, j = 0;
2080         struct stripe_head *head_sh = sh;
2081         int last_stripe;
2082         int synflags;
2083         unsigned long txflags;
2084
2085         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
2086
2087         for (i = 0; i < sh->disks; i++) {
2088                 if (sh->pd_idx == i || sh->qd_idx == i)
2089                         continue;
2090                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
2091                         break;
2092         }
2093         if (i >= sh->disks) {
2094                 atomic_inc(&sh->count);
2095                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2096                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2097                 ops_complete_reconstruct(sh);
2098                 return;
2099         }
2100
2101 again:
2102         blocks = to_addr_page(percpu, j);
2103         offs = to_addr_offs(sh, percpu);
2104
2105         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
2106                 synflags = SYNDROME_SRC_WRITTEN;
2107                 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
2108         } else {
2109                 synflags = SYNDROME_SRC_ALL;
2110                 txflags = ASYNC_TX_ACK;
2111         }
2112
2113         count = set_syndrome_sources(blocks, offs, sh, synflags);
2114         last_stripe = !head_sh->batch_head ||
2115                 list_first_entry(&sh->batch_list,
2116                                  struct stripe_head, batch_list) == head_sh;
2117
2118         if (last_stripe) {
2119                 atomic_inc(&head_sh->count);
2120                 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
2121                                   head_sh, to_addr_conv(sh, percpu, j));
2122         } else
2123                 init_async_submit(&submit, 0, tx, NULL, NULL,
2124                                   to_addr_conv(sh, percpu, j));
2125         tx = async_gen_syndrome(blocks, offs, count+2,
2126                         RAID5_STRIPE_SIZE(sh->raid_conf),  &submit);
2127         if (!last_stripe) {
2128                 j++;
2129                 sh = list_first_entry(&sh->batch_list, struct stripe_head,
2130                                       batch_list);
2131                 goto again;
2132         }
2133 }
2134
2135 static void ops_complete_check(void *stripe_head_ref)
2136 {
2137         struct stripe_head *sh = stripe_head_ref;
2138
2139         pr_debug("%s: stripe %llu\n", __func__,
2140                 (unsigned long long)sh->sector);
2141
2142         sh->check_state = check_state_check_result;
2143         set_bit(STRIPE_HANDLE, &sh->state);
2144         raid5_release_stripe(sh);
2145 }
2146
2147 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2148 {
2149         int disks = sh->disks;
2150         int pd_idx = sh->pd_idx;
2151         int qd_idx = sh->qd_idx;
2152         struct page *xor_dest;
2153         unsigned int off_dest;
2154         struct page **xor_srcs = to_addr_page(percpu, 0);
2155         unsigned int *off_srcs = to_addr_offs(sh, percpu);
2156         struct dma_async_tx_descriptor *tx;
2157         struct async_submit_ctl submit;
2158         int count;
2159         int i;
2160
2161         pr_debug("%s: stripe %llu\n", __func__,
2162                 (unsigned long long)sh->sector);
2163
2164         BUG_ON(sh->batch_head);
2165         count = 0;
2166         xor_dest = sh->dev[pd_idx].page;
2167         off_dest = sh->dev[pd_idx].offset;
2168         off_srcs[count] = off_dest;
2169         xor_srcs[count++] = xor_dest;
2170         for (i = disks; i--; ) {
2171                 if (i == pd_idx || i == qd_idx)
2172                         continue;
2173                 off_srcs[count] = sh->dev[i].offset;
2174                 xor_srcs[count++] = sh->dev[i].page;
2175         }
2176
2177         init_async_submit(&submit, 0, NULL, NULL, NULL,
2178                           to_addr_conv(sh, percpu, 0));
2179         tx = async_xor_val_offs(xor_dest, off_dest, xor_srcs, off_srcs, count,
2180                            RAID5_STRIPE_SIZE(sh->raid_conf),
2181                            &sh->ops.zero_sum_result, &submit);
2182
2183         atomic_inc(&sh->count);
2184         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2185         tx = async_trigger_callback(&submit);
2186 }
2187
2188 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2189 {
2190         struct page **srcs = to_addr_page(percpu, 0);
2191         unsigned int *offs = to_addr_offs(sh, percpu);
2192         struct async_submit_ctl submit;
2193         int count;
2194
2195         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2196                 (unsigned long long)sh->sector, checkp);
2197
2198         BUG_ON(sh->batch_head);
2199         count = set_syndrome_sources(srcs, offs, sh, SYNDROME_SRC_ALL);
2200         if (!checkp)
2201                 srcs[count] = NULL;
2202
2203         atomic_inc(&sh->count);
2204         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2205                           sh, to_addr_conv(sh, percpu, 0));
2206         async_syndrome_val(srcs, offs, count+2,
2207                            RAID5_STRIPE_SIZE(sh->raid_conf),
2208                            &sh->ops.zero_sum_result, percpu->spare_page, 0, &submit);
2209 }
2210
2211 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2212 {
2213         int overlap_clear = 0, i, disks = sh->disks;
2214         struct dma_async_tx_descriptor *tx = NULL;
2215         struct r5conf *conf = sh->raid_conf;
2216         int level = conf->level;
2217         struct raid5_percpu *percpu;
2218         unsigned long cpu;
2219
2220         cpu = get_cpu();
2221         percpu = per_cpu_ptr(conf->percpu, cpu);
2222         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2223                 ops_run_biofill(sh);
2224                 overlap_clear++;
2225         }
2226
2227         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2228                 if (level < 6)
2229                         tx = ops_run_compute5(sh, percpu);
2230                 else {
2231                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
2232                                 tx = ops_run_compute6_1(sh, percpu);
2233                         else
2234                                 tx = ops_run_compute6_2(sh, percpu);
2235                 }
2236                 /* terminate the chain if reconstruct is not set to be run */
2237                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2238                         async_tx_ack(tx);
2239         }
2240
2241         if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2242                 if (level < 6)
2243                         tx = ops_run_prexor5(sh, percpu, tx);
2244                 else
2245                         tx = ops_run_prexor6(sh, percpu, tx);
2246         }
2247
2248         if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2249                 tx = ops_run_partial_parity(sh, percpu, tx);
2250
2251         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2252                 tx = ops_run_biodrain(sh, tx);
2253                 overlap_clear++;
2254         }
2255
2256         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2257                 if (level < 6)
2258                         ops_run_reconstruct5(sh, percpu, tx);
2259                 else
2260                         ops_run_reconstruct6(sh, percpu, tx);
2261         }
2262
2263         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2264                 if (sh->check_state == check_state_run)
2265                         ops_run_check_p(sh, percpu);
2266                 else if (sh->check_state == check_state_run_q)
2267                         ops_run_check_pq(sh, percpu, 0);
2268                 else if (sh->check_state == check_state_run_pq)
2269                         ops_run_check_pq(sh, percpu, 1);
2270                 else
2271                         BUG();
2272         }
2273
2274         if (overlap_clear && !sh->batch_head)
2275                 for (i = disks; i--; ) {
2276                         struct r5dev *dev = &sh->dev[i];
2277                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
2278                                 wake_up(&sh->raid_conf->wait_for_overlap);
2279                 }
2280         put_cpu();
2281 }
2282
2283 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2284 {
2285 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2286         kfree(sh->pages);
2287 #endif
2288         if (sh->ppl_page)
2289                 __free_page(sh->ppl_page);
2290         kmem_cache_free(sc, sh);
2291 }
2292
2293 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2294         int disks, struct r5conf *conf)
2295 {
2296         struct stripe_head *sh;
2297         int i;
2298
2299         sh = kmem_cache_zalloc(sc, gfp);
2300         if (sh) {
2301                 spin_lock_init(&sh->stripe_lock);
2302                 spin_lock_init(&sh->batch_lock);
2303                 INIT_LIST_HEAD(&sh->batch_list);
2304                 INIT_LIST_HEAD(&sh->lru);
2305                 INIT_LIST_HEAD(&sh->r5c);
2306                 INIT_LIST_HEAD(&sh->log_list);
2307                 atomic_set(&sh->count, 1);
2308                 sh->raid_conf = conf;
2309                 sh->log_start = MaxSector;
2310                 for (i = 0; i < disks; i++) {
2311                         struct r5dev *dev = &sh->dev[i];
2312
2313                         bio_init(&dev->req, &dev->vec, 1);
2314                         bio_init(&dev->rreq, &dev->rvec, 1);
2315                 }
2316
2317                 if (raid5_has_ppl(conf)) {
2318                         sh->ppl_page = alloc_page(gfp);
2319                         if (!sh->ppl_page) {
2320                                 free_stripe(sc, sh);
2321                                 return NULL;
2322                         }
2323                 }
2324 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2325                 if (init_stripe_shared_pages(sh, conf, disks)) {
2326                         free_stripe(sc, sh);
2327                         return NULL;
2328                 }
2329 #endif
2330         }
2331         return sh;
2332 }
2333 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2334 {
2335         struct stripe_head *sh;
2336
2337         sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2338         if (!sh)
2339                 return 0;
2340
2341         if (grow_buffers(sh, gfp)) {
2342                 shrink_buffers(sh);
2343                 free_stripe(conf->slab_cache, sh);
2344                 return 0;
2345         }
2346         sh->hash_lock_index =
2347                 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2348         /* we just created an active stripe so... */
2349         atomic_inc(&conf->active_stripes);
2350
2351         raid5_release_stripe(sh);
2352         conf->max_nr_stripes++;
2353         return 1;
2354 }
2355
2356 static int grow_stripes(struct r5conf *conf, int num)
2357 {
2358         struct kmem_cache *sc;
2359         size_t namelen = sizeof(conf->cache_name[0]);
2360         int devs = max(conf->raid_disks, conf->previous_raid_disks);
2361
2362         if (conf->mddev->gendisk)
2363                 snprintf(conf->cache_name[0], namelen,
2364                         "raid%d-%s", conf->level, mdname(conf->mddev));
2365         else
2366                 snprintf(conf->cache_name[0], namelen,
2367                         "raid%d-%p", conf->level, conf->mddev);
2368         snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2369
2370         conf->active_name = 0;
2371         sc = kmem_cache_create(conf->cache_name[conf->active_name],
2372                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2373                                0, 0, NULL);
2374         if (!sc)
2375                 return 1;
2376         conf->slab_cache = sc;
2377         conf->pool_size = devs;
2378         while (num--)
2379                 if (!grow_one_stripe(conf, GFP_KERNEL))
2380                         return 1;
2381
2382         return 0;
2383 }
2384
2385 /**
2386  * scribble_alloc - allocate percpu scribble buffer for required size
2387  *                  of the scribble region
2388  * @percpu: from for_each_present_cpu() of the caller
2389  * @num: total number of disks in the array
2390  * @cnt: scribble objs count for required size of the scribble region
2391  *
2392  * The scribble buffer size must be enough to contain:
2393  * 1/ a struct page pointer for each device in the array +2
2394  * 2/ room to convert each entry in (1) to its corresponding dma
2395  *    (dma_map_page()) or page (page_address()) address.
2396  *
2397  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2398  * calculate over all devices (not just the data blocks), using zeros in place
2399  * of the P and Q blocks.
2400  */
2401 static int scribble_alloc(struct raid5_percpu *percpu,
2402                           int num, int cnt)
2403 {
2404         size_t obj_size =
2405                 sizeof(struct page *) * (num + 2) +
2406                 sizeof(addr_conv_t) * (num + 2) +
2407                 sizeof(unsigned int) * (num + 2);
2408         void *scribble;
2409
2410         /*
2411          * If here is in raid array suspend context, it is in memalloc noio
2412          * context as well, there is no potential recursive memory reclaim
2413          * I/Os with the GFP_KERNEL flag.
2414          */
2415         scribble = kvmalloc_array(cnt, obj_size, GFP_KERNEL);
2416         if (!scribble)
2417                 return -ENOMEM;
2418
2419         kvfree(percpu->scribble);
2420
2421         percpu->scribble = scribble;
2422         percpu->scribble_obj_size = obj_size;
2423         return 0;
2424 }
2425
2426 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2427 {
2428         unsigned long cpu;
2429         int err = 0;
2430
2431         /*
2432          * Never shrink. And mddev_suspend() could deadlock if this is called
2433          * from raid5d. In that case, scribble_disks and scribble_sectors
2434          * should equal to new_disks and new_sectors
2435          */
2436         if (conf->scribble_disks >= new_disks &&
2437             conf->scribble_sectors >= new_sectors)
2438                 return 0;
2439         mddev_suspend(conf->mddev);
2440         cpus_read_lock();
2441
2442         for_each_present_cpu(cpu) {
2443                 struct raid5_percpu *percpu;
2444
2445                 percpu = per_cpu_ptr(conf->percpu, cpu);
2446                 err = scribble_alloc(percpu, new_disks,
2447                                      new_sectors / RAID5_STRIPE_SECTORS(conf));
2448                 if (err)
2449                         break;
2450         }
2451
2452         cpus_read_unlock();
2453         mddev_resume(conf->mddev);
2454         if (!err) {
2455                 conf->scribble_disks = new_disks;
2456                 conf->scribble_sectors = new_sectors;
2457         }
2458         return err;
2459 }
2460
2461 static int resize_stripes(struct r5conf *conf, int newsize)
2462 {
2463         /* Make all the stripes able to hold 'newsize' devices.
2464          * New slots in each stripe get 'page' set to a new page.
2465          *
2466          * This happens in stages:
2467          * 1/ create a new kmem_cache and allocate the required number of
2468          *    stripe_heads.
2469          * 2/ gather all the old stripe_heads and transfer the pages across
2470          *    to the new stripe_heads.  This will have the side effect of
2471          *    freezing the array as once all stripe_heads have been collected,
2472          *    no IO will be possible.  Old stripe heads are freed once their
2473          *    pages have been transferred over, and the old kmem_cache is
2474          *    freed when all stripes are done.
2475          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2476          *    we simple return a failure status - no need to clean anything up.
2477          * 4/ allocate new pages for the new slots in the new stripe_heads.
2478          *    If this fails, we don't bother trying the shrink the
2479          *    stripe_heads down again, we just leave them as they are.
2480          *    As each stripe_head is processed the new one is released into
2481          *    active service.
2482          *
2483          * Once step2 is started, we cannot afford to wait for a write,
2484          * so we use GFP_NOIO allocations.
2485          */
2486         struct stripe_head *osh, *nsh;
2487         LIST_HEAD(newstripes);
2488         struct disk_info *ndisks;
2489         int err = 0;
2490         struct kmem_cache *sc;
2491         int i;
2492         int hash, cnt;
2493
2494         md_allow_write(conf->mddev);
2495
2496         /* Step 1 */
2497         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2498                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2499                                0, 0, NULL);
2500         if (!sc)
2501                 return -ENOMEM;
2502
2503         /* Need to ensure auto-resizing doesn't interfere */
2504         mutex_lock(&conf->cache_size_mutex);
2505
2506         for (i = conf->max_nr_stripes; i; i--) {
2507                 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2508                 if (!nsh)
2509                         break;
2510
2511                 list_add(&nsh->lru, &newstripes);
2512         }
2513         if (i) {
2514                 /* didn't get enough, give up */
2515                 while (!list_empty(&newstripes)) {
2516                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
2517                         list_del(&nsh->lru);
2518                         free_stripe(sc, nsh);
2519                 }
2520                 kmem_cache_destroy(sc);
2521                 mutex_unlock(&conf->cache_size_mutex);
2522                 return -ENOMEM;
2523         }
2524         /* Step 2 - Must use GFP_NOIO now.
2525          * OK, we have enough stripes, start collecting inactive
2526          * stripes and copying them over
2527          */
2528         hash = 0;
2529         cnt = 0;
2530         list_for_each_entry(nsh, &newstripes, lru) {
2531                 lock_device_hash_lock(conf, hash);
2532                 wait_event_cmd(conf->wait_for_stripe,
2533                                     !list_empty(conf->inactive_list + hash),
2534                                     unlock_device_hash_lock(conf, hash),
2535                                     lock_device_hash_lock(conf, hash));
2536                 osh = get_free_stripe(conf, hash);
2537                 unlock_device_hash_lock(conf, hash);
2538
2539 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2540         for (i = 0; i < osh->nr_pages; i++) {
2541                 nsh->pages[i] = osh->pages[i];
2542                 osh->pages[i] = NULL;
2543         }
2544 #endif
2545                 for(i=0; i<conf->pool_size; i++) {
2546                         nsh->dev[i].page = osh->dev[i].page;
2547                         nsh->dev[i].orig_page = osh->dev[i].page;
2548                         nsh->dev[i].offset = osh->dev[i].offset;
2549                 }
2550                 nsh->hash_lock_index = hash;
2551                 free_stripe(conf->slab_cache, osh);
2552                 cnt++;
2553                 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2554                     !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2555                         hash++;
2556                         cnt = 0;
2557                 }
2558         }
2559         kmem_cache_destroy(conf->slab_cache);
2560
2561         /* Step 3.
2562          * At this point, we are holding all the stripes so the array
2563          * is completely stalled, so now is a good time to resize
2564          * conf->disks and the scribble region
2565          */
2566         ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2567         if (ndisks) {
2568                 for (i = 0; i < conf->pool_size; i++)
2569                         ndisks[i] = conf->disks[i];
2570
2571                 for (i = conf->pool_size; i < newsize; i++) {
2572                         ndisks[i].extra_page = alloc_page(GFP_NOIO);
2573                         if (!ndisks[i].extra_page)
2574                                 err = -ENOMEM;
2575                 }
2576
2577                 if (err) {
2578                         for (i = conf->pool_size; i < newsize; i++)
2579                                 if (ndisks[i].extra_page)
2580                                         put_page(ndisks[i].extra_page);
2581                         kfree(ndisks);
2582                 } else {
2583                         kfree(conf->disks);
2584                         conf->disks = ndisks;
2585                 }
2586         } else
2587                 err = -ENOMEM;
2588
2589         conf->slab_cache = sc;
2590         conf->active_name = 1-conf->active_name;
2591
2592         /* Step 4, return new stripes to service */
2593         while(!list_empty(&newstripes)) {
2594                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2595                 list_del_init(&nsh->lru);
2596
2597 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
2598                 for (i = 0; i < nsh->nr_pages; i++) {
2599                         if (nsh->pages[i])
2600                                 continue;
2601                         nsh->pages[i] = alloc_page(GFP_NOIO);
2602                         if (!nsh->pages[i])
2603                                 err = -ENOMEM;
2604                 }
2605
2606                 for (i = conf->raid_disks; i < newsize; i++) {
2607                         if (nsh->dev[i].page)
2608                                 continue;
2609                         nsh->dev[i].page = raid5_get_dev_page(nsh, i);
2610                         nsh->dev[i].orig_page = nsh->dev[i].page;
2611                         nsh->dev[i].offset = raid5_get_page_offset(nsh, i);
2612                 }
2613 #else
2614                 for (i=conf->raid_disks; i < newsize; i++)
2615                         if (nsh->dev[i].page == NULL) {
2616                                 struct page *p = alloc_page(GFP_NOIO);
2617                                 nsh->dev[i].page = p;
2618                                 nsh->dev[i].orig_page = p;
2619                                 nsh->dev[i].offset = 0;
2620                                 if (!p)
2621                                         err = -ENOMEM;
2622                         }
2623 #endif
2624                 raid5_release_stripe(nsh);
2625         }
2626         /* critical section pass, GFP_NOIO no longer needed */
2627
2628         if (!err)
2629                 conf->pool_size = newsize;
2630         mutex_unlock(&conf->cache_size_mutex);
2631
2632         return err;
2633 }
2634
2635 static int drop_one_stripe(struct r5conf *conf)
2636 {
2637         struct stripe_head *sh;
2638         int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2639
2640         spin_lock_irq(conf->hash_locks + hash);
2641         sh = get_free_stripe(conf, hash);
2642         spin_unlock_irq(conf->hash_locks + hash);
2643         if (!sh)
2644                 return 0;
2645         BUG_ON(atomic_read(&sh->count));
2646         shrink_buffers(sh);
2647         free_stripe(conf->slab_cache, sh);
2648         atomic_dec(&conf->active_stripes);
2649         conf->max_nr_stripes--;
2650         return 1;
2651 }
2652
2653 static void shrink_stripes(struct r5conf *conf)
2654 {
2655         while (conf->max_nr_stripes &&
2656                drop_one_stripe(conf))
2657                 ;
2658
2659         kmem_cache_destroy(conf->slab_cache);
2660         conf->slab_cache = NULL;
2661 }
2662
2663 static void raid5_end_read_request(struct bio * bi)
2664 {
2665         struct stripe_head *sh = bi->bi_private;
2666         struct r5conf *conf = sh->raid_conf;
2667         int disks = sh->disks, i;
2668         char b[BDEVNAME_SIZE];
2669         struct md_rdev *rdev = NULL;
2670         sector_t s;
2671
2672         for (i=0 ; i<disks; i++)
2673                 if (bi == &sh->dev[i].req)
2674                         break;
2675
2676         pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2677                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2678                 bi->bi_status);
2679         if (i == disks) {
2680                 bio_reset(bi);
2681                 BUG();
2682                 return;
2683         }
2684         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2685                 /* If replacement finished while this request was outstanding,
2686                  * 'replacement' might be NULL already.
2687                  * In that case it moved down to 'rdev'.
2688                  * rdev is not removed until all requests are finished.
2689                  */
2690                 rdev = conf->disks[i].replacement;
2691         if (!rdev)
2692                 rdev = conf->disks[i].rdev;
2693
2694         if (use_new_offset(conf, sh))
2695                 s = sh->sector + rdev->new_data_offset;
2696         else
2697                 s = sh->sector + rdev->data_offset;
2698         if (!bi->bi_status) {
2699                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2700                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2701                         /* Note that this cannot happen on a
2702                          * replacement device.  We just fail those on
2703                          * any error
2704                          */
2705                         pr_info_ratelimited(
2706                                 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2707                                 mdname(conf->mddev), RAID5_STRIPE_SECTORS(conf),
2708                                 (unsigned long long)s,
2709                                 bdevname(rdev->bdev, b));
2710                         atomic_add(RAID5_STRIPE_SECTORS(conf), &rdev->corrected_errors);
2711                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2712                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2713                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2714                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2715
2716                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2717                         /*
2718                          * end read for a page in journal, this
2719                          * must be preparing for prexor in rmw
2720                          */
2721                         set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2722
2723                 if (atomic_read(&rdev->read_errors))
2724                         atomic_set(&rdev->read_errors, 0);
2725         } else {
2726                 const char *bdn = bdevname(rdev->bdev, b);
2727                 int retry = 0;
2728                 int set_bad = 0;
2729
2730                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2731                 if (!(bi->bi_status == BLK_STS_PROTECTION))
2732                         atomic_inc(&rdev->read_errors);
2733                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2734                         pr_warn_ratelimited(
2735                                 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2736                                 mdname(conf->mddev),
2737                                 (unsigned long long)s,
2738                                 bdn);
2739                 else if (conf->mddev->degraded >= conf->max_degraded) {
2740                         set_bad = 1;
2741                         pr_warn_ratelimited(
2742                                 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2743                                 mdname(conf->mddev),
2744                                 (unsigned long long)s,
2745                                 bdn);
2746                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2747                         /* Oh, no!!! */
2748                         set_bad = 1;
2749                         pr_warn_ratelimited(
2750                                 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2751                                 mdname(conf->mddev),
2752                                 (unsigned long long)s,
2753                                 bdn);
2754                 } else if (atomic_read(&rdev->read_errors)
2755                          > conf->max_nr_stripes) {
2756                         if (!test_bit(Faulty, &rdev->flags)) {
2757                                 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2758                                     mdname(conf->mddev),
2759                                     atomic_read(&rdev->read_errors),
2760                                     conf->max_nr_stripes);
2761                                 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2762                                     mdname(conf->mddev), bdn);
2763                         }
2764                 } else
2765                         retry = 1;
2766                 if (set_bad && test_bit(In_sync, &rdev->flags)
2767                     && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2768                         retry = 1;
2769                 if (retry)
2770                         if (sh->qd_idx >= 0 && sh->pd_idx == i)
2771                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2772                         else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2773                                 set_bit(R5_ReadError, &sh->dev[i].flags);
2774                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2775                         } else
2776                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2777                 else {
2778                         clear_bit(R5_ReadError, &sh->dev[i].flags);
2779                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
2780                         if (!(set_bad
2781                               && test_bit(In_sync, &rdev->flags)
2782                               && rdev_set_badblocks(
2783                                       rdev, sh->sector, RAID5_STRIPE_SECTORS(conf), 0)))
2784                                 md_error(conf->mddev, rdev);
2785                 }
2786         }
2787         rdev_dec_pending(rdev, conf->mddev);
2788         bio_reset(bi);
2789         clear_bit(R5_LOCKED, &sh->dev[i].flags);
2790         set_bit(STRIPE_HANDLE, &sh->state);
2791         raid5_release_stripe(sh);
2792 }
2793
2794 static void raid5_end_write_request(struct bio *bi)
2795 {
2796         struct stripe_head *sh = bi->bi_private;
2797         struct r5conf *conf = sh->raid_conf;
2798         int disks = sh->disks, i;
2799         struct md_rdev *rdev;
2800         sector_t first_bad;
2801         int bad_sectors;
2802         int replacement = 0;
2803
2804         for (i = 0 ; i < disks; i++) {
2805                 if (bi == &sh->dev[i].req) {
2806                         rdev = conf->disks[i].rdev;
2807                         break;
2808                 }
2809                 if (bi == &sh->dev[i].rreq) {
2810                         rdev = conf->disks[i].replacement;
2811                         if (rdev)
2812                                 replacement = 1;
2813                         else
2814                                 /* rdev was removed and 'replacement'
2815                                  * replaced it.  rdev is not removed
2816                                  * until all requests are finished.
2817                                  */
2818                                 rdev = conf->disks[i].rdev;
2819                         break;
2820                 }
2821         }
2822         pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2823                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2824                 bi->bi_status);
2825         if (i == disks) {
2826                 bio_reset(bi);
2827                 BUG();
2828                 return;
2829         }
2830
2831         if (replacement) {
2832                 if (bi->bi_status)
2833                         md_error(conf->mddev, rdev);
2834                 else if (is_badblock(rdev, sh->sector,
2835                                      RAID5_STRIPE_SECTORS(conf),
2836                                      &first_bad, &bad_sectors))
2837                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2838         } else {
2839                 if (bi->bi_status) {
2840                         set_bit(STRIPE_DEGRADED, &sh->state);
2841                         set_bit(WriteErrorSeen, &rdev->flags);
2842                         set_bit(R5_WriteError, &sh->dev[i].flags);
2843                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2844                                 set_bit(MD_RECOVERY_NEEDED,
2845                                         &rdev->mddev->recovery);
2846                 } else if (is_badblock(rdev, sh->sector,
2847                                        RAID5_STRIPE_SECTORS(conf),
2848                                        &first_bad, &bad_sectors)) {
2849                         set_bit(R5_MadeGood, &sh->dev[i].flags);
2850                         if (test_bit(R5_ReadError, &sh->dev[i].flags))
2851                                 /* That was a successful write so make
2852                                  * sure it looks like we already did
2853                                  * a re-write.
2854                                  */
2855                                 set_bit(R5_ReWrite, &sh->dev[i].flags);
2856                 }
2857         }
2858         rdev_dec_pending(rdev, conf->mddev);
2859
2860         if (sh->batch_head && bi->bi_status && !replacement)
2861                 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2862
2863         bio_reset(bi);
2864         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2865                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2866         set_bit(STRIPE_HANDLE, &sh->state);
2867         raid5_release_stripe(sh);
2868
2869         if (sh->batch_head && sh != sh->batch_head)
2870                 raid5_release_stripe(sh->batch_head);
2871 }
2872
2873 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2874 {
2875         char b[BDEVNAME_SIZE];
2876         struct r5conf *conf = mddev->private;
2877         unsigned long flags;
2878         pr_debug("raid456: error called\n");
2879
2880         spin_lock_irqsave(&conf->device_lock, flags);
2881
2882         if (test_bit(In_sync, &rdev->flags) &&
2883             mddev->degraded == conf->max_degraded) {
2884                 /*
2885                  * Don't allow to achieve failed state
2886                  * Don't try to recover this device
2887                  */
2888                 conf->recovery_disabled = mddev->recovery_disabled;
2889                 spin_unlock_irqrestore(&conf->device_lock, flags);
2890                 return;
2891         }
2892
2893         set_bit(Faulty, &rdev->flags);
2894         clear_bit(In_sync, &rdev->flags);
2895         mddev->degraded = raid5_calc_degraded(conf);
2896         spin_unlock_irqrestore(&conf->device_lock, flags);
2897         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2898
2899         set_bit(Blocked, &rdev->flags);
2900         set_mask_bits(&mddev->sb_flags, 0,
2901                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2902         pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2903                 "md/raid:%s: Operation continuing on %d devices.\n",
2904                 mdname(mddev),
2905                 bdevname(rdev->bdev, b),
2906                 mdname(mddev),
2907                 conf->raid_disks - mddev->degraded);
2908         r5c_update_on_rdev_error(mddev, rdev);
2909 }
2910
2911 /*
2912  * Input: a 'big' sector number,
2913  * Output: index of the data and parity disk, and the sector # in them.
2914  */
2915 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2916                               int previous, int *dd_idx,
2917                               struct stripe_head *sh)
2918 {
2919         sector_t stripe, stripe2;
2920         sector_t chunk_number;
2921         unsigned int chunk_offset;
2922         int pd_idx, qd_idx;
2923         int ddf_layout = 0;
2924         sector_t new_sector;
2925         int algorithm = previous ? conf->prev_algo
2926                                  : conf->algorithm;
2927         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2928                                          : conf->chunk_sectors;
2929         int raid_disks = previous ? conf->previous_raid_disks
2930                                   : conf->raid_disks;
2931         int data_disks = raid_disks - conf->max_degraded;
2932
2933         /* First compute the information on this sector */
2934
2935         /*
2936          * Compute the chunk number and the sector offset inside the chunk
2937          */
2938         chunk_offset = sector_div(r_sector, sectors_per_chunk);
2939         chunk_number = r_sector;
2940
2941         /*
2942          * Compute the stripe number
2943          */
2944         stripe = chunk_number;
2945         *dd_idx = sector_div(stripe, data_disks);
2946         stripe2 = stripe;
2947         /*
2948          * Select the parity disk based on the user selected algorithm.
2949          */
2950         pd_idx = qd_idx = -1;
2951         switch(conf->level) {
2952         case 4:
2953                 pd_idx = data_disks;
2954                 break;
2955         case 5:
2956                 switch (algorithm) {
2957                 case ALGORITHM_LEFT_ASYMMETRIC:
2958                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2959                         if (*dd_idx >= pd_idx)
2960                                 (*dd_idx)++;
2961                         break;
2962                 case ALGORITHM_RIGHT_ASYMMETRIC:
2963                         pd_idx = sector_div(stripe2, raid_disks);
2964                         if (*dd_idx >= pd_idx)
2965                                 (*dd_idx)++;
2966                         break;
2967                 case ALGORITHM_LEFT_SYMMETRIC:
2968                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2969                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2970                         break;
2971                 case ALGORITHM_RIGHT_SYMMETRIC:
2972                         pd_idx = sector_div(stripe2, raid_disks);
2973                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2974                         break;
2975                 case ALGORITHM_PARITY_0:
2976                         pd_idx = 0;
2977                         (*dd_idx)++;
2978                         break;
2979                 case ALGORITHM_PARITY_N:
2980                         pd_idx = data_disks;
2981                         break;
2982                 default:
2983                         BUG();
2984                 }
2985                 break;
2986         case 6:
2987
2988                 switch (algorithm) {
2989                 case ALGORITHM_LEFT_ASYMMETRIC:
2990                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2991                         qd_idx = pd_idx + 1;
2992                         if (pd_idx == raid_disks-1) {
2993                                 (*dd_idx)++;    /* Q D D D P */
2994                                 qd_idx = 0;
2995                         } else if (*dd_idx >= pd_idx)
2996                                 (*dd_idx) += 2; /* D D P Q D */
2997                         break;
2998                 case ALGORITHM_RIGHT_ASYMMETRIC:
2999                         pd_idx = sector_div(stripe2, raid_disks);
3000                         qd_idx = pd_idx + 1;
3001                         if (pd_idx == raid_disks-1) {
3002                                 (*dd_idx)++;    /* Q D D D P */
3003                                 qd_idx = 0;
3004                         } else if (*dd_idx >= pd_idx)
3005                                 (*dd_idx) += 2; /* D D P Q D */
3006                         break;
3007                 case ALGORITHM_LEFT_SYMMETRIC:
3008                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3009                         qd_idx = (pd_idx + 1) % raid_disks;
3010                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3011                         break;
3012                 case ALGORITHM_RIGHT_SYMMETRIC:
3013                         pd_idx = sector_div(stripe2, raid_disks);
3014                         qd_idx = (pd_idx + 1) % raid_disks;
3015                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
3016                         break;
3017
3018                 case ALGORITHM_PARITY_0:
3019                         pd_idx = 0;
3020                         qd_idx = 1;
3021                         (*dd_idx) += 2;
3022                         break;
3023                 case ALGORITHM_PARITY_N:
3024                         pd_idx = data_disks;
3025                         qd_idx = data_disks + 1;
3026                         break;
3027
3028                 case ALGORITHM_ROTATING_ZERO_RESTART:
3029                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
3030                          * of blocks for computing Q is different.
3031                          */
3032                         pd_idx = sector_div(stripe2, raid_disks);
3033                         qd_idx = pd_idx + 1;
3034                         if (pd_idx == raid_disks-1) {
3035                                 (*dd_idx)++;    /* Q D D D P */
3036                                 qd_idx = 0;
3037                         } else if (*dd_idx >= pd_idx)
3038                                 (*dd_idx) += 2; /* D D P Q D */
3039                         ddf_layout = 1;
3040                         break;
3041
3042                 case ALGORITHM_ROTATING_N_RESTART:
3043                         /* Same a left_asymmetric, by first stripe is
3044                          * D D D P Q  rather than
3045                          * Q D D D P
3046                          */
3047                         stripe2 += 1;
3048                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3049                         qd_idx = pd_idx + 1;
3050                         if (pd_idx == raid_disks-1) {
3051                                 (*dd_idx)++;    /* Q D D D P */
3052                                 qd_idx = 0;
3053                         } else if (*dd_idx >= pd_idx)
3054                                 (*dd_idx) += 2; /* D D P Q D */
3055                         ddf_layout = 1;
3056                         break;
3057
3058                 case ALGORITHM_ROTATING_N_CONTINUE:
3059                         /* Same as left_symmetric but Q is before P */
3060                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
3061                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
3062                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
3063                         ddf_layout = 1;
3064                         break;
3065
3066                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3067                         /* RAID5 left_asymmetric, with Q on last device */
3068                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3069                         if (*dd_idx >= pd_idx)
3070                                 (*dd_idx)++;
3071                         qd_idx = raid_disks - 1;
3072                         break;
3073
3074                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3075                         pd_idx = sector_div(stripe2, raid_disks-1);
3076                         if (*dd_idx >= pd_idx)
3077                                 (*dd_idx)++;
3078                         qd_idx = raid_disks - 1;
3079                         break;
3080
3081                 case ALGORITHM_LEFT_SYMMETRIC_6:
3082                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
3083                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3084                         qd_idx = raid_disks - 1;
3085                         break;
3086
3087                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3088                         pd_idx = sector_div(stripe2, raid_disks-1);
3089                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
3090                         qd_idx = raid_disks - 1;
3091                         break;
3092
3093                 case ALGORITHM_PARITY_0_6:
3094                         pd_idx = 0;
3095                         (*dd_idx)++;
3096                         qd_idx = raid_disks - 1;
3097                         break;
3098
3099                 default:
3100                         BUG();
3101                 }
3102                 break;
3103         }
3104
3105         if (sh) {
3106                 sh->pd_idx = pd_idx;
3107                 sh->qd_idx = qd_idx;
3108                 sh->ddf_layout = ddf_layout;
3109         }
3110         /*
3111          * Finally, compute the new sector number
3112          */
3113         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
3114         return new_sector;
3115 }
3116
3117 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
3118 {
3119         struct r5conf *conf = sh->raid_conf;
3120         int raid_disks = sh->disks;
3121         int data_disks = raid_disks - conf->max_degraded;
3122         sector_t new_sector = sh->sector, check;
3123         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
3124                                          : conf->chunk_sectors;
3125         int algorithm = previous ? conf->prev_algo
3126                                  : conf->algorithm;
3127         sector_t stripe;
3128         int chunk_offset;
3129         sector_t chunk_number;
3130         int dummy1, dd_idx = i;
3131         sector_t r_sector;
3132         struct stripe_head sh2;
3133
3134         chunk_offset = sector_div(new_sector, sectors_per_chunk);
3135         stripe = new_sector;
3136
3137         if (i == sh->pd_idx)
3138                 return 0;
3139         switch(conf->level) {
3140         case 4: break;
3141         case 5:
3142                 switch (algorithm) {
3143                 case ALGORITHM_LEFT_ASYMMETRIC:
3144                 case ALGORITHM_RIGHT_ASYMMETRIC:
3145                         if (i > sh->pd_idx)
3146                                 i--;
3147                         break;
3148                 case ALGORITHM_LEFT_SYMMETRIC:
3149                 case ALGORITHM_RIGHT_SYMMETRIC:
3150                         if (i < sh->pd_idx)
3151                                 i += raid_disks;
3152                         i -= (sh->pd_idx + 1);
3153                         break;
3154                 case ALGORITHM_PARITY_0:
3155                         i -= 1;
3156                         break;
3157                 case ALGORITHM_PARITY_N:
3158                         break;
3159                 default:
3160                         BUG();
3161                 }
3162                 break;
3163         case 6:
3164                 if (i == sh->qd_idx)
3165                         return 0; /* It is the Q disk */
3166                 switch (algorithm) {
3167                 case ALGORITHM_LEFT_ASYMMETRIC:
3168                 case ALGORITHM_RIGHT_ASYMMETRIC:
3169                 case ALGORITHM_ROTATING_ZERO_RESTART:
3170                 case ALGORITHM_ROTATING_N_RESTART:
3171                         if (sh->pd_idx == raid_disks-1)
3172                                 i--;    /* Q D D D P */
3173                         else if (i > sh->pd_idx)
3174                                 i -= 2; /* D D P Q D */
3175                         break;
3176                 case ALGORITHM_LEFT_SYMMETRIC:
3177                 case ALGORITHM_RIGHT_SYMMETRIC:
3178                         if (sh->pd_idx == raid_disks-1)
3179                                 i--; /* Q D D D P */
3180                         else {
3181                                 /* D D P Q D */
3182                                 if (i < sh->pd_idx)
3183                                         i += raid_disks;
3184                                 i -= (sh->pd_idx + 2);
3185                         }
3186                         break;
3187                 case ALGORITHM_PARITY_0:
3188                         i -= 2;
3189                         break;
3190                 case ALGORITHM_PARITY_N:
3191                         break;
3192                 case ALGORITHM_ROTATING_N_CONTINUE:
3193                         /* Like left_symmetric, but P is before Q */
3194                         if (sh->pd_idx == 0)
3195                                 i--;    /* P D D D Q */
3196                         else {
3197                                 /* D D Q P D */
3198                                 if (i < sh->pd_idx)
3199                                         i += raid_disks;
3200                                 i -= (sh->pd_idx + 1);
3201                         }
3202                         break;
3203                 case ALGORITHM_LEFT_ASYMMETRIC_6:
3204                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3205                         if (i > sh->pd_idx)
3206                                 i--;
3207                         break;
3208                 case ALGORITHM_LEFT_SYMMETRIC_6:
3209                 case ALGORITHM_RIGHT_SYMMETRIC_6:
3210                         if (i < sh->pd_idx)
3211                                 i += data_disks + 1;
3212                         i -= (sh->pd_idx + 1);
3213                         break;
3214                 case ALGORITHM_PARITY_0_6:
3215                         i -= 1;
3216                         break;
3217                 default:
3218                         BUG();
3219                 }
3220                 break;
3221         }
3222
3223         chunk_number = stripe * data_disks + i;
3224         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3225
3226         check = raid5_compute_sector(conf, r_sector,
3227                                      previous, &dummy1, &sh2);
3228         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3229                 || sh2.qd_idx != sh->qd_idx) {
3230                 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3231                         mdname(conf->mddev));
3232                 return 0;
3233         }
3234         return r_sector;
3235 }
3236
3237 /*
3238  * There are cases where we want handle_stripe_dirtying() and
3239  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3240  *
3241  * This function checks whether we want to delay the towrite. Specifically,
3242  * we delay the towrite when:
3243  *
3244  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3245  *      stripe has data in journal (for other devices).
3246  *
3247  *      In this case, when reading data for the non-overwrite dev, it is
3248  *      necessary to handle complex rmw of write back cache (prexor with
3249  *      orig_page, and xor with page). To keep read path simple, we would
3250  *      like to flush data in journal to RAID disks first, so complex rmw
3251  *      is handled in the write patch (handle_stripe_dirtying).
3252  *
3253  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3254  *
3255  *      It is important to be able to flush all stripes in raid5-cache.
3256  *      Therefore, we need reserve some space on the journal device for
3257  *      these flushes. If flush operation includes pending writes to the
3258  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3259  *      for the flush out. If we exclude these pending writes from flush
3260  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3261  *      Therefore, excluding pending writes in these cases enables more
3262  *      efficient use of the journal device.
3263  *
3264  *      Note: To make sure the stripe makes progress, we only delay
3265  *      towrite for stripes with data already in journal (injournal > 0).
3266  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3267  *      no_space_stripes list.
3268  *
3269  *   3. during journal failure
3270  *      In journal failure, we try to flush all cached data to raid disks
3271  *      based on data in stripe cache. The array is read-only to upper
3272  *      layers, so we would skip all pending writes.
3273  *
3274  */
3275 static inline bool delay_towrite(struct r5conf *conf,
3276                                  struct r5dev *dev,
3277                                  struct stripe_head_state *s)
3278 {
3279         /* case 1 above */
3280         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3281             !test_bit(R5_Insync, &dev->flags) && s->injournal)
3282                 return true;
3283         /* case 2 above */
3284         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3285             s->injournal > 0)
3286                 return true;
3287         /* case 3 above */
3288         if (s->log_failed && s->injournal)
3289                 return true;
3290         return false;
3291 }
3292
3293 static void
3294 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3295                          int rcw, int expand)
3296 {
3297         int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3298         struct r5conf *conf = sh->raid_conf;
3299         int level = conf->level;
3300
3301         if (rcw) {
3302                 /*
3303                  * In some cases, handle_stripe_dirtying initially decided to
3304                  * run rmw and allocates extra page for prexor. However, rcw is
3305                  * cheaper later on. We need to free the extra page now,
3306                  * because we won't be able to do that in ops_complete_prexor().
3307                  */
3308                 r5c_release_extra_page(sh);
3309
3310                 for (i = disks; i--; ) {
3311                         struct r5dev *dev = &sh->dev[i];
3312
3313                         if (dev->towrite && !delay_towrite(conf, dev, s)) {
3314                                 set_bit(R5_LOCKED, &dev->flags);
3315                                 set_bit(R5_Wantdrain, &dev->flags);
3316                                 if (!expand)
3317                                         clear_bit(R5_UPTODATE, &dev->flags);
3318                                 s->locked++;
3319                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3320                                 set_bit(R5_LOCKED, &dev->flags);
3321                                 s->locked++;
3322                         }
3323                 }
3324                 /* if we are not expanding this is a proper write request, and
3325                  * there will be bios with new data to be drained into the
3326                  * stripe cache
3327                  */
3328                 if (!expand) {
3329                         if (!s->locked)
3330                                 /* False alarm, nothing to do */
3331                                 return;
3332                         sh->reconstruct_state = reconstruct_state_drain_run;
3333                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3334                 } else
3335                         sh->reconstruct_state = reconstruct_state_run;
3336
3337                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3338
3339                 if (s->locked + conf->max_degraded == disks)
3340                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3341                                 atomic_inc(&conf->pending_full_writes);
3342         } else {
3343                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3344                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3345                 BUG_ON(level == 6 &&
3346                         (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3347                            test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3348
3349                 for (i = disks; i--; ) {
3350                         struct r5dev *dev = &sh->dev[i];
3351                         if (i == pd_idx || i == qd_idx)
3352                                 continue;
3353
3354                         if (dev->towrite &&
3355                             (test_bit(R5_UPTODATE, &dev->flags) ||
3356                              test_bit(R5_Wantcompute, &dev->flags))) {
3357                                 set_bit(R5_Wantdrain, &dev->flags);
3358                                 set_bit(R5_LOCKED, &dev->flags);
3359                                 clear_bit(R5_UPTODATE, &dev->flags);
3360                                 s->locked++;
3361                         } else if (test_bit(R5_InJournal, &dev->flags)) {
3362                                 set_bit(R5_LOCKED, &dev->flags);
3363                                 s->locked++;
3364                         }
3365                 }
3366                 if (!s->locked)
3367                         /* False alarm - nothing to do */
3368                         return;
3369                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3370                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3371                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3372                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3373         }
3374
3375         /* keep the parity disk(s) locked while asynchronous operations
3376          * are in flight
3377          */
3378         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3379         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3380         s->locked++;
3381
3382         if (level == 6) {
3383                 int qd_idx = sh->qd_idx;
3384                 struct r5dev *dev = &sh->dev[qd_idx];
3385
3386                 set_bit(R5_LOCKED, &dev->flags);
3387                 clear_bit(R5_UPTODATE, &dev->flags);
3388                 s->locked++;
3389         }
3390
3391         if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3392             test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3393             !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3394             test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3395                 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3396
3397         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3398                 __func__, (unsigned long long)sh->sector,
3399                 s->locked, s->ops_request);
3400 }
3401
3402 /*
3403  * Each stripe/dev can have one or more bion attached.
3404  * toread/towrite point to the first in a chain.
3405  * The bi_next chain must be in order.
3406  */
3407 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3408                           int forwrite, int previous)
3409 {
3410         struct bio **bip;
3411         struct r5conf *conf = sh->raid_conf;
3412         int firstwrite=0;
3413
3414         pr_debug("adding bi b#%llu to stripe s#%llu\n",
3415                 (unsigned long long)bi->bi_iter.bi_sector,
3416                 (unsigned long long)sh->sector);
3417
3418         spin_lock_irq(&sh->stripe_lock);
3419         sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3420         /* Don't allow new IO added to stripes in batch list */
3421         if (sh->batch_head)
3422                 goto overlap;
3423         if (forwrite) {
3424                 bip = &sh->dev[dd_idx].towrite;
3425                 if (*bip == NULL)
3426                         firstwrite = 1;
3427         } else
3428                 bip = &sh->dev[dd_idx].toread;
3429         while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3430                 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3431                         goto overlap;
3432                 bip = & (*bip)->bi_next;
3433         }
3434         if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3435                 goto overlap;
3436
3437         if (forwrite && raid5_has_ppl(conf)) {
3438                 /*
3439                  * With PPL only writes to consecutive data chunks within a
3440                  * stripe are allowed because for a single stripe_head we can
3441                  * only have one PPL entry at a time, which describes one data
3442                  * range. Not really an overlap, but wait_for_overlap can be
3443                  * used to handle this.
3444                  */
3445                 sector_t sector;
3446                 sector_t first = 0;
3447                 sector_t last = 0;
3448                 int count = 0;
3449                 int i;
3450
3451                 for (i = 0; i < sh->disks; i++) {
3452                         if (i != sh->pd_idx &&
3453                             (i == dd_idx || sh->dev[i].towrite)) {
3454                                 sector = sh->dev[i].sector;
3455                                 if (count == 0 || sector < first)
3456                                         first = sector;
3457                                 if (sector > last)
3458                                         last = sector;
3459                                 count++;
3460                         }
3461                 }
3462
3463                 if (first + conf->chunk_sectors * (count - 1) != last)
3464                         goto overlap;
3465         }
3466
3467         if (!forwrite || previous)
3468                 clear_bit(STRIPE_BATCH_READY, &sh->state);
3469
3470         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3471         if (*bip)
3472                 bi->bi_next = *bip;
3473         *bip = bi;
3474         bio_inc_remaining(bi);
3475         md_write_inc(conf->mddev, bi);
3476
3477         if (forwrite) {
3478                 /* check if page is covered */
3479                 sector_t sector = sh->dev[dd_idx].sector;
3480                 for (bi=sh->dev[dd_idx].towrite;
3481                      sector < sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf) &&
3482                              bi && bi->bi_iter.bi_sector <= sector;
3483                      bi = r5_next_bio(conf, bi, sh->dev[dd_idx].sector)) {
3484                         if (bio_end_sector(bi) >= sector)
3485                                 sector = bio_end_sector(bi);
3486                 }
3487                 if (sector >= sh->dev[dd_idx].sector + RAID5_STRIPE_SECTORS(conf))
3488                         if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3489                                 sh->overwrite_disks++;
3490         }
3491
3492         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3493                 (unsigned long long)(*bip)->bi_iter.bi_sector,
3494                 (unsigned long long)sh->sector, dd_idx);
3495
3496         if (conf->mddev->bitmap && firstwrite) {
3497                 /* Cannot hold spinlock over bitmap_startwrite,
3498                  * but must ensure this isn't added to a batch until
3499                  * we have added to the bitmap and set bm_seq.
3500                  * So set STRIPE_BITMAP_PENDING to prevent
3501                  * batching.
3502                  * If multiple add_stripe_bio() calls race here they
3503                  * much all set STRIPE_BITMAP_PENDING.  So only the first one
3504                  * to complete "bitmap_startwrite" gets to set
3505                  * STRIPE_BIT_DELAY.  This is important as once a stripe
3506                  * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3507                  * any more.
3508                  */
3509                 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3510                 spin_unlock_irq(&sh->stripe_lock);
3511                 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3512                                      RAID5_STRIPE_SECTORS(conf), 0);
3513                 spin_lock_irq(&sh->stripe_lock);
3514                 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3515                 if (!sh->batch_head) {
3516                         sh->bm_seq = conf->seq_flush+1;
3517                         set_bit(STRIPE_BIT_DELAY, &sh->state);
3518                 }
3519         }
3520         spin_unlock_irq(&sh->stripe_lock);
3521
3522         if (stripe_can_batch(sh))
3523                 stripe_add_to_batch_list(conf, sh);
3524         return 1;
3525
3526  overlap:
3527         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3528         spin_unlock_irq(&sh->stripe_lock);
3529         return 0;
3530 }
3531
3532 static void end_reshape(struct r5conf *conf);
3533
3534 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3535                             struct stripe_head *sh)
3536 {
3537         int sectors_per_chunk =
3538                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3539         int dd_idx;
3540         int chunk_offset = sector_div(stripe, sectors_per_chunk);
3541         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3542
3543         raid5_compute_sector(conf,
3544                              stripe * (disks - conf->max_degraded)
3545                              *sectors_per_chunk + chunk_offset,
3546                              previous,
3547                              &dd_idx, sh);
3548 }
3549
3550 static void
3551 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3552                      struct stripe_head_state *s, int disks)
3553 {
3554         int i;
3555         BUG_ON(sh->batch_head);
3556         for (i = disks; i--; ) {
3557                 struct bio *bi;
3558                 int bitmap_end = 0;
3559
3560                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3561                         struct md_rdev *rdev;
3562                         rcu_read_lock();
3563                         rdev = rcu_dereference(conf->disks[i].rdev);
3564                         if (rdev && test_bit(In_sync, &rdev->flags) &&
3565                             !test_bit(Faulty, &rdev->flags))
3566                                 atomic_inc(&rdev->nr_pending);
3567                         else
3568                                 rdev = NULL;
3569                         rcu_read_unlock();
3570                         if (rdev) {
3571                                 if (!rdev_set_badblocks(
3572                                             rdev,
3573                                             sh->sector,
3574                                             RAID5_STRIPE_SECTORS(conf), 0))
3575                                         md_error(conf->mddev, rdev);
3576                                 rdev_dec_pending(rdev, conf->mddev);
3577                         }
3578                 }
3579                 spin_lock_irq(&sh->stripe_lock);
3580                 /* fail all writes first */
3581                 bi = sh->dev[i].towrite;
3582                 sh->dev[i].towrite = NULL;
3583                 sh->overwrite_disks = 0;
3584                 spin_unlock_irq(&sh->stripe_lock);
3585                 if (bi)
3586                         bitmap_end = 1;
3587
3588                 log_stripe_write_finished(sh);
3589
3590                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3591                         wake_up(&conf->wait_for_overlap);
3592
3593                 while (bi && bi->bi_iter.bi_sector <
3594                         sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3595                         struct bio *nextbi = r5_next_bio(conf, bi, sh->dev[i].sector);
3596
3597                         md_write_end(conf->mddev);
3598                         bio_io_error(bi);
3599                         bi = nextbi;
3600                 }
3601                 if (bitmap_end)
3602                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3603                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3604                 bitmap_end = 0;
3605                 /* and fail all 'written' */
3606                 bi = sh->dev[i].written;
3607                 sh->dev[i].written = NULL;
3608                 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3609                         WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3610                         sh->dev[i].page = sh->dev[i].orig_page;
3611                 }
3612
3613                 if (bi) bitmap_end = 1;
3614                 while (bi && bi->bi_iter.bi_sector <
3615                        sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3616                         struct bio *bi2 = r5_next_bio(conf, bi, sh->dev[i].sector);
3617
3618                         md_write_end(conf->mddev);
3619                         bio_io_error(bi);
3620                         bi = bi2;
3621                 }
3622
3623                 /* fail any reads if this device is non-operational and
3624                  * the data has not reached the cache yet.
3625                  */
3626                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3627                     s->failed > conf->max_degraded &&
3628                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3629                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
3630                         spin_lock_irq(&sh->stripe_lock);
3631                         bi = sh->dev[i].toread;
3632                         sh->dev[i].toread = NULL;
3633                         spin_unlock_irq(&sh->stripe_lock);
3634                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3635                                 wake_up(&conf->wait_for_overlap);
3636                         if (bi)
3637                                 s->to_read--;
3638                         while (bi && bi->bi_iter.bi_sector <
3639                                sh->dev[i].sector + RAID5_STRIPE_SECTORS(conf)) {
3640                                 struct bio *nextbi =
3641                                         r5_next_bio(conf, bi, sh->dev[i].sector);
3642
3643                                 bio_io_error(bi);
3644                                 bi = nextbi;
3645                         }
3646                 }
3647                 if (bitmap_end)
3648                         md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3649                                            RAID5_STRIPE_SECTORS(conf), 0, 0);
3650                 /* If we were in the middle of a write the parity block might
3651                  * still be locked - so just clear all R5_LOCKED flags
3652                  */
3653                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3654         }
3655         s->to_write = 0;
3656         s->written = 0;
3657
3658         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3659                 if (atomic_dec_and_test(&conf->pending_full_writes))
3660                         md_wakeup_thread(conf->mddev->thread);
3661 }
3662
3663 static void
3664 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3665                    struct stripe_head_state *s)
3666 {
3667         int abort = 0;
3668         int i;
3669
3670         BUG_ON(sh->batch_head);
3671         clear_bit(STRIPE_SYNCING, &sh->state);
3672         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3673                 wake_up(&conf->wait_for_overlap);
3674         s->syncing = 0;
3675         s->replacing = 0;
3676         /* There is nothing more to do for sync/check/repair.
3677          * Don't even need to abort as that is handled elsewhere
3678          * if needed, and not always wanted e.g. if there is a known
3679          * bad block here.
3680          * For recover/replace we need to record a bad block on all
3681          * non-sync devices, or abort the recovery
3682          */
3683         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3684                 /* During recovery devices cannot be removed, so
3685                  * locking and refcounting of rdevs is not needed
3686                  */
3687                 rcu_read_lock();
3688                 for (i = 0; i < conf->raid_disks; i++) {
3689                         struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3690                         if (rdev
3691                             && !test_bit(Faulty, &rdev->flags)
3692                             && !test_bit(In_sync, &rdev->flags)
3693                             && !rdev_set_badblocks(rdev, sh->sector,
3694                                                    RAID5_STRIPE_SECTORS(conf), 0))
3695                                 abort = 1;
3696                         rdev = rcu_dereference(conf->disks[i].replacement);
3697                         if (rdev
3698                             && !test_bit(Faulty, &rdev->flags)
3699                             && !test_bit(In_sync, &rdev->flags)
3700                             && !rdev_set_badblocks(rdev, sh->sector,
3701                                                    RAID5_STRIPE_SECTORS(conf), 0))
3702                                 abort = 1;
3703                 }
3704                 rcu_read_unlock();
3705                 if (abort)
3706                         conf->recovery_disabled =
3707                                 conf->mddev->recovery_disabled;
3708         }
3709         md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), !abort);
3710 }
3711
3712 static int want_replace(struct stripe_head *sh, int disk_idx)
3713 {
3714         struct md_rdev *rdev;
3715         int rv = 0;
3716
3717         rcu_read_lock();
3718         rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3719         if (rdev
3720             && !test_bit(Faulty, &rdev->flags)
3721             && !test_bit(In_sync, &rdev->flags)
3722             && (rdev->recovery_offset <= sh->sector
3723                 || rdev->mddev->recovery_cp <= sh->sector))
3724                 rv = 1;
3725         rcu_read_unlock();
3726         return rv;
3727 }
3728
3729 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3730                            int disk_idx, int disks)
3731 {
3732         struct r5dev *dev = &sh->dev[disk_idx];
3733         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3734                                   &sh->dev[s->failed_num[1]] };
3735         int i;
3736         bool force_rcw = (sh->raid_conf->rmw_level == PARITY_DISABLE_RMW);
3737
3738
3739         if (test_bit(R5_LOCKED, &dev->flags) ||
3740             test_bit(R5_UPTODATE, &dev->flags))
3741                 /* No point reading this as we already have it or have
3742                  * decided to get it.
3743                  */
3744                 return 0;
3745
3746         if (dev->toread ||
3747             (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3748                 /* We need this block to directly satisfy a request */
3749                 return 1;
3750
3751         if (s->syncing || s->expanding ||
3752             (s->replacing && want_replace(sh, disk_idx)))
3753                 /* When syncing, or expanding we read everything.
3754                  * When replacing, we need the replaced block.
3755                  */
3756                 return 1;
3757
3758         if ((s->failed >= 1 && fdev[0]->toread) ||
3759             (s->failed >= 2 && fdev[1]->toread))
3760                 /* If we want to read from a failed device, then
3761                  * we need to actually read every other device.
3762                  */
3763                 return 1;
3764
3765         /* Sometimes neither read-modify-write nor reconstruct-write
3766          * cycles can work.  In those cases we read every block we
3767          * can.  Then the parity-update is certain to have enough to
3768          * work with.
3769          * This can only be a problem when we need to write something,
3770          * and some device has failed.  If either of those tests
3771          * fail we need look no further.
3772          */
3773         if (!s->failed || !s->to_write)
3774                 return 0;
3775
3776         if (test_bit(R5_Insync, &dev->flags) &&
3777             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3778                 /* Pre-reads at not permitted until after short delay
3779                  * to gather multiple requests.  However if this
3780                  * device is no Insync, the block could only be computed
3781                  * and there is no need to delay that.
3782                  */
3783                 return 0;
3784
3785         for (i = 0; i < s->failed && i < 2; i++) {
3786                 if (fdev[i]->towrite &&
3787                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3788                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3789                         /* If we have a partial write to a failed
3790                          * device, then we will need to reconstruct
3791                          * the content of that device, so all other
3792                          * devices must be read.
3793                          */
3794                         return 1;
3795
3796                 if (s->failed >= 2 &&
3797                     (fdev[i]->towrite ||
3798                      s->failed_num[i] == sh->pd_idx ||
3799                      s->failed_num[i] == sh->qd_idx) &&
3800                     !test_bit(R5_UPTODATE, &fdev[i]->flags))
3801                         /* In max degraded raid6, If the failed disk is P, Q,
3802                          * or we want to read the failed disk, we need to do
3803                          * reconstruct-write.
3804                          */
3805                         force_rcw = true;
3806         }
3807
3808         /* If we are forced to do a reconstruct-write, because parity
3809          * cannot be trusted and we are currently recovering it, there
3810          * is extra need to be careful.
3811          * If one of the devices that we would need to read, because
3812          * it is not being overwritten (and maybe not written at all)
3813          * is missing/faulty, then we need to read everything we can.
3814          */
3815         if (!force_rcw &&
3816             sh->sector < sh->raid_conf->mddev->recovery_cp)
3817                 /* reconstruct-write isn't being forced */
3818                 return 0;
3819         for (i = 0; i < s->failed && i < 2; i++) {
3820                 if (s->failed_num[i] != sh->pd_idx &&
3821                     s->failed_num[i] != sh->qd_idx &&
3822                     !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3823                     !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3824                         return 1;
3825         }
3826
3827         return 0;
3828 }
3829
3830 /* fetch_block - checks the given member device to see if its data needs
3831  * to be read or computed to satisfy a request.
3832  *
3833  * Returns 1 when no more member devices need to be checked, otherwise returns
3834  * 0 to tell the loop in handle_stripe_fill to continue
3835  */
3836 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3837                        int disk_idx, int disks)
3838 {
3839         struct r5dev *dev = &sh->dev[disk_idx];
3840
3841         /* is the data in this block needed, and can we get it? */
3842         if (need_this_block(sh, s, disk_idx, disks)) {
3843                 /* we would like to get this block, possibly by computing it,
3844                  * otherwise read it if the backing disk is insync
3845                  */
3846                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3847                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3848                 BUG_ON(sh->batch_head);
3849
3850                 /*
3851                  * In the raid6 case if the only non-uptodate disk is P
3852                  * then we already trusted P to compute the other failed
3853                  * drives. It is safe to compute rather than re-read P.
3854                  * In other cases we only compute blocks from failed
3855                  * devices, otherwise check/repair might fail to detect
3856                  * a real inconsistency.
3857                  */
3858
3859                 if ((s->uptodate == disks - 1) &&
3860                     ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3861                     (s->failed && (disk_idx == s->failed_num[0] ||
3862                                    disk_idx == s->failed_num[1])))) {
3863                         /* have disk failed, and we're requested to fetch it;
3864                          * do compute it
3865                          */
3866                         pr_debug("Computing stripe %llu block %d\n",
3867                                (unsigned long long)sh->sector, disk_idx);
3868                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3869                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3870                         set_bit(R5_Wantcompute, &dev->flags);
3871                         sh->ops.target = disk_idx;
3872                         sh->ops.target2 = -1; /* no 2nd target */
3873                         s->req_compute = 1;
3874                         /* Careful: from this point on 'uptodate' is in the eye
3875                          * of raid_run_ops which services 'compute' operations
3876                          * before writes. R5_Wantcompute flags a block that will
3877                          * be R5_UPTODATE by the time it is needed for a
3878                          * subsequent operation.
3879                          */
3880                         s->uptodate++;
3881                         return 1;
3882                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3883                         /* Computing 2-failure is *very* expensive; only
3884                          * do it if failed >= 2
3885                          */
3886                         int other;
3887                         for (other = disks; other--; ) {
3888                                 if (other == disk_idx)
3889                                         continue;
3890                                 if (!test_bit(R5_UPTODATE,
3891                                       &sh->dev[other].flags))
3892                                         break;
3893                         }
3894                         BUG_ON(other < 0);
3895                         pr_debug("Computing stripe %llu blocks %d,%d\n",
3896                                (unsigned long long)sh->sector,
3897                                disk_idx, other);
3898                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3899                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3900                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3901                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
3902                         sh->ops.target = disk_idx;
3903                         sh->ops.target2 = other;
3904                         s->uptodate += 2;
3905                         s->req_compute = 1;
3906                         return 1;
3907                 } else if (test_bit(R5_Insync, &dev->flags)) {
3908                         set_bit(R5_LOCKED, &dev->flags);
3909                         set_bit(R5_Wantread, &dev->flags);
3910                         s->locked++;
3911                         pr_debug("Reading block %d (sync=%d)\n",
3912                                 disk_idx, s->syncing);
3913                 }
3914         }
3915
3916         return 0;
3917 }
3918
3919 /*
3920  * handle_stripe_fill - read or compute data to satisfy pending requests.
3921  */
3922 static void handle_stripe_fill(struct stripe_head *sh,
3923                                struct stripe_head_state *s,
3924                                int disks)
3925 {
3926         int i;
3927
3928         /* look for blocks to read/compute, skip this if a compute
3929          * is already in flight, or if the stripe contents are in the
3930          * midst of changing due to a write
3931          */
3932         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3933             !sh->reconstruct_state) {
3934
3935                 /*
3936                  * For degraded stripe with data in journal, do not handle
3937                  * read requests yet, instead, flush the stripe to raid
3938                  * disks first, this avoids handling complex rmw of write
3939                  * back cache (prexor with orig_page, and then xor with
3940                  * page) in the read path
3941                  */
3942                 if (s->injournal && s->failed) {
3943                         if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3944                                 r5c_make_stripe_write_out(sh);
3945                         goto out;
3946                 }
3947
3948                 for (i = disks; i--; )
3949                         if (fetch_block(sh, s, i, disks))
3950                                 break;
3951         }
3952 out:
3953         set_bit(STRIPE_HANDLE, &sh->state);
3954 }
3955
3956 static void break_stripe_batch_list(struct stripe_head *head_sh,
3957                                     unsigned long handle_flags);
3958 /* handle_stripe_clean_event
3959  * any written block on an uptodate or failed drive can be returned.
3960  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3961  * never LOCKED, so we don't need to test 'failed' directly.
3962  */
3963 static void handle_stripe_clean_event(struct r5conf *conf,
3964         struct stripe_head *sh, int disks)
3965 {
3966         int i;
3967         struct r5dev *dev;
3968         int discard_pending = 0;
3969         struct stripe_head *head_sh = sh;
3970         bool do_endio = false;
3971
3972         for (i = disks; i--; )
3973                 if (sh->dev[i].written) {
3974                         dev = &sh->dev[i];
3975                         if (!test_bit(R5_LOCKED, &dev->flags) &&
3976                             (test_bit(R5_UPTODATE, &dev->flags) ||
3977                              test_bit(R5_Discard, &dev->flags) ||
3978                              test_bit(R5_SkipCopy, &dev->flags))) {
3979                                 /* We can return any write requests */
3980                                 struct bio *wbi, *wbi2;
3981                                 pr_debug("Return write for disc %d\n", i);
3982                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
3983                                         clear_bit(R5_UPTODATE, &dev->flags);
3984                                 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3985                                         WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3986                                 }
3987                                 do_endio = true;
3988
3989 returnbi:
3990                                 dev->page = dev->orig_page;
3991                                 wbi = dev->written;
3992                                 dev->written = NULL;
3993                                 while (wbi && wbi->bi_iter.bi_sector <
3994                                         dev->sector + RAID5_STRIPE_SECTORS(conf)) {
3995                                         wbi2 = r5_next_bio(conf, wbi, dev->sector);
3996                                         md_write_end(conf->mddev);
3997                                         bio_endio(wbi);
3998                                         wbi = wbi2;
3999                                 }
4000                                 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
4001                                                    RAID5_STRIPE_SECTORS(conf),
4002                                                    !test_bit(STRIPE_DEGRADED, &sh->state),
4003                                                    0);
4004                                 if (head_sh->batch_head) {
4005                                         sh = list_first_entry(&sh->batch_list,
4006                                                               struct stripe_head,
4007                                                               batch_list);
4008                                         if (sh != head_sh) {
4009                                                 dev = &sh->dev[i];
4010                                                 goto returnbi;
4011                                         }
4012                                 }
4013                                 sh = head_sh;
4014                                 dev = &sh->dev[i];
4015                         } else if (test_bit(R5_Discard, &dev->flags))
4016                                 discard_pending = 1;
4017                 }
4018
4019         log_stripe_write_finished(sh);
4020
4021         if (!discard_pending &&
4022             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
4023                 int hash;
4024                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
4025                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4026                 if (sh->qd_idx >= 0) {
4027                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
4028                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
4029                 }
4030                 /* now that discard is done we can proceed with any sync */
4031                 clear_bit(STRIPE_DISCARD, &sh->state);
4032                 /*
4033                  * SCSI discard will change some bio fields and the stripe has
4034                  * no updated data, so remove it from hash list and the stripe
4035                  * will be reinitialized
4036                  */
4037 unhash:
4038                 hash = sh->hash_lock_index;
4039                 spin_lock_irq(conf->hash_locks + hash);
4040                 remove_hash(sh);
4041                 spin_unlock_irq(conf->hash_locks + hash);
4042                 if (head_sh->batch_head) {
4043                         sh = list_first_entry(&sh->batch_list,
4044                                               struct stripe_head, batch_list);
4045                         if (sh != head_sh)
4046                                         goto unhash;
4047                 }
4048                 sh = head_sh;
4049
4050                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
4051                         set_bit(STRIPE_HANDLE, &sh->state);
4052
4053         }
4054
4055         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
4056                 if (atomic_dec_and_test(&conf->pending_full_writes))
4057                         md_wakeup_thread(conf->mddev->thread);
4058
4059         if (head_sh->batch_head && do_endio)
4060                 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
4061 }
4062
4063 /*
4064  * For RMW in write back cache, we need extra page in prexor to store the
4065  * old data. This page is stored in dev->orig_page.
4066  *
4067  * This function checks whether we have data for prexor. The exact logic
4068  * is:
4069  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
4070  */
4071 static inline bool uptodate_for_rmw(struct r5dev *dev)
4072 {
4073         return (test_bit(R5_UPTODATE, &dev->flags)) &&
4074                 (!test_bit(R5_InJournal, &dev->flags) ||
4075                  test_bit(R5_OrigPageUPTDODATE, &dev->flags));
4076 }
4077
4078 static int handle_stripe_dirtying(struct r5conf *conf,
4079                                   struct stripe_head *sh,
4080                                   struct stripe_head_state *s,
4081                                   int disks)
4082 {
4083         int rmw = 0, rcw = 0, i;
4084         sector_t recovery_cp = conf->mddev->recovery_cp;
4085
4086         /* Check whether resync is now happening or should start.
4087          * If yes, then the array is dirty (after unclean shutdown or
4088          * initial creation), so parity in some stripes might be inconsistent.
4089          * In this case, we need to always do reconstruct-write, to ensure
4090          * that in case of drive failure or read-error correction, we
4091          * generate correct data from the parity.
4092          */
4093         if (conf->rmw_level == PARITY_DISABLE_RMW ||
4094             (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
4095              s->failed == 0)) {
4096                 /* Calculate the real rcw later - for now make it
4097                  * look like rcw is cheaper
4098                  */
4099                 rcw = 1; rmw = 2;
4100                 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
4101                          conf->rmw_level, (unsigned long long)recovery_cp,
4102                          (unsigned long long)sh->sector);
4103         } else for (i = disks; i--; ) {
4104                 /* would I have to read this buffer for read_modify_write */
4105                 struct r5dev *dev = &sh->dev[i];
4106                 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4107                      i == sh->pd_idx || i == sh->qd_idx ||
4108                      test_bit(R5_InJournal, &dev->flags)) &&
4109                     !test_bit(R5_LOCKED, &dev->flags) &&
4110                     !(uptodate_for_rmw(dev) ||
4111                       test_bit(R5_Wantcompute, &dev->flags))) {
4112                         if (test_bit(R5_Insync, &dev->flags))
4113                                 rmw++;
4114                         else
4115                                 rmw += 2*disks;  /* cannot read it */
4116                 }
4117                 /* Would I have to read this buffer for reconstruct_write */
4118                 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4119                     i != sh->pd_idx && i != sh->qd_idx &&
4120                     !test_bit(R5_LOCKED, &dev->flags) &&
4121                     !(test_bit(R5_UPTODATE, &dev->flags) ||
4122                       test_bit(R5_Wantcompute, &dev->flags))) {
4123                         if (test_bit(R5_Insync, &dev->flags))
4124                                 rcw++;
4125                         else
4126                                 rcw += 2*disks;
4127                 }
4128         }
4129
4130         pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
4131                  (unsigned long long)sh->sector, sh->state, rmw, rcw);
4132         set_bit(STRIPE_HANDLE, &sh->state);
4133         if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
4134                 /* prefer read-modify-write, but need to get some data */
4135                 if (conf->mddev->queue)
4136                         blk_add_trace_msg(conf->mddev->queue,
4137                                           "raid5 rmw %llu %d",
4138                                           (unsigned long long)sh->sector, rmw);
4139                 for (i = disks; i--; ) {
4140                         struct r5dev *dev = &sh->dev[i];
4141                         if (test_bit(R5_InJournal, &dev->flags) &&
4142                             dev->page == dev->orig_page &&
4143                             !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
4144                                 /* alloc page for prexor */
4145                                 struct page *p = alloc_page(GFP_NOIO);
4146
4147                                 if (p) {
4148                                         dev->orig_page = p;
4149                                         continue;
4150                                 }
4151
4152                                 /*
4153                                  * alloc_page() failed, try use
4154                                  * disk_info->extra_page
4155                                  */
4156                                 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
4157                                                       &conf->cache_state)) {
4158                                         r5c_use_extra_page(sh);
4159                                         break;
4160                                 }
4161
4162                                 /* extra_page in use, add to delayed_list */
4163                                 set_bit(STRIPE_DELAYED, &sh->state);
4164                                 s->waiting_extra_page = 1;
4165                                 return -EAGAIN;
4166                         }
4167                 }
4168
4169                 for (i = disks; i--; ) {
4170                         struct r5dev *dev = &sh->dev[i];
4171                         if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
4172                              i == sh->pd_idx || i == sh->qd_idx ||
4173                              test_bit(R5_InJournal, &dev->flags)) &&
4174                             !test_bit(R5_LOCKED, &dev->flags) &&
4175                             !(uptodate_for_rmw(dev) ||
4176                               test_bit(R5_Wantcompute, &dev->flags)) &&
4177                             test_bit(R5_Insync, &dev->flags)) {
4178                                 if (test_bit(STRIPE_PREREAD_ACTIVE,
4179                                              &sh->state)) {
4180                                         pr_debug("Read_old block %d for r-m-w\n",
4181                                                  i);
4182                                         set_bit(R5_LOCKED, &dev->flags);
4183                                         set_bit(R5_Wantread, &dev->flags);
4184                                         s->locked++;
4185                                 } else
4186                                         set_bit(STRIPE_DELAYED, &sh->state);
4187                         }
4188                 }
4189         }
4190         if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
4191                 /* want reconstruct write, but need to get some data */
4192                 int qread =0;
4193                 rcw = 0;
4194                 for (i = disks; i--; ) {
4195                         struct r5dev *dev = &sh->dev[i];
4196                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
4197                             i != sh->pd_idx && i != sh->qd_idx &&
4198                             !test_bit(R5_LOCKED, &dev->flags) &&
4199                             !(test_bit(R5_UPTODATE, &dev->flags) ||
4200                               test_bit(R5_Wantcompute, &dev->flags))) {
4201                                 rcw++;
4202                                 if (test_bit(R5_Insync, &dev->flags) &&
4203                                     test_bit(STRIPE_PREREAD_ACTIVE,
4204                                              &sh->state)) {
4205                                         pr_debug("Read_old block "
4206                                                 "%d for Reconstruct\n", i);
4207                                         set_bit(R5_LOCKED, &dev->flags);
4208                                         set_bit(R5_Wantread, &dev->flags);
4209                                         s->locked++;
4210                                         qread++;
4211                                 } else
4212                                         set_bit(STRIPE_DELAYED, &sh->state);
4213                         }
4214                 }
4215                 if (rcw && conf->mddev->queue)
4216                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4217                                           (unsigned long long)sh->sector,
4218                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4219         }
4220
4221         if (rcw > disks && rmw > disks &&
4222             !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4223                 set_bit(STRIPE_DELAYED, &sh->state);
4224
4225         /* now if nothing is locked, and if we have enough data,
4226          * we can start a write request
4227          */
4228         /* since handle_stripe can be called at any time we need to handle the
4229          * case where a compute block operation has been submitted and then a
4230          * subsequent call wants to start a write request.  raid_run_ops only
4231          * handles the case where compute block and reconstruct are requested
4232          * simultaneously.  If this is not the case then new writes need to be
4233          * held off until the compute completes.
4234          */
4235         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4236             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4237              !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4238                 schedule_reconstruction(sh, s, rcw == 0, 0);
4239         return 0;
4240 }
4241
4242 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4243                                 struct stripe_head_state *s, int disks)
4244 {
4245         struct r5dev *dev = NULL;
4246
4247         BUG_ON(sh->batch_head);
4248         set_bit(STRIPE_HANDLE, &sh->state);
4249
4250         switch (sh->check_state) {
4251         case check_state_idle:
4252                 /* start a new check operation if there are no failures */
4253                 if (s->failed == 0) {
4254                         BUG_ON(s->uptodate != disks);
4255                         sh->check_state = check_state_run;
4256                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4257                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4258                         s->uptodate--;
4259                         break;
4260                 }
4261                 dev = &sh->dev[s->failed_num[0]];
4262                 fallthrough;
4263         case check_state_compute_result:
4264                 sh->check_state = check_state_idle;
4265                 if (!dev)
4266                         dev = &sh->dev[sh->pd_idx];
4267
4268                 /* check that a write has not made the stripe insync */
4269                 if (test_bit(STRIPE_INSYNC, &sh->state))
4270                         break;
4271
4272                 /* either failed parity check, or recovery is happening */
4273                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4274                 BUG_ON(s->uptodate != disks);
4275
4276                 set_bit(R5_LOCKED, &dev->flags);
4277                 s->locked++;
4278                 set_bit(R5_Wantwrite, &dev->flags);
4279
4280                 clear_bit(STRIPE_DEGRADED, &sh->state);
4281                 set_bit(STRIPE_INSYNC, &sh->state);
4282                 break;
4283         case check_state_run:
4284                 break; /* we will be called again upon completion */
4285         case check_state_check_result:
4286                 sh->check_state = check_state_idle;
4287
4288                 /* if a failure occurred during the check operation, leave
4289                  * STRIPE_INSYNC not set and let the stripe be handled again
4290                  */
4291                 if (s->failed)
4292                         break;
4293
4294                 /* handle a successful check operation, if parity is correct
4295                  * we are done.  Otherwise update the mismatch count and repair
4296                  * parity if !MD_RECOVERY_CHECK
4297                  */
4298                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4299                         /* parity is correct (on disc,
4300                          * not in buffer any more)
4301                          */
4302                         set_bit(STRIPE_INSYNC, &sh->state);
4303                 else {
4304                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4305                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4306                                 /* don't try to repair!! */
4307                                 set_bit(STRIPE_INSYNC, &sh->state);
4308                                 pr_warn_ratelimited("%s: mismatch sector in range "
4309                                                     "%llu-%llu\n", mdname(conf->mddev),
4310                                                     (unsigned long long) sh->sector,
4311                                                     (unsigned long long) sh->sector +
4312                                                     RAID5_STRIPE_SECTORS(conf));
4313                         } else {
4314                                 sh->check_state = check_state_compute_run;
4315                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4316                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4317                                 set_bit(R5_Wantcompute,
4318                                         &sh->dev[sh->pd_idx].flags);
4319                                 sh->ops.target = sh->pd_idx;
4320                                 sh->ops.target2 = -1;
4321                                 s->uptodate++;
4322                         }
4323                 }
4324                 break;
4325         case check_state_compute_run:
4326                 break;
4327         default:
4328                 pr_err("%s: unknown check_state: %d sector: %llu\n",
4329                        __func__, sh->check_state,
4330                        (unsigned long long) sh->sector);
4331                 BUG();
4332         }
4333 }
4334
4335 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4336                                   struct stripe_head_state *s,
4337                                   int disks)
4338 {
4339         int pd_idx = sh->pd_idx;
4340         int qd_idx = sh->qd_idx;
4341         struct r5dev *dev;
4342
4343         BUG_ON(sh->batch_head);
4344         set_bit(STRIPE_HANDLE, &sh->state);
4345
4346         BUG_ON(s->failed > 2);
4347
4348         /* Want to check and possibly repair P and Q.
4349          * However there could be one 'failed' device, in which
4350          * case we can only check one of them, possibly using the
4351          * other to generate missing data
4352          */
4353
4354         switch (sh->check_state) {
4355         case check_state_idle:
4356                 /* start a new check operation if there are < 2 failures */
4357                 if (s->failed == s->q_failed) {
4358                         /* The only possible failed device holds Q, so it
4359                          * makes sense to check P (If anything else were failed,
4360                          * we would have used P to recreate it).
4361                          */
4362                         sh->check_state = check_state_run;
4363                 }
4364                 if (!s->q_failed && s->failed < 2) {
4365                         /* Q is not failed, and we didn't use it to generate
4366                          * anything, so it makes sense to check it
4367                          */
4368                         if (sh->check_state == check_state_run)
4369                                 sh->check_state = check_state_run_pq;
4370                         else
4371                                 sh->check_state = check_state_run_q;
4372                 }
4373
4374                 /* discard potentially stale zero_sum_result */
4375                 sh->ops.zero_sum_result = 0;
4376
4377                 if (sh->check_state == check_state_run) {
4378                         /* async_xor_zero_sum destroys the contents of P */
4379                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4380                         s->uptodate--;
4381                 }
4382                 if (sh->check_state >= check_state_run &&
4383                     sh->check_state <= check_state_run_pq) {
4384                         /* async_syndrome_zero_sum preserves P and Q, so
4385                          * no need to mark them !uptodate here
4386                          */
4387                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
4388                         break;
4389                 }
4390
4391                 /* we have 2-disk failure */
4392                 BUG_ON(s->failed != 2);
4393                 fallthrough;
4394         case check_state_compute_result:
4395                 sh->check_state = check_state_idle;
4396
4397                 /* check that a write has not made the stripe insync */
4398                 if (test_bit(STRIPE_INSYNC, &sh->state))
4399                         break;
4400
4401                 /* now write out any block on a failed drive,
4402                  * or P or Q if they were recomputed
4403                  */
4404                 dev = NULL;
4405                 if (s->failed == 2) {
4406                         dev = &sh->dev[s->failed_num[1]];
4407                         s->locked++;
4408                         set_bit(R5_LOCKED, &dev->flags);
4409                         set_bit(R5_Wantwrite, &dev->flags);
4410                 }
4411                 if (s->failed >= 1) {
4412                         dev = &sh->dev[s->failed_num[0]];
4413                         s->locked++;
4414                         set_bit(R5_LOCKED, &dev->flags);
4415                         set_bit(R5_Wantwrite, &dev->flags);
4416                 }
4417                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4418                         dev = &sh->dev[pd_idx];
4419                         s->locked++;
4420                         set_bit(R5_LOCKED, &dev->flags);
4421                         set_bit(R5_Wantwrite, &dev->flags);
4422                 }
4423                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4424                         dev = &sh->dev[qd_idx];
4425                         s->locked++;
4426                         set_bit(R5_LOCKED, &dev->flags);
4427                         set_bit(R5_Wantwrite, &dev->flags);
4428                 }
4429                 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4430                               "%s: disk%td not up to date\n",
4431                               mdname(conf->mddev),
4432                               dev - (struct r5dev *) &sh->dev)) {
4433                         clear_bit(R5_LOCKED, &dev->flags);
4434                         clear_bit(R5_Wantwrite, &dev->flags);
4435                         s->locked--;
4436                 }
4437                 clear_bit(STRIPE_DEGRADED, &sh->state);
4438
4439                 set_bit(STRIPE_INSYNC, &sh->state);
4440                 break;
4441         case check_state_run:
4442         case check_state_run_q:
4443         case check_state_run_pq:
4444                 break; /* we will be called again upon completion */
4445         case check_state_check_result:
4446                 sh->check_state = check_state_idle;
4447
4448                 /* handle a successful check operation, if parity is correct
4449                  * we are done.  Otherwise update the mismatch count and repair
4450                  * parity if !MD_RECOVERY_CHECK
4451                  */
4452                 if (sh->ops.zero_sum_result == 0) {
4453                         /* both parities are correct */
4454                         if (!s->failed)
4455                                 set_bit(STRIPE_INSYNC, &sh->state);
4456                         else {
4457                                 /* in contrast to the raid5 case we can validate
4458                                  * parity, but still have a failure to write
4459                                  * back
4460                                  */
4461                                 sh->check_state = check_state_compute_result;
4462                                 /* Returning at this point means that we may go
4463                                  * off and bring p and/or q uptodate again so
4464                                  * we make sure to check zero_sum_result again
4465                                  * to verify if p or q need writeback
4466                                  */
4467                         }
4468                 } else {
4469                         atomic64_add(RAID5_STRIPE_SECTORS(conf), &conf->mddev->resync_mismatches);
4470                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4471                                 /* don't try to repair!! */
4472                                 set_bit(STRIPE_INSYNC, &sh->state);
4473                                 pr_warn_ratelimited("%s: mismatch sector in range "
4474                                                     "%llu-%llu\n", mdname(conf->mddev),
4475                                                     (unsigned long long) sh->sector,
4476                                                     (unsigned long long) sh->sector +
4477                                                     RAID5_STRIPE_SECTORS(conf));
4478                         } else {
4479                                 int *target = &sh->ops.target;
4480
4481                                 sh->ops.target = -1;
4482                                 sh->ops.target2 = -1;
4483                                 sh->check_state = check_state_compute_run;
4484                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4485                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4486                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4487                                         set_bit(R5_Wantcompute,
4488                                                 &sh->dev[pd_idx].flags);
4489                                         *target = pd_idx;
4490                                         target = &sh->ops.target2;
4491                                         s->uptodate++;
4492                                 }
4493                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4494                                         set_bit(R5_Wantcompute,
4495                                                 &sh->dev[qd_idx].flags);
4496                                         *target = qd_idx;
4497                                         s->uptodate++;
4498                                 }
4499                         }
4500                 }
4501                 break;
4502         case check_state_compute_run:
4503                 break;
4504         default:
4505                 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4506                         __func__, sh->check_state,
4507                         (unsigned long long) sh->sector);
4508                 BUG();
4509         }
4510 }
4511
4512 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4513 {
4514         int i;
4515
4516         /* We have read all the blocks in this stripe and now we need to
4517          * copy some of them into a target stripe for expand.
4518          */
4519         struct dma_async_tx_descriptor *tx = NULL;
4520         BUG_ON(sh->batch_head);
4521         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4522         for (i = 0; i < sh->disks; i++)
4523                 if (i != sh->pd_idx && i != sh->qd_idx) {
4524                         int dd_idx, j;
4525                         struct stripe_head *sh2;
4526                         struct async_submit_ctl submit;
4527
4528                         sector_t bn = raid5_compute_blocknr(sh, i, 1);
4529                         sector_t s = raid5_compute_sector(conf, bn, 0,
4530                                                           &dd_idx, NULL);
4531                         sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4532                         if (sh2 == NULL)
4533                                 /* so far only the early blocks of this stripe
4534                                  * have been requested.  When later blocks
4535                                  * get requested, we will try again
4536                                  */
4537                                 continue;
4538                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4539                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4540                                 /* must have already done this block */
4541                                 raid5_release_stripe(sh2);
4542                                 continue;
4543                         }
4544
4545                         /* place all the copies on one channel */
4546                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4547                         tx = async_memcpy(sh2->dev[dd_idx].page,
4548                                           sh->dev[i].page, sh2->dev[dd_idx].offset,
4549                                           sh->dev[i].offset, RAID5_STRIPE_SIZE(conf),
4550                                           &submit);
4551
4552                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4553                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4554                         for (j = 0; j < conf->raid_disks; j++)
4555                                 if (j != sh2->pd_idx &&
4556                                     j != sh2->qd_idx &&
4557                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
4558                                         break;
4559                         if (j == conf->raid_disks) {
4560                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4561                                 set_bit(STRIPE_HANDLE, &sh2->state);
4562                         }
4563                         raid5_release_stripe(sh2);
4564
4565                 }
4566         /* done submitting copies, wait for them to complete */
4567         async_tx_quiesce(&tx);
4568 }
4569
4570 /*
4571  * handle_stripe - do things to a stripe.
4572  *
4573  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4574  * state of various bits to see what needs to be done.
4575  * Possible results:
4576  *    return some read requests which now have data
4577  *    return some write requests which are safely on storage
4578  *    schedule a read on some buffers
4579  *    schedule a write of some buffers
4580  *    return confirmation of parity correctness
4581  *
4582  */
4583
4584 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4585 {
4586         struct r5conf *conf = sh->raid_conf;
4587         int disks = sh->disks;
4588         struct r5dev *dev;
4589         int i;
4590         int do_recovery = 0;
4591
4592         memset(s, 0, sizeof(*s));
4593
4594         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4595         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4596         s->failed_num[0] = -1;
4597         s->failed_num[1] = -1;
4598         s->log_failed = r5l_log_disk_error(conf);
4599
4600         /* Now to look around and see what can be done */
4601         rcu_read_lock();
4602         for (i=disks; i--; ) {
4603                 struct md_rdev *rdev;
4604                 sector_t first_bad;
4605                 int bad_sectors;
4606                 int is_bad = 0;
4607
4608                 dev = &sh->dev[i];
4609
4610                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4611                          i, dev->flags,
4612                          dev->toread, dev->towrite, dev->written);
4613                 /* maybe we can reply to a read
4614                  *
4615                  * new wantfill requests are only permitted while
4616                  * ops_complete_biofill is guaranteed to be inactive
4617                  */
4618                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4619                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4620                         set_bit(R5_Wantfill, &dev->flags);
4621
4622                 /* now count some things */
4623                 if (test_bit(R5_LOCKED, &dev->flags))
4624                         s->locked++;
4625                 if (test_bit(R5_UPTODATE, &dev->flags))
4626                         s->uptodate++;
4627                 if (test_bit(R5_Wantcompute, &dev->flags)) {
4628                         s->compute++;
4629                         BUG_ON(s->compute > 2);
4630                 }
4631
4632                 if (test_bit(R5_Wantfill, &dev->flags))
4633                         s->to_fill++;
4634                 else if (dev->toread)
4635                         s->to_read++;
4636                 if (dev->towrite) {
4637                         s->to_write++;
4638                         if (!test_bit(R5_OVERWRITE, &dev->flags))
4639                                 s->non_overwrite++;
4640                 }
4641                 if (dev->written)
4642                         s->written++;
4643                 /* Prefer to use the replacement for reads, but only
4644                  * if it is recovered enough and has no bad blocks.
4645                  */
4646                 rdev = rcu_dereference(conf->disks[i].replacement);
4647                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4648                     rdev->recovery_offset >= sh->sector + RAID5_STRIPE_SECTORS(conf) &&
4649                     !is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4650                                  &first_bad, &bad_sectors))
4651                         set_bit(R5_ReadRepl, &dev->flags);
4652                 else {
4653                         if (rdev && !test_bit(Faulty, &rdev->flags))
4654                                 set_bit(R5_NeedReplace, &dev->flags);
4655                         else
4656                                 clear_bit(R5_NeedReplace, &dev->flags);
4657                         rdev = rcu_dereference(conf->disks[i].rdev);
4658                         clear_bit(R5_ReadRepl, &dev->flags);
4659                 }
4660                 if (rdev && test_bit(Faulty, &rdev->flags))
4661                         rdev = NULL;
4662                 if (rdev) {
4663                         is_bad = is_badblock(rdev, sh->sector, RAID5_STRIPE_SECTORS(conf),
4664                                              &first_bad, &bad_sectors);
4665                         if (s->blocked_rdev == NULL
4666                             && (test_bit(Blocked, &rdev->flags)
4667                                 || is_bad < 0)) {
4668                                 if (is_bad < 0)
4669                                         set_bit(BlockedBadBlocks,
4670                                                 &rdev->flags);
4671                                 s->blocked_rdev = rdev;
4672                                 atomic_inc(&rdev->nr_pending);
4673                         }
4674                 }
4675                 clear_bit(R5_Insync, &dev->flags);
4676                 if (!rdev)
4677                         /* Not in-sync */;
4678                 else if (is_bad) {
4679                         /* also not in-sync */
4680                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4681                             test_bit(R5_UPTODATE, &dev->flags)) {
4682                                 /* treat as in-sync, but with a read error
4683                                  * which we can now try to correct
4684                                  */
4685                                 set_bit(R5_Insync, &dev->flags);
4686                                 set_bit(R5_ReadError, &dev->flags);
4687                         }
4688                 } else if (test_bit(In_sync, &rdev->flags))
4689                         set_bit(R5_Insync, &dev->flags);
4690                 else if (sh->sector + RAID5_STRIPE_SECTORS(conf) <= rdev->recovery_offset)
4691                         /* in sync if before recovery_offset */
4692                         set_bit(R5_Insync, &dev->flags);
4693                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4694                          test_bit(R5_Expanded, &dev->flags))
4695                         /* If we've reshaped into here, we assume it is Insync.
4696                          * We will shortly update recovery_offset to make
4697                          * it official.
4698                          */
4699                         set_bit(R5_Insync, &dev->flags);
4700
4701                 if (test_bit(R5_WriteError, &dev->flags)) {
4702                         /* This flag does not apply to '.replacement'
4703                          * only to .rdev, so make sure to check that*/
4704                         struct md_rdev *rdev2 = rcu_dereference(
4705                                 conf->disks[i].rdev);
4706                         if (rdev2 == rdev)
4707                                 clear_bit(R5_Insync, &dev->flags);
4708                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4709                                 s->handle_bad_blocks = 1;
4710                                 atomic_inc(&rdev2->nr_pending);
4711                         } else
4712                                 clear_bit(R5_WriteError, &dev->flags);
4713                 }
4714                 if (test_bit(R5_MadeGood, &dev->flags)) {
4715                         /* This flag does not apply to '.replacement'
4716                          * only to .rdev, so make sure to check that*/
4717                         struct md_rdev *rdev2 = rcu_dereference(
4718                                 conf->disks[i].rdev);
4719                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4720                                 s->handle_bad_blocks = 1;
4721                                 atomic_inc(&rdev2->nr_pending);
4722                         } else
4723                                 clear_bit(R5_MadeGood, &dev->flags);
4724                 }
4725                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4726                         struct md_rdev *rdev2 = rcu_dereference(
4727                                 conf->disks[i].replacement);
4728                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4729                                 s->handle_bad_blocks = 1;
4730                                 atomic_inc(&rdev2->nr_pending);
4731                         } else
4732                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
4733                 }
4734                 if (!test_bit(R5_Insync, &dev->flags)) {
4735                         /* The ReadError flag will just be confusing now */
4736                         clear_bit(R5_ReadError, &dev->flags);
4737                         clear_bit(R5_ReWrite, &dev->flags);
4738                 }
4739                 if (test_bit(R5_ReadError, &dev->flags))
4740                         clear_bit(R5_Insync, &dev->flags);
4741                 if (!test_bit(R5_Insync, &dev->flags)) {
4742                         if (s->failed < 2)
4743                                 s->failed_num[s->failed] = i;
4744                         s->failed++;
4745                         if (rdev && !test_bit(Faulty, &rdev->flags))
4746                                 do_recovery = 1;
4747                         else if (!rdev) {
4748                                 rdev = rcu_dereference(
4749                                     conf->disks[i].replacement);
4750                                 if (rdev && !test_bit(Faulty, &rdev->flags))
4751                                         do_recovery = 1;
4752                         }
4753                 }
4754
4755                 if (test_bit(R5_InJournal, &dev->flags))
4756                         s->injournal++;
4757                 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4758                         s->just_cached++;
4759         }
4760         if (test_bit(STRIPE_SYNCING, &sh->state)) {
4761                 /* If there is a failed device being replaced,
4762                  *     we must be recovering.
4763                  * else if we are after recovery_cp, we must be syncing
4764                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4765                  * else we can only be replacing
4766                  * sync and recovery both need to read all devices, and so
4767                  * use the same flag.
4768                  */
4769                 if (do_recovery ||
4770                     sh->sector >= conf->mddev->recovery_cp ||
4771                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4772                         s->syncing = 1;
4773                 else
4774                         s->replacing = 1;
4775         }
4776         rcu_read_unlock();
4777 }
4778
4779 /*
4780  * Return '1' if this is a member of batch, or '0' if it is a lone stripe or
4781  * a head which can now be handled.
4782  */
4783 static int clear_batch_ready(struct stripe_head *sh)
4784 {
4785         struct stripe_head *tmp;
4786         if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4787                 return (sh->batch_head && sh->batch_head != sh);
4788         spin_lock(&sh->stripe_lock);
4789         if (!sh->batch_head) {
4790                 spin_unlock(&sh->stripe_lock);
4791                 return 0;
4792         }
4793
4794         /*
4795          * this stripe could be added to a batch list before we check
4796          * BATCH_READY, skips it
4797          */
4798         if (sh->batch_head != sh) {
4799                 spin_unlock(&sh->stripe_lock);
4800                 return 1;
4801         }
4802         spin_lock(&sh->batch_lock);
4803         list_for_each_entry(tmp, &sh->batch_list, batch_list)
4804                 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4805         spin_unlock(&sh->batch_lock);
4806         spin_unlock(&sh->stripe_lock);
4807
4808         /*
4809          * BATCH_READY is cleared, no new stripes can be added.
4810          * batch_list can be accessed without lock
4811          */
4812         return 0;
4813 }
4814
4815 static void break_stripe_batch_list(struct stripe_head *head_sh,
4816                                     unsigned long handle_flags)
4817 {
4818         struct stripe_head *sh, *next;
4819         int i;
4820         int do_wakeup = 0;
4821
4822         list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4823
4824                 list_del_init(&sh->batch_list);
4825
4826                 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4827                                           (1 << STRIPE_SYNCING) |
4828                                           (1 << STRIPE_REPLACED) |
4829                                           (1 << STRIPE_DELAYED) |
4830                                           (1 << STRIPE_BIT_DELAY) |
4831                                           (1 << STRIPE_FULL_WRITE) |
4832                                           (1 << STRIPE_BIOFILL_RUN) |
4833                                           (1 << STRIPE_COMPUTE_RUN)  |
4834                                           (1 << STRIPE_DISCARD) |
4835                                           (1 << STRIPE_BATCH_READY) |
4836                                           (1 << STRIPE_BATCH_ERR) |
4837                                           (1 << STRIPE_BITMAP_PENDING)),
4838                         "stripe state: %lx\n", sh->state);
4839                 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4840                                               (1 << STRIPE_REPLACED)),
4841                         "head stripe state: %lx\n", head_sh->state);
4842
4843                 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4844                                             (1 << STRIPE_PREREAD_ACTIVE) |
4845                                             (1 << STRIPE_DEGRADED) |
4846                                             (1 << STRIPE_ON_UNPLUG_LIST)),
4847                               head_sh->state & (1 << STRIPE_INSYNC));
4848
4849                 sh->check_state = head_sh->check_state;
4850                 sh->reconstruct_state = head_sh->reconstruct_state;
4851                 spin_lock_irq(&sh->stripe_lock);
4852                 sh->batch_head = NULL;
4853                 spin_unlock_irq(&sh->stripe_lock);
4854                 for (i = 0; i < sh->disks; i++) {
4855                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4856                                 do_wakeup = 1;
4857                         sh->dev[i].flags = head_sh->dev[i].flags &
4858                                 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4859                 }
4860                 if (handle_flags == 0 ||
4861                     sh->state & handle_flags)
4862                         set_bit(STRIPE_HANDLE, &sh->state);
4863                 raid5_release_stripe(sh);
4864         }
4865         spin_lock_irq(&head_sh->stripe_lock);
4866         head_sh->batch_head = NULL;
4867         spin_unlock_irq(&head_sh->stripe_lock);
4868         for (i = 0; i < head_sh->disks; i++)
4869                 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4870                         do_wakeup = 1;
4871         if (head_sh->state & handle_flags)
4872                 set_bit(STRIPE_HANDLE, &head_sh->state);
4873
4874         if (do_wakeup)
4875                 wake_up(&head_sh->raid_conf->wait_for_overlap);
4876 }
4877
4878 static void handle_stripe(struct stripe_head *sh)
4879 {
4880         struct stripe_head_state s;
4881         struct r5conf *conf = sh->raid_conf;
4882         int i;
4883         int prexor;
4884         int disks = sh->disks;
4885         struct r5dev *pdev, *qdev;
4886
4887         clear_bit(STRIPE_HANDLE, &sh->state);
4888
4889         /*
4890          * handle_stripe should not continue handle the batched stripe, only
4891          * the head of batch list or lone stripe can continue. Otherwise we
4892          * could see break_stripe_batch_list warns about the STRIPE_ACTIVE
4893          * is set for the batched stripe.
4894          */
4895         if (clear_batch_ready(sh))
4896                 return;
4897
4898         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4899                 /* already being handled, ensure it gets handled
4900                  * again when current action finishes */
4901                 set_bit(STRIPE_HANDLE, &sh->state);
4902                 return;
4903         }
4904
4905         if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4906                 break_stripe_batch_list(sh, 0);
4907
4908         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4909                 spin_lock(&sh->stripe_lock);
4910                 /*
4911                  * Cannot process 'sync' concurrently with 'discard'.
4912                  * Flush data in r5cache before 'sync'.
4913                  */
4914                 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4915                     !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4916                     !test_bit(STRIPE_DISCARD, &sh->state) &&
4917                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4918                         set_bit(STRIPE_SYNCING, &sh->state);
4919                         clear_bit(STRIPE_INSYNC, &sh->state);
4920                         clear_bit(STRIPE_REPLACED, &sh->state);
4921                 }
4922                 spin_unlock(&sh->stripe_lock);
4923         }
4924         clear_bit(STRIPE_DELAYED, &sh->state);
4925
4926         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4927                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4928                (unsigned long long)sh->sector, sh->state,
4929                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4930                sh->check_state, sh->reconstruct_state);
4931
4932         analyse_stripe(sh, &s);
4933
4934         if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4935                 goto finish;
4936
4937         if (s.handle_bad_blocks ||
4938             test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4939                 set_bit(STRIPE_HANDLE, &sh->state);
4940                 goto finish;
4941         }
4942
4943         if (unlikely(s.blocked_rdev)) {
4944                 if (s.syncing || s.expanding || s.expanded ||
4945                     s.replacing || s.to_write || s.written) {
4946                         set_bit(STRIPE_HANDLE, &sh->state);
4947                         goto finish;
4948                 }
4949                 /* There is nothing for the blocked_rdev to block */
4950                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4951                 s.blocked_rdev = NULL;
4952         }
4953
4954         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4955                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4956                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4957         }
4958
4959         pr_debug("locked=%d uptodate=%d to_read=%d"
4960                " to_write=%d failed=%d failed_num=%d,%d\n",
4961                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4962                s.failed_num[0], s.failed_num[1]);
4963         /*
4964          * check if the array has lost more than max_degraded devices and,
4965          * if so, some requests might need to be failed.
4966          *
4967          * When journal device failed (log_failed), we will only process
4968          * the stripe if there is data need write to raid disks
4969          */
4970         if (s.failed > conf->max_degraded ||
4971             (s.log_failed && s.injournal == 0)) {
4972                 sh->check_state = 0;
4973                 sh->reconstruct_state = 0;
4974                 break_stripe_batch_list(sh, 0);
4975                 if (s.to_read+s.to_write+s.written)
4976                         handle_failed_stripe(conf, sh, &s, disks);
4977                 if (s.syncing + s.replacing)
4978                         handle_failed_sync(conf, sh, &s);
4979         }
4980
4981         /* Now we check to see if any write operations have recently
4982          * completed
4983          */
4984         prexor = 0;
4985         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4986                 prexor = 1;
4987         if (sh->reconstruct_state == reconstruct_state_drain_result ||
4988             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4989                 sh->reconstruct_state = reconstruct_state_idle;
4990
4991                 /* All the 'written' buffers and the parity block are ready to
4992                  * be written back to disk
4993                  */
4994                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4995                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4996                 BUG_ON(sh->qd_idx >= 0 &&
4997                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4998                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4999                 for (i = disks; i--; ) {
5000                         struct r5dev *dev = &sh->dev[i];
5001                         if (test_bit(R5_LOCKED, &dev->flags) &&
5002                                 (i == sh->pd_idx || i == sh->qd_idx ||
5003                                  dev->written || test_bit(R5_InJournal,
5004                                                           &dev->flags))) {
5005                                 pr_debug("Writing block %d\n", i);
5006                                 set_bit(R5_Wantwrite, &dev->flags);
5007                                 if (prexor)
5008                                         continue;
5009                                 if (s.failed > 1)
5010                                         continue;
5011                                 if (!test_bit(R5_Insync, &dev->flags) ||
5012                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
5013                                      s.failed == 0))
5014                                         set_bit(STRIPE_INSYNC, &sh->state);
5015                         }
5016                 }
5017                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5018                         s.dec_preread_active = 1;
5019         }
5020
5021         /*
5022          * might be able to return some write requests if the parity blocks
5023          * are safe, or on a failed drive
5024          */
5025         pdev = &sh->dev[sh->pd_idx];
5026         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
5027                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
5028         qdev = &sh->dev[sh->qd_idx];
5029         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
5030                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
5031                 || conf->level < 6;
5032
5033         if (s.written &&
5034             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
5035                              && !test_bit(R5_LOCKED, &pdev->flags)
5036                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
5037                                  test_bit(R5_Discard, &pdev->flags))))) &&
5038             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
5039                              && !test_bit(R5_LOCKED, &qdev->flags)
5040                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
5041                                  test_bit(R5_Discard, &qdev->flags))))))
5042                 handle_stripe_clean_event(conf, sh, disks);
5043
5044         if (s.just_cached)
5045                 r5c_handle_cached_data_endio(conf, sh, disks);
5046         log_stripe_write_finished(sh);
5047
5048         /* Now we might consider reading some blocks, either to check/generate
5049          * parity, or to satisfy requests
5050          * or to load a block that is being partially written.
5051          */
5052         if (s.to_read || s.non_overwrite
5053             || (s.to_write && s.failed)
5054             || (s.syncing && (s.uptodate + s.compute < disks))
5055             || s.replacing
5056             || s.expanding)
5057                 handle_stripe_fill(sh, &s, disks);
5058
5059         /*
5060          * When the stripe finishes full journal write cycle (write to journal
5061          * and raid disk), this is the clean up procedure so it is ready for
5062          * next operation.
5063          */
5064         r5c_finish_stripe_write_out(conf, sh, &s);
5065
5066         /*
5067          * Now to consider new write requests, cache write back and what else,
5068          * if anything should be read.  We do not handle new writes when:
5069          * 1/ A 'write' operation (copy+xor) is already in flight.
5070          * 2/ A 'check' operation is in flight, as it may clobber the parity
5071          *    block.
5072          * 3/ A r5c cache log write is in flight.
5073          */
5074
5075         if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
5076                 if (!r5c_is_writeback(conf->log)) {
5077                         if (s.to_write)
5078                                 handle_stripe_dirtying(conf, sh, &s, disks);
5079                 } else { /* write back cache */
5080                         int ret = 0;
5081
5082                         /* First, try handle writes in caching phase */
5083                         if (s.to_write)
5084                                 ret = r5c_try_caching_write(conf, sh, &s,
5085                                                             disks);
5086                         /*
5087                          * If caching phase failed: ret == -EAGAIN
5088                          *    OR
5089                          * stripe under reclaim: !caching && injournal
5090                          *
5091                          * fall back to handle_stripe_dirtying()
5092                          */
5093                         if (ret == -EAGAIN ||
5094                             /* stripe under reclaim: !caching && injournal */
5095                             (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
5096                              s.injournal > 0)) {
5097                                 ret = handle_stripe_dirtying(conf, sh, &s,
5098                                                              disks);
5099                                 if (ret == -EAGAIN)
5100                                         goto finish;
5101                         }
5102                 }
5103         }
5104
5105         /* maybe we need to check and possibly fix the parity for this stripe
5106          * Any reads will already have been scheduled, so we just see if enough
5107          * data is available.  The parity check is held off while parity
5108          * dependent operations are in flight.
5109          */
5110         if (sh->check_state ||
5111             (s.syncing && s.locked == 0 &&
5112              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5113              !test_bit(STRIPE_INSYNC, &sh->state))) {
5114                 if (conf->level == 6)
5115                         handle_parity_checks6(conf, sh, &s, disks);
5116                 else
5117                         handle_parity_checks5(conf, sh, &s, disks);
5118         }
5119
5120         if ((s.replacing || s.syncing) && s.locked == 0
5121             && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
5122             && !test_bit(STRIPE_REPLACED, &sh->state)) {
5123                 /* Write out to replacement devices where possible */
5124                 for (i = 0; i < conf->raid_disks; i++)
5125                         if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
5126                                 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
5127                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
5128                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
5129                                 s.locked++;
5130                         }
5131                 if (s.replacing)
5132                         set_bit(STRIPE_INSYNC, &sh->state);
5133                 set_bit(STRIPE_REPLACED, &sh->state);
5134         }
5135         if ((s.syncing || s.replacing) && s.locked == 0 &&
5136             !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
5137             test_bit(STRIPE_INSYNC, &sh->state)) {
5138                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5139                 clear_bit(STRIPE_SYNCING, &sh->state);
5140                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
5141                         wake_up(&conf->wait_for_overlap);
5142         }
5143
5144         /* If the failed drives are just a ReadError, then we might need
5145          * to progress the repair/check process
5146          */
5147         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
5148                 for (i = 0; i < s.failed; i++) {
5149                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
5150                         if (test_bit(R5_ReadError, &dev->flags)
5151                             && !test_bit(R5_LOCKED, &dev->flags)
5152                             && test_bit(R5_UPTODATE, &dev->flags)
5153                                 ) {
5154                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
5155                                         set_bit(R5_Wantwrite, &dev->flags);
5156                                         set_bit(R5_ReWrite, &dev->flags);
5157                                 } else
5158                                         /* let's read it back */
5159                                         set_bit(R5_Wantread, &dev->flags);
5160                                 set_bit(R5_LOCKED, &dev->flags);
5161                                 s.locked++;
5162                         }
5163                 }
5164
5165         /* Finish reconstruct operations initiated by the expansion process */
5166         if (sh->reconstruct_state == reconstruct_state_result) {
5167                 struct stripe_head *sh_src
5168                         = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
5169                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
5170                         /* sh cannot be written until sh_src has been read.
5171                          * so arrange for sh to be delayed a little
5172                          */
5173                         set_bit(STRIPE_DELAYED, &sh->state);
5174                         set_bit(STRIPE_HANDLE, &sh->state);
5175                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
5176                                               &sh_src->state))
5177                                 atomic_inc(&conf->preread_active_stripes);
5178                         raid5_release_stripe(sh_src);
5179                         goto finish;
5180                 }
5181                 if (sh_src)
5182                         raid5_release_stripe(sh_src);
5183
5184                 sh->reconstruct_state = reconstruct_state_idle;
5185                 clear_bit(STRIPE_EXPANDING, &sh->state);
5186                 for (i = conf->raid_disks; i--; ) {
5187                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
5188                         set_bit(R5_LOCKED, &sh->dev[i].flags);
5189                         s.locked++;
5190                 }
5191         }
5192
5193         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
5194             !sh->reconstruct_state) {
5195                 /* Need to write out all blocks after computing parity */
5196                 sh->disks = conf->raid_disks;
5197                 stripe_set_idx(sh->sector, conf, 0, sh);
5198                 schedule_reconstruction(sh, &s, 1, 1);
5199         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
5200                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
5201                 atomic_dec(&conf->reshape_stripes);
5202                 wake_up(&conf->wait_for_overlap);
5203                 md_done_sync(conf->mddev, RAID5_STRIPE_SECTORS(conf), 1);
5204         }
5205
5206         if (s.expanding && s.locked == 0 &&
5207             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
5208                 handle_stripe_expansion(conf, sh);
5209
5210 finish:
5211         /* wait for this device to become unblocked */
5212         if (unlikely(s.blocked_rdev)) {
5213                 if (conf->mddev->external)
5214                         md_wait_for_blocked_rdev(s.blocked_rdev,
5215                                                  conf->mddev);
5216                 else
5217                         /* Internal metadata will immediately
5218                          * be written by raid5d, so we don't
5219                          * need to wait here.
5220                          */
5221                         rdev_dec_pending(s.blocked_rdev,
5222                                          conf->mddev);
5223         }
5224
5225         if (s.handle_bad_blocks)
5226                 for (i = disks; i--; ) {
5227                         struct md_rdev *rdev;
5228                         struct r5dev *dev = &sh->dev[i];
5229                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5230                                 /* We own a safe reference to the rdev */
5231                                 rdev = conf->disks[i].rdev;
5232                                 if (!rdev_set_badblocks(rdev, sh->sector,
5233                                                         RAID5_STRIPE_SECTORS(conf), 0))
5234                                         md_error(conf->mddev, rdev);
5235                                 rdev_dec_pending(rdev, conf->mddev);
5236                         }
5237                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5238                                 rdev = conf->disks[i].rdev;
5239                                 rdev_clear_badblocks(rdev, sh->sector,
5240                                                      RAID5_STRIPE_SECTORS(conf), 0);
5241                                 rdev_dec_pending(rdev, conf->mddev);
5242                         }
5243                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5244                                 rdev = conf->disks[i].replacement;
5245                                 if (!rdev)
5246                                         /* rdev have been moved down */
5247                                         rdev = conf->disks[i].rdev;
5248                                 rdev_clear_badblocks(rdev, sh->sector,
5249                                                      RAID5_STRIPE_SECTORS(conf), 0);
5250                                 rdev_dec_pending(rdev, conf->mddev);
5251                         }
5252                 }
5253
5254         if (s.ops_request)
5255                 raid_run_ops(sh, s.ops_request);
5256
5257         ops_run_io(sh, &s);
5258
5259         if (s.dec_preread_active) {
5260                 /* We delay this until after ops_run_io so that if make_request
5261                  * is waiting on a flush, it won't continue until the writes
5262                  * have actually been submitted.
5263                  */
5264                 atomic_dec(&conf->preread_active_stripes);
5265                 if (atomic_read(&conf->preread_active_stripes) <
5266                     IO_THRESHOLD)
5267                         md_wakeup_thread(conf->mddev->thread);
5268         }
5269
5270         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5271 }
5272
5273 static void raid5_activate_delayed(struct r5conf *conf)
5274 {
5275         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5276                 while (!list_empty(&conf->delayed_list)) {
5277                         struct list_head *l = conf->delayed_list.next;
5278                         struct stripe_head *sh;
5279                         sh = list_entry(l, struct stripe_head, lru);
5280                         list_del_init(l);
5281                         clear_bit(STRIPE_DELAYED, &sh->state);
5282                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5283                                 atomic_inc(&conf->preread_active_stripes);
5284                         list_add_tail(&sh->lru, &conf->hold_list);
5285                         raid5_wakeup_stripe_thread(sh);
5286                 }
5287         }
5288 }
5289
5290 static void activate_bit_delay(struct r5conf *conf,
5291         struct list_head *temp_inactive_list)
5292 {
5293         /* device_lock is held */
5294         struct list_head head;
5295         list_add(&head, &conf->bitmap_list);
5296         list_del_init(&conf->bitmap_list);
5297         while (!list_empty(&head)) {
5298                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5299                 int hash;
5300                 list_del_init(&sh->lru);
5301                 atomic_inc(&sh->count);
5302                 hash = sh->hash_lock_index;
5303                 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5304         }
5305 }
5306
5307 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5308 {
5309         struct r5conf *conf = mddev->private;
5310         sector_t sector = bio->bi_iter.bi_sector;
5311         unsigned int chunk_sectors;
5312         unsigned int bio_sectors = bio_sectors(bio);
5313
5314         chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5315         return  chunk_sectors >=
5316                 ((sector & (chunk_sectors - 1)) + bio_sectors);
5317 }
5318
5319 /*
5320  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5321  *  later sampled by raid5d.
5322  */
5323 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5324 {
5325         unsigned long flags;
5326
5327         spin_lock_irqsave(&conf->device_lock, flags);
5328
5329         bi->bi_next = conf->retry_read_aligned_list;
5330         conf->retry_read_aligned_list = bi;
5331
5332         spin_unlock_irqrestore(&conf->device_lock, flags);
5333         md_wakeup_thread(conf->mddev->thread);
5334 }
5335
5336 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5337                                          unsigned int *offset)
5338 {
5339         struct bio *bi;
5340
5341         bi = conf->retry_read_aligned;
5342         if (bi) {
5343                 *offset = conf->retry_read_offset;
5344                 conf->retry_read_aligned = NULL;
5345                 return bi;
5346         }
5347         bi = conf->retry_read_aligned_list;
5348         if(bi) {
5349                 conf->retry_read_aligned_list = bi->bi_next;
5350                 bi->bi_next = NULL;
5351                 *offset = 0;
5352         }
5353
5354         return bi;
5355 }
5356
5357 /*
5358  *  The "raid5_align_endio" should check if the read succeeded and if it
5359  *  did, call bio_endio on the original bio (having bio_put the new bio
5360  *  first).
5361  *  If the read failed..
5362  */
5363 static void raid5_align_endio(struct bio *bi)
5364 {
5365         struct md_io_acct *md_io_acct = bi->bi_private;
5366         struct bio *raid_bi = md_io_acct->orig_bio;
5367         struct mddev *mddev;
5368         struct r5conf *conf;
5369         struct md_rdev *rdev;
5370         blk_status_t error = bi->bi_status;
5371         unsigned long start_time = md_io_acct->start_time;
5372
5373         bio_put(bi);
5374
5375         rdev = (void*)raid_bi->bi_next;
5376         raid_bi->bi_next = NULL;
5377         mddev = rdev->mddev;
5378         conf = mddev->private;
5379
5380         rdev_dec_pending(rdev, conf->mddev);
5381
5382         if (!error) {
5383                 if (blk_queue_io_stat(raid_bi->bi_bdev->bd_disk->queue))
5384                         bio_end_io_acct(raid_bi, start_time);
5385                 bio_endio(raid_bi);
5386                 if (atomic_dec_and_test(&conf->active_aligned_reads))
5387                         wake_up(&conf->wait_for_quiescent);
5388                 return;
5389         }
5390
5391         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5392
5393         add_bio_to_retry(raid_bi, conf);
5394 }
5395
5396 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5397 {
5398         struct r5conf *conf = mddev->private;
5399         struct bio *align_bio;
5400         struct md_rdev *rdev;
5401         sector_t sector, end_sector, first_bad;
5402         int bad_sectors, dd_idx;
5403         struct md_io_acct *md_io_acct;
5404         bool did_inc;
5405
5406         if (!in_chunk_boundary(mddev, raid_bio)) {
5407                 pr_debug("%s: non aligned\n", __func__);
5408                 return 0;
5409         }
5410
5411         sector = raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector, 0,
5412                                       &dd_idx, NULL);
5413         end_sector = bio_end_sector(raid_bio);
5414
5415         rcu_read_lock();
5416         if (r5c_big_stripe_cached(conf, sector))
5417                 goto out_rcu_unlock;
5418
5419         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5420         if (!rdev || test_bit(Faulty, &rdev->flags) ||
5421             rdev->recovery_offset < end_sector) {
5422                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5423                 if (!rdev)
5424                         goto out_rcu_unlock;
5425                 if (test_bit(Faulty, &rdev->flags) ||
5426                     !(test_bit(In_sync, &rdev->flags) ||
5427                       rdev->recovery_offset >= end_sector))
5428                         goto out_rcu_unlock;
5429         }
5430
5431         atomic_inc(&rdev->nr_pending);
5432         rcu_read_unlock();
5433
5434         if (is_badblock(rdev, sector, bio_sectors(raid_bio), &first_bad,
5435                         &bad_sectors)) {
5436                 bio_put(raid_bio);
5437                 rdev_dec_pending(rdev, mddev);
5438                 return 0;
5439         }
5440
5441         align_bio = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->io_acct_set);
5442         md_io_acct = container_of(align_bio, struct md_io_acct, bio_clone);
5443         raid_bio->bi_next = (void *)rdev;
5444         if (blk_queue_io_stat(raid_bio->bi_bdev->bd_disk->queue))
5445                 md_io_acct->start_time = bio_start_io_acct(raid_bio);
5446         md_io_acct->orig_bio = raid_bio;
5447
5448         bio_set_dev(align_bio, rdev->bdev);
5449         align_bio->bi_end_io = raid5_align_endio;
5450         align_bio->bi_private = md_io_acct;
5451         align_bio->bi_iter.bi_sector = sector;
5452
5453         /* No reshape active, so we can trust rdev->data_offset */
5454         align_bio->bi_iter.bi_sector += rdev->data_offset;
5455
5456         did_inc = false;
5457         if (conf->quiesce == 0) {
5458                 atomic_inc(&conf->active_aligned_reads);
5459                 did_inc = true;
5460         }
5461         /* need a memory barrier to detect the race with raid5_quiesce() */
5462         if (!did_inc || smp_load_acquire(&conf->quiesce) != 0) {
5463                 /* quiesce is in progress, so we need to undo io activation and wait
5464                  * for it to finish
5465                  */
5466                 if (did_inc && atomic_dec_and_test(&conf->active_aligned_reads))
5467                         wake_up(&conf->wait_for_quiescent);
5468                 spin_lock_irq(&conf->device_lock);
5469                 wait_event_lock_irq(conf->wait_for_quiescent, conf->quiesce == 0,
5470                                     conf->device_lock);
5471                 atomic_inc(&conf->active_aligned_reads);
5472                 spin_unlock_irq(&conf->device_lock);
5473         }
5474
5475         if (mddev->gendisk)
5476                 trace_block_bio_remap(align_bio, disk_devt(mddev->gendisk),
5477                                       raid_bio->bi_iter.bi_sector);
5478         submit_bio_noacct(align_bio);
5479         return 1;
5480
5481 out_rcu_unlock:
5482         rcu_read_unlock();
5483         return 0;
5484 }
5485
5486 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5487 {
5488         struct bio *split;
5489         sector_t sector = raid_bio->bi_iter.bi_sector;
5490         unsigned chunk_sects = mddev->chunk_sectors;
5491         unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5492
5493         if (sectors < bio_sectors(raid_bio)) {
5494                 struct r5conf *conf = mddev->private;
5495                 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5496                 bio_chain(split, raid_bio);
5497                 submit_bio_noacct(raid_bio);
5498                 raid_bio = split;
5499         }
5500
5501         if (!raid5_read_one_chunk(mddev, raid_bio))
5502                 return raid_bio;
5503
5504         return NULL;
5505 }
5506
5507 /* __get_priority_stripe - get the next stripe to process
5508  *
5509  * Full stripe writes are allowed to pass preread active stripes up until
5510  * the bypass_threshold is exceeded.  In general the bypass_count
5511  * increments when the handle_list is handled before the hold_list; however, it
5512  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5513  * stripe with in flight i/o.  The bypass_count will be reset when the
5514  * head of the hold_list has changed, i.e. the head was promoted to the
5515  * handle_list.
5516  */
5517 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5518 {
5519         struct stripe_head *sh, *tmp;
5520         struct list_head *handle_list = NULL;
5521         struct r5worker_group *wg;
5522         bool second_try = !r5c_is_writeback(conf->log) &&
5523                 !r5l_log_disk_error(conf);
5524         bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5525                 r5l_log_disk_error(conf);
5526
5527 again:
5528         wg = NULL;
5529         sh = NULL;
5530         if (conf->worker_cnt_per_group == 0) {
5531                 handle_list = try_loprio ? &conf->loprio_list :
5532                                         &conf->handle_list;
5533         } else if (group != ANY_GROUP) {
5534                 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5535                                 &conf->worker_groups[group].handle_list;
5536                 wg = &conf->worker_groups[group];
5537         } else {
5538                 int i;
5539                 for (i = 0; i < conf->group_cnt; i++) {
5540                         handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5541                                 &conf->worker_groups[i].handle_list;
5542                         wg = &conf->worker_groups[i];
5543                         if (!list_empty(handle_list))
5544                                 break;
5545                 }
5546         }
5547
5548         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5549                   __func__,
5550                   list_empty(handle_list) ? "empty" : "busy",
5551                   list_empty(&conf->hold_list) ? "empty" : "busy",
5552                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
5553
5554         if (!list_empty(handle_list)) {
5555                 sh = list_entry(handle_list->next, typeof(*sh), lru);
5556
5557                 if (list_empty(&conf->hold_list))
5558                         conf->bypass_count = 0;
5559                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5560                         if (conf->hold_list.next == conf->last_hold)
5561                                 conf->bypass_count++;
5562                         else {
5563                                 conf->last_hold = conf->hold_list.next;
5564                                 conf->bypass_count -= conf->bypass_threshold;
5565                                 if (conf->bypass_count < 0)
5566                                         conf->bypass_count = 0;
5567                         }
5568                 }
5569         } else if (!list_empty(&conf->hold_list) &&
5570                    ((conf->bypass_threshold &&
5571                      conf->bypass_count > conf->bypass_threshold) ||
5572                     atomic_read(&conf->pending_full_writes) == 0)) {
5573
5574                 list_for_each_entry(tmp, &conf->hold_list,  lru) {
5575                         if (conf->worker_cnt_per_group == 0 ||
5576                             group == ANY_GROUP ||
5577                             !cpu_online(tmp->cpu) ||
5578                             cpu_to_group(tmp->cpu) == group) {
5579                                 sh = tmp;
5580                                 break;
5581                         }
5582                 }
5583
5584                 if (sh) {
5585                         conf->bypass_count -= conf->bypass_threshold;
5586                         if (conf->bypass_count < 0)
5587                                 conf->bypass_count = 0;
5588                 }
5589                 wg = NULL;
5590         }
5591
5592         if (!sh) {
5593                 if (second_try)
5594                         return NULL;
5595                 second_try = true;
5596                 try_loprio = !try_loprio;
5597                 goto again;
5598         }
5599
5600         if (wg) {
5601                 wg->stripes_cnt--;
5602                 sh->group = NULL;
5603         }
5604         list_del_init(&sh->lru);
5605         BUG_ON(atomic_inc_return(&sh->count) != 1);
5606         return sh;
5607 }
5608
5609 struct raid5_plug_cb {
5610         struct blk_plug_cb      cb;
5611         struct list_head        list;
5612         struct list_head        temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5613 };
5614
5615 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5616 {
5617         struct raid5_plug_cb *cb = container_of(
5618                 blk_cb, struct raid5_plug_cb, cb);
5619         struct stripe_head *sh;
5620         struct mddev *mddev = cb->cb.data;
5621         struct r5conf *conf = mddev->private;
5622         int cnt = 0;
5623         int hash;
5624
5625         if (cb->list.next && !list_empty(&cb->list)) {
5626                 spin_lock_irq(&conf->device_lock);
5627                 while (!list_empty(&cb->list)) {
5628                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
5629                         list_del_init(&sh->lru);
5630                         /*
5631                          * avoid race release_stripe_plug() sees
5632                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
5633                          * is still in our list
5634                          */
5635                         smp_mb__before_atomic();
5636                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5637                         /*
5638                          * STRIPE_ON_RELEASE_LIST could be set here. In that
5639                          * case, the count is always > 1 here
5640                          */
5641                         hash = sh->hash_lock_index;
5642                         __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5643                         cnt++;
5644                 }
5645                 spin_unlock_irq(&conf->device_lock);
5646         }
5647         release_inactive_stripe_list(conf, cb->temp_inactive_list,
5648                                      NR_STRIPE_HASH_LOCKS);
5649         if (mddev->queue)
5650                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5651         kfree(cb);
5652 }
5653
5654 static void release_stripe_plug(struct mddev *mddev,
5655                                 struct stripe_head *sh)
5656 {
5657         struct blk_plug_cb *blk_cb = blk_check_plugged(
5658                 raid5_unplug, mddev,
5659                 sizeof(struct raid5_plug_cb));
5660         struct raid5_plug_cb *cb;
5661
5662         if (!blk_cb) {
5663                 raid5_release_stripe(sh);
5664                 return;
5665         }
5666
5667         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5668
5669         if (cb->list.next == NULL) {
5670                 int i;
5671                 INIT_LIST_HEAD(&cb->list);
5672                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5673                         INIT_LIST_HEAD(cb->temp_inactive_list + i);
5674         }
5675
5676         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5677                 list_add_tail(&sh->lru, &cb->list);
5678         else
5679                 raid5_release_stripe(sh);
5680 }
5681
5682 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5683 {
5684         struct r5conf *conf = mddev->private;
5685         sector_t logical_sector, last_sector;
5686         struct stripe_head *sh;
5687         int stripe_sectors;
5688
5689         if (mddev->reshape_position != MaxSector)
5690                 /* Skip discard while reshape is happening */
5691                 return;
5692
5693         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5694         last_sector = bio_end_sector(bi);
5695
5696         bi->bi_next = NULL;
5697
5698         stripe_sectors = conf->chunk_sectors *
5699                 (conf->raid_disks - conf->max_degraded);
5700         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5701                                                stripe_sectors);
5702         sector_div(last_sector, stripe_sectors);
5703
5704         logical_sector *= conf->chunk_sectors;
5705         last_sector *= conf->chunk_sectors;
5706
5707         for (; logical_sector < last_sector;
5708              logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5709                 DEFINE_WAIT(w);
5710                 int d;
5711         again:
5712                 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5713                 prepare_to_wait(&conf->wait_for_overlap, &w,
5714                                 TASK_UNINTERRUPTIBLE);
5715                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5716                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5717                         raid5_release_stripe(sh);
5718                         schedule();
5719                         goto again;
5720                 }
5721                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5722                 spin_lock_irq(&sh->stripe_lock);
5723                 for (d = 0; d < conf->raid_disks; d++) {
5724                         if (d == sh->pd_idx || d == sh->qd_idx)
5725                                 continue;
5726                         if (sh->dev[d].towrite || sh->dev[d].toread) {
5727                                 set_bit(R5_Overlap, &sh->dev[d].flags);
5728                                 spin_unlock_irq(&sh->stripe_lock);
5729                                 raid5_release_stripe(sh);
5730                                 schedule();
5731                                 goto again;
5732                         }
5733                 }
5734                 set_bit(STRIPE_DISCARD, &sh->state);
5735                 finish_wait(&conf->wait_for_overlap, &w);
5736                 sh->overwrite_disks = 0;
5737                 for (d = 0; d < conf->raid_disks; d++) {
5738                         if (d == sh->pd_idx || d == sh->qd_idx)
5739                                 continue;
5740                         sh->dev[d].towrite = bi;
5741                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5742                         bio_inc_remaining(bi);
5743                         md_write_inc(mddev, bi);
5744                         sh->overwrite_disks++;
5745                 }
5746                 spin_unlock_irq(&sh->stripe_lock);
5747                 if (conf->mddev->bitmap) {
5748                         for (d = 0;
5749                              d < conf->raid_disks - conf->max_degraded;
5750                              d++)
5751                                 md_bitmap_startwrite(mddev->bitmap,
5752                                                      sh->sector,
5753                                                      RAID5_STRIPE_SECTORS(conf),
5754                                                      0);
5755                         sh->bm_seq = conf->seq_flush + 1;
5756                         set_bit(STRIPE_BIT_DELAY, &sh->state);
5757                 }
5758
5759                 set_bit(STRIPE_HANDLE, &sh->state);
5760                 clear_bit(STRIPE_DELAYED, &sh->state);
5761                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5762                         atomic_inc(&conf->preread_active_stripes);
5763                 release_stripe_plug(mddev, sh);
5764         }
5765
5766         bio_endio(bi);
5767 }
5768
5769 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5770 {
5771         struct r5conf *conf = mddev->private;
5772         int dd_idx;
5773         sector_t new_sector;
5774         sector_t logical_sector, last_sector;
5775         struct stripe_head *sh;
5776         const int rw = bio_data_dir(bi);
5777         DEFINE_WAIT(w);
5778         bool do_prepare;
5779         bool do_flush = false;
5780
5781         if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5782                 int ret = log_handle_flush_request(conf, bi);
5783
5784                 if (ret == 0)
5785                         return true;
5786                 if (ret == -ENODEV) {
5787                         if (md_flush_request(mddev, bi))
5788                                 return true;
5789                 }
5790                 /* ret == -EAGAIN, fallback */
5791                 /*
5792                  * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5793                  * we need to flush journal device
5794                  */
5795                 do_flush = bi->bi_opf & REQ_PREFLUSH;
5796         }
5797
5798         if (!md_write_start(mddev, bi))
5799                 return false;
5800         /*
5801          * If array is degraded, better not do chunk aligned read because
5802          * later we might have to read it again in order to reconstruct
5803          * data on failed drives.
5804          */
5805         if (rw == READ && mddev->degraded == 0 &&
5806             mddev->reshape_position == MaxSector) {
5807                 bi = chunk_aligned_read(mddev, bi);
5808                 if (!bi)
5809                         return true;
5810         }
5811
5812         if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5813                 make_discard_request(mddev, bi);
5814                 md_write_end(mddev);
5815                 return true;
5816         }
5817
5818         logical_sector = bi->bi_iter.bi_sector & ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
5819         last_sector = bio_end_sector(bi);
5820         bi->bi_next = NULL;
5821
5822         md_account_bio(mddev, &bi);
5823         prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5824         for (; logical_sector < last_sector; logical_sector += RAID5_STRIPE_SECTORS(conf)) {
5825                 int previous;
5826                 int seq;
5827
5828                 do_prepare = false;
5829         retry:
5830                 seq = read_seqcount_begin(&conf->gen_lock);
5831                 previous = 0;
5832                 if (do_prepare)
5833                         prepare_to_wait(&conf->wait_for_overlap, &w,
5834                                 TASK_UNINTERRUPTIBLE);
5835                 if (unlikely(conf->reshape_progress != MaxSector)) {
5836                         /* spinlock is needed as reshape_progress may be
5837                          * 64bit on a 32bit platform, and so it might be
5838                          * possible to see a half-updated value
5839                          * Of course reshape_progress could change after
5840                          * the lock is dropped, so once we get a reference
5841                          * to the stripe that we think it is, we will have
5842                          * to check again.
5843                          */
5844                         spin_lock_irq(&conf->device_lock);
5845                         if (mddev->reshape_backwards
5846                             ? logical_sector < conf->reshape_progress
5847                             : logical_sector >= conf->reshape_progress) {
5848                                 previous = 1;
5849                         } else {
5850                                 if (mddev->reshape_backwards
5851                                     ? logical_sector < conf->reshape_safe
5852                                     : logical_sector >= conf->reshape_safe) {
5853                                         spin_unlock_irq(&conf->device_lock);
5854                                         schedule();
5855                                         do_prepare = true;
5856                                         goto retry;
5857                                 }
5858                         }
5859                         spin_unlock_irq(&conf->device_lock);
5860                 }
5861
5862                 new_sector = raid5_compute_sector(conf, logical_sector,
5863                                                   previous,
5864                                                   &dd_idx, NULL);
5865                 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5866                         (unsigned long long)new_sector,
5867                         (unsigned long long)logical_sector);
5868
5869                 sh = raid5_get_active_stripe(conf, new_sector, previous,
5870                                        (bi->bi_opf & REQ_RAHEAD), 0);
5871                 if (sh) {
5872                         if (unlikely(previous)) {
5873                                 /* expansion might have moved on while waiting for a
5874                                  * stripe, so we must do the range check again.
5875                                  * Expansion could still move past after this
5876                                  * test, but as we are holding a reference to
5877                                  * 'sh', we know that if that happens,
5878                                  *  STRIPE_EXPANDING will get set and the expansion
5879                                  * won't proceed until we finish with the stripe.
5880                                  */
5881                                 int must_retry = 0;
5882                                 spin_lock_irq(&conf->device_lock);
5883                                 if (mddev->reshape_backwards
5884                                     ? logical_sector >= conf->reshape_progress
5885                                     : logical_sector < conf->reshape_progress)
5886                                         /* mismatch, need to try again */
5887                                         must_retry = 1;
5888                                 spin_unlock_irq(&conf->device_lock);
5889                                 if (must_retry) {
5890                                         raid5_release_stripe(sh);
5891                                         schedule();
5892                                         do_prepare = true;
5893                                         goto retry;
5894                                 }
5895                         }
5896                         if (read_seqcount_retry(&conf->gen_lock, seq)) {
5897                                 /* Might have got the wrong stripe_head
5898                                  * by accident
5899                                  */
5900                                 raid5_release_stripe(sh);
5901                                 goto retry;
5902                         }
5903
5904                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5905                             !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5906                                 /* Stripe is busy expanding or
5907                                  * add failed due to overlap.  Flush everything
5908                                  * and wait a while
5909                                  */
5910                                 md_wakeup_thread(mddev->thread);
5911                                 raid5_release_stripe(sh);
5912                                 schedule();
5913                                 do_prepare = true;
5914                                 goto retry;
5915                         }
5916                         if (do_flush) {
5917                                 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5918                                 /* we only need flush for one stripe */
5919                                 do_flush = false;
5920                         }
5921
5922                         set_bit(STRIPE_HANDLE, &sh->state);
5923                         clear_bit(STRIPE_DELAYED, &sh->state);
5924                         if ((!sh->batch_head || sh == sh->batch_head) &&
5925                             (bi->bi_opf & REQ_SYNC) &&
5926                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5927                                 atomic_inc(&conf->preread_active_stripes);
5928                         release_stripe_plug(mddev, sh);
5929                 } else {
5930                         /* cannot get stripe for read-ahead, just give-up */
5931                         bi->bi_status = BLK_STS_IOERR;
5932                         break;
5933                 }
5934         }
5935         finish_wait(&conf->wait_for_overlap, &w);
5936
5937         if (rw == WRITE)
5938                 md_write_end(mddev);
5939         bio_endio(bi);
5940         return true;
5941 }
5942
5943 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5944
5945 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5946 {
5947         /* reshaping is quite different to recovery/resync so it is
5948          * handled quite separately ... here.
5949          *
5950          * On each call to sync_request, we gather one chunk worth of
5951          * destination stripes and flag them as expanding.
5952          * Then we find all the source stripes and request reads.
5953          * As the reads complete, handle_stripe will copy the data
5954          * into the destination stripe and release that stripe.
5955          */
5956         struct r5conf *conf = mddev->private;
5957         struct stripe_head *sh;
5958         struct md_rdev *rdev;
5959         sector_t first_sector, last_sector;
5960         int raid_disks = conf->previous_raid_disks;
5961         int data_disks = raid_disks - conf->max_degraded;
5962         int new_data_disks = conf->raid_disks - conf->max_degraded;
5963         int i;
5964         int dd_idx;
5965         sector_t writepos, readpos, safepos;
5966         sector_t stripe_addr;
5967         int reshape_sectors;
5968         struct list_head stripes;
5969         sector_t retn;
5970
5971         if (sector_nr == 0) {
5972                 /* If restarting in the middle, skip the initial sectors */
5973                 if (mddev->reshape_backwards &&
5974                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5975                         sector_nr = raid5_size(mddev, 0, 0)
5976                                 - conf->reshape_progress;
5977                 } else if (mddev->reshape_backwards &&
5978                            conf->reshape_progress == MaxSector) {
5979                         /* shouldn't happen, but just in case, finish up.*/
5980                         sector_nr = MaxSector;
5981                 } else if (!mddev->reshape_backwards &&
5982                            conf->reshape_progress > 0)
5983                         sector_nr = conf->reshape_progress;
5984                 sector_div(sector_nr, new_data_disks);
5985                 if (sector_nr) {
5986                         mddev->curr_resync_completed = sector_nr;
5987                         sysfs_notify_dirent_safe(mddev->sysfs_completed);
5988                         *skipped = 1;
5989                         retn = sector_nr;
5990                         goto finish;
5991                 }
5992         }
5993
5994         /* We need to process a full chunk at a time.
5995          * If old and new chunk sizes differ, we need to process the
5996          * largest of these
5997          */
5998
5999         reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
6000
6001         /* We update the metadata at least every 10 seconds, or when
6002          * the data about to be copied would over-write the source of
6003          * the data at the front of the range.  i.e. one new_stripe
6004          * along from reshape_progress new_maps to after where
6005          * reshape_safe old_maps to
6006          */
6007         writepos = conf->reshape_progress;
6008         sector_div(writepos, new_data_disks);
6009         readpos = conf->reshape_progress;
6010         sector_div(readpos, data_disks);
6011         safepos = conf->reshape_safe;
6012         sector_div(safepos, data_disks);
6013         if (mddev->reshape_backwards) {
6014                 BUG_ON(writepos < reshape_sectors);
6015                 writepos -= reshape_sectors;
6016                 readpos += reshape_sectors;
6017                 safepos += reshape_sectors;
6018         } else {
6019                 writepos += reshape_sectors;
6020                 /* readpos and safepos are worst-case calculations.
6021                  * A negative number is overly pessimistic, and causes
6022                  * obvious problems for unsigned storage.  So clip to 0.
6023                  */
6024                 readpos -= min_t(sector_t, reshape_sectors, readpos);
6025                 safepos -= min_t(sector_t, reshape_sectors, safepos);
6026         }
6027
6028         /* Having calculated the 'writepos' possibly use it
6029          * to set 'stripe_addr' which is where we will write to.
6030          */
6031         if (mddev->reshape_backwards) {
6032                 BUG_ON(conf->reshape_progress == 0);
6033                 stripe_addr = writepos;
6034                 BUG_ON((mddev->dev_sectors &
6035                         ~((sector_t)reshape_sectors - 1))
6036                        - reshape_sectors - stripe_addr
6037                        != sector_nr);
6038         } else {
6039                 BUG_ON(writepos != sector_nr + reshape_sectors);
6040                 stripe_addr = sector_nr;
6041         }
6042
6043         /* 'writepos' is the most advanced device address we might write.
6044          * 'readpos' is the least advanced device address we might read.
6045          * 'safepos' is the least address recorded in the metadata as having
6046          *     been reshaped.
6047          * If there is a min_offset_diff, these are adjusted either by
6048          * increasing the safepos/readpos if diff is negative, or
6049          * increasing writepos if diff is positive.
6050          * If 'readpos' is then behind 'writepos', there is no way that we can
6051          * ensure safety in the face of a crash - that must be done by userspace
6052          * making a backup of the data.  So in that case there is no particular
6053          * rush to update metadata.
6054          * Otherwise if 'safepos' is behind 'writepos', then we really need to
6055          * update the metadata to advance 'safepos' to match 'readpos' so that
6056          * we can be safe in the event of a crash.
6057          * So we insist on updating metadata if safepos is behind writepos and
6058          * readpos is beyond writepos.
6059          * In any case, update the metadata every 10 seconds.
6060          * Maybe that number should be configurable, but I'm not sure it is
6061          * worth it.... maybe it could be a multiple of safemode_delay???
6062          */
6063         if (conf->min_offset_diff < 0) {
6064                 safepos += -conf->min_offset_diff;
6065                 readpos += -conf->min_offset_diff;
6066         } else
6067                 writepos += conf->min_offset_diff;
6068
6069         if ((mddev->reshape_backwards
6070              ? (safepos > writepos && readpos < writepos)
6071              : (safepos < writepos && readpos > writepos)) ||
6072             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
6073                 /* Cannot proceed until we've updated the superblock... */
6074                 wait_event(conf->wait_for_overlap,
6075                            atomic_read(&conf->reshape_stripes)==0
6076                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6077                 if (atomic_read(&conf->reshape_stripes) != 0)
6078                         return 0;
6079                 mddev->reshape_position = conf->reshape_progress;
6080                 mddev->curr_resync_completed = sector_nr;
6081                 if (!mddev->reshape_backwards)
6082                         /* Can update recovery_offset */
6083                         rdev_for_each(rdev, mddev)
6084                                 if (rdev->raid_disk >= 0 &&
6085                                     !test_bit(Journal, &rdev->flags) &&
6086                                     !test_bit(In_sync, &rdev->flags) &&
6087                                     rdev->recovery_offset < sector_nr)
6088                                         rdev->recovery_offset = sector_nr;
6089
6090                 conf->reshape_checkpoint = jiffies;
6091                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6092                 md_wakeup_thread(mddev->thread);
6093                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
6094                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6095                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6096                         return 0;
6097                 spin_lock_irq(&conf->device_lock);
6098                 conf->reshape_safe = mddev->reshape_position;
6099                 spin_unlock_irq(&conf->device_lock);
6100                 wake_up(&conf->wait_for_overlap);
6101                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6102         }
6103
6104         INIT_LIST_HEAD(&stripes);
6105         for (i = 0; i < reshape_sectors; i += RAID5_STRIPE_SECTORS(conf)) {
6106                 int j;
6107                 int skipped_disk = 0;
6108                 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
6109                 set_bit(STRIPE_EXPANDING, &sh->state);
6110                 atomic_inc(&conf->reshape_stripes);
6111                 /* If any of this stripe is beyond the end of the old
6112                  * array, then we need to zero those blocks
6113                  */
6114                 for (j=sh->disks; j--;) {
6115                         sector_t s;
6116                         if (j == sh->pd_idx)
6117                                 continue;
6118                         if (conf->level == 6 &&
6119                             j == sh->qd_idx)
6120                                 continue;
6121                         s = raid5_compute_blocknr(sh, j, 0);
6122                         if (s < raid5_size(mddev, 0, 0)) {
6123                                 skipped_disk = 1;
6124                                 continue;
6125                         }
6126                         memset(page_address(sh->dev[j].page), 0, RAID5_STRIPE_SIZE(conf));
6127                         set_bit(R5_Expanded, &sh->dev[j].flags);
6128                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
6129                 }
6130                 if (!skipped_disk) {
6131                         set_bit(STRIPE_EXPAND_READY, &sh->state);
6132                         set_bit(STRIPE_HANDLE, &sh->state);
6133                 }
6134                 list_add(&sh->lru, &stripes);
6135         }
6136         spin_lock_irq(&conf->device_lock);
6137         if (mddev->reshape_backwards)
6138                 conf->reshape_progress -= reshape_sectors * new_data_disks;
6139         else
6140                 conf->reshape_progress += reshape_sectors * new_data_disks;
6141         spin_unlock_irq(&conf->device_lock);
6142         /* Ok, those stripe are ready. We can start scheduling
6143          * reads on the source stripes.
6144          * The source stripes are determined by mapping the first and last
6145          * block on the destination stripes.
6146          */
6147         first_sector =
6148                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
6149                                      1, &dd_idx, NULL);
6150         last_sector =
6151                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
6152                                             * new_data_disks - 1),
6153                                      1, &dd_idx, NULL);
6154         if (last_sector >= mddev->dev_sectors)
6155                 last_sector = mddev->dev_sectors - 1;
6156         while (first_sector <= last_sector) {
6157                 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
6158                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
6159                 set_bit(STRIPE_HANDLE, &sh->state);
6160                 raid5_release_stripe(sh);
6161                 first_sector += RAID5_STRIPE_SECTORS(conf);
6162         }
6163         /* Now that the sources are clearly marked, we can release
6164          * the destination stripes
6165          */
6166         while (!list_empty(&stripes)) {
6167                 sh = list_entry(stripes.next, struct stripe_head, lru);
6168                 list_del_init(&sh->lru);
6169                 raid5_release_stripe(sh);
6170         }
6171         /* If this takes us to the resync_max point where we have to pause,
6172          * then we need to write out the superblock.
6173          */
6174         sector_nr += reshape_sectors;
6175         retn = reshape_sectors;
6176 finish:
6177         if (mddev->curr_resync_completed > mddev->resync_max ||
6178             (sector_nr - mddev->curr_resync_completed) * 2
6179             >= mddev->resync_max - mddev->curr_resync_completed) {
6180                 /* Cannot proceed until we've updated the superblock... */
6181                 wait_event(conf->wait_for_overlap,
6182                            atomic_read(&conf->reshape_stripes) == 0
6183                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6184                 if (atomic_read(&conf->reshape_stripes) != 0)
6185                         goto ret;
6186                 mddev->reshape_position = conf->reshape_progress;
6187                 mddev->curr_resync_completed = sector_nr;
6188                 if (!mddev->reshape_backwards)
6189                         /* Can update recovery_offset */
6190                         rdev_for_each(rdev, mddev)
6191                                 if (rdev->raid_disk >= 0 &&
6192                                     !test_bit(Journal, &rdev->flags) &&
6193                                     !test_bit(In_sync, &rdev->flags) &&
6194                                     rdev->recovery_offset < sector_nr)
6195                                         rdev->recovery_offset = sector_nr;
6196                 conf->reshape_checkpoint = jiffies;
6197                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6198                 md_wakeup_thread(mddev->thread);
6199                 wait_event(mddev->sb_wait,
6200                            !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6201                            || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6202                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6203                         goto ret;
6204                 spin_lock_irq(&conf->device_lock);
6205                 conf->reshape_safe = mddev->reshape_position;
6206                 spin_unlock_irq(&conf->device_lock);
6207                 wake_up(&conf->wait_for_overlap);
6208                 sysfs_notify_dirent_safe(mddev->sysfs_completed);
6209         }
6210 ret:
6211         return retn;
6212 }
6213
6214 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6215                                           int *skipped)
6216 {
6217         struct r5conf *conf = mddev->private;
6218         struct stripe_head *sh;
6219         sector_t max_sector = mddev->dev_sectors;
6220         sector_t sync_blocks;
6221         int still_degraded = 0;
6222         int i;
6223
6224         if (sector_nr >= max_sector) {
6225                 /* just being told to finish up .. nothing much to do */
6226
6227                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6228                         end_reshape(conf);
6229                         return 0;
6230                 }
6231
6232                 if (mddev->curr_resync < max_sector) /* aborted */
6233                         md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6234                                            &sync_blocks, 1);
6235                 else /* completed sync */
6236                         conf->fullsync = 0;
6237                 md_bitmap_close_sync(mddev->bitmap);
6238
6239                 return 0;
6240         }
6241
6242         /* Allow raid5_quiesce to complete */
6243         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6244
6245         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6246                 return reshape_request(mddev, sector_nr, skipped);
6247
6248         /* No need to check resync_max as we never do more than one
6249          * stripe, and as resync_max will always be on a chunk boundary,
6250          * if the check in md_do_sync didn't fire, there is no chance
6251          * of overstepping resync_max here
6252          */
6253
6254         /* if there is too many failed drives and we are trying
6255          * to resync, then assert that we are finished, because there is
6256          * nothing we can do.
6257          */
6258         if (mddev->degraded >= conf->max_degraded &&
6259             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6260                 sector_t rv = mddev->dev_sectors - sector_nr;
6261                 *skipped = 1;
6262                 return rv;
6263         }
6264         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6265             !conf->fullsync &&
6266             !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6267             sync_blocks >= RAID5_STRIPE_SECTORS(conf)) {
6268                 /* we can skip this block, and probably more */
6269                 do_div(sync_blocks, RAID5_STRIPE_SECTORS(conf));
6270                 *skipped = 1;
6271                 /* keep things rounded to whole stripes */
6272                 return sync_blocks * RAID5_STRIPE_SECTORS(conf);
6273         }
6274
6275         md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6276
6277         sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6278         if (sh == NULL) {
6279                 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6280                 /* make sure we don't swamp the stripe cache if someone else
6281                  * is trying to get access
6282                  */
6283                 schedule_timeout_uninterruptible(1);
6284         }
6285         /* Need to check if array will still be degraded after recovery/resync
6286          * Note in case of > 1 drive failures it's possible we're rebuilding
6287          * one drive while leaving another faulty drive in array.
6288          */
6289         rcu_read_lock();
6290         for (i = 0; i < conf->raid_disks; i++) {
6291                 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6292
6293                 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6294                         still_degraded = 1;
6295         }
6296         rcu_read_unlock();
6297
6298         md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6299
6300         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6301         set_bit(STRIPE_HANDLE, &sh->state);
6302
6303         raid5_release_stripe(sh);
6304
6305         return RAID5_STRIPE_SECTORS(conf);
6306 }
6307
6308 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6309                                unsigned int offset)
6310 {
6311         /* We may not be able to submit a whole bio at once as there
6312          * may not be enough stripe_heads available.
6313          * We cannot pre-allocate enough stripe_heads as we may need
6314          * more than exist in the cache (if we allow ever large chunks).
6315          * So we do one stripe head at a time and record in
6316          * ->bi_hw_segments how many have been done.
6317          *
6318          * We *know* that this entire raid_bio is in one chunk, so
6319          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6320          */
6321         struct stripe_head *sh;
6322         int dd_idx;
6323         sector_t sector, logical_sector, last_sector;
6324         int scnt = 0;
6325         int handled = 0;
6326
6327         logical_sector = raid_bio->bi_iter.bi_sector &
6328                 ~((sector_t)RAID5_STRIPE_SECTORS(conf)-1);
6329         sector = raid5_compute_sector(conf, logical_sector,
6330                                       0, &dd_idx, NULL);
6331         last_sector = bio_end_sector(raid_bio);
6332
6333         for (; logical_sector < last_sector;
6334              logical_sector += RAID5_STRIPE_SECTORS(conf),
6335                      sector += RAID5_STRIPE_SECTORS(conf),
6336                      scnt++) {
6337
6338                 if (scnt < offset)
6339                         /* already done this stripe */
6340                         continue;
6341
6342                 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6343
6344                 if (!sh) {
6345                         /* failed to get a stripe - must wait */
6346                         conf->retry_read_aligned = raid_bio;
6347                         conf->retry_read_offset = scnt;
6348                         return handled;
6349                 }
6350
6351                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6352                         raid5_release_stripe(sh);
6353                         conf->retry_read_aligned = raid_bio;
6354                         conf->retry_read_offset = scnt;
6355                         return handled;
6356                 }
6357
6358                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6359                 handle_stripe(sh);
6360                 raid5_release_stripe(sh);
6361                 handled++;
6362         }
6363
6364         bio_endio(raid_bio);
6365
6366         if (atomic_dec_and_test(&conf->active_aligned_reads))
6367                 wake_up(&conf->wait_for_quiescent);
6368         return handled;
6369 }
6370
6371 static int handle_active_stripes(struct r5conf *conf, int group,
6372                                  struct r5worker *worker,
6373                                  struct list_head *temp_inactive_list)
6374                 __releases(&conf->device_lock)
6375                 __acquires(&conf->device_lock)
6376 {
6377         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6378         int i, batch_size = 0, hash;
6379         bool release_inactive = false;
6380
6381         while (batch_size < MAX_STRIPE_BATCH &&
6382                         (sh = __get_priority_stripe(conf, group)) != NULL)
6383                 batch[batch_size++] = sh;
6384
6385         if (batch_size == 0) {
6386                 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6387                         if (!list_empty(temp_inactive_list + i))
6388                                 break;
6389                 if (i == NR_STRIPE_HASH_LOCKS) {
6390                         spin_unlock_irq(&conf->device_lock);
6391                         log_flush_stripe_to_raid(conf);
6392                         spin_lock_irq(&conf->device_lock);
6393                         return batch_size;
6394                 }
6395                 release_inactive = true;
6396         }
6397         spin_unlock_irq(&conf->device_lock);
6398
6399         release_inactive_stripe_list(conf, temp_inactive_list,
6400                                      NR_STRIPE_HASH_LOCKS);
6401
6402         r5l_flush_stripe_to_raid(conf->log);
6403         if (release_inactive) {
6404                 spin_lock_irq(&conf->device_lock);
6405                 return 0;
6406         }
6407
6408         for (i = 0; i < batch_size; i++)
6409                 handle_stripe(batch[i]);
6410         log_write_stripe_run(conf);
6411
6412         cond_resched();
6413
6414         spin_lock_irq(&conf->device_lock);
6415         for (i = 0; i < batch_size; i++) {
6416                 hash = batch[i]->hash_lock_index;
6417                 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6418         }
6419         return batch_size;
6420 }
6421
6422 static void raid5_do_work(struct work_struct *work)
6423 {
6424         struct r5worker *worker = container_of(work, struct r5worker, work);
6425         struct r5worker_group *group = worker->group;
6426         struct r5conf *conf = group->conf;
6427         struct mddev *mddev = conf->mddev;
6428         int group_id = group - conf->worker_groups;
6429         int handled;
6430         struct blk_plug plug;
6431
6432         pr_debug("+++ raid5worker active\n");
6433
6434         blk_start_plug(&plug);
6435         handled = 0;
6436         spin_lock_irq(&conf->device_lock);
6437         while (1) {
6438                 int batch_size, released;
6439
6440                 released = release_stripe_list(conf, worker->temp_inactive_list);
6441
6442                 batch_size = handle_active_stripes(conf, group_id, worker,
6443                                                    worker->temp_inactive_list);
6444                 worker->working = false;
6445                 if (!batch_size && !released)
6446                         break;
6447                 handled += batch_size;
6448                 wait_event_lock_irq(mddev->sb_wait,
6449                         !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6450                         conf->device_lock);
6451         }
6452         pr_debug("%d stripes handled\n", handled);
6453
6454         spin_unlock_irq(&conf->device_lock);
6455
6456         flush_deferred_bios(conf);
6457
6458         r5l_flush_stripe_to_raid(conf->log);
6459
6460         async_tx_issue_pending_all();
6461         blk_finish_plug(&plug);
6462
6463         pr_debug("--- raid5worker inactive\n");
6464 }
6465
6466 /*
6467  * This is our raid5 kernel thread.
6468  *
6469  * We scan the hash table for stripes which can be handled now.
6470  * During the scan, completed stripes are saved for us by the interrupt
6471  * handler, so that they will not have to wait for our next wakeup.
6472  */
6473 static void raid5d(struct md_thread *thread)
6474 {
6475         struct mddev *mddev = thread->mddev;
6476         struct r5conf *conf = mddev->private;
6477         int handled;
6478         struct blk_plug plug;
6479
6480         pr_debug("+++ raid5d active\n");
6481
6482         md_check_recovery(mddev);
6483
6484         blk_start_plug(&plug);
6485         handled = 0;
6486         spin_lock_irq(&conf->device_lock);
6487         while (1) {
6488                 struct bio *bio;
6489                 int batch_size, released;
6490                 unsigned int offset;
6491
6492                 released = release_stripe_list(conf, conf->temp_inactive_list);
6493                 if (released)
6494                         clear_bit(R5_DID_ALLOC, &conf->cache_state);
6495
6496                 if (
6497                     !list_empty(&conf->bitmap_list)) {
6498                         /* Now is a good time to flush some bitmap updates */
6499                         conf->seq_flush++;
6500                         spin_unlock_irq(&conf->device_lock);
6501                         md_bitmap_unplug(mddev->bitmap);
6502                         spin_lock_irq(&conf->device_lock);
6503                         conf->seq_write = conf->seq_flush;
6504                         activate_bit_delay(conf, conf->temp_inactive_list);
6505                 }
6506                 raid5_activate_delayed(conf);
6507
6508                 while ((bio = remove_bio_from_retry(conf, &offset))) {
6509                         int ok;
6510                         spin_unlock_irq(&conf->device_lock);
6511                         ok = retry_aligned_read(conf, bio, offset);
6512                         spin_lock_irq(&conf->device_lock);
6513                         if (!ok)
6514                                 break;
6515                         handled++;
6516                 }
6517
6518                 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6519                                                    conf->temp_inactive_list);
6520                 if (!batch_size && !released)
6521                         break;
6522                 handled += batch_size;
6523
6524                 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6525                         spin_unlock_irq(&conf->device_lock);
6526                         md_check_recovery(mddev);
6527                         spin_lock_irq(&conf->device_lock);
6528                 }
6529         }
6530         pr_debug("%d stripes handled\n", handled);
6531
6532         spin_unlock_irq(&conf->device_lock);
6533         if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6534             mutex_trylock(&conf->cache_size_mutex)) {
6535                 grow_one_stripe(conf, __GFP_NOWARN);
6536                 /* Set flag even if allocation failed.  This helps
6537                  * slow down allocation requests when mem is short
6538                  */
6539                 set_bit(R5_DID_ALLOC, &conf->cache_state);
6540                 mutex_unlock(&conf->cache_size_mutex);
6541         }
6542
6543         flush_deferred_bios(conf);
6544
6545         r5l_flush_stripe_to_raid(conf->log);
6546
6547         async_tx_issue_pending_all();
6548         blk_finish_plug(&plug);
6549
6550         pr_debug("--- raid5d inactive\n");
6551 }
6552
6553 static ssize_t
6554 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6555 {
6556         struct r5conf *conf;
6557         int ret = 0;
6558         spin_lock(&mddev->lock);
6559         conf = mddev->private;
6560         if (conf)
6561                 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6562         spin_unlock(&mddev->lock);
6563         return ret;
6564 }
6565
6566 int
6567 raid5_set_cache_size(struct mddev *mddev, int size)
6568 {
6569         int result = 0;
6570         struct r5conf *conf = mddev->private;
6571
6572         if (size <= 16 || size > 32768)
6573                 return -EINVAL;
6574
6575         conf->min_nr_stripes = size;
6576         mutex_lock(&conf->cache_size_mutex);
6577         while (size < conf->max_nr_stripes &&
6578                drop_one_stripe(conf))
6579                 ;
6580         mutex_unlock(&conf->cache_size_mutex);
6581
6582         md_allow_write(mddev);
6583
6584         mutex_lock(&conf->cache_size_mutex);
6585         while (size > conf->max_nr_stripes)
6586                 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6587                         conf->min_nr_stripes = conf->max_nr_stripes;
6588                         result = -ENOMEM;
6589                         break;
6590                 }
6591         mutex_unlock(&conf->cache_size_mutex);
6592
6593         return result;
6594 }
6595 EXPORT_SYMBOL(raid5_set_cache_size);
6596
6597 static ssize_t
6598 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6599 {
6600         struct r5conf *conf;
6601         unsigned long new;
6602         int err;
6603
6604         if (len >= PAGE_SIZE)
6605                 return -EINVAL;
6606         if (kstrtoul(page, 10, &new))
6607                 return -EINVAL;
6608         err = mddev_lock(mddev);
6609         if (err)
6610                 return err;
6611         conf = mddev->private;
6612         if (!conf)
6613                 err = -ENODEV;
6614         else
6615                 err = raid5_set_cache_size(mddev, new);
6616         mddev_unlock(mddev);
6617
6618         return err ?: len;
6619 }
6620
6621 static struct md_sysfs_entry
6622 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6623                                 raid5_show_stripe_cache_size,
6624                                 raid5_store_stripe_cache_size);
6625
6626 static ssize_t
6627 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6628 {
6629         struct r5conf *conf = mddev->private;
6630         if (conf)
6631                 return sprintf(page, "%d\n", conf->rmw_level);
6632         else
6633                 return 0;
6634 }
6635
6636 static ssize_t
6637 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6638 {
6639         struct r5conf *conf = mddev->private;
6640         unsigned long new;
6641
6642         if (!conf)
6643                 return -ENODEV;
6644
6645         if (len >= PAGE_SIZE)
6646                 return -EINVAL;
6647
6648         if (kstrtoul(page, 10, &new))
6649                 return -EINVAL;
6650
6651         if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6652                 return -EINVAL;
6653
6654         if (new != PARITY_DISABLE_RMW &&
6655             new != PARITY_ENABLE_RMW &&
6656             new != PARITY_PREFER_RMW)
6657                 return -EINVAL;
6658
6659         conf->rmw_level = new;
6660         return len;
6661 }
6662
6663 static struct md_sysfs_entry
6664 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6665                          raid5_show_rmw_level,
6666                          raid5_store_rmw_level);
6667
6668 static ssize_t
6669 raid5_show_stripe_size(struct mddev  *mddev, char *page)
6670 {
6671         struct r5conf *conf;
6672         int ret = 0;
6673
6674         spin_lock(&mddev->lock);
6675         conf = mddev->private;
6676         if (conf)
6677                 ret = sprintf(page, "%lu\n", RAID5_STRIPE_SIZE(conf));
6678         spin_unlock(&mddev->lock);
6679         return ret;
6680 }
6681
6682 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
6683 static ssize_t
6684 raid5_store_stripe_size(struct mddev  *mddev, const char *page, size_t len)
6685 {
6686         struct r5conf *conf;
6687         unsigned long new;
6688         int err;
6689         int size;
6690
6691         if (len >= PAGE_SIZE)
6692                 return -EINVAL;
6693         if (kstrtoul(page, 10, &new))
6694                 return -EINVAL;
6695
6696         /*
6697          * The value should not be bigger than PAGE_SIZE. It requires to
6698          * be multiple of DEFAULT_STRIPE_SIZE and the value should be power
6699          * of two.
6700          */
6701         if (new % DEFAULT_STRIPE_SIZE != 0 ||
6702                         new > PAGE_SIZE || new == 0 ||
6703                         new != roundup_pow_of_two(new))
6704                 return -EINVAL;
6705
6706         err = mddev_lock(mddev);
6707         if (err)
6708                 return err;
6709
6710         conf = mddev->private;
6711         if (!conf) {
6712                 err = -ENODEV;
6713                 goto out_unlock;
6714         }
6715
6716         if (new == conf->stripe_size)
6717                 goto out_unlock;
6718
6719         pr_debug("md/raid: change stripe_size from %lu to %lu\n",
6720                         conf->stripe_size, new);
6721
6722         if (mddev->sync_thread ||
6723                 test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
6724                 mddev->reshape_position != MaxSector ||
6725                 mddev->sysfs_active) {
6726                 err = -EBUSY;
6727                 goto out_unlock;
6728         }
6729
6730         mddev_suspend(mddev);
6731         mutex_lock(&conf->cache_size_mutex);
6732         size = conf->max_nr_stripes;
6733
6734         shrink_stripes(conf);
6735
6736         conf->stripe_size = new;
6737         conf->stripe_shift = ilog2(new) - 9;
6738         conf->stripe_sectors = new >> 9;
6739         if (grow_stripes(conf, size)) {
6740                 pr_warn("md/raid:%s: couldn't allocate buffers\n",
6741                                 mdname(mddev));
6742                 err = -ENOMEM;
6743         }
6744         mutex_unlock(&conf->cache_size_mutex);
6745         mddev_resume(mddev);
6746
6747 out_unlock:
6748         mddev_unlock(mddev);
6749         return err ?: len;
6750 }
6751
6752 static struct md_sysfs_entry
6753 raid5_stripe_size = __ATTR(stripe_size, 0644,
6754                          raid5_show_stripe_size,
6755                          raid5_store_stripe_size);
6756 #else
6757 static struct md_sysfs_entry
6758 raid5_stripe_size = __ATTR(stripe_size, 0444,
6759                          raid5_show_stripe_size,
6760                          NULL);
6761 #endif
6762
6763 static ssize_t
6764 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6765 {
6766         struct r5conf *conf;
6767         int ret = 0;
6768         spin_lock(&mddev->lock);
6769         conf = mddev->private;
6770         if (conf)
6771                 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6772         spin_unlock(&mddev->lock);
6773         return ret;
6774 }
6775
6776 static ssize_t
6777 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6778 {
6779         struct r5conf *conf;
6780         unsigned long new;
6781         int err;
6782
6783         if (len >= PAGE_SIZE)
6784                 return -EINVAL;
6785         if (kstrtoul(page, 10, &new))
6786                 return -EINVAL;
6787
6788         err = mddev_lock(mddev);
6789         if (err)
6790                 return err;
6791         conf = mddev->private;
6792         if (!conf)
6793                 err = -ENODEV;
6794         else if (new > conf->min_nr_stripes)
6795                 err = -EINVAL;
6796         else
6797                 conf->bypass_threshold = new;
6798         mddev_unlock(mddev);
6799         return err ?: len;
6800 }
6801
6802 static struct md_sysfs_entry
6803 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6804                                         S_IRUGO | S_IWUSR,
6805                                         raid5_show_preread_threshold,
6806                                         raid5_store_preread_threshold);
6807
6808 static ssize_t
6809 raid5_show_skip_copy(struct mddev *mddev, char *page)
6810 {
6811         struct r5conf *conf;
6812         int ret = 0;
6813         spin_lock(&mddev->lock);
6814         conf = mddev->private;
6815         if (conf)
6816                 ret = sprintf(page, "%d\n", conf->skip_copy);
6817         spin_unlock(&mddev->lock);
6818         return ret;
6819 }
6820
6821 static ssize_t
6822 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6823 {
6824         struct r5conf *conf;
6825         unsigned long new;
6826         int err;
6827
6828         if (len >= PAGE_SIZE)
6829                 return -EINVAL;
6830         if (kstrtoul(page, 10, &new))
6831                 return -EINVAL;
6832         new = !!new;
6833
6834         err = mddev_lock(mddev);
6835         if (err)
6836                 return err;
6837         conf = mddev->private;
6838         if (!conf)
6839                 err = -ENODEV;
6840         else if (new != conf->skip_copy) {
6841                 struct request_queue *q = mddev->queue;
6842
6843                 mddev_suspend(mddev);
6844                 conf->skip_copy = new;
6845                 if (new)
6846                         blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, q);
6847                 else
6848                         blk_queue_flag_clear(QUEUE_FLAG_STABLE_WRITES, q);
6849                 mddev_resume(mddev);
6850         }
6851         mddev_unlock(mddev);
6852         return err ?: len;
6853 }
6854
6855 static struct md_sysfs_entry
6856 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6857                                         raid5_show_skip_copy,
6858                                         raid5_store_skip_copy);
6859
6860 static ssize_t
6861 stripe_cache_active_show(struct mddev *mddev, char *page)
6862 {
6863         struct r5conf *conf = mddev->private;
6864         if (conf)
6865                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6866         else
6867                 return 0;
6868 }
6869
6870 static struct md_sysfs_entry
6871 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6872
6873 static ssize_t
6874 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6875 {
6876         struct r5conf *conf;
6877         int ret = 0;
6878         spin_lock(&mddev->lock);
6879         conf = mddev->private;
6880         if (conf)
6881                 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6882         spin_unlock(&mddev->lock);
6883         return ret;
6884 }
6885
6886 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6887                                int *group_cnt,
6888                                struct r5worker_group **worker_groups);
6889 static ssize_t
6890 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6891 {
6892         struct r5conf *conf;
6893         unsigned int new;
6894         int err;
6895         struct r5worker_group *new_groups, *old_groups;
6896         int group_cnt;
6897
6898         if (len >= PAGE_SIZE)
6899                 return -EINVAL;
6900         if (kstrtouint(page, 10, &new))
6901                 return -EINVAL;
6902         /* 8192 should be big enough */
6903         if (new > 8192)
6904                 return -EINVAL;
6905
6906         err = mddev_lock(mddev);
6907         if (err)
6908                 return err;
6909         conf = mddev->private;
6910         if (!conf)
6911                 err = -ENODEV;
6912         else if (new != conf->worker_cnt_per_group) {
6913                 mddev_suspend(mddev);
6914
6915                 old_groups = conf->worker_groups;
6916                 if (old_groups)
6917                         flush_workqueue(raid5_wq);
6918
6919                 err = alloc_thread_groups(conf, new, &group_cnt, &new_groups);
6920                 if (!err) {
6921                         spin_lock_irq(&conf->device_lock);
6922                         conf->group_cnt = group_cnt;
6923                         conf->worker_cnt_per_group = new;
6924                         conf->worker_groups = new_groups;
6925                         spin_unlock_irq(&conf->device_lock);
6926
6927                         if (old_groups)
6928                                 kfree(old_groups[0].workers);
6929                         kfree(old_groups);
6930                 }
6931                 mddev_resume(mddev);
6932         }
6933         mddev_unlock(mddev);
6934
6935         return err ?: len;
6936 }
6937
6938 static struct md_sysfs_entry
6939 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6940                                 raid5_show_group_thread_cnt,
6941                                 raid5_store_group_thread_cnt);
6942
6943 static struct attribute *raid5_attrs[] =  {
6944         &raid5_stripecache_size.attr,
6945         &raid5_stripecache_active.attr,
6946         &raid5_preread_bypass_threshold.attr,
6947         &raid5_group_thread_cnt.attr,
6948         &raid5_skip_copy.attr,
6949         &raid5_rmw_level.attr,
6950         &raid5_stripe_size.attr,
6951         &r5c_journal_mode.attr,
6952         &ppl_write_hint.attr,
6953         NULL,
6954 };
6955 static const struct attribute_group raid5_attrs_group = {
6956         .name = NULL,
6957         .attrs = raid5_attrs,
6958 };
6959
6960 static int alloc_thread_groups(struct r5conf *conf, int cnt, int *group_cnt,
6961                                struct r5worker_group **worker_groups)
6962 {
6963         int i, j, k;
6964         ssize_t size;
6965         struct r5worker *workers;
6966
6967         if (cnt == 0) {
6968                 *group_cnt = 0;
6969                 *worker_groups = NULL;
6970                 return 0;
6971         }
6972         *group_cnt = num_possible_nodes();
6973         size = sizeof(struct r5worker) * cnt;
6974         workers = kcalloc(size, *group_cnt, GFP_NOIO);
6975         *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6976                                  GFP_NOIO);
6977         if (!*worker_groups || !workers) {
6978                 kfree(workers);
6979                 kfree(*worker_groups);
6980                 return -ENOMEM;
6981         }
6982
6983         for (i = 0; i < *group_cnt; i++) {
6984                 struct r5worker_group *group;
6985
6986                 group = &(*worker_groups)[i];
6987                 INIT_LIST_HEAD(&group->handle_list);
6988                 INIT_LIST_HEAD(&group->loprio_list);
6989                 group->conf = conf;
6990                 group->workers = workers + i * cnt;
6991
6992                 for (j = 0; j < cnt; j++) {
6993                         struct r5worker *worker = group->workers + j;
6994                         worker->group = group;
6995                         INIT_WORK(&worker->work, raid5_do_work);
6996
6997                         for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6998                                 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6999                 }
7000         }
7001
7002         return 0;
7003 }
7004
7005 static void free_thread_groups(struct r5conf *conf)
7006 {
7007         if (conf->worker_groups)
7008                 kfree(conf->worker_groups[0].workers);
7009         kfree(conf->worker_groups);
7010         conf->worker_groups = NULL;
7011 }
7012
7013 static sector_t
7014 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
7015 {
7016         struct r5conf *conf = mddev->private;
7017
7018         if (!sectors)
7019                 sectors = mddev->dev_sectors;
7020         if (!raid_disks)
7021                 /* size is defined by the smallest of previous and new size */
7022                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
7023
7024         sectors &= ~((sector_t)conf->chunk_sectors - 1);
7025         sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
7026         return sectors * (raid_disks - conf->max_degraded);
7027 }
7028
7029 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7030 {
7031         safe_put_page(percpu->spare_page);
7032         percpu->spare_page = NULL;
7033         kvfree(percpu->scribble);
7034         percpu->scribble = NULL;
7035 }
7036
7037 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
7038 {
7039         if (conf->level == 6 && !percpu->spare_page) {
7040                 percpu->spare_page = alloc_page(GFP_KERNEL);
7041                 if (!percpu->spare_page)
7042                         return -ENOMEM;
7043         }
7044
7045         if (scribble_alloc(percpu,
7046                            max(conf->raid_disks,
7047                                conf->previous_raid_disks),
7048                            max(conf->chunk_sectors,
7049                                conf->prev_chunk_sectors)
7050                            / RAID5_STRIPE_SECTORS(conf))) {
7051                 free_scratch_buffer(conf, percpu);
7052                 return -ENOMEM;
7053         }
7054
7055         return 0;
7056 }
7057
7058 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
7059 {
7060         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7061
7062         free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
7063         return 0;
7064 }
7065
7066 static void raid5_free_percpu(struct r5conf *conf)
7067 {
7068         if (!conf->percpu)
7069                 return;
7070
7071         cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7072         free_percpu(conf->percpu);
7073 }
7074
7075 static void free_conf(struct r5conf *conf)
7076 {
7077         int i;
7078
7079         log_exit(conf);
7080
7081         unregister_shrinker(&conf->shrinker);
7082         free_thread_groups(conf);
7083         shrink_stripes(conf);
7084         raid5_free_percpu(conf);
7085         for (i = 0; i < conf->pool_size; i++)
7086                 if (conf->disks[i].extra_page)
7087                         put_page(conf->disks[i].extra_page);
7088         kfree(conf->disks);
7089         bioset_exit(&conf->bio_split);
7090         kfree(conf->stripe_hashtbl);
7091         kfree(conf->pending_data);
7092         kfree(conf);
7093 }
7094
7095 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
7096 {
7097         struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
7098         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
7099
7100         if (alloc_scratch_buffer(conf, percpu)) {
7101                 pr_warn("%s: failed memory allocation for cpu%u\n",
7102                         __func__, cpu);
7103                 return -ENOMEM;
7104         }
7105         return 0;
7106 }
7107
7108 static int raid5_alloc_percpu(struct r5conf *conf)
7109 {
7110         int err = 0;
7111
7112         conf->percpu = alloc_percpu(struct raid5_percpu);
7113         if (!conf->percpu)
7114                 return -ENOMEM;
7115
7116         err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
7117         if (!err) {
7118                 conf->scribble_disks = max(conf->raid_disks,
7119                         conf->previous_raid_disks);
7120                 conf->scribble_sectors = max(conf->chunk_sectors,
7121                         conf->prev_chunk_sectors);
7122         }
7123         return err;
7124 }
7125
7126 static unsigned long raid5_cache_scan(struct shrinker *shrink,
7127                                       struct shrink_control *sc)
7128 {
7129         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7130         unsigned long ret = SHRINK_STOP;
7131
7132         if (mutex_trylock(&conf->cache_size_mutex)) {
7133                 ret= 0;
7134                 while (ret < sc->nr_to_scan &&
7135                        conf->max_nr_stripes > conf->min_nr_stripes) {
7136                         if (drop_one_stripe(conf) == 0) {
7137                                 ret = SHRINK_STOP;
7138                                 break;
7139                         }
7140                         ret++;
7141                 }
7142                 mutex_unlock(&conf->cache_size_mutex);
7143         }
7144         return ret;
7145 }
7146
7147 static unsigned long raid5_cache_count(struct shrinker *shrink,
7148                                        struct shrink_control *sc)
7149 {
7150         struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
7151
7152         if (conf->max_nr_stripes < conf->min_nr_stripes)
7153                 /* unlikely, but not impossible */
7154                 return 0;
7155         return conf->max_nr_stripes - conf->min_nr_stripes;
7156 }
7157
7158 static struct r5conf *setup_conf(struct mddev *mddev)
7159 {
7160         struct r5conf *conf;
7161         int raid_disk, memory, max_disks;
7162         struct md_rdev *rdev;
7163         struct disk_info *disk;
7164         char pers_name[6];
7165         int i;
7166         int group_cnt;
7167         struct r5worker_group *new_group;
7168         int ret;
7169
7170         if (mddev->new_level != 5
7171             && mddev->new_level != 4
7172             && mddev->new_level != 6) {
7173                 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
7174                         mdname(mddev), mddev->new_level);
7175                 return ERR_PTR(-EIO);
7176         }
7177         if ((mddev->new_level == 5
7178              && !algorithm_valid_raid5(mddev->new_layout)) ||
7179             (mddev->new_level == 6
7180              && !algorithm_valid_raid6(mddev->new_layout))) {
7181                 pr_warn("md/raid:%s: layout %d not supported\n",
7182                         mdname(mddev), mddev->new_layout);
7183                 return ERR_PTR(-EIO);
7184         }
7185         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
7186                 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
7187                         mdname(mddev), mddev->raid_disks);
7188                 return ERR_PTR(-EINVAL);
7189         }
7190
7191         if (!mddev->new_chunk_sectors ||
7192             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
7193             !is_power_of_2(mddev->new_chunk_sectors)) {
7194                 pr_warn("md/raid:%s: invalid chunk size %d\n",
7195                         mdname(mddev), mddev->new_chunk_sectors << 9);
7196                 return ERR_PTR(-EINVAL);
7197         }
7198
7199         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
7200         if (conf == NULL)
7201                 goto abort;
7202
7203 #if PAGE_SIZE != DEFAULT_STRIPE_SIZE
7204         conf->stripe_size = DEFAULT_STRIPE_SIZE;
7205         conf->stripe_shift = ilog2(DEFAULT_STRIPE_SIZE) - 9;
7206         conf->stripe_sectors = DEFAULT_STRIPE_SIZE >> 9;
7207 #endif
7208         INIT_LIST_HEAD(&conf->free_list);
7209         INIT_LIST_HEAD(&conf->pending_list);
7210         conf->pending_data = kcalloc(PENDING_IO_MAX,
7211                                      sizeof(struct r5pending_data),
7212                                      GFP_KERNEL);
7213         if (!conf->pending_data)
7214                 goto abort;
7215         for (i = 0; i < PENDING_IO_MAX; i++)
7216                 list_add(&conf->pending_data[i].sibling, &conf->free_list);
7217         /* Don't enable multi-threading by default*/
7218         if (!alloc_thread_groups(conf, 0, &group_cnt, &new_group)) {
7219                 conf->group_cnt = group_cnt;
7220                 conf->worker_cnt_per_group = 0;
7221                 conf->worker_groups = new_group;
7222         } else
7223                 goto abort;
7224         spin_lock_init(&conf->device_lock);
7225         seqcount_spinlock_init(&conf->gen_lock, &conf->device_lock);
7226         mutex_init(&conf->cache_size_mutex);
7227         init_waitqueue_head(&conf->wait_for_quiescent);
7228         init_waitqueue_head(&conf->wait_for_stripe);
7229         init_waitqueue_head(&conf->wait_for_overlap);
7230         INIT_LIST_HEAD(&conf->handle_list);
7231         INIT_LIST_HEAD(&conf->loprio_list);
7232         INIT_LIST_HEAD(&conf->hold_list);
7233         INIT_LIST_HEAD(&conf->delayed_list);
7234         INIT_LIST_HEAD(&conf->bitmap_list);
7235         init_llist_head(&conf->released_stripes);
7236         atomic_set(&conf->active_stripes, 0);
7237         atomic_set(&conf->preread_active_stripes, 0);
7238         atomic_set(&conf->active_aligned_reads, 0);
7239         spin_lock_init(&conf->pending_bios_lock);
7240         conf->batch_bio_dispatch = true;
7241         rdev_for_each(rdev, mddev) {
7242                 if (test_bit(Journal, &rdev->flags))
7243                         continue;
7244                 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
7245                         conf->batch_bio_dispatch = false;
7246                         break;
7247                 }
7248         }
7249
7250         conf->bypass_threshold = BYPASS_THRESHOLD;
7251         conf->recovery_disabled = mddev->recovery_disabled - 1;
7252
7253         conf->raid_disks = mddev->raid_disks;
7254         if (mddev->reshape_position == MaxSector)
7255                 conf->previous_raid_disks = mddev->raid_disks;
7256         else
7257                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
7258         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
7259
7260         conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
7261                               GFP_KERNEL);
7262
7263         if (!conf->disks)
7264                 goto abort;
7265
7266         for (i = 0; i < max_disks; i++) {
7267                 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
7268                 if (!conf->disks[i].extra_page)
7269                         goto abort;
7270         }
7271
7272         ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
7273         if (ret)
7274                 goto abort;
7275         conf->mddev = mddev;
7276
7277         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
7278                 goto abort;
7279
7280         /* We init hash_locks[0] separately to that it can be used
7281          * as the reference lock in the spin_lock_nest_lock() call
7282          * in lock_all_device_hash_locks_irq in order to convince
7283          * lockdep that we know what we are doing.
7284          */
7285         spin_lock_init(conf->hash_locks);
7286         for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7287                 spin_lock_init(conf->hash_locks + i);
7288
7289         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7290                 INIT_LIST_HEAD(conf->inactive_list + i);
7291
7292         for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7293                 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7294
7295         atomic_set(&conf->r5c_cached_full_stripes, 0);
7296         INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7297         atomic_set(&conf->r5c_cached_partial_stripes, 0);
7298         INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7299         atomic_set(&conf->r5c_flushing_full_stripes, 0);
7300         atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7301
7302         conf->level = mddev->new_level;
7303         conf->chunk_sectors = mddev->new_chunk_sectors;
7304         if (raid5_alloc_percpu(conf) != 0)
7305                 goto abort;
7306
7307         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7308
7309         rdev_for_each(rdev, mddev) {
7310                 raid_disk = rdev->raid_disk;
7311                 if (raid_disk >= max_disks
7312                     || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7313                         continue;
7314                 disk = conf->disks + raid_disk;
7315
7316                 if (test_bit(Replacement, &rdev->flags)) {
7317                         if (disk->replacement)
7318                                 goto abort;
7319                         disk->replacement = rdev;
7320                 } else {
7321                         if (disk->rdev)
7322                                 goto abort;
7323                         disk->rdev = rdev;
7324                 }
7325
7326                 if (test_bit(In_sync, &rdev->flags)) {
7327                         char b[BDEVNAME_SIZE];
7328                         pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7329                                 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7330                 } else if (rdev->saved_raid_disk != raid_disk)
7331                         /* Cannot rely on bitmap to complete recovery */
7332                         conf->fullsync = 1;
7333         }
7334
7335         conf->level = mddev->new_level;
7336         if (conf->level == 6) {
7337                 conf->max_degraded = 2;
7338                 if (raid6_call.xor_syndrome)
7339                         conf->rmw_level = PARITY_ENABLE_RMW;
7340                 else
7341                         conf->rmw_level = PARITY_DISABLE_RMW;
7342         } else {
7343                 conf->max_degraded = 1;
7344                 conf->rmw_level = PARITY_ENABLE_RMW;
7345         }
7346         conf->algorithm = mddev->new_layout;
7347         conf->reshape_progress = mddev->reshape_position;
7348         if (conf->reshape_progress != MaxSector) {
7349                 conf->prev_chunk_sectors = mddev->chunk_sectors;
7350                 conf->prev_algo = mddev->layout;
7351         } else {
7352                 conf->prev_chunk_sectors = conf->chunk_sectors;
7353                 conf->prev_algo = conf->algorithm;
7354         }
7355
7356         conf->min_nr_stripes = NR_STRIPES;
7357         if (mddev->reshape_position != MaxSector) {
7358                 int stripes = max_t(int,
7359                         ((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4,
7360                         ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4);
7361                 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7362                 if (conf->min_nr_stripes != NR_STRIPES)
7363                         pr_info("md/raid:%s: force stripe size %d for reshape\n",
7364                                 mdname(mddev), conf->min_nr_stripes);
7365         }
7366         memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7367                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7368         atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7369         if (grow_stripes(conf, conf->min_nr_stripes)) {
7370                 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7371                         mdname(mddev), memory);
7372                 goto abort;
7373         } else
7374                 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7375         /*
7376          * Losing a stripe head costs more than the time to refill it,
7377          * it reduces the queue depth and so can hurt throughput.
7378          * So set it rather large, scaled by number of devices.
7379          */
7380         conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7381         conf->shrinker.scan_objects = raid5_cache_scan;
7382         conf->shrinker.count_objects = raid5_cache_count;
7383         conf->shrinker.batch = 128;
7384         conf->shrinker.flags = 0;
7385         if (register_shrinker(&conf->shrinker)) {
7386                 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7387                         mdname(mddev));
7388                 goto abort;
7389         }
7390
7391         sprintf(pers_name, "raid%d", mddev->new_level);
7392         conf->thread = md_register_thread(raid5d, mddev, pers_name);
7393         if (!conf->thread) {
7394                 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7395                         mdname(mddev));
7396                 goto abort;
7397         }
7398
7399         return conf;
7400
7401  abort:
7402         if (conf) {
7403                 free_conf(conf);
7404                 return ERR_PTR(-EIO);
7405         } else
7406                 return ERR_PTR(-ENOMEM);
7407 }
7408
7409 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7410 {
7411         switch (algo) {
7412         case ALGORITHM_PARITY_0:
7413                 if (raid_disk < max_degraded)
7414                         return 1;
7415                 break;
7416         case ALGORITHM_PARITY_N:
7417                 if (raid_disk >= raid_disks - max_degraded)
7418                         return 1;
7419                 break;
7420         case ALGORITHM_PARITY_0_6:
7421                 if (raid_disk == 0 ||
7422                     raid_disk == raid_disks - 1)
7423                         return 1;
7424                 break;
7425         case ALGORITHM_LEFT_ASYMMETRIC_6:
7426         case ALGORITHM_RIGHT_ASYMMETRIC_6:
7427         case ALGORITHM_LEFT_SYMMETRIC_6:
7428         case ALGORITHM_RIGHT_SYMMETRIC_6:
7429                 if (raid_disk == raid_disks - 1)
7430                         return 1;
7431         }
7432         return 0;
7433 }
7434
7435 static void raid5_set_io_opt(struct r5conf *conf)
7436 {
7437         blk_queue_io_opt(conf->mddev->queue, (conf->chunk_sectors << 9) *
7438                          (conf->raid_disks - conf->max_degraded));
7439 }
7440
7441 static int raid5_run(struct mddev *mddev)
7442 {
7443         struct r5conf *conf;
7444         int working_disks = 0;
7445         int dirty_parity_disks = 0;
7446         struct md_rdev *rdev;
7447         struct md_rdev *journal_dev = NULL;
7448         sector_t reshape_offset = 0;
7449         int i;
7450         long long min_offset_diff = 0;
7451         int first = 1;
7452
7453         if (mddev_init_writes_pending(mddev) < 0)
7454                 return -ENOMEM;
7455
7456         if (mddev->recovery_cp != MaxSector)
7457                 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7458                           mdname(mddev));
7459
7460         rdev_for_each(rdev, mddev) {
7461                 long long diff;
7462
7463                 if (test_bit(Journal, &rdev->flags)) {
7464                         journal_dev = rdev;
7465                         continue;
7466                 }
7467                 if (rdev->raid_disk < 0)
7468                         continue;
7469                 diff = (rdev->new_data_offset - rdev->data_offset);
7470                 if (first) {
7471                         min_offset_diff = diff;
7472                         first = 0;
7473                 } else if (mddev->reshape_backwards &&
7474                          diff < min_offset_diff)
7475                         min_offset_diff = diff;
7476                 else if (!mddev->reshape_backwards &&
7477                          diff > min_offset_diff)
7478                         min_offset_diff = diff;
7479         }
7480
7481         if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7482             (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7483                 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7484                           mdname(mddev));
7485                 return -EINVAL;
7486         }
7487
7488         if (mddev->reshape_position != MaxSector) {
7489                 /* Check that we can continue the reshape.
7490                  * Difficulties arise if the stripe we would write to
7491                  * next is at or after the stripe we would read from next.
7492                  * For a reshape that changes the number of devices, this
7493                  * is only possible for a very short time, and mdadm makes
7494                  * sure that time appears to have past before assembling
7495                  * the array.  So we fail if that time hasn't passed.
7496                  * For a reshape that keeps the number of devices the same
7497                  * mdadm must be monitoring the reshape can keeping the
7498                  * critical areas read-only and backed up.  It will start
7499                  * the array in read-only mode, so we check for that.
7500                  */
7501                 sector_t here_new, here_old;
7502                 int old_disks;
7503                 int max_degraded = (mddev->level == 6 ? 2 : 1);
7504                 int chunk_sectors;
7505                 int new_data_disks;
7506
7507                 if (journal_dev) {
7508                         pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7509                                 mdname(mddev));
7510                         return -EINVAL;
7511                 }
7512
7513                 if (mddev->new_level != mddev->level) {
7514                         pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7515                                 mdname(mddev));
7516                         return -EINVAL;
7517                 }
7518                 old_disks = mddev->raid_disks - mddev->delta_disks;
7519                 /* reshape_position must be on a new-stripe boundary, and one
7520                  * further up in new geometry must map after here in old
7521                  * geometry.
7522                  * If the chunk sizes are different, then as we perform reshape
7523                  * in units of the largest of the two, reshape_position needs
7524                  * be a multiple of the largest chunk size times new data disks.
7525                  */
7526                 here_new = mddev->reshape_position;
7527                 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7528                 new_data_disks = mddev->raid_disks - max_degraded;
7529                 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7530                         pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7531                                 mdname(mddev));
7532                         return -EINVAL;
7533                 }
7534                 reshape_offset = here_new * chunk_sectors;
7535                 /* here_new is the stripe we will write to */
7536                 here_old = mddev->reshape_position;
7537                 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7538                 /* here_old is the first stripe that we might need to read
7539                  * from */
7540                 if (mddev->delta_disks == 0) {
7541                         /* We cannot be sure it is safe to start an in-place
7542                          * reshape.  It is only safe if user-space is monitoring
7543                          * and taking constant backups.
7544                          * mdadm always starts a situation like this in
7545                          * readonly mode so it can take control before
7546                          * allowing any writes.  So just check for that.
7547                          */
7548                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7549                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
7550                                 /* not really in-place - so OK */;
7551                         else if (mddev->ro == 0) {
7552                                 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7553                                         mdname(mddev));
7554                                 return -EINVAL;
7555                         }
7556                 } else if (mddev->reshape_backwards
7557                     ? (here_new * chunk_sectors + min_offset_diff <=
7558                        here_old * chunk_sectors)
7559                     : (here_new * chunk_sectors >=
7560                        here_old * chunk_sectors + (-min_offset_diff))) {
7561                         /* Reading from the same stripe as writing to - bad */
7562                         pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7563                                 mdname(mddev));
7564                         return -EINVAL;
7565                 }
7566                 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7567                 /* OK, we should be able to continue; */
7568         } else {
7569                 BUG_ON(mddev->level != mddev->new_level);
7570                 BUG_ON(mddev->layout != mddev->new_layout);
7571                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7572                 BUG_ON(mddev->delta_disks != 0);
7573         }
7574
7575         if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7576             test_bit(MD_HAS_PPL, &mddev->flags)) {
7577                 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7578                         mdname(mddev));
7579                 clear_bit(MD_HAS_PPL, &mddev->flags);
7580                 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7581         }
7582
7583         if (mddev->private == NULL)
7584                 conf = setup_conf(mddev);
7585         else
7586                 conf = mddev->private;
7587
7588         if (IS_ERR(conf))
7589                 return PTR_ERR(conf);
7590
7591         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7592                 if (!journal_dev) {
7593                         pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7594                                 mdname(mddev));
7595                         mddev->ro = 1;
7596                         set_disk_ro(mddev->gendisk, 1);
7597                 } else if (mddev->recovery_cp == MaxSector)
7598                         set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7599         }
7600
7601         conf->min_offset_diff = min_offset_diff;
7602         mddev->thread = conf->thread;
7603         conf->thread = NULL;
7604         mddev->private = conf;
7605
7606         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7607              i++) {
7608                 rdev = conf->disks[i].rdev;
7609                 if (!rdev && conf->disks[i].replacement) {
7610                         /* The replacement is all we have yet */
7611                         rdev = conf->disks[i].replacement;
7612                         conf->disks[i].replacement = NULL;
7613                         clear_bit(Replacement, &rdev->flags);
7614                         conf->disks[i].rdev = rdev;
7615                 }
7616                 if (!rdev)
7617                         continue;
7618                 if (conf->disks[i].replacement &&
7619                     conf->reshape_progress != MaxSector) {
7620                         /* replacements and reshape simply do not mix. */
7621                         pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7622                         goto abort;
7623                 }
7624                 if (test_bit(In_sync, &rdev->flags)) {
7625                         working_disks++;
7626                         continue;
7627                 }
7628                 /* This disc is not fully in-sync.  However if it
7629                  * just stored parity (beyond the recovery_offset),
7630                  * when we don't need to be concerned about the
7631                  * array being dirty.
7632                  * When reshape goes 'backwards', we never have
7633                  * partially completed devices, so we only need
7634                  * to worry about reshape going forwards.
7635                  */
7636                 /* Hack because v0.91 doesn't store recovery_offset properly. */
7637                 if (mddev->major_version == 0 &&
7638                     mddev->minor_version > 90)
7639                         rdev->recovery_offset = reshape_offset;
7640
7641                 if (rdev->recovery_offset < reshape_offset) {
7642                         /* We need to check old and new layout */
7643                         if (!only_parity(rdev->raid_disk,
7644                                          conf->algorithm,
7645                                          conf->raid_disks,
7646                                          conf->max_degraded))
7647                                 continue;
7648                 }
7649                 if (!only_parity(rdev->raid_disk,
7650                                  conf->prev_algo,
7651                                  conf->previous_raid_disks,
7652                                  conf->max_degraded))
7653                         continue;
7654                 dirty_parity_disks++;
7655         }
7656
7657         /*
7658          * 0 for a fully functional array, 1 or 2 for a degraded array.
7659          */
7660         mddev->degraded = raid5_calc_degraded(conf);
7661
7662         if (has_failed(conf)) {
7663                 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7664                         mdname(mddev), mddev->degraded, conf->raid_disks);
7665                 goto abort;
7666         }
7667
7668         /* device size must be a multiple of chunk size */
7669         mddev->dev_sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7670         mddev->resync_max_sectors = mddev->dev_sectors;
7671
7672         if (mddev->degraded > dirty_parity_disks &&
7673             mddev->recovery_cp != MaxSector) {
7674                 if (test_bit(MD_HAS_PPL, &mddev->flags))
7675                         pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7676                                 mdname(mddev));
7677                 else if (mddev->ok_start_degraded)
7678                         pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7679                                 mdname(mddev));
7680                 else {
7681                         pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7682                                 mdname(mddev));
7683                         goto abort;
7684                 }
7685         }
7686
7687         pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7688                 mdname(mddev), conf->level,
7689                 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7690                 mddev->new_layout);
7691
7692         print_raid5_conf(conf);
7693
7694         if (conf->reshape_progress != MaxSector) {
7695                 conf->reshape_safe = conf->reshape_progress;
7696                 atomic_set(&conf->reshape_stripes, 0);
7697                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7698                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7699                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7700                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7701                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7702                                                         "reshape");
7703                 if (!mddev->sync_thread)
7704                         goto abort;
7705         }
7706
7707         /* Ok, everything is just fine now */
7708         if (mddev->to_remove == &raid5_attrs_group)
7709                 mddev->to_remove = NULL;
7710         else if (mddev->kobj.sd &&
7711             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7712                 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7713                         mdname(mddev));
7714         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7715
7716         if (mddev->queue) {
7717                 int chunk_size;
7718                 /* read-ahead size must cover two whole stripes, which
7719                  * is 2 * (datadisks) * chunksize where 'n' is the
7720                  * number of raid devices
7721                  */
7722                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7723                 int stripe = data_disks *
7724                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7725
7726                 chunk_size = mddev->chunk_sectors << 9;
7727                 blk_queue_io_min(mddev->queue, chunk_size);
7728                 raid5_set_io_opt(conf);
7729                 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7730                 /*
7731                  * We can only discard a whole stripe. It doesn't make sense to
7732                  * discard data disk but write parity disk
7733                  */
7734                 stripe = stripe * PAGE_SIZE;
7735                 /* Round up to power of 2, as discard handling
7736                  * currently assumes that */
7737                 while ((stripe-1) & stripe)
7738                         stripe = (stripe | (stripe-1)) + 1;
7739                 mddev->queue->limits.discard_alignment = stripe;
7740                 mddev->queue->limits.discard_granularity = stripe;
7741
7742                 blk_queue_max_write_same_sectors(mddev->queue, 0);
7743                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7744
7745                 rdev_for_each(rdev, mddev) {
7746                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7747                                           rdev->data_offset << 9);
7748                         disk_stack_limits(mddev->gendisk, rdev->bdev,
7749                                           rdev->new_data_offset << 9);
7750                 }
7751
7752                 /*
7753                  * zeroing is required, otherwise data
7754                  * could be lost. Consider a scenario: discard a stripe
7755                  * (the stripe could be inconsistent if
7756                  * discard_zeroes_data is 0); write one disk of the
7757                  * stripe (the stripe could be inconsistent again
7758                  * depending on which disks are used to calculate
7759                  * parity); the disk is broken; The stripe data of this
7760                  * disk is lost.
7761                  *
7762                  * We only allow DISCARD if the sysadmin has confirmed that
7763                  * only safe devices are in use by setting a module parameter.
7764                  * A better idea might be to turn DISCARD into WRITE_ZEROES
7765                  * requests, as that is required to be safe.
7766                  */
7767                 if (devices_handle_discard_safely &&
7768                     mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7769                     mddev->queue->limits.discard_granularity >= stripe)
7770                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7771                                                 mddev->queue);
7772                 else
7773                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7774                                                 mddev->queue);
7775
7776                 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7777         }
7778
7779         if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7780                 goto abort;
7781
7782         return 0;
7783 abort:
7784         md_unregister_thread(&mddev->thread);
7785         print_raid5_conf(conf);
7786         free_conf(conf);
7787         mddev->private = NULL;
7788         pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7789         return -EIO;
7790 }
7791
7792 static void raid5_free(struct mddev *mddev, void *priv)
7793 {
7794         struct r5conf *conf = priv;
7795
7796         free_conf(conf);
7797         mddev->to_remove = &raid5_attrs_group;
7798 }
7799
7800 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7801 {
7802         struct r5conf *conf = mddev->private;
7803         int i;
7804
7805         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7806                 conf->chunk_sectors / 2, mddev->layout);
7807         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7808         rcu_read_lock();
7809         for (i = 0; i < conf->raid_disks; i++) {
7810                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7811                 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7812         }
7813         rcu_read_unlock();
7814         seq_printf (seq, "]");
7815 }
7816
7817 static void print_raid5_conf (struct r5conf *conf)
7818 {
7819         int i;
7820         struct disk_info *tmp;
7821
7822         pr_debug("RAID conf printout:\n");
7823         if (!conf) {
7824                 pr_debug("(conf==NULL)\n");
7825                 return;
7826         }
7827         pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7828                conf->raid_disks,
7829                conf->raid_disks - conf->mddev->degraded);
7830
7831         for (i = 0; i < conf->raid_disks; i++) {
7832                 char b[BDEVNAME_SIZE];
7833                 tmp = conf->disks + i;
7834                 if (tmp->rdev)
7835                         pr_debug(" disk %d, o:%d, dev:%s\n",
7836                                i, !test_bit(Faulty, &tmp->rdev->flags),
7837                                bdevname(tmp->rdev->bdev, b));
7838         }
7839 }
7840
7841 static int raid5_spare_active(struct mddev *mddev)
7842 {
7843         int i;
7844         struct r5conf *conf = mddev->private;
7845         struct disk_info *tmp;
7846         int count = 0;
7847         unsigned long flags;
7848
7849         for (i = 0; i < conf->raid_disks; i++) {
7850                 tmp = conf->disks + i;
7851                 if (tmp->replacement
7852                     && tmp->replacement->recovery_offset == MaxSector
7853                     && !test_bit(Faulty, &tmp->replacement->flags)
7854                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7855                         /* Replacement has just become active. */
7856                         if (!tmp->rdev
7857                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7858                                 count++;
7859                         if (tmp->rdev) {
7860                                 /* Replaced device not technically faulty,
7861                                  * but we need to be sure it gets removed
7862                                  * and never re-added.
7863                                  */
7864                                 set_bit(Faulty, &tmp->rdev->flags);
7865                                 sysfs_notify_dirent_safe(
7866                                         tmp->rdev->sysfs_state);
7867                         }
7868                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7869                 } else if (tmp->rdev
7870                     && tmp->rdev->recovery_offset == MaxSector
7871                     && !test_bit(Faulty, &tmp->rdev->flags)
7872                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7873                         count++;
7874                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7875                 }
7876         }
7877         spin_lock_irqsave(&conf->device_lock, flags);
7878         mddev->degraded = raid5_calc_degraded(conf);
7879         spin_unlock_irqrestore(&conf->device_lock, flags);
7880         print_raid5_conf(conf);
7881         return count;
7882 }
7883
7884 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7885 {
7886         struct r5conf *conf = mddev->private;
7887         int err = 0;
7888         int number = rdev->raid_disk;
7889         struct md_rdev **rdevp;
7890         struct disk_info *p = conf->disks + number;
7891
7892         print_raid5_conf(conf);
7893         if (test_bit(Journal, &rdev->flags) && conf->log) {
7894                 /*
7895                  * we can't wait pending write here, as this is called in
7896                  * raid5d, wait will deadlock.
7897                  * neilb: there is no locking about new writes here,
7898                  * so this cannot be safe.
7899                  */
7900                 if (atomic_read(&conf->active_stripes) ||
7901                     atomic_read(&conf->r5c_cached_full_stripes) ||
7902                     atomic_read(&conf->r5c_cached_partial_stripes)) {
7903                         return -EBUSY;
7904                 }
7905                 log_exit(conf);
7906                 return 0;
7907         }
7908         if (rdev == p->rdev)
7909                 rdevp = &p->rdev;
7910         else if (rdev == p->replacement)
7911                 rdevp = &p->replacement;
7912         else
7913                 return 0;
7914
7915         if (number >= conf->raid_disks &&
7916             conf->reshape_progress == MaxSector)
7917                 clear_bit(In_sync, &rdev->flags);
7918
7919         if (test_bit(In_sync, &rdev->flags) ||
7920             atomic_read(&rdev->nr_pending)) {
7921                 err = -EBUSY;
7922                 goto abort;
7923         }
7924         /* Only remove non-faulty devices if recovery
7925          * isn't possible.
7926          */
7927         if (!test_bit(Faulty, &rdev->flags) &&
7928             mddev->recovery_disabled != conf->recovery_disabled &&
7929             !has_failed(conf) &&
7930             (!p->replacement || p->replacement == rdev) &&
7931             number < conf->raid_disks) {
7932                 err = -EBUSY;
7933                 goto abort;
7934         }
7935         *rdevp = NULL;
7936         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7937                 synchronize_rcu();
7938                 if (atomic_read(&rdev->nr_pending)) {
7939                         /* lost the race, try later */
7940                         err = -EBUSY;
7941                         *rdevp = rdev;
7942                 }
7943         }
7944         if (!err) {
7945                 err = log_modify(conf, rdev, false);
7946                 if (err)
7947                         goto abort;
7948         }
7949         if (p->replacement) {
7950                 /* We must have just cleared 'rdev' */
7951                 p->rdev = p->replacement;
7952                 clear_bit(Replacement, &p->replacement->flags);
7953                 smp_mb(); /* Make sure other CPUs may see both as identical
7954                            * but will never see neither - if they are careful
7955                            */
7956                 p->replacement = NULL;
7957
7958                 if (!err)
7959                         err = log_modify(conf, p->rdev, true);
7960         }
7961
7962         clear_bit(WantReplacement, &rdev->flags);
7963 abort:
7964
7965         print_raid5_conf(conf);
7966         return err;
7967 }
7968
7969 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7970 {
7971         struct r5conf *conf = mddev->private;
7972         int ret, err = -EEXIST;
7973         int disk;
7974         struct disk_info *p;
7975         int first = 0;
7976         int last = conf->raid_disks - 1;
7977
7978         if (test_bit(Journal, &rdev->flags)) {
7979                 if (conf->log)
7980                         return -EBUSY;
7981
7982                 rdev->raid_disk = 0;
7983                 /*
7984                  * The array is in readonly mode if journal is missing, so no
7985                  * write requests running. We should be safe
7986                  */
7987                 ret = log_init(conf, rdev, false);
7988                 if (ret)
7989                         return ret;
7990
7991                 ret = r5l_start(conf->log);
7992                 if (ret)
7993                         return ret;
7994
7995                 return 0;
7996         }
7997         if (mddev->recovery_disabled == conf->recovery_disabled)
7998                 return -EBUSY;
7999
8000         if (rdev->saved_raid_disk < 0 && has_failed(conf))
8001                 /* no point adding a device */
8002                 return -EINVAL;
8003
8004         if (rdev->raid_disk >= 0)
8005                 first = last = rdev->raid_disk;
8006
8007         /*
8008          * find the disk ... but prefer rdev->saved_raid_disk
8009          * if possible.
8010          */
8011         if (rdev->saved_raid_disk >= 0 &&
8012             rdev->saved_raid_disk >= first &&
8013             conf->disks[rdev->saved_raid_disk].rdev == NULL)
8014                 first = rdev->saved_raid_disk;
8015
8016         for (disk = first; disk <= last; disk++) {
8017                 p = conf->disks + disk;
8018                 if (p->rdev == NULL) {
8019                         clear_bit(In_sync, &rdev->flags);
8020                         rdev->raid_disk = disk;
8021                         if (rdev->saved_raid_disk != disk)
8022                                 conf->fullsync = 1;
8023                         rcu_assign_pointer(p->rdev, rdev);
8024
8025                         err = log_modify(conf, rdev, true);
8026
8027                         goto out;
8028                 }
8029         }
8030         for (disk = first; disk <= last; disk++) {
8031                 p = conf->disks + disk;
8032                 if (test_bit(WantReplacement, &p->rdev->flags) &&
8033                     p->replacement == NULL) {
8034                         clear_bit(In_sync, &rdev->flags);
8035                         set_bit(Replacement, &rdev->flags);
8036                         rdev->raid_disk = disk;
8037                         err = 0;
8038                         conf->fullsync = 1;
8039                         rcu_assign_pointer(p->replacement, rdev);
8040                         break;
8041                 }
8042         }
8043 out:
8044         print_raid5_conf(conf);
8045         return err;
8046 }
8047
8048 static int raid5_resize(struct mddev *mddev, sector_t sectors)
8049 {
8050         /* no resync is happening, and there is enough space
8051          * on all devices, so we can resize.
8052          * We need to make sure resync covers any new space.
8053          * If the array is shrinking we should possibly wait until
8054          * any io in the removed space completes, but it hardly seems
8055          * worth it.
8056          */
8057         sector_t newsize;
8058         struct r5conf *conf = mddev->private;
8059
8060         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8061                 return -EINVAL;
8062         sectors &= ~((sector_t)conf->chunk_sectors - 1);
8063         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
8064         if (mddev->external_size &&
8065             mddev->array_sectors > newsize)
8066                 return -EINVAL;
8067         if (mddev->bitmap) {
8068                 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
8069                 if (ret)
8070                         return ret;
8071         }
8072         md_set_array_sectors(mddev, newsize);
8073         if (sectors > mddev->dev_sectors &&
8074             mddev->recovery_cp > mddev->dev_sectors) {
8075                 mddev->recovery_cp = mddev->dev_sectors;
8076                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
8077         }
8078         mddev->dev_sectors = sectors;
8079         mddev->resync_max_sectors = sectors;
8080         return 0;
8081 }
8082
8083 static int check_stripe_cache(struct mddev *mddev)
8084 {
8085         /* Can only proceed if there are plenty of stripe_heads.
8086          * We need a minimum of one full stripe,, and for sensible progress
8087          * it is best to have about 4 times that.
8088          * If we require 4 times, then the default 256 4K stripe_heads will
8089          * allow for chunk sizes up to 256K, which is probably OK.
8090          * If the chunk size is greater, user-space should request more
8091          * stripe_heads first.
8092          */
8093         struct r5conf *conf = mddev->private;
8094         if (((mddev->chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8095             > conf->min_nr_stripes ||
8096             ((mddev->new_chunk_sectors << 9) / RAID5_STRIPE_SIZE(conf)) * 4
8097             > conf->min_nr_stripes) {
8098                 pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
8099                         mdname(mddev),
8100                         ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
8101                          / RAID5_STRIPE_SIZE(conf))*4);
8102                 return 0;
8103         }
8104         return 1;
8105 }
8106
8107 static int check_reshape(struct mddev *mddev)
8108 {
8109         struct r5conf *conf = mddev->private;
8110
8111         if (raid5_has_log(conf) || raid5_has_ppl(conf))
8112                 return -EINVAL;
8113         if (mddev->delta_disks == 0 &&
8114             mddev->new_layout == mddev->layout &&
8115             mddev->new_chunk_sectors == mddev->chunk_sectors)
8116                 return 0; /* nothing to do */
8117         if (has_failed(conf))
8118                 return -EINVAL;
8119         if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
8120                 /* We might be able to shrink, but the devices must
8121                  * be made bigger first.
8122                  * For raid6, 4 is the minimum size.
8123                  * Otherwise 2 is the minimum
8124                  */
8125                 int min = 2;
8126                 if (mddev->level == 6)
8127                         min = 4;
8128                 if (mddev->raid_disks + mddev->delta_disks < min)
8129                         return -EINVAL;
8130         }
8131
8132         if (!check_stripe_cache(mddev))
8133                 return -ENOSPC;
8134
8135         if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
8136             mddev->delta_disks > 0)
8137                 if (resize_chunks(conf,
8138                                   conf->previous_raid_disks
8139                                   + max(0, mddev->delta_disks),
8140                                   max(mddev->new_chunk_sectors,
8141                                       mddev->chunk_sectors)
8142                             ) < 0)
8143                         return -ENOMEM;
8144
8145         if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
8146                 return 0; /* never bother to shrink */
8147         return resize_stripes(conf, (conf->previous_raid_disks
8148                                      + mddev->delta_disks));
8149 }
8150
8151 static int raid5_start_reshape(struct mddev *mddev)
8152 {
8153         struct r5conf *conf = mddev->private;
8154         struct md_rdev *rdev;
8155         int spares = 0;
8156         unsigned long flags;
8157
8158         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
8159                 return -EBUSY;
8160
8161         if (!check_stripe_cache(mddev))
8162                 return -ENOSPC;
8163
8164         if (has_failed(conf))
8165                 return -EINVAL;
8166
8167         rdev_for_each(rdev, mddev) {
8168                 if (!test_bit(In_sync, &rdev->flags)
8169                     && !test_bit(Faulty, &rdev->flags))
8170                         spares++;
8171         }
8172
8173         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
8174                 /* Not enough devices even to make a degraded array
8175                  * of that size
8176                  */
8177                 return -EINVAL;
8178
8179         /* Refuse to reduce size of the array.  Any reductions in
8180          * array size must be through explicit setting of array_size
8181          * attribute.
8182          */
8183         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
8184             < mddev->array_sectors) {
8185                 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
8186                         mdname(mddev));
8187                 return -EINVAL;
8188         }
8189
8190         atomic_set(&conf->reshape_stripes, 0);
8191         spin_lock_irq(&conf->device_lock);
8192         write_seqcount_begin(&conf->gen_lock);
8193         conf->previous_raid_disks = conf->raid_disks;
8194         conf->raid_disks += mddev->delta_disks;
8195         conf->prev_chunk_sectors = conf->chunk_sectors;
8196         conf->chunk_sectors = mddev->new_chunk_sectors;
8197         conf->prev_algo = conf->algorithm;
8198         conf->algorithm = mddev->new_layout;
8199         conf->generation++;
8200         /* Code that selects data_offset needs to see the generation update
8201          * if reshape_progress has been set - so a memory barrier needed.
8202          */
8203         smp_mb();
8204         if (mddev->reshape_backwards)
8205                 conf->reshape_progress = raid5_size(mddev, 0, 0);
8206         else
8207                 conf->reshape_progress = 0;
8208         conf->reshape_safe = conf->reshape_progress;
8209         write_seqcount_end(&conf->gen_lock);
8210         spin_unlock_irq(&conf->device_lock);
8211
8212         /* Now make sure any requests that proceeded on the assumption
8213          * the reshape wasn't running - like Discard or Read - have
8214          * completed.
8215          */
8216         mddev_suspend(mddev);
8217         mddev_resume(mddev);
8218
8219         /* Add some new drives, as many as will fit.
8220          * We know there are enough to make the newly sized array work.
8221          * Don't add devices if we are reducing the number of
8222          * devices in the array.  This is because it is not possible
8223          * to correctly record the "partially reconstructed" state of
8224          * such devices during the reshape and confusion could result.
8225          */
8226         if (mddev->delta_disks >= 0) {
8227                 rdev_for_each(rdev, mddev)
8228                         if (rdev->raid_disk < 0 &&
8229                             !test_bit(Faulty, &rdev->flags)) {
8230                                 if (raid5_add_disk(mddev, rdev) == 0) {
8231                                         if (rdev->raid_disk
8232                                             >= conf->previous_raid_disks)
8233                                                 set_bit(In_sync, &rdev->flags);
8234                                         else
8235                                                 rdev->recovery_offset = 0;
8236
8237                                         /* Failure here is OK */
8238                                         sysfs_link_rdev(mddev, rdev);
8239                                 }
8240                         } else if (rdev->raid_disk >= conf->previous_raid_disks
8241                                    && !test_bit(Faulty, &rdev->flags)) {
8242                                 /* This is a spare that was manually added */
8243                                 set_bit(In_sync, &rdev->flags);
8244                         }
8245
8246                 /* When a reshape changes the number of devices,
8247                  * ->degraded is measured against the larger of the
8248                  * pre and post number of devices.
8249                  */
8250                 spin_lock_irqsave(&conf->device_lock, flags);
8251                 mddev->degraded = raid5_calc_degraded(conf);
8252                 spin_unlock_irqrestore(&conf->device_lock, flags);
8253         }
8254         mddev->raid_disks = conf->raid_disks;
8255         mddev->reshape_position = conf->reshape_progress;
8256         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8257
8258         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
8259         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
8260         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
8261         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
8262         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
8263         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
8264                                                 "reshape");
8265         if (!mddev->sync_thread) {
8266                 mddev->recovery = 0;
8267                 spin_lock_irq(&conf->device_lock);
8268                 write_seqcount_begin(&conf->gen_lock);
8269                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
8270                 mddev->new_chunk_sectors =
8271                         conf->chunk_sectors = conf->prev_chunk_sectors;
8272                 mddev->new_layout = conf->algorithm = conf->prev_algo;
8273                 rdev_for_each(rdev, mddev)
8274                         rdev->new_data_offset = rdev->data_offset;
8275                 smp_wmb();
8276                 conf->generation --;
8277                 conf->reshape_progress = MaxSector;
8278                 mddev->reshape_position = MaxSector;
8279                 write_seqcount_end(&conf->gen_lock);
8280                 spin_unlock_irq(&conf->device_lock);
8281                 return -EAGAIN;
8282         }
8283         conf->reshape_checkpoint = jiffies;
8284         md_wakeup_thread(mddev->sync_thread);
8285         md_new_event(mddev);
8286         return 0;
8287 }
8288
8289 /* This is called from the reshape thread and should make any
8290  * changes needed in 'conf'
8291  */
8292 static void end_reshape(struct r5conf *conf)
8293 {
8294
8295         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8296                 struct md_rdev *rdev;
8297
8298                 spin_lock_irq(&conf->device_lock);
8299                 conf->previous_raid_disks = conf->raid_disks;
8300                 md_finish_reshape(conf->mddev);
8301                 smp_wmb();
8302                 conf->reshape_progress = MaxSector;
8303                 conf->mddev->reshape_position = MaxSector;
8304                 rdev_for_each(rdev, conf->mddev)
8305                         if (rdev->raid_disk >= 0 &&
8306                             !test_bit(Journal, &rdev->flags) &&
8307                             !test_bit(In_sync, &rdev->flags))
8308                                 rdev->recovery_offset = MaxSector;
8309                 spin_unlock_irq(&conf->device_lock);
8310                 wake_up(&conf->wait_for_overlap);
8311
8312                 if (conf->mddev->queue)
8313                         raid5_set_io_opt(conf);
8314         }
8315 }
8316
8317 /* This is called from the raid5d thread with mddev_lock held.
8318  * It makes config changes to the device.
8319  */
8320 static void raid5_finish_reshape(struct mddev *mddev)
8321 {
8322         struct r5conf *conf = mddev->private;
8323
8324         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8325
8326                 if (mddev->delta_disks <= 0) {
8327                         int d;
8328                         spin_lock_irq(&conf->device_lock);
8329                         mddev->degraded = raid5_calc_degraded(conf);
8330                         spin_unlock_irq(&conf->device_lock);
8331                         for (d = conf->raid_disks ;
8332                              d < conf->raid_disks - mddev->delta_disks;
8333                              d++) {
8334                                 struct md_rdev *rdev = conf->disks[d].rdev;
8335                                 if (rdev)
8336                                         clear_bit(In_sync, &rdev->flags);
8337                                 rdev = conf->disks[d].replacement;
8338                                 if (rdev)
8339                                         clear_bit(In_sync, &rdev->flags);
8340                         }
8341                 }
8342                 mddev->layout = conf->algorithm;
8343                 mddev->chunk_sectors = conf->chunk_sectors;
8344                 mddev->reshape_position = MaxSector;
8345                 mddev->delta_disks = 0;
8346                 mddev->reshape_backwards = 0;
8347         }
8348 }
8349
8350 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8351 {
8352         struct r5conf *conf = mddev->private;
8353
8354         if (quiesce) {
8355                 /* stop all writes */
8356                 lock_all_device_hash_locks_irq(conf);
8357                 /* '2' tells resync/reshape to pause so that all
8358                  * active stripes can drain
8359                  */
8360                 r5c_flush_cache(conf, INT_MAX);
8361                 /* need a memory barrier to make sure read_one_chunk() sees
8362                  * quiesce started and reverts to slow (locked) path.
8363                  */
8364                 smp_store_release(&conf->quiesce, 2);
8365                 wait_event_cmd(conf->wait_for_quiescent,
8366                                     atomic_read(&conf->active_stripes) == 0 &&
8367                                     atomic_read(&conf->active_aligned_reads) == 0,
8368                                     unlock_all_device_hash_locks_irq(conf),
8369                                     lock_all_device_hash_locks_irq(conf));
8370                 conf->quiesce = 1;
8371                 unlock_all_device_hash_locks_irq(conf);
8372                 /* allow reshape to continue */
8373                 wake_up(&conf->wait_for_overlap);
8374         } else {
8375                 /* re-enable writes */
8376                 lock_all_device_hash_locks_irq(conf);
8377                 conf->quiesce = 0;
8378                 wake_up(&conf->wait_for_quiescent);
8379                 wake_up(&conf->wait_for_overlap);
8380                 unlock_all_device_hash_locks_irq(conf);
8381         }
8382         log_quiesce(conf, quiesce);
8383 }
8384
8385 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8386 {
8387         struct r0conf *raid0_conf = mddev->private;
8388         sector_t sectors;
8389
8390         /* for raid0 takeover only one zone is supported */
8391         if (raid0_conf->nr_strip_zones > 1) {
8392                 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8393                         mdname(mddev));
8394                 return ERR_PTR(-EINVAL);
8395         }
8396
8397         sectors = raid0_conf->strip_zone[0].zone_end;
8398         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8399         mddev->dev_sectors = sectors;
8400         mddev->new_level = level;
8401         mddev->new_layout = ALGORITHM_PARITY_N;
8402         mddev->new_chunk_sectors = mddev->chunk_sectors;
8403         mddev->raid_disks += 1;
8404         mddev->delta_disks = 1;
8405         /* make sure it will be not marked as dirty */
8406         mddev->recovery_cp = MaxSector;
8407
8408         return setup_conf(mddev);
8409 }
8410
8411 static void *raid5_takeover_raid1(struct mddev *mddev)
8412 {
8413         int chunksect;
8414         void *ret;
8415
8416         if (mddev->raid_disks != 2 ||
8417             mddev->degraded > 1)
8418                 return ERR_PTR(-EINVAL);
8419
8420         /* Should check if there are write-behind devices? */
8421
8422         chunksect = 64*2; /* 64K by default */
8423
8424         /* The array must be an exact multiple of chunksize */
8425         while (chunksect && (mddev->array_sectors & (chunksect-1)))
8426                 chunksect >>= 1;
8427
8428         if ((chunksect<<9) < RAID5_STRIPE_SIZE((struct r5conf *)mddev->private))
8429                 /* array size does not allow a suitable chunk size */
8430                 return ERR_PTR(-EINVAL);
8431
8432         mddev->new_level = 5;
8433         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8434         mddev->new_chunk_sectors = chunksect;
8435
8436         ret = setup_conf(mddev);
8437         if (!IS_ERR(ret))
8438                 mddev_clear_unsupported_flags(mddev,
8439                         UNSUPPORTED_MDDEV_FLAGS);
8440         return ret;
8441 }
8442
8443 static void *raid5_takeover_raid6(struct mddev *mddev)
8444 {
8445         int new_layout;
8446
8447         switch (mddev->layout) {
8448         case ALGORITHM_LEFT_ASYMMETRIC_6:
8449                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8450                 break;
8451         case ALGORITHM_RIGHT_ASYMMETRIC_6:
8452                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8453                 break;
8454         case ALGORITHM_LEFT_SYMMETRIC_6:
8455                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8456                 break;
8457         case ALGORITHM_RIGHT_SYMMETRIC_6:
8458                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8459                 break;
8460         case ALGORITHM_PARITY_0_6:
8461                 new_layout = ALGORITHM_PARITY_0;
8462                 break;
8463         case ALGORITHM_PARITY_N:
8464                 new_layout = ALGORITHM_PARITY_N;
8465                 break;
8466         default:
8467                 return ERR_PTR(-EINVAL);
8468         }
8469         mddev->new_level = 5;
8470         mddev->new_layout = new_layout;
8471         mddev->delta_disks = -1;
8472         mddev->raid_disks -= 1;
8473         return setup_conf(mddev);
8474 }
8475
8476 static int raid5_check_reshape(struct mddev *mddev)
8477 {
8478         /* For a 2-drive array, the layout and chunk size can be changed
8479          * immediately as not restriping is needed.
8480          * For larger arrays we record the new value - after validation
8481          * to be used by a reshape pass.
8482          */
8483         struct r5conf *conf = mddev->private;
8484         int new_chunk = mddev->new_chunk_sectors;
8485
8486         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8487                 return -EINVAL;
8488         if (new_chunk > 0) {
8489                 if (!is_power_of_2(new_chunk))
8490                         return -EINVAL;
8491                 if (new_chunk < (PAGE_SIZE>>9))
8492                         return -EINVAL;
8493                 if (mddev->array_sectors & (new_chunk-1))
8494                         /* not factor of array size */
8495                         return -EINVAL;
8496         }
8497
8498         /* They look valid */
8499
8500         if (mddev->raid_disks == 2) {
8501                 /* can make the change immediately */
8502                 if (mddev->new_layout >= 0) {
8503                         conf->algorithm = mddev->new_layout;
8504                         mddev->layout = mddev->new_layout;
8505                 }
8506                 if (new_chunk > 0) {
8507                         conf->chunk_sectors = new_chunk ;
8508                         mddev->chunk_sectors = new_chunk;
8509                 }
8510                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8511                 md_wakeup_thread(mddev->thread);
8512         }
8513         return check_reshape(mddev);
8514 }
8515
8516 static int raid6_check_reshape(struct mddev *mddev)
8517 {
8518         int new_chunk = mddev->new_chunk_sectors;
8519
8520         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8521                 return -EINVAL;
8522         if (new_chunk > 0) {
8523                 if (!is_power_of_2(new_chunk))
8524                         return -EINVAL;
8525                 if (new_chunk < (PAGE_SIZE >> 9))
8526                         return -EINVAL;
8527                 if (mddev->array_sectors & (new_chunk-1))
8528                         /* not factor of array size */
8529                         return -EINVAL;
8530         }
8531
8532         /* They look valid */
8533         return check_reshape(mddev);
8534 }
8535
8536 static void *raid5_takeover(struct mddev *mddev)
8537 {
8538         /* raid5 can take over:
8539          *  raid0 - if there is only one strip zone - make it a raid4 layout
8540          *  raid1 - if there are two drives.  We need to know the chunk size
8541          *  raid4 - trivial - just use a raid4 layout.
8542          *  raid6 - Providing it is a *_6 layout
8543          */
8544         if (mddev->level == 0)
8545                 return raid45_takeover_raid0(mddev, 5);
8546         if (mddev->level == 1)
8547                 return raid5_takeover_raid1(mddev);
8548         if (mddev->level == 4) {
8549                 mddev->new_layout = ALGORITHM_PARITY_N;
8550                 mddev->new_level = 5;
8551                 return setup_conf(mddev);
8552         }
8553         if (mddev->level == 6)
8554                 return raid5_takeover_raid6(mddev);
8555
8556         return ERR_PTR(-EINVAL);
8557 }
8558
8559 static void *raid4_takeover(struct mddev *mddev)
8560 {
8561         /* raid4 can take over:
8562          *  raid0 - if there is only one strip zone
8563          *  raid5 - if layout is right
8564          */
8565         if (mddev->level == 0)
8566                 return raid45_takeover_raid0(mddev, 4);
8567         if (mddev->level == 5 &&
8568             mddev->layout == ALGORITHM_PARITY_N) {
8569                 mddev->new_layout = 0;
8570                 mddev->new_level = 4;
8571                 return setup_conf(mddev);
8572         }
8573         return ERR_PTR(-EINVAL);
8574 }
8575
8576 static struct md_personality raid5_personality;
8577
8578 static void *raid6_takeover(struct mddev *mddev)
8579 {
8580         /* Currently can only take over a raid5.  We map the
8581          * personality to an equivalent raid6 personality
8582          * with the Q block at the end.
8583          */
8584         int new_layout;
8585
8586         if (mddev->pers != &raid5_personality)
8587                 return ERR_PTR(-EINVAL);
8588         if (mddev->degraded > 1)
8589                 return ERR_PTR(-EINVAL);
8590         if (mddev->raid_disks > 253)
8591                 return ERR_PTR(-EINVAL);
8592         if (mddev->raid_disks < 3)
8593                 return ERR_PTR(-EINVAL);
8594
8595         switch (mddev->layout) {
8596         case ALGORITHM_LEFT_ASYMMETRIC:
8597                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8598                 break;
8599         case ALGORITHM_RIGHT_ASYMMETRIC:
8600                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8601                 break;
8602         case ALGORITHM_LEFT_SYMMETRIC:
8603                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8604                 break;
8605         case ALGORITHM_RIGHT_SYMMETRIC:
8606                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8607                 break;
8608         case ALGORITHM_PARITY_0:
8609                 new_layout = ALGORITHM_PARITY_0_6;
8610                 break;
8611         case ALGORITHM_PARITY_N:
8612                 new_layout = ALGORITHM_PARITY_N;
8613                 break;
8614         default:
8615                 return ERR_PTR(-EINVAL);
8616         }
8617         mddev->new_level = 6;
8618         mddev->new_layout = new_layout;
8619         mddev->delta_disks = 1;
8620         mddev->raid_disks += 1;
8621         return setup_conf(mddev);
8622 }
8623
8624 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8625 {
8626         struct r5conf *conf;
8627         int err;
8628
8629         err = mddev_lock(mddev);
8630         if (err)
8631                 return err;
8632         conf = mddev->private;
8633         if (!conf) {
8634                 mddev_unlock(mddev);
8635                 return -ENODEV;
8636         }
8637
8638         if (strncmp(buf, "ppl", 3) == 0) {
8639                 /* ppl only works with RAID 5 */
8640                 if (!raid5_has_ppl(conf) && conf->level == 5) {
8641                         err = log_init(conf, NULL, true);
8642                         if (!err) {
8643                                 err = resize_stripes(conf, conf->pool_size);
8644                                 if (err)
8645                                         log_exit(conf);
8646                         }
8647                 } else
8648                         err = -EINVAL;
8649         } else if (strncmp(buf, "resync", 6) == 0) {
8650                 if (raid5_has_ppl(conf)) {
8651                         mddev_suspend(mddev);
8652                         log_exit(conf);
8653                         mddev_resume(mddev);
8654                         err = resize_stripes(conf, conf->pool_size);
8655                 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8656                            r5l_log_disk_error(conf)) {
8657                         bool journal_dev_exists = false;
8658                         struct md_rdev *rdev;
8659
8660                         rdev_for_each(rdev, mddev)
8661                                 if (test_bit(Journal, &rdev->flags)) {
8662                                         journal_dev_exists = true;
8663                                         break;
8664                                 }
8665
8666                         if (!journal_dev_exists) {
8667                                 mddev_suspend(mddev);
8668                                 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8669                                 mddev_resume(mddev);
8670                         } else  /* need remove journal device first */
8671                                 err = -EBUSY;
8672                 } else
8673                         err = -EINVAL;
8674         } else {
8675                 err = -EINVAL;
8676         }
8677
8678         if (!err)
8679                 md_update_sb(mddev, 1);
8680
8681         mddev_unlock(mddev);
8682
8683         return err;
8684 }
8685
8686 static int raid5_start(struct mddev *mddev)
8687 {
8688         struct r5conf *conf = mddev->private;
8689
8690         return r5l_start(conf->log);
8691 }
8692
8693 static struct md_personality raid6_personality =
8694 {
8695         .name           = "raid6",
8696         .level          = 6,
8697         .owner          = THIS_MODULE,
8698         .make_request   = raid5_make_request,
8699         .run            = raid5_run,
8700         .start          = raid5_start,
8701         .free           = raid5_free,
8702         .status         = raid5_status,
8703         .error_handler  = raid5_error,
8704         .hot_add_disk   = raid5_add_disk,
8705         .hot_remove_disk= raid5_remove_disk,
8706         .spare_active   = raid5_spare_active,
8707         .sync_request   = raid5_sync_request,
8708         .resize         = raid5_resize,
8709         .size           = raid5_size,
8710         .check_reshape  = raid6_check_reshape,
8711         .start_reshape  = raid5_start_reshape,
8712         .finish_reshape = raid5_finish_reshape,
8713         .quiesce        = raid5_quiesce,
8714         .takeover       = raid6_takeover,
8715         .change_consistency_policy = raid5_change_consistency_policy,
8716 };
8717 static struct md_personality raid5_personality =
8718 {
8719         .name           = "raid5",
8720         .level          = 5,
8721         .owner          = THIS_MODULE,
8722         .make_request   = raid5_make_request,
8723         .run            = raid5_run,
8724         .start          = raid5_start,
8725         .free           = raid5_free,
8726         .status         = raid5_status,
8727         .error_handler  = raid5_error,
8728         .hot_add_disk   = raid5_add_disk,
8729         .hot_remove_disk= raid5_remove_disk,
8730         .spare_active   = raid5_spare_active,
8731         .sync_request   = raid5_sync_request,
8732         .resize         = raid5_resize,
8733         .size           = raid5_size,
8734         .check_reshape  = raid5_check_reshape,
8735         .start_reshape  = raid5_start_reshape,
8736         .finish_reshape = raid5_finish_reshape,
8737         .quiesce        = raid5_quiesce,
8738         .takeover       = raid5_takeover,
8739         .change_consistency_policy = raid5_change_consistency_policy,
8740 };
8741
8742 static struct md_personality raid4_personality =
8743 {
8744         .name           = "raid4",
8745         .level          = 4,
8746         .owner          = THIS_MODULE,
8747         .make_request   = raid5_make_request,
8748         .run            = raid5_run,
8749         .start          = raid5_start,
8750         .free           = raid5_free,
8751         .status         = raid5_status,
8752         .error_handler  = raid5_error,
8753         .hot_add_disk   = raid5_add_disk,
8754         .hot_remove_disk= raid5_remove_disk,
8755         .spare_active   = raid5_spare_active,
8756         .sync_request   = raid5_sync_request,
8757         .resize         = raid5_resize,
8758         .size           = raid5_size,
8759         .check_reshape  = raid5_check_reshape,
8760         .start_reshape  = raid5_start_reshape,
8761         .finish_reshape = raid5_finish_reshape,
8762         .quiesce        = raid5_quiesce,
8763         .takeover       = raid4_takeover,
8764         .change_consistency_policy = raid5_change_consistency_policy,
8765 };
8766
8767 static int __init raid5_init(void)
8768 {
8769         int ret;
8770
8771         raid5_wq = alloc_workqueue("raid5wq",
8772                 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8773         if (!raid5_wq)
8774                 return -ENOMEM;
8775
8776         ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8777                                       "md/raid5:prepare",
8778                                       raid456_cpu_up_prepare,
8779                                       raid456_cpu_dead);
8780         if (ret) {
8781                 destroy_workqueue(raid5_wq);
8782                 return ret;
8783         }
8784         register_md_personality(&raid6_personality);
8785         register_md_personality(&raid5_personality);
8786         register_md_personality(&raid4_personality);
8787         return 0;
8788 }
8789
8790 static void raid5_exit(void)
8791 {
8792         unregister_md_personality(&raid6_personality);
8793         unregister_md_personality(&raid5_personality);
8794         unregister_md_personality(&raid4_personality);
8795         cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8796         destroy_workqueue(raid5_wq);
8797 }
8798
8799 module_init(raid5_init);
8800 module_exit(raid5_exit);
8801 MODULE_LICENSE("GPL");
8802 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8803 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8804 MODULE_ALIAS("md-raid5");
8805 MODULE_ALIAS("md-raid4");
8806 MODULE_ALIAS("md-level-5");
8807 MODULE_ALIAS("md-level-4");
8808 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8809 MODULE_ALIAS("md-raid6");
8810 MODULE_ALIAS("md-level-6");
8811
8812 /* This used to be two separate modules, they were: */
8813 MODULE_ALIAS("raid5");
8814 MODULE_ALIAS("raid6");