drivers/md/raid5.c

   1 /*
   2  * raid5.c : Multiple Devices driver for Linux
   3  *         Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   4  *         Copyright (C) 1999, 2000 Ingo Molnar
   5  *         Copyright (C) 2002, 2003 H. Peter Anvin
   6  *
   7  * RAID-4/5/6 management functions.
   8  * Thanks to Penguin Computing for making the RAID-6 development possible
   9  * by donating a test server!
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * BITMAP UNPLUGGING:
  23  *
  24  * The sequencing for updating the bitmap reliably is a little
  25  * subtle (and I got it wrong the first time) so it deserves some
  26  * explanation.
  27  *
  28  * We group bitmap updates into batches.  Each batch has a number.
  29  * We may write out several batches at once, but that isn't very important.
  30  * conf->seq_write is the number of the last batch successfully written.
  31  * conf->seq_flush is the number of the last batch that was closed to
  32  *    new additions.
  33  * When we discover that we will need to write to any block in a stripe
  34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
  35  * the number of the batch it will be in. This is seq_flush+1.
  36  * When we are ready to do a write, if that batch hasn't been written yet,
  37  *   we plug the array and queue the stripe for later.
  38  * When an unplug happens, we increment bm_flush, thus closing the current
  39  *   batch.
  40  * When we notice that bm_flush > bm_write, we write out all pending updates
  41  * to the bitmap, and advance bm_write to where bm_flush was.
  42  * This may occasionally write a bit out twice, but is sure never to
  43  * miss any bits.
  44  */
  45
  46 #include <linux/blkdev.h>
  47 #include <linux/kthread.h>
  48 #include <linux/raid/pq.h>
  49 #include <linux/async_tx.h>
  50 #include <linux/module.h>
  51 #include <linux/async.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/cpu.h>
  54 #include <linux/slab.h>
  55 #include <linux/ratelimit.h>
  56 #include <trace/events/block.h>
  57
  58 #include "md.h"
  59 #include "raid5.h"
  60 #include "raid0.h"
  61 #include "bitmap.h"
  62
  63 /*
  64  * Stripe cache
  65  */
  66
  67 #define NR_STRIPES              256
  68 #define STRIPE_SIZE             PAGE_SIZE
  69 #define STRIPE_SHIFT            (PAGE_SHIFT - 9)
  70 #define STRIPE_SECTORS          (STRIPE_SIZE>>9)
  71 #define IO_THRESHOLD            1
  72 #define BYPASS_THRESHOLD        1
  73 #define NR_HASH                 (PAGE_SIZE / sizeof(struct hlist_head))
  74 #define HASH_MASK               (NR_HASH - 1)
  75
  76 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
  77 {
  78         int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
  79         return &conf->stripe_hashtbl[hash];
  80 }
  81
  82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
  83  * order without overlap.  There may be several bio's per stripe+device, and
  84  * a bio could span several devices.
  85  * When walking this list for a particular stripe+device, we must never proceed
  86  * beyond a bio that extends past this device, as the next bio might no longer
  87  * be valid.
  88  * This function is used to determine the 'next' bio in the list, given the sector
  89  * of the current stripe+device
  90  */
  91 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
  92 {
  93         int sectors = bio->bi_size >> 9;
  94         if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
  95                 return bio->bi_next;
  96         else
  97                 return NULL;
  98 }
  99
 100 /*
 101  * We maintain a biased count of active stripes in the bottom 16 bits of
 102  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
 103  */
 104 static inline int raid5_bi_processed_stripes(struct bio *bio)
 105 {
 106         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 107         return (atomic_read(segments) >> 16) & 0xffff;
 108 }
 109
 110 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
 111 {
 112         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 113         return atomic_sub_return(1, segments) & 0xffff;
 114 }
 115
 116 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
 117 {
 118         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 119         atomic_inc(segments);
 120 }
 121
 122 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
 123         unsigned int cnt)
 124 {
 125         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 126         int old, new;
 127
 128         do {
 129                 old = atomic_read(segments);
 130                 new = (old & 0xffff) | (cnt << 16);
 131         } while (atomic_cmpxchg(segments, old, new) != old);
 132 }
 133
 134 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
 135 {
 136         atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
 137         atomic_set(segments, cnt);
 138 }
 139
 140 /* Find first data disk in a raid6 stripe */
 141 static inline int raid6_d0(struct stripe_head *sh)
 142 {
 143         if (sh->ddf_layout)
 144                 /* ddf always start from first device */
 145                 return 0;
 146         /* md starts just after Q block */
 147         if (sh->qd_idx == sh->disks - 1)
 148                 return 0;
 149         else
 150                 return sh->qd_idx + 1;
 151 }
 152 static inline int raid6_next_disk(int disk, int raid_disks)
 153 {
 154         disk++;
 155         return (disk < raid_disks) ? disk : 0;
 156 }
 157
 158 /* When walking through the disks in a raid5, starting at raid6_d0,
 159  * We need to map each disk to a 'slot', where the data disks are slot
 160  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
 161  * is raid_disks-1.  This help does that mapping.
 162  */
 163 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
 164                              int *count, int syndrome_disks)
 165 {
 166         int slot = *count;
 167
 168         if (sh->ddf_layout)
 169                 (*count)++;
 170         if (idx == sh->pd_idx)
 171                 return syndrome_disks;
 172         if (idx == sh->qd_idx)
 173                 return syndrome_disks + 1;
 174         if (!sh->ddf_layout)
 175                 (*count)++;
 176         return slot;
 177 }
 178
 179 static void return_io(struct bio *return_bi)
 180 {
 181         struct bio *bi = return_bi;
 182         while (bi) {
 183
 184                 return_bi = bi->bi_next;
 185                 bi->bi_next = NULL;
 186                 bi->bi_size = 0;
 187                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
 188                                          bi, 0);
 189                 bio_endio(bi, 0);
 190                 bi = return_bi;
 191         }
 192 }
 193
 194 static void print_raid5_conf (struct r5conf *conf);
 195
 196 static int stripe_operations_active(struct stripe_head *sh)
 197 {
 198         return sh->check_state || sh->reconstruct_state ||
 199                test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
 200                test_bit(STRIPE_COMPUTE_RUN, &sh->state);
 201 }
 202
 203 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
 204 {
 205         BUG_ON(!list_empty(&sh->lru));
 206         BUG_ON(atomic_read(&conf->active_stripes)==0);
 207         if (test_bit(STRIPE_HANDLE, &sh->state)) {
 208                 if (test_bit(STRIPE_DELAYED, &sh->state) &&
 209                     !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 210                         list_add_tail(&sh->lru, &conf->delayed_list);
 211                 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
 212                            sh->bm_seq - conf->seq_write > 0)
 213                         list_add_tail(&sh->lru, &conf->bitmap_list);
 214                 else {
 215                         clear_bit(STRIPE_DELAYED, &sh->state);
 216                         clear_bit(STRIPE_BIT_DELAY, &sh->state);
 217                         list_add_tail(&sh->lru, &conf->handle_list);
 218                 }
 219                 md_wakeup_thread(conf->mddev->thread);
 220         } else {
 221                 BUG_ON(stripe_operations_active(sh));
 222                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 223                         if (atomic_dec_return(&conf->preread_active_stripes)
 224                             < IO_THRESHOLD)
 225                                 md_wakeup_thread(conf->mddev->thread);
 226                 atomic_dec(&conf->active_stripes);
 227                 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
 228                         list_add_tail(&sh->lru, &conf->inactive_list);
 229                         wake_up(&conf->wait_for_stripe);
 230                         if (conf->retry_read_aligned)
 231                                 md_wakeup_thread(conf->mddev->thread);
 232                 }
 233         }
 234 }
 235
 236 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
 237 {
 238         if (atomic_dec_and_test(&sh->count))
 239                 do_release_stripe(conf, sh);
 240 }
 241
 242 static void release_stripe(struct stripe_head *sh)
 243 {
 244         struct r5conf *conf = sh->raid_conf;
 245         unsigned long flags;
 246
 247         local_irq_save(flags);
 248         if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
 249                 do_release_stripe(conf, sh);
 250                 spin_unlock(&conf->device_lock);
 251         }
 252         local_irq_restore(flags);
 253 }
 254
 255 static inline void remove_hash(struct stripe_head *sh)
 256 {
 257         pr_debug("remove_hash(), stripe %llu\n",
 258                 (unsigned long long)sh->sector);
 259
 260         hlist_del_init(&sh->hash);
 261 }
 262
 263 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
 264 {
 265         struct hlist_head *hp = stripe_hash(conf, sh->sector);
 266
 267         pr_debug("insert_hash(), stripe %llu\n",
 268                 (unsigned long long)sh->sector);
 269
 270         hlist_add_head(&sh->hash, hp);
 271 }
 272
 273
 274 /* find an idle stripe, make sure it is unhashed, and return it. */
 275 static struct stripe_head *get_free_stripe(struct r5conf *conf)
 276 {
 277         struct stripe_head *sh = NULL;
 278         struct list_head *first;
 279
 280         if (list_empty(&conf->inactive_list))
 281                 goto out;
 282         first = conf->inactive_list.next;
 283         sh = list_entry(first, struct stripe_head, lru);
 284         list_del_init(first);
 285         remove_hash(sh);
 286         atomic_inc(&conf->active_stripes);
 287 out:
 288         return sh;
 289 }
 290
 291 static void shrink_buffers(struct stripe_head *sh)
 292 {
 293         struct page *p;
 294         int i;
 295         int num = sh->raid_conf->pool_size;
 296
 297         for (i = 0; i < num ; i++) {
 298                 p = sh->dev[i].page;
 299                 if (!p)
 300                         continue;
 301                 sh->dev[i].page = NULL;
 302                 put_page(p);
 303         }
 304 }
 305
 306 static int grow_buffers(struct stripe_head *sh)
 307 {
 308         int i;
 309         int num = sh->raid_conf->pool_size;
 310
 311         for (i = 0; i < num; i++) {
 312                 struct page *page;
 313
 314                 if (!(page = alloc_page(GFP_KERNEL))) {
 315                         return 1;
 316                 }
 317                 sh->dev[i].page = page;
 318         }
 319         return 0;
 320 }
 321
 322 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
 323 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
 324                             struct stripe_head *sh);
 325
 326 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
 327 {
 328         struct r5conf *conf = sh->raid_conf;
 329         int i;
 330
 331         BUG_ON(atomic_read(&sh->count) != 0);
 332         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
 333         BUG_ON(stripe_operations_active(sh));
 334
 335         pr_debug("init_stripe called, stripe %llu\n",
 336                 (unsigned long long)sh->sector);
 337
 338         remove_hash(sh);
 339
 340         sh->generation = conf->generation - previous;
 341         sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
 342         sh->sector = sector;
 343         stripe_set_idx(sector, conf, previous, sh);
 344         sh->state = 0;
 345
 346
 347         for (i = sh->disks; i--; ) {
 348                 struct r5dev *dev = &sh->dev[i];
 349
 350                 if (dev->toread || dev->read || dev->towrite || dev->written ||
 351                     test_bit(R5_LOCKED, &dev->flags)) {
 352                         printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
 353                                (unsigned long long)sh->sector, i, dev->toread,
 354                                dev->read, dev->towrite, dev->written,
 355                                test_bit(R5_LOCKED, &dev->flags));
 356                         WARN_ON(1);
 357                 }
 358                 dev->flags = 0;
 359                 raid5_build_block(sh, i, previous);
 360         }
 361         insert_hash(conf, sh);
 362 }
 363
 364 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
 365                                          short generation)
 366 {
 367         struct stripe_head *sh;
 368
 369         pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
 370         hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
 371                 if (sh->sector == sector && sh->generation == generation)
 372                         return sh;
 373         pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
 374         return NULL;
 375 }
 376
 377 /*
 378  * Need to check if array has failed when deciding whether to:
 379  *  - start an array
 380  *  - remove non-faulty devices
 381  *  - add a spare
 382  *  - allow a reshape
 383  * This determination is simple when no reshape is happening.
 384  * However if there is a reshape, we need to carefully check
 385  * both the before and after sections.
 386  * This is because some failed devices may only affect one
 387  * of the two sections, and some non-in_sync devices may
 388  * be insync in the section most affected by failed devices.
 389  */
 390 static int calc_degraded(struct r5conf *conf)
 391 {
 392         int degraded, degraded2;
 393         int i;
 394
 395         rcu_read_lock();
 396         degraded = 0;
 397         for (i = 0; i < conf->previous_raid_disks; i++) {
 398                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 399                 if (rdev && test_bit(Faulty, &rdev->flags))
 400                         rdev = rcu_dereference(conf->disks[i].replacement);
 401                 if (!rdev || test_bit(Faulty, &rdev->flags))
 402                         degraded++;
 403                 else if (test_bit(In_sync, &rdev->flags))
 404                         ;
 405                 else
 406                         /* not in-sync or faulty.
 407                          * If the reshape increases the number of devices,
 408                          * this is being recovered by the reshape, so
 409                          * this 'previous' section is not in_sync.
 410                          * If the number of devices is being reduced however,
 411                          * the device can only be part of the array if
 412                          * we are reverting a reshape, so this section will
 413                          * be in-sync.
 414                          */
 415                         if (conf->raid_disks >= conf->previous_raid_disks)
 416                                 degraded++;
 417         }
 418         rcu_read_unlock();
 419         if (conf->raid_disks == conf->previous_raid_disks)
 420                 return degraded;
 421         rcu_read_lock();
 422         degraded2 = 0;
 423         for (i = 0; i < conf->raid_disks; i++) {
 424                 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
 425                 if (rdev && test_bit(Faulty, &rdev->flags))
 426                         rdev = rcu_dereference(conf->disks[i].replacement);
 427                 if (!rdev || test_bit(Faulty, &rdev->flags))
 428                         degraded2++;
 429                 else if (test_bit(In_sync, &rdev->flags))
 430                         ;
 431                 else
 432                         /* not in-sync or faulty.
 433                          * If reshape increases the number of devices, this
 434                          * section has already been recovered, else it
 435                          * almost certainly hasn't.
 436                          */
 437                         if (conf->raid_disks <= conf->previous_raid_disks)
 438                                 degraded2++;
 439         }
 440         rcu_read_unlock();
 441         if (degraded2 > degraded)
 442                 return degraded2;
 443         return degraded;
 444 }
 445
 446 static int has_failed(struct r5conf *conf)
 447 {
 448         int degraded;
 449
 450         if (conf->mddev->reshape_position == MaxSector)
 451                 return conf->mddev->degraded > conf->max_degraded;
 452
 453         degraded = calc_degraded(conf);
 454         if (degraded > conf->max_degraded)
 455                 return 1;
 456         return 0;
 457 }
 458
 459 static struct stripe_head *
 460 get_active_stripe(struct r5conf *conf, sector_t sector,
 461                   int previous, int noblock, int noquiesce)
 462 {
 463         struct stripe_head *sh;
 464
 465         pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
 466
 467         spin_lock_irq(&conf->device_lock);
 468
 469         do {
 470                 wait_event_lock_irq(conf->wait_for_stripe,
 471                                     conf->quiesce == 0 || noquiesce,
 472                                     conf->device_lock);
 473                 sh = __find_stripe(conf, sector, conf->generation - previous);
 474                 if (!sh) {
 475                         if (!conf->inactive_blocked)
 476                                 sh = get_free_stripe(conf);
 477                         if (noblock && sh == NULL)
 478                                 break;
 479                         if (!sh) {
 480                                 conf->inactive_blocked = 1;
 481                                 wait_event_lock_irq(conf->wait_for_stripe,
 482                                                     !list_empty(&conf->inactive_list) &&
 483                                                     (atomic_read(&conf->active_stripes)
 484                                                      < (conf->max_nr_stripes *3/4)
 485                                                      || !conf->inactive_blocked),
 486                                                     conf->device_lock);
 487                                 conf->inactive_blocked = 0;
 488                         } else
 489                                 init_stripe(sh, sector, previous);
 490                 } else {
 491                         if (atomic_read(&sh->count)) {
 492                                 BUG_ON(!list_empty(&sh->lru)
 493                                     && !test_bit(STRIPE_EXPANDING, &sh->state)
 494                                     && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
 495                         } else {
 496                                 if (!test_bit(STRIPE_HANDLE, &sh->state))
 497                                         atomic_inc(&conf->active_stripes);
 498                                 if (list_empty(&sh->lru) &&
 499                                     !test_bit(STRIPE_EXPANDING, &sh->state))
 500                                         BUG();
 501                                 list_del_init(&sh->lru);
 502                         }
 503                 }
 504         } while (sh == NULL);
 505
 506         if (sh)
 507                 atomic_inc(&sh->count);
 508
 509         spin_unlock_irq(&conf->device_lock);
 510         return sh;
 511 }
 512
 513 /* Determine if 'data_offset' or 'new_data_offset' should be used
 514  * in this stripe_head.
 515  */
 516 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
 517 {
 518         sector_t progress = conf->reshape_progress;
 519         /* Need a memory barrier to make sure we see the value
 520          * of conf->generation, or ->data_offset that was set before
 521          * reshape_progress was updated.
 522          */
 523         smp_rmb();
 524         if (progress == MaxSector)
 525                 return 0;
 526         if (sh->generation == conf->generation - 1)
 527                 return 0;
 528         /* We are in a reshape, and this is a new-generation stripe,
 529          * so use new_data_offset.
 530          */
 531         return 1;
 532 }
 533
 534 static void
 535 raid5_end_read_request(struct bio *bi, int error);
 536 static void
 537 raid5_end_write_request(struct bio *bi, int error);
 538
 539 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
 540 {
 541         struct r5conf *conf = sh->raid_conf;
 542         int i, disks = sh->disks;
 543
 544         might_sleep();
 545
 546         for (i = disks; i--; ) {
 547                 int rw;
 548                 int replace_only = 0;
 549                 struct bio *bi, *rbi;
 550                 struct md_rdev *rdev, *rrdev = NULL;
 551                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 552                         if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
 553                                 rw = WRITE_FUA;
 554                         else
 555                                 rw = WRITE;
 556                         if (test_bit(R5_Discard, &sh->dev[i].flags))
 557                                 rw |= REQ_DISCARD;
 558                 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
 559                         rw = READ;
 560                 else if (test_and_clear_bit(R5_WantReplace,
 561                                             &sh->dev[i].flags)) {
 562                         rw = WRITE;
 563                         replace_only = 1;
 564                 } else
 565                         continue;
 566                 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
 567                         rw |= REQ_SYNC;
 568
 569                 bi = &sh->dev[i].req;
 570                 rbi = &sh->dev[i].rreq; /* For writing to replacement */
 571
 572                 bi->bi_rw = rw;
 573                 rbi->bi_rw = rw;
 574                 if (rw & WRITE) {
 575                         bi->bi_end_io = raid5_end_write_request;
 576                         rbi->bi_end_io = raid5_end_write_request;
 577                 } else
 578                         bi->bi_end_io = raid5_end_read_request;
 579
 580                 rcu_read_lock();
 581                 rrdev = rcu_dereference(conf->disks[i].replacement);
 582                 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
 583                 rdev = rcu_dereference(conf->disks[i].rdev);
 584                 if (!rdev) {
 585                         rdev = rrdev;
 586                         rrdev = NULL;
 587                 }
 588                 if (rw & WRITE) {
 589                         if (replace_only)
 590                                 rdev = NULL;
 591                         if (rdev == rrdev)
 592                                 /* We raced and saw duplicates */
 593                                 rrdev = NULL;
 594                 } else {
 595                         if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
 596                                 rdev = rrdev;
 597                         rrdev = NULL;
 598                 }
 599
 600                 if (rdev && test_bit(Faulty, &rdev->flags))
 601                         rdev = NULL;
 602                 if (rdev)
 603                         atomic_inc(&rdev->nr_pending);
 604                 if (rrdev && test_bit(Faulty, &rrdev->flags))
 605                         rrdev = NULL;
 606                 if (rrdev)
 607                         atomic_inc(&rrdev->nr_pending);
 608                 rcu_read_unlock();
 609
 610                 /* We have already checked bad blocks for reads.  Now
 611                  * need to check for writes.  We never accept write errors
 612                  * on the replacement, so we don't to check rrdev.
 613                  */
 614                 while ((rw & WRITE) && rdev &&
 615                        test_bit(WriteErrorSeen, &rdev->flags)) {
 616                         sector_t first_bad;
 617                         int bad_sectors;
 618                         int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
 619                                               &first_bad, &bad_sectors);
 620                         if (!bad)
 621                                 break;
 622
 623                         if (bad < 0) {
 624                                 set_bit(BlockedBadBlocks, &rdev->flags);
 625                                 if (!conf->mddev->external &&
 626                                     conf->mddev->flags) {
 627                                         /* It is very unlikely, but we might
 628                                          * still need to write out the
 629                                          * bad block log - better give it
 630                                          * a chance*/
 631                                         md_check_recovery(conf->mddev);
 632                                 }
 633                                 /*
 634                                  * Because md_wait_for_blocked_rdev
 635                                  * will dec nr_pending, we must
 636                                  * increment it first.
 637                                  */
 638                                 atomic_inc(&rdev->nr_pending);
 639                                 md_wait_for_blocked_rdev(rdev, conf->mddev);
 640                         } else {
 641                                 /* Acknowledged bad block - skip the write */
 642                                 rdev_dec_pending(rdev, conf->mddev);
 643                                 rdev = NULL;
 644                         }
 645                 }
 646
 647                 if (rdev) {
 648                         if (s->syncing || s->expanding || s->expanded
 649                             || s->replacing)
 650                                 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
 651
 652                         set_bit(STRIPE_IO_STARTED, &sh->state);
 653
 654                         bi->bi_bdev = rdev->bdev;
 655                         pr_debug("%s: for %llu schedule op %ld on disc %d\n",
 656                                 __func__, (unsigned long long)sh->sector,
 657                                 bi->bi_rw, i);
 658                         atomic_inc(&sh->count);
 659                         if (use_new_offset(conf, sh))
 660                                 bi->bi_sector = (sh->sector
 661                                                  + rdev->new_data_offset);
 662                         else
 663                                 bi->bi_sector = (sh->sector
 664                                                  + rdev->data_offset);
 665                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
 666                                 bi->bi_rw |= REQ_FLUSH;
 667
 668                         bi->bi_flags = 1 << BIO_UPTODATE;
 669                         bi->bi_idx = 0;
 670                         bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 671                         bi->bi_io_vec[0].bv_offset = 0;
 672                         bi->bi_size = STRIPE_SIZE;
 673                         bi->bi_next = NULL;
 674                         if (rrdev)
 675                                 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
 676
 677                         if (conf->mddev->gendisk)
 678                                 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
 679                                                       bi, disk_devt(conf->mddev->gendisk),
 680                                                       sh->dev[i].sector);
 681                         generic_make_request(bi);
 682                 }
 683                 if (rrdev) {
 684                         if (s->syncing || s->expanding || s->expanded
 685                             || s->replacing)
 686                                 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
 687
 688                         set_bit(STRIPE_IO_STARTED, &sh->state);
 689
 690                         rbi->bi_bdev = rrdev->bdev;
 691                         pr_debug("%s: for %llu schedule op %ld on "
 692                                  "replacement disc %d\n",
 693                                 __func__, (unsigned long long)sh->sector,
 694                                 rbi->bi_rw, i);
 695                         atomic_inc(&sh->count);
 696                         if (use_new_offset(conf, sh))
 697                                 rbi->bi_sector = (sh->sector
 698                                                   + rrdev->new_data_offset);
 699                         else
 700                                 rbi->bi_sector = (sh->sector
 701                                                   + rrdev->data_offset);
 702                         rbi->bi_flags = 1 << BIO_UPTODATE;
 703                         rbi->bi_idx = 0;
 704                         rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
 705                         rbi->bi_io_vec[0].bv_offset = 0;
 706                         rbi->bi_size = STRIPE_SIZE;
 707                         rbi->bi_next = NULL;
 708                         if (conf->mddev->gendisk)
 709                                 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
 710                                                       rbi, disk_devt(conf->mddev->gendisk),
 711                                                       sh->dev[i].sector);
 712                         generic_make_request(rbi);
 713                 }
 714                 if (!rdev && !rrdev) {
 715                         if (rw & WRITE)
 716                                 set_bit(STRIPE_DEGRADED, &sh->state);
 717                         pr_debug("skip op %ld on disc %d for sector %llu\n",
 718                                 bi->bi_rw, i, (unsigned long long)sh->sector);
 719                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 720                         set_bit(STRIPE_HANDLE, &sh->state);
 721                 }
 722         }
 723 }
 724
 725 static struct dma_async_tx_descriptor *
 726 async_copy_data(int frombio, struct bio *bio, struct page *page,
 727         sector_t sector, struct dma_async_tx_descriptor *tx)
 728 {
 729         struct bio_vec *bvl;
 730         struct page *bio_page;
 731         int i;
 732         int page_offset;
 733         struct async_submit_ctl submit;
 734         enum async_tx_flags flags = 0;
 735
 736         if (bio->bi_sector >= sector)
 737                 page_offset = (signed)(bio->bi_sector - sector) * 512;
 738         else
 739                 page_offset = (signed)(sector - bio->bi_sector) * -512;
 740
 741         if (frombio)
 742                 flags |= ASYNC_TX_FENCE;
 743         init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
 744
 745         bio_for_each_segment(bvl, bio, i) {
 746                 int len = bvl->bv_len;
 747                 int clen;
 748                 int b_offset = 0;
 749
 750                 if (page_offset < 0) {
 751                         b_offset = -page_offset;
 752                         page_offset += b_offset;
 753                         len -= b_offset;
 754                 }
 755
 756                 if (len > 0 && page_offset + len > STRIPE_SIZE)
 757                         clen = STRIPE_SIZE - page_offset;
 758                 else
 759                         clen = len;
 760
 761                 if (clen > 0) {
 762                         b_offset += bvl->bv_offset;
 763                         bio_page = bvl->bv_page;
 764                         if (frombio)
 765                                 tx = async_memcpy(page, bio_page, page_offset,
 766                                                   b_offset, clen, &submit);
 767                         else
 768                                 tx = async_memcpy(bio_page, page, b_offset,
 769                                                   page_offset, clen, &submit);
 770                 }
 771                 /* chain the operations */
 772                 submit.depend_tx = tx;
 773
 774                 if (clen < len) /* hit end of page */
 775                         break;
 776                 page_offset +=  len;
 777         }
 778
 779         return tx;
 780 }
 781
 782 static void ops_complete_biofill(void *stripe_head_ref)
 783 {
 784         struct stripe_head *sh = stripe_head_ref;
 785         struct bio *return_bi = NULL;
 786         int i;
 787
 788         pr_debug("%s: stripe %llu\n", __func__,
 789                 (unsigned long long)sh->sector);
 790
 791         /* clear completed biofills */
 792         for (i = sh->disks; i--; ) {
 793                 struct r5dev *dev = &sh->dev[i];
 794
 795                 /* acknowledge completion of a biofill operation */
 796                 /* and check if we need to reply to a read request,
 797                  * new R5_Wantfill requests are held off until
 798                  * !STRIPE_BIOFILL_RUN
 799                  */
 800                 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
 801                         struct bio *rbi, *rbi2;
 802
 803                         BUG_ON(!dev->read);
 804                         rbi = dev->read;
 805                         dev->read = NULL;
 806                         while (rbi && rbi->bi_sector <
 807                                 dev->sector + STRIPE_SECTORS) {
 808                                 rbi2 = r5_next_bio(rbi, dev->sector);
 809                                 if (!raid5_dec_bi_active_stripes(rbi)) {
 810                                         rbi->bi_next = return_bi;
 811                                         return_bi = rbi;
 812                                 }
 813                                 rbi = rbi2;
 814                         }
 815                 }
 816         }
 817         clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
 818
 819         return_io(return_bi);
 820
 821         set_bit(STRIPE_HANDLE, &sh->state);
 822         release_stripe(sh);
 823 }
 824
 825 static void ops_run_biofill(struct stripe_head *sh)
 826 {
 827         struct dma_async_tx_descriptor *tx = NULL;
 828         struct async_submit_ctl submit;
 829         int i;
 830
 831         pr_debug("%s: stripe %llu\n", __func__,
 832                 (unsigned long long)sh->sector);
 833
 834         for (i = sh->disks; i--; ) {
 835                 struct r5dev *dev = &sh->dev[i];
 836                 if (test_bit(R5_Wantfill, &dev->flags)) {
 837                         struct bio *rbi;
 838                         spin_lock_irq(&sh->stripe_lock);
 839                         dev->read = rbi = dev->toread;
 840                         dev->toread = NULL;
 841                         spin_unlock_irq(&sh->stripe_lock);
 842                         while (rbi && rbi->bi_sector <
 843                                 dev->sector + STRIPE_SECTORS) {
 844                                 tx = async_copy_data(0, rbi, dev->page,
 845                                         dev->sector, tx);
 846                                 rbi = r5_next_bio(rbi, dev->sector);
 847                         }
 848                 }
 849         }
 850
 851         atomic_inc(&sh->count);
 852         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
 853         async_trigger_callback(&submit);
 854 }
 855
 856 static void mark_target_uptodate(struct stripe_head *sh, int target)
 857 {
 858         struct r5dev *tgt;
 859
 860         if (target < 0)
 861                 return;
 862
 863         tgt = &sh->dev[target];
 864         set_bit(R5_UPTODATE, &tgt->flags);
 865         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 866         clear_bit(R5_Wantcompute, &tgt->flags);
 867 }
 868
 869 static void ops_complete_compute(void *stripe_head_ref)
 870 {
 871         struct stripe_head *sh = stripe_head_ref;
 872
 873         pr_debug("%s: stripe %llu\n", __func__,
 874                 (unsigned long long)sh->sector);
 875
 876         /* mark the computed target(s) as uptodate */
 877         mark_target_uptodate(sh, sh->ops.target);
 878         mark_target_uptodate(sh, sh->ops.target2);
 879
 880         clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
 881         if (sh->check_state == check_state_compute_run)
 882                 sh->check_state = check_state_compute_result;
 883         set_bit(STRIPE_HANDLE, &sh->state);
 884         release_stripe(sh);
 885 }
 886
 887 /* return a pointer to the address conversion region of the scribble buffer */
 888 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
 889                                  struct raid5_percpu *percpu)
 890 {
 891         return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
 892 }
 893
 894 static struct dma_async_tx_descriptor *
 895 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
 896 {
 897         int disks = sh->disks;
 898         struct page **xor_srcs = percpu->scribble;
 899         int target = sh->ops.target;
 900         struct r5dev *tgt = &sh->dev[target];
 901         struct page *xor_dest = tgt->page;
 902         int count = 0;
 903         struct dma_async_tx_descriptor *tx;
 904         struct async_submit_ctl submit;
 905         int i;
 906
 907         pr_debug("%s: stripe %llu block: %d\n",
 908                 __func__, (unsigned long long)sh->sector, target);
 909         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 910
 911         for (i = disks; i--; )
 912                 if (i != target)
 913                         xor_srcs[count++] = sh->dev[i].page;
 914
 915         atomic_inc(&sh->count);
 916
 917         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
 918                           ops_complete_compute, sh, to_addr_conv(sh, percpu));
 919         if (unlikely(count == 1))
 920                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
 921         else
 922                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
 923
 924         return tx;
 925 }
 926
 927 /* set_syndrome_sources - populate source buffers for gen_syndrome
 928  * @srcs - (struct page *) array of size sh->disks
 929  * @sh - stripe_head to parse
 930  *
 931  * Populates srcs in proper layout order for the stripe and returns the
 932  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
 933  * destination buffer is recorded in srcs[count] and the Q destination
 934  * is recorded in srcs[count+1]].
 935  */
 936 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
 937 {
 938         int disks = sh->disks;
 939         int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
 940         int d0_idx = raid6_d0(sh);
 941         int count;
 942         int i;
 943
 944         for (i = 0; i < disks; i++)
 945                 srcs[i] = NULL;
 946
 947         count = 0;
 948         i = d0_idx;
 949         do {
 950                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
 951
 952                 srcs[slot] = sh->dev[i].page;
 953                 i = raid6_next_disk(i, disks);
 954         } while (i != d0_idx);
 955
 956         return syndrome_disks;
 957 }
 958
 959 static struct dma_async_tx_descriptor *
 960 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
 961 {
 962         int disks = sh->disks;
 963         struct page **blocks = percpu->scribble;
 964         int target;
 965         int qd_idx = sh->qd_idx;
 966         struct dma_async_tx_descriptor *tx;
 967         struct async_submit_ctl submit;
 968         struct r5dev *tgt;
 969         struct page *dest;
 970         int i;
 971         int count;
 972
 973         if (sh->ops.target < 0)
 974                 target = sh->ops.target2;
 975         else if (sh->ops.target2 < 0)
 976                 target = sh->ops.target;
 977         else
 978                 /* we should only have one valid target */
 979                 BUG();
 980         BUG_ON(target < 0);
 981         pr_debug("%s: stripe %llu block: %d\n",
 982                 __func__, (unsigned long long)sh->sector, target);
 983
 984         tgt = &sh->dev[target];
 985         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
 986         dest = tgt->page;
 987
 988         atomic_inc(&sh->count);
 989
 990         if (target == qd_idx) {
 991                 count = set_syndrome_sources(blocks, sh);
 992                 blocks[count] = NULL; /* regenerating p is not necessary */
 993                 BUG_ON(blocks[count+1] != dest); /* q should already be set */
 994                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
 995                                   ops_complete_compute, sh,
 996                                   to_addr_conv(sh, percpu));
 997                 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
 998         } else {
 999                 /* Compute any data- or p-drive using XOR */
1000                 count = 0;
1001                 for (i = disks; i-- ; ) {
1002                         if (i == target || i == qd_idx)
1003                                 continue;
1004                         blocks[count++] = sh->dev[i].page;
1005                 }
1006
1007                 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1008                                   NULL, ops_complete_compute, sh,
1009                                   to_addr_conv(sh, percpu));
1010                 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1011         }
1012
1013         return tx;
1014 }
1015
1016 static struct dma_async_tx_descriptor *
1017 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1018 {
1019         int i, count, disks = sh->disks;
1020         int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1021         int d0_idx = raid6_d0(sh);
1022         int faila = -1, failb = -1;
1023         int target = sh->ops.target;
1024         int target2 = sh->ops.target2;
1025         struct r5dev *tgt = &sh->dev[target];
1026         struct r5dev *tgt2 = &sh->dev[target2];
1027         struct dma_async_tx_descriptor *tx;
1028         struct page **blocks = percpu->scribble;
1029         struct async_submit_ctl submit;
1030
1031         pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1032                  __func__, (unsigned long long)sh->sector, target, target2);
1033         BUG_ON(target < 0 || target2 < 0);
1034         BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1035         BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1036
1037         /* we need to open-code set_syndrome_sources to handle the
1038          * slot number conversion for 'faila' and 'failb'
1039          */
1040         for (i = 0; i < disks ; i++)
1041                 blocks[i] = NULL;
1042         count = 0;
1043         i = d0_idx;
1044         do {
1045                 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1046
1047                 blocks[slot] = sh->dev[i].page;
1048
1049                 if (i == target)
1050                         faila = slot;
1051                 if (i == target2)
1052                         failb = slot;
1053                 i = raid6_next_disk(i, disks);
1054         } while (i != d0_idx);
1055
1056         BUG_ON(faila == failb);
1057         if (failb < faila)
1058                 swap(faila, failb);
1059         pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1060                  __func__, (unsigned long long)sh->sector, faila, failb);
1061
1062         atomic_inc(&sh->count);
1063
1064         if (failb == syndrome_disks+1) {
1065                 /* Q disk is one of the missing disks */
1066                 if (faila == syndrome_disks) {
1067                         /* Missing P+Q, just recompute */
1068                         init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1069                                           ops_complete_compute, sh,
1070                                           to_addr_conv(sh, percpu));
1071                         return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1072                                                   STRIPE_SIZE, &submit);
1073                 } else {
1074                         struct page *dest;
1075                         int data_target;
1076                         int qd_idx = sh->qd_idx;
1077
1078                         /* Missing D+Q: recompute D from P, then recompute Q */
1079                         if (target == qd_idx)
1080                                 data_target = target2;
1081                         else
1082                                 data_target = target;
1083
1084                         count = 0;
1085                         for (i = disks; i-- ; ) {
1086                                 if (i == data_target || i == qd_idx)
1087                                         continue;
1088                                 blocks[count++] = sh->dev[i].page;
1089                         }
1090                         dest = sh->dev[data_target].page;
1091                         init_async_submit(&submit,
1092                                           ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1093                                           NULL, NULL, NULL,
1094                                           to_addr_conv(sh, percpu));
1095                         tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1096                                        &submit);
1097
1098                         count = set_syndrome_sources(blocks, sh);
1099                         init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1100                                           ops_complete_compute, sh,
1101                                           to_addr_conv(sh, percpu));
1102                         return async_gen_syndrome(blocks, 0, count+2,
1103                                                   STRIPE_SIZE, &submit);
1104                 }
1105         } else {
1106                 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1107                                   ops_complete_compute, sh,
1108                                   to_addr_conv(sh, percpu));
1109                 if (failb == syndrome_disks) {
1110                         /* We're missing D+P. */
1111                         return async_raid6_datap_recov(syndrome_disks+2,
1112                                                        STRIPE_SIZE, faila,
1113                                                        blocks, &submit);
1114                 } else {
1115                         /* We're missing D+D. */
1116                         return async_raid6_2data_recov(syndrome_disks+2,
1117                                                        STRIPE_SIZE, faila, failb,
1118                                                        blocks, &submit);
1119                 }
1120         }
1121 }
1122
1123
1124 static void ops_complete_prexor(void *stripe_head_ref)
1125 {
1126         struct stripe_head *sh = stripe_head_ref;
1127
1128         pr_debug("%s: stripe %llu\n", __func__,
1129                 (unsigned long long)sh->sector);
1130 }
1131
1132 static struct dma_async_tx_descriptor *
1133 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1134                struct dma_async_tx_descriptor *tx)
1135 {
1136         int disks = sh->disks;
1137         struct page **xor_srcs = percpu->scribble;
1138         int count = 0, pd_idx = sh->pd_idx, i;
1139         struct async_submit_ctl submit;
1140
1141         /* existing parity data subtracted */
1142         struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1143
1144         pr_debug("%s: stripe %llu\n", __func__,
1145                 (unsigned long long)sh->sector);
1146
1147         for (i = disks; i--; ) {
1148                 struct r5dev *dev = &sh->dev[i];
1149                 /* Only process blocks that are known to be uptodate */
1150                 if (test_bit(R5_Wantdrain, &dev->flags))
1151                         xor_srcs[count++] = dev->page;
1152         }
1153
1154         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1155                           ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1156         tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1157
1158         return tx;
1159 }
1160
1161 static struct dma_async_tx_descriptor *
1162 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1163 {
1164         int disks = sh->disks;
1165         int i;
1166
1167         pr_debug("%s: stripe %llu\n", __func__,
1168                 (unsigned long long)sh->sector);
1169
1170         for (i = disks; i--; ) {
1171                 struct r5dev *dev = &sh->dev[i];
1172                 struct bio *chosen;
1173
1174                 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1175                         struct bio *wbi;
1176
1177                         spin_lock_irq(&sh->stripe_lock);
1178                         chosen = dev->towrite;
1179                         dev->towrite = NULL;
1180                         BUG_ON(dev->written);
1181                         wbi = dev->written = chosen;
1182                         spin_unlock_irq(&sh->stripe_lock);
1183
1184                         while (wbi && wbi->bi_sector <
1185                                 dev->sector + STRIPE_SECTORS) {
1186                                 if (wbi->bi_rw & REQ_FUA)
1187                                         set_bit(R5_WantFUA, &dev->flags);
1188                                 if (wbi->bi_rw & REQ_SYNC)
1189                                         set_bit(R5_SyncIO, &dev->flags);
1190                                 if (wbi->bi_rw & REQ_DISCARD)
1191                                         set_bit(R5_Discard, &dev->flags);
1192                                 else
1193                                         tx = async_copy_data(1, wbi, dev->page,
1194                                                 dev->sector, tx);
1195                                 wbi = r5_next_bio(wbi, dev->sector);
1196                         }
1197                 }
1198         }
1199
1200         return tx;
1201 }
1202
1203 static void ops_complete_reconstruct(void *stripe_head_ref)
1204 {
1205         struct stripe_head *sh = stripe_head_ref;
1206         int disks = sh->disks;
1207         int pd_idx = sh->pd_idx;
1208         int qd_idx = sh->qd_idx;
1209         int i;
1210         bool fua = false, sync = false, discard = false;
1211
1212         pr_debug("%s: stripe %llu\n", __func__,
1213                 (unsigned long long)sh->sector);
1214
1215         for (i = disks; i--; ) {
1216                 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1217                 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1218                 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1219         }
1220
1221         for (i = disks; i--; ) {
1222                 struct r5dev *dev = &sh->dev[i];
1223
1224                 if (dev->written || i == pd_idx || i == qd_idx) {
1225                         if (!discard)
1226                                 set_bit(R5_UPTODATE, &dev->flags);
1227                         if (fua)
1228                                 set_bit(R5_WantFUA, &dev->flags);
1229                         if (sync)
1230                                 set_bit(R5_SyncIO, &dev->flags);
1231                 }
1232         }
1233
1234         if (sh->reconstruct_state == reconstruct_state_drain_run)
1235                 sh->reconstruct_state = reconstruct_state_drain_result;
1236         else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1237                 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1238         else {
1239                 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1240                 sh->reconstruct_state = reconstruct_state_result;
1241         }
1242
1243         set_bit(STRIPE_HANDLE, &sh->state);
1244         release_stripe(sh);
1245 }
1246
1247 static void
1248 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1249                      struct dma_async_tx_descriptor *tx)
1250 {
1251         int disks = sh->disks;
1252         struct page **xor_srcs = percpu->scribble;
1253         struct async_submit_ctl submit;
1254         int count = 0, pd_idx = sh->pd_idx, i;
1255         struct page *xor_dest;
1256         int prexor = 0;
1257         unsigned long flags;
1258
1259         pr_debug("%s: stripe %llu\n", __func__,
1260                 (unsigned long long)sh->sector);
1261
1262         for (i = 0; i < sh->disks; i++) {
1263                 if (pd_idx == i)
1264                         continue;
1265                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1266                         break;
1267         }
1268         if (i >= sh->disks) {
1269                 atomic_inc(&sh->count);
1270                 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1271                 ops_complete_reconstruct(sh);
1272                 return;
1273         }
1274         /* check if prexor is active which means only process blocks
1275          * that are part of a read-modify-write (written)
1276          */
1277         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1278                 prexor = 1;
1279                 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1280                 for (i = disks; i--; ) {
1281                         struct r5dev *dev = &sh->dev[i];
1282                         if (dev->written)
1283                                 xor_srcs[count++] = dev->page;
1284                 }
1285         } else {
1286                 xor_dest = sh->dev[pd_idx].page;
1287                 for (i = disks; i--; ) {
1288                         struct r5dev *dev = &sh->dev[i];
1289                         if (i != pd_idx)
1290                                 xor_srcs[count++] = dev->page;
1291                 }
1292         }
1293
1294         /* 1/ if we prexor'd then the dest is reused as a source
1295          * 2/ if we did not prexor then we are redoing the parity
1296          * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1297          * for the synchronous xor case
1298          */
1299         flags = ASYNC_TX_ACK |
1300                 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1301
1302         atomic_inc(&sh->count);
1303
1304         init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1305                           to_addr_conv(sh, percpu));
1306         if (unlikely(count == 1))
1307                 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1308         else
1309                 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1310 }
1311
1312 static void
1313 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1314                      struct dma_async_tx_descriptor *tx)
1315 {
1316         struct async_submit_ctl submit;
1317         struct page **blocks = percpu->scribble;
1318         int count, i;
1319
1320         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1321
1322         for (i = 0; i < sh->disks; i++) {
1323                 if (sh->pd_idx == i || sh->qd_idx == i)
1324                         continue;
1325                 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1326                         break;
1327         }
1328         if (i >= sh->disks) {
1329                 atomic_inc(&sh->count);
1330                 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1331                 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1332                 ops_complete_reconstruct(sh);
1333                 return;
1334         }
1335
1336         count = set_syndrome_sources(blocks, sh);
1337
1338         atomic_inc(&sh->count);
1339
1340         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1341                           sh, to_addr_conv(sh, percpu));
1342         async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1343 }
1344
1345 static void ops_complete_check(void *stripe_head_ref)
1346 {
1347         struct stripe_head *sh = stripe_head_ref;
1348
1349         pr_debug("%s: stripe %llu\n", __func__,
1350                 (unsigned long long)sh->sector);
1351
1352         sh->check_state = check_state_check_result;
1353         set_bit(STRIPE_HANDLE, &sh->state);
1354         release_stripe(sh);
1355 }
1356
1357 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1358 {
1359         int disks = sh->disks;
1360         int pd_idx = sh->pd_idx;
1361         int qd_idx = sh->qd_idx;
1362         struct page *xor_dest;
1363         struct page **xor_srcs = percpu->scribble;
1364         struct dma_async_tx_descriptor *tx;
1365         struct async_submit_ctl submit;
1366         int count;
1367         int i;
1368
1369         pr_debug("%s: stripe %llu\n", __func__,
1370                 (unsigned long long)sh->sector);
1371
1372         count = 0;
1373         xor_dest = sh->dev[pd_idx].page;
1374         xor_srcs[count++] = xor_dest;
1375         for (i = disks; i--; ) {
1376                 if (i == pd_idx || i == qd_idx)
1377                         continue;
1378                 xor_srcs[count++] = sh->dev[i].page;
1379         }
1380
1381         init_async_submit(&submit, 0, NULL, NULL, NULL,
1382                           to_addr_conv(sh, percpu));
1383         tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1384                            &sh->ops.zero_sum_result, &submit);
1385
1386         atomic_inc(&sh->count);
1387         init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1388         tx = async_trigger_callback(&submit);
1389 }
1390
1391 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1392 {
1393         struct page **srcs = percpu->scribble;
1394         struct async_submit_ctl submit;
1395         int count;
1396
1397         pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1398                 (unsigned long long)sh->sector, checkp);
1399
1400         count = set_syndrome_sources(srcs, sh);
1401         if (!checkp)
1402                 srcs[count] = NULL;
1403
1404         atomic_inc(&sh->count);
1405         init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1406                           sh, to_addr_conv(sh, percpu));
1407         async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1408                            &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1409 }
1410
1411 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1412 {
1413         int overlap_clear = 0, i, disks = sh->disks;
1414         struct dma_async_tx_descriptor *tx = NULL;
1415         struct r5conf *conf = sh->raid_conf;
1416         int level = conf->level;
1417         struct raid5_percpu *percpu;
1418         unsigned long cpu;
1419
1420         cpu = get_cpu();
1421         percpu = per_cpu_ptr(conf->percpu, cpu);
1422         if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1423                 ops_run_biofill(sh);
1424                 overlap_clear++;
1425         }
1426
1427         if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1428                 if (level < 6)
1429                         tx = ops_run_compute5(sh, percpu);
1430                 else {
1431                         if (sh->ops.target2 < 0 || sh->ops.target < 0)
1432                                 tx = ops_run_compute6_1(sh, percpu);
1433                         else
1434                                 tx = ops_run_compute6_2(sh, percpu);
1435                 }
1436                 /* terminate the chain if reconstruct is not set to be run */
1437                 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1438                         async_tx_ack(tx);
1439         }
1440
1441         if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1442                 tx = ops_run_prexor(sh, percpu, tx);
1443
1444         if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1445                 tx = ops_run_biodrain(sh, tx);
1446                 overlap_clear++;
1447         }
1448
1449         if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1450                 if (level < 6)
1451                         ops_run_reconstruct5(sh, percpu, tx);
1452                 else
1453                         ops_run_reconstruct6(sh, percpu, tx);
1454         }
1455
1456         if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1457                 if (sh->check_state == check_state_run)
1458                         ops_run_check_p(sh, percpu);
1459                 else if (sh->check_state == check_state_run_q)
1460                         ops_run_check_pq(sh, percpu, 0);
1461                 else if (sh->check_state == check_state_run_pq)
1462                         ops_run_check_pq(sh, percpu, 1);
1463                 else
1464                         BUG();
1465         }
1466
1467         if (overlap_clear)
1468                 for (i = disks; i--; ) {
1469                         struct r5dev *dev = &sh->dev[i];
1470                         if (test_and_clear_bit(R5_Overlap, &dev->flags))
1471                                 wake_up(&sh->raid_conf->wait_for_overlap);
1472                 }
1473         put_cpu();
1474 }
1475
1476 static int grow_one_stripe(struct r5conf *conf)
1477 {
1478         struct stripe_head *sh;
1479         sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1480         if (!sh)
1481                 return 0;
1482
1483         sh->raid_conf = conf;
1484
1485         spin_lock_init(&sh->stripe_lock);
1486
1487         if (grow_buffers(sh)) {
1488                 shrink_buffers(sh);
1489                 kmem_cache_free(conf->slab_cache, sh);
1490                 return 0;
1491         }
1492         /* we just created an active stripe so... */
1493         atomic_set(&sh->count, 1);
1494         atomic_inc(&conf->active_stripes);
1495         INIT_LIST_HEAD(&sh->lru);
1496         release_stripe(sh);
1497         return 1;
1498 }
1499
1500 static int grow_stripes(struct r5conf *conf, int num)
1501 {
1502         struct kmem_cache *sc;
1503         int devs = max(conf->raid_disks, conf->previous_raid_disks);
1504
1505         if (conf->mddev->gendisk)
1506                 sprintf(conf->cache_name[0],
1507                         "raid%d-%s", conf->level, mdname(conf->mddev));
1508         else
1509                 sprintf(conf->cache_name[0],
1510                         "raid%d-%p", conf->level, conf->mddev);
1511         sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1512
1513         conf->active_name = 0;
1514         sc = kmem_cache_create(conf->cache_name[conf->active_name],
1515                                sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1516                                0, 0, NULL);
1517         if (!sc)
1518                 return 1;
1519         conf->slab_cache = sc;
1520         conf->pool_size = devs;
1521         while (num--)
1522                 if (!grow_one_stripe(conf))
1523                         return 1;
1524         return 0;
1525 }
1526
1527 /**
1528  * scribble_len - return the required size of the scribble region
1529  * @num - total number of disks in the array
1530  *
1531  * The size must be enough to contain:
1532  * 1/ a struct page pointer for each device in the array +2
1533  * 2/ room to convert each entry in (1) to its corresponding dma
1534  *    (dma_map_page()) or page (page_address()) address.
1535  *
1536  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1537  * calculate over all devices (not just the data blocks), using zeros in place
1538  * of the P and Q blocks.
1539  */
1540 static size_t scribble_len(int num)
1541 {
1542         size_t len;
1543
1544         len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1545
1546         return len;
1547 }
1548
1549 static int resize_stripes(struct r5conf *conf, int newsize)
1550 {
1551         /* Make all the stripes able to hold 'newsize' devices.
1552          * New slots in each stripe get 'page' set to a new page.
1553          *
1554          * This happens in stages:
1555          * 1/ create a new kmem_cache and allocate the required number of
1556          *    stripe_heads.
1557          * 2/ gather all the old stripe_heads and transfer the pages across
1558          *    to the new stripe_heads.  This will have the side effect of
1559          *    freezing the array as once all stripe_heads have been collected,
1560          *    no IO will be possible.  Old stripe heads are freed once their
1561          *    pages have been transferred over, and the old kmem_cache is
1562          *    freed when all stripes are done.
1563          * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1564          *    we simple return a failre status - no need to clean anything up.
1565          * 4/ allocate new pages for the new slots in the new stripe_heads.
1566          *    If this fails, we don't bother trying the shrink the
1567          *    stripe_heads down again, we just leave them as they are.
1568          *    As each stripe_head is processed the new one is released into
1569          *    active service.
1570          *
1571          * Once step2 is started, we cannot afford to wait for a write,
1572          * so we use GFP_NOIO allocations.
1573          */
1574         struct stripe_head *osh, *nsh;
1575         LIST_HEAD(newstripes);
1576         struct disk_info *ndisks;
1577         unsigned long cpu;
1578         int err;
1579         struct kmem_cache *sc;
1580         int i;
1581
1582         if (newsize <= conf->pool_size)
1583                 return 0; /* never bother to shrink */
1584
1585         err = md_allow_write(conf->mddev);
1586         if (err)
1587                 return err;
1588
1589         /* Step 1 */
1590         sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1591                                sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1592                                0, 0, NULL);
1593         if (!sc)
1594                 return -ENOMEM;
1595
1596         for (i = conf->max_nr_stripes; i; i--) {
1597                 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1598                 if (!nsh)
1599                         break;
1600
1601                 nsh->raid_conf = conf;
1602                 spin_lock_init(&nsh->stripe_lock);
1603
1604                 list_add(&nsh->lru, &newstripes);
1605         }
1606         if (i) {
1607                 /* didn't get enough, give up */
1608                 while (!list_empty(&newstripes)) {
1609                         nsh = list_entry(newstripes.next, struct stripe_head, lru);
1610                         list_del(&nsh->lru);
1611                         kmem_cache_free(sc, nsh);
1612                 }
1613                 kmem_cache_destroy(sc);
1614                 return -ENOMEM;
1615         }
1616         /* Step 2 - Must use GFP_NOIO now.
1617          * OK, we have enough stripes, start collecting inactive
1618          * stripes and copying them over
1619          */
1620         list_for_each_entry(nsh, &newstripes, lru) {
1621                 spin_lock_irq(&conf->device_lock);
1622                 wait_event_lock_irq(conf->wait_for_stripe,
1623                                     !list_empty(&conf->inactive_list),
1624                                     conf->device_lock);
1625                 osh = get_free_stripe(conf);
1626                 spin_unlock_irq(&conf->device_lock);
1627                 atomic_set(&nsh->count, 1);
1628                 for(i=0; i<conf->pool_size; i++)
1629                         nsh->dev[i].page = osh->dev[i].page;
1630                 for( ; i<newsize; i++)
1631                         nsh->dev[i].page = NULL;
1632                 kmem_cache_free(conf->slab_cache, osh);
1633         }
1634         kmem_cache_destroy(conf->slab_cache);
1635
1636         /* Step 3.
1637          * At this point, we are holding all the stripes so the array
1638          * is completely stalled, so now is a good time to resize
1639          * conf->disks and the scribble region
1640          */
1641         ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1642         if (ndisks) {
1643                 for (i=0; i<conf->raid_disks; i++)
1644                         ndisks[i] = conf->disks[i];
1645                 kfree(conf->disks);
1646                 conf->disks = ndisks;
1647         } else
1648                 err = -ENOMEM;
1649
1650         get_online_cpus();
1651         conf->scribble_len = scribble_len(newsize);
1652         for_each_present_cpu(cpu) {
1653                 struct raid5_percpu *percpu;
1654                 void *scribble;
1655
1656                 percpu = per_cpu_ptr(conf->percpu, cpu);
1657                 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1658
1659                 if (scribble) {
1660                         kfree(percpu->scribble);
1661                         percpu->scribble = scribble;
1662                 } else {
1663                         err = -ENOMEM;
1664                         break;
1665                 }
1666         }
1667         put_online_cpus();
1668
1669         /* Step 4, return new stripes to service */
1670         while(!list_empty(&newstripes)) {
1671                 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1672                 list_del_init(&nsh->lru);
1673
1674                 for (i=conf->raid_disks; i < newsize; i++)
1675                         if (nsh->dev[i].page == NULL) {
1676                                 struct page *p = alloc_page(GFP_NOIO);
1677                                 nsh->dev[i].page = p;
1678                                 if (!p)
1679                                         err = -ENOMEM;
1680                         }
1681                 release_stripe(nsh);
1682         }
1683         /* critical section pass, GFP_NOIO no longer needed */
1684
1685         conf->slab_cache = sc;
1686         conf->active_name = 1-conf->active_name;
1687         conf->pool_size = newsize;
1688         return err;
1689 }
1690
1691 static int drop_one_stripe(struct r5conf *conf)
1692 {
1693         struct stripe_head *sh;
1694
1695         spin_lock_irq(&conf->device_lock);
1696         sh = get_free_stripe(conf);
1697         spin_unlock_irq(&conf->device_lock);
1698         if (!sh)
1699                 return 0;
1700         BUG_ON(atomic_read(&sh->count));
1701         shrink_buffers(sh);
1702         kmem_cache_free(conf->slab_cache, sh);
1703         atomic_dec(&conf->active_stripes);
1704         return 1;
1705 }
1706
1707 static void shrink_stripes(struct r5conf *conf)
1708 {
1709         while (drop_one_stripe(conf))
1710                 ;
1711
1712         if (conf->slab_cache)
1713                 kmem_cache_destroy(conf->slab_cache);
1714         conf->slab_cache = NULL;
1715 }
1716
1717 static void raid5_end_read_request(struct bio * bi, int error)
1718 {
1719         struct stripe_head *sh = bi->bi_private;
1720         struct r5conf *conf = sh->raid_conf;
1721         int disks = sh->disks, i;
1722         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1723         char b[BDEVNAME_SIZE];
1724         struct md_rdev *rdev = NULL;
1725         sector_t s;
1726
1727         for (i=0 ; i<disks; i++)
1728                 if (bi == &sh->dev[i].req)
1729                         break;
1730
1731         pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1732                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1733                 uptodate);
1734         if (i == disks) {
1735                 BUG();
1736                 return;
1737         }
1738         if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1739                 /* If replacement finished while this request was outstanding,
1740                  * 'replacement' might be NULL already.
1741                  * In that case it moved down to 'rdev'.
1742                  * rdev is not removed until all requests are finished.
1743                  */
1744                 rdev = conf->disks[i].replacement;
1745         if (!rdev)
1746                 rdev = conf->disks[i].rdev;
1747
1748         if (use_new_offset(conf, sh))
1749                 s = sh->sector + rdev->new_data_offset;
1750         else
1751                 s = sh->sector + rdev->data_offset;
1752         if (uptodate) {
1753                 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1754                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1755                         /* Note that this cannot happen on a
1756                          * replacement device.  We just fail those on
1757                          * any error
1758                          */
1759                         printk_ratelimited(
1760                                 KERN_INFO
1761                                 "md/raid:%s: read error corrected"
1762                                 " (%lu sectors at %llu on %s)\n",
1763                                 mdname(conf->mddev), STRIPE_SECTORS,
1764                                 (unsigned long long)s,
1765                                 bdevname(rdev->bdev, b));
1766                         atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1767                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1768                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1769                 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1770                         clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1771
1772                 if (atomic_read(&rdev->read_errors))
1773                         atomic_set(&rdev->read_errors, 0);
1774         } else {
1775                 const char *bdn = bdevname(rdev->bdev, b);
1776                 int retry = 0;
1777                 int set_bad = 0;
1778
1779                 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1780                 atomic_inc(&rdev->read_errors);
1781                 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1782                         printk_ratelimited(
1783                                 KERN_WARNING
1784                                 "md/raid:%s: read error on replacement device "
1785                                 "(sector %llu on %s).\n",
1786                                 mdname(conf->mddev),
1787                                 (unsigned long long)s,
1788                                 bdn);
1789                 else if (conf->mddev->degraded >= conf->max_degraded) {
1790                         set_bad = 1;
1791                         printk_ratelimited(
1792                                 KERN_WARNING
1793                                 "md/raid:%s: read error not correctable "
1794                                 "(sector %llu on %s).\n",
1795                                 mdname(conf->mddev),
1796                                 (unsigned long long)s,
1797                                 bdn);
1798                 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1799                         /* Oh, no!!! */
1800                         set_bad = 1;
1801                         printk_ratelimited(
1802                                 KERN_WARNING
1803                                 "md/raid:%s: read error NOT corrected!! "
1804                                 "(sector %llu on %s).\n",
1805                                 mdname(conf->mddev),
1806                                 (unsigned long long)s,
1807                                 bdn);
1808                 } else if (atomic_read(&rdev->read_errors)
1809                          > conf->max_nr_stripes)
1810                         printk(KERN_WARNING
1811                                "md/raid:%s: Too many read errors, failing device %s.\n",
1812                                mdname(conf->mddev), bdn);
1813                 else
1814                         retry = 1;
1815                 if (retry)
1816                         if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1817                                 set_bit(R5_ReadError, &sh->dev[i].flags);
1818                                 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1819                         } else
1820                                 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1821                 else {
1822                         clear_bit(R5_ReadError, &sh->dev[i].flags);
1823                         clear_bit(R5_ReWrite, &sh->dev[i].flags);
1824                         if (!(set_bad
1825                               && test_bit(In_sync, &rdev->flags)
1826                               && rdev_set_badblocks(
1827                                       rdev, sh->sector, STRIPE_SECTORS, 0)))
1828                                 md_error(conf->mddev, rdev);
1829                 }
1830         }
1831         rdev_dec_pending(rdev, conf->mddev);
1832         clear_bit(R5_LOCKED, &sh->dev[i].flags);
1833         set_bit(STRIPE_HANDLE, &sh->state);
1834         release_stripe(sh);
1835 }
1836
1837 static void raid5_end_write_request(struct bio *bi, int error)
1838 {
1839         struct stripe_head *sh = bi->bi_private;
1840         struct r5conf *conf = sh->raid_conf;
1841         int disks = sh->disks, i;
1842         struct md_rdev *uninitialized_var(rdev);
1843         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1844         sector_t first_bad;
1845         int bad_sectors;
1846         int replacement = 0;
1847
1848         for (i = 0 ; i < disks; i++) {
1849                 if (bi == &sh->dev[i].req) {
1850                         rdev = conf->disks[i].rdev;
1851                         break;
1852                 }
1853                 if (bi == &sh->dev[i].rreq) {
1854                         rdev = conf->disks[i].replacement;
1855                         if (rdev)
1856                                 replacement = 1;
1857                         else
1858                                 /* rdev was removed and 'replacement'
1859                                  * replaced it.  rdev is not removed
1860                                  * until all requests are finished.
1861                                  */
1862                                 rdev = conf->disks[i].rdev;
1863                         break;
1864                 }
1865         }
1866         pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1867                 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1868                 uptodate);
1869         if (i == disks) {
1870                 BUG();
1871                 return;
1872         }
1873
1874         if (replacement) {
1875                 if (!uptodate)
1876                         md_error(conf->mddev, rdev);
1877                 else if (is_badblock(rdev, sh->sector,
1878                                      STRIPE_SECTORS,
1879                                      &first_bad, &bad_sectors))
1880                         set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1881         } else {
1882                 if (!uptodate) {
1883                         set_bit(WriteErrorSeen, &rdev->flags);
1884                         set_bit(R5_WriteError, &sh->dev[i].flags);
1885                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1886                                 set_bit(MD_RECOVERY_NEEDED,
1887                                         &rdev->mddev->recovery);
1888                 } else if (is_badblock(rdev, sh->sector,
1889                                        STRIPE_SECTORS,
1890                                        &first_bad, &bad_sectors))
1891                         set_bit(R5_MadeGood, &sh->dev[i].flags);
1892         }
1893         rdev_dec_pending(rdev, conf->mddev);
1894
1895         if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1896                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1897         set_bit(STRIPE_HANDLE, &sh->state);
1898         release_stripe(sh);
1899 }
1900
1901 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1902
1903 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1904 {
1905         struct r5dev *dev = &sh->dev[i];
1906
1907         bio_init(&dev->req);
1908         dev->req.bi_io_vec = &dev->vec;
1909         dev->req.bi_vcnt++;
1910         dev->req.bi_max_vecs++;
1911         dev->req.bi_private = sh;
1912         dev->vec.bv_page = dev->page;
1913
1914         bio_init(&dev->rreq);
1915         dev->rreq.bi_io_vec = &dev->rvec;
1916         dev->rreq.bi_vcnt++;
1917         dev->rreq.bi_max_vecs++;
1918         dev->rreq.bi_private = sh;
1919         dev->rvec.bv_page = dev->page;
1920
1921         dev->flags = 0;
1922         dev->sector = compute_blocknr(sh, i, previous);
1923 }
1924
1925 static void error(struct mddev *mddev, struct md_rdev *rdev)
1926 {
1927         char b[BDEVNAME_SIZE];
1928         struct r5conf *conf = mddev->private;
1929         unsigned long flags;
1930         pr_debug("raid456: error called\n");
1931
1932         spin_lock_irqsave(&conf->device_lock, flags);
1933         clear_bit(In_sync, &rdev->flags);
1934         mddev->degraded = calc_degraded(conf);
1935         spin_unlock_irqrestore(&conf->device_lock, flags);
1936         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1937
1938         set_bit(Blocked, &rdev->flags);
1939         set_bit(Faulty, &rdev->flags);
1940         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1941         printk(KERN_ALERT
1942                "md/raid:%s: Disk failure on %s, disabling device.\n"
1943                "md/raid:%s: Operation continuing on %d devices.\n",
1944                mdname(mddev),
1945                bdevname(rdev->bdev, b),
1946                mdname(mddev),
1947                conf->raid_disks - mddev->degraded);
1948 }
1949
1950 /*
1951  * Input: a 'big' sector number,
1952  * Output: index of the data and parity disk, and the sector # in them.
1953  */
1954 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1955                                      int previous, int *dd_idx,
1956                                      struct stripe_head *sh)
1957 {
1958         sector_t stripe, stripe2;
1959         sector_t chunk_number;
1960         unsigned int chunk_offset;
1961         int pd_idx, qd_idx;
1962         int ddf_layout = 0;
1963         sector_t new_sector;
1964         int algorithm = previous ? conf->prev_algo
1965                                  : conf->algorithm;
1966         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1967                                          : conf->chunk_sectors;
1968         int raid_disks = previous ? conf->previous_raid_disks
1969                                   : conf->raid_disks;
1970         int data_disks = raid_disks - conf->max_degraded;
1971
1972         /* First compute the information on this sector */
1973
1974         /*
1975          * Compute the chunk number and the sector offset inside the chunk
1976          */
1977         chunk_offset = sector_div(r_sector, sectors_per_chunk);
1978         chunk_number = r_sector;
1979
1980         /*
1981          * Compute the stripe number
1982          */
1983         stripe = chunk_number;
1984         *dd_idx = sector_div(stripe, data_disks);
1985         stripe2 = stripe;
1986         /*
1987          * Select the parity disk based on the user selected algorithm.
1988          */
1989         pd_idx = qd_idx = -1;
1990         switch(conf->level) {
1991         case 4:
1992                 pd_idx = data_disks;
1993                 break;
1994         case 5:
1995                 switch (algorithm) {
1996                 case ALGORITHM_LEFT_ASYMMETRIC:
1997                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
1998                         if (*dd_idx >= pd_idx)
1999                                 (*dd_idx)++;
2000                         break;
2001                 case ALGORITHM_RIGHT_ASYMMETRIC:
2002                         pd_idx = sector_div(stripe2, raid_disks);
2003                         if (*dd_idx >= pd_idx)
2004                                 (*dd_idx)++;
2005                         break;
2006                 case ALGORITHM_LEFT_SYMMETRIC:
2007                         pd_idx = data_disks - sector_div(stripe2, raid_disks);
2008                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2009                         break;
2010                 case ALGORITHM_RIGHT_SYMMETRIC:
2011                         pd_idx = sector_div(stripe2, raid_disks);
2012                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2013                         break;
2014                 case ALGORITHM_PARITY_0:
2015                         pd_idx = 0;
2016                         (*dd_idx)++;
2017                         break;
2018                 case ALGORITHM_PARITY_N:
2019                         pd_idx = data_disks;
2020                         break;
2021                 default:
2022                         BUG();
2023                 }
2024                 break;
2025         case 6:
2026
2027                 switch (algorithm) {
2028                 case ALGORITHM_LEFT_ASYMMETRIC:
2029                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2030                         qd_idx = pd_idx + 1;
2031                         if (pd_idx == raid_disks-1) {
2032                                 (*dd_idx)++;    /* Q D D D P */
2033                                 qd_idx = 0;
2034                         } else if (*dd_idx >= pd_idx)
2035                                 (*dd_idx) += 2; /* D D P Q D */
2036                         break;
2037                 case ALGORITHM_RIGHT_ASYMMETRIC:
2038                         pd_idx = sector_div(stripe2, raid_disks);
2039                         qd_idx = pd_idx + 1;
2040                         if (pd_idx == raid_disks-1) {
2041                                 (*dd_idx)++;    /* Q D D D P */
2042                                 qd_idx = 0;
2043                         } else if (*dd_idx >= pd_idx)
2044                                 (*dd_idx) += 2; /* D D P Q D */
2045                         break;
2046                 case ALGORITHM_LEFT_SYMMETRIC:
2047                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2048                         qd_idx = (pd_idx + 1) % raid_disks;
2049                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2050                         break;
2051                 case ALGORITHM_RIGHT_SYMMETRIC:
2052                         pd_idx = sector_div(stripe2, raid_disks);
2053                         qd_idx = (pd_idx + 1) % raid_disks;
2054                         *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2055                         break;
2056
2057                 case ALGORITHM_PARITY_0:
2058                         pd_idx = 0;
2059                         qd_idx = 1;
2060                         (*dd_idx) += 2;
2061                         break;
2062                 case ALGORITHM_PARITY_N:
2063                         pd_idx = data_disks;
2064                         qd_idx = data_disks + 1;
2065                         break;
2066
2067                 case ALGORITHM_ROTATING_ZERO_RESTART:
2068                         /* Exactly the same as RIGHT_ASYMMETRIC, but or
2069                          * of blocks for computing Q is different.
2070                          */
2071                         pd_idx = sector_div(stripe2, raid_disks);
2072                         qd_idx = pd_idx + 1;
2073                         if (pd_idx == raid_disks-1) {
2074                                 (*dd_idx)++;    /* Q D D D P */
2075                                 qd_idx = 0;
2076                         } else if (*dd_idx >= pd_idx)
2077                                 (*dd_idx) += 2; /* D D P Q D */
2078                         ddf_layout = 1;
2079                         break;
2080
2081                 case ALGORITHM_ROTATING_N_RESTART:
2082                         /* Same a left_asymmetric, by first stripe is
2083                          * D D D P Q  rather than
2084                          * Q D D D P
2085                          */
2086                         stripe2 += 1;
2087                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2088                         qd_idx = pd_idx + 1;
2089                         if (pd_idx == raid_disks-1) {
2090                                 (*dd_idx)++;    /* Q D D D P */
2091                                 qd_idx = 0;
2092                         } else if (*dd_idx >= pd_idx)
2093                                 (*dd_idx) += 2; /* D D P Q D */
2094                         ddf_layout = 1;
2095                         break;
2096
2097                 case ALGORITHM_ROTATING_N_CONTINUE:
2098                         /* Same as left_symmetric but Q is before P */
2099                         pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2100                         qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2101                         *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2102                         ddf_layout = 1;
2103                         break;
2104
2105                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2106                         /* RAID5 left_asymmetric, with Q on last device */
2107                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2108                         if (*dd_idx >= pd_idx)
2109                                 (*dd_idx)++;
2110                         qd_idx = raid_disks - 1;
2111                         break;
2112
2113                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2114                         pd_idx = sector_div(stripe2, raid_disks-1);
2115                         if (*dd_idx >= pd_idx)
2116                                 (*dd_idx)++;
2117                         qd_idx = raid_disks - 1;
2118                         break;
2119
2120                 case ALGORITHM_LEFT_SYMMETRIC_6:
2121                         pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2122                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2123                         qd_idx = raid_disks - 1;
2124                         break;
2125
2126                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2127                         pd_idx = sector_div(stripe2, raid_disks-1);
2128                         *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2129                         qd_idx = raid_disks - 1;
2130                         break;
2131
2132                 case ALGORITHM_PARITY_0_6:
2133                         pd_idx = 0;
2134                         (*dd_idx)++;
2135                         qd_idx = raid_disks - 1;
2136                         break;
2137
2138                 default:
2139                         BUG();
2140                 }
2141                 break;
2142         }
2143
2144         if (sh) {
2145                 sh->pd_idx = pd_idx;
2146                 sh->qd_idx = qd_idx;
2147                 sh->ddf_layout = ddf_layout;
2148         }
2149         /*
2150          * Finally, compute the new sector number
2151          */
2152         new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2153         return new_sector;
2154 }
2155
2156
2157 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2158 {
2159         struct r5conf *conf = sh->raid_conf;
2160         int raid_disks = sh->disks;
2161         int data_disks = raid_disks - conf->max_degraded;
2162         sector_t new_sector = sh->sector, check;
2163         int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2164                                          : conf->chunk_sectors;
2165         int algorithm = previous ? conf->prev_algo
2166                                  : conf->algorithm;
2167         sector_t stripe;
2168         int chunk_offset;
2169         sector_t chunk_number;
2170         int dummy1, dd_idx = i;
2171         sector_t r_sector;
2172         struct stripe_head sh2;
2173
2174
2175         chunk_offset = sector_div(new_sector, sectors_per_chunk);
2176         stripe = new_sector;
2177
2178         if (i == sh->pd_idx)
2179                 return 0;
2180         switch(conf->level) {
2181         case 4: break;
2182         case 5:
2183                 switch (algorithm) {
2184                 case ALGORITHM_LEFT_ASYMMETRIC:
2185                 case ALGORITHM_RIGHT_ASYMMETRIC:
2186                         if (i > sh->pd_idx)
2187                                 i--;
2188                         break;
2189                 case ALGORITHM_LEFT_SYMMETRIC:
2190                 case ALGORITHM_RIGHT_SYMMETRIC:
2191                         if (i < sh->pd_idx)
2192                                 i += raid_disks;
2193                         i -= (sh->pd_idx + 1);
2194                         break;
2195                 case ALGORITHM_PARITY_0:
2196                         i -= 1;
2197                         break;
2198                 case ALGORITHM_PARITY_N:
2199                         break;
2200                 default:
2201                         BUG();
2202                 }
2203                 break;
2204         case 6:
2205                 if (i == sh->qd_idx)
2206                         return 0; /* It is the Q disk */
2207                 switch (algorithm) {
2208                 case ALGORITHM_LEFT_ASYMMETRIC:
2209                 case ALGORITHM_RIGHT_ASYMMETRIC:
2210                 case ALGORITHM_ROTATING_ZERO_RESTART:
2211                 case ALGORITHM_ROTATING_N_RESTART:
2212                         if (sh->pd_idx == raid_disks-1)
2213                                 i--;    /* Q D D D P */
2214                         else if (i > sh->pd_idx)
2215                                 i -= 2; /* D D P Q D */
2216                         break;
2217                 case ALGORITHM_LEFT_SYMMETRIC:
2218                 case ALGORITHM_RIGHT_SYMMETRIC:
2219                         if (sh->pd_idx == raid_disks-1)
2220                                 i--; /* Q D D D P */
2221                         else {
2222                                 /* D D P Q D */
2223                                 if (i < sh->pd_idx)
2224                                         i += raid_disks;
2225                                 i -= (sh->pd_idx + 2);
2226                         }
2227                         break;
2228                 case ALGORITHM_PARITY_0:
2229                         i -= 2;
2230                         break;
2231                 case ALGORITHM_PARITY_N:
2232                         break;
2233                 case ALGORITHM_ROTATING_N_CONTINUE:
2234                         /* Like left_symmetric, but P is before Q */
2235                         if (sh->pd_idx == 0)
2236                                 i--;    /* P D D D Q */
2237                         else {
2238                                 /* D D Q P D */
2239                                 if (i < sh->pd_idx)
2240                                         i += raid_disks;
2241                                 i -= (sh->pd_idx + 1);
2242                         }
2243                         break;
2244                 case ALGORITHM_LEFT_ASYMMETRIC_6:
2245                 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2246                         if (i > sh->pd_idx)
2247                                 i--;
2248                         break;
2249                 case ALGORITHM_LEFT_SYMMETRIC_6:
2250                 case ALGORITHM_RIGHT_SYMMETRIC_6:
2251                         if (i < sh->pd_idx)
2252                                 i += data_disks + 1;
2253                         i -= (sh->pd_idx + 1);
2254                         break;
2255                 case ALGORITHM_PARITY_0_6:
2256                         i -= 1;
2257                         break;
2258                 default:
2259                         BUG();
2260                 }
2261                 break;
2262         }
2263
2264         chunk_number = stripe * data_disks + i;
2265         r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2266
2267         check = raid5_compute_sector(conf, r_sector,
2268                                      previous, &dummy1, &sh2);
2269         if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2270                 || sh2.qd_idx != sh->qd_idx) {
2271                 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2272                        mdname(conf->mddev));
2273                 return 0;
2274         }
2275         return r_sector;
2276 }
2277
2278
2279 static void
2280 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2281                          int rcw, int expand)
2282 {
2283         int i, pd_idx = sh->pd_idx, disks = sh->disks;
2284         struct r5conf *conf = sh->raid_conf;
2285         int level = conf->level;
2286
2287         if (rcw) {
2288
2289                 for (i = disks; i--; ) {
2290                         struct r5dev *dev = &sh->dev[i];
2291
2292                         if (dev->towrite) {
2293                                 set_bit(R5_LOCKED, &dev->flags);
2294                                 set_bit(R5_Wantdrain, &dev->flags);
2295                                 if (!expand)
2296                                         clear_bit(R5_UPTODATE, &dev->flags);
2297                                 s->locked++;
2298                         }
2299                 }
2300                 /* if we are not expanding this is a proper write request, and
2301                  * there will be bios with new data to be drained into the
2302                  * stripe cache
2303                  */
2304                 if (!expand) {
2305                         if (!s->locked)
2306                                 /* False alarm, nothing to do */
2307                                 return;
2308                         sh->reconstruct_state = reconstruct_state_drain_run;
2309                         set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2310                 } else
2311                         sh->reconstruct_state = reconstruct_state_run;
2312
2313                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2314
2315                 if (s->locked + conf->max_degraded == disks)
2316                         if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2317                                 atomic_inc(&conf->pending_full_writes);
2318         } else {
2319                 BUG_ON(level == 6);
2320                 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2321                         test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2322
2323                 for (i = disks; i--; ) {
2324                         struct r5dev *dev = &sh->dev[i];
2325                         if (i == pd_idx)
2326                                 continue;
2327
2328                         if (dev->towrite &&
2329                             (test_bit(R5_UPTODATE, &dev->flags) ||
2330                              test_bit(R5_Wantcompute, &dev->flags))) {
2331                                 set_bit(R5_Wantdrain, &dev->flags);
2332                                 set_bit(R5_LOCKED, &dev->flags);
2333                                 clear_bit(R5_UPTODATE, &dev->flags);
2334                                 s->locked++;
2335                         }
2336                 }
2337                 if (!s->locked)
2338                         /* False alarm - nothing to do */
2339                         return;
2340                 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2341                 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2342                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2343                 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2344         }
2345
2346         /* keep the parity disk(s) locked while asynchronous operations
2347          * are in flight
2348          */
2349         set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2350         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2351         s->locked++;
2352
2353         if (level == 6) {
2354                 int qd_idx = sh->qd_idx;
2355                 struct r5dev *dev = &sh->dev[qd_idx];
2356
2357                 set_bit(R5_LOCKED, &dev->flags);
2358                 clear_bit(R5_UPTODATE, &dev->flags);
2359                 s->locked++;
2360         }
2361
2362         pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2363                 __func__, (unsigned long long)sh->sector,
2364                 s->locked, s->ops_request);
2365 }
2366
2367 /*
2368  * Each stripe/dev can have one or more bion attached.
2369  * toread/towrite point to the first in a chain.
2370  * The bi_next chain must be in order.
2371  */
2372 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2373 {
2374         struct bio **bip;
2375         struct r5conf *conf = sh->raid_conf;
2376         int firstwrite=0;
2377
2378         pr_debug("adding bi b#%llu to stripe s#%llu\n",
2379                 (unsigned long long)bi->bi_sector,
2380                 (unsigned long long)sh->sector);
2381
2382         /*
2383          * If several bio share a stripe. The bio bi_phys_segments acts as a
2384          * reference count to avoid race. The reference count should already be
2385          * increased before this function is called (for example, in
2386          * make_request()), so other bio sharing this stripe will not free the
2387          * stripe. If a stripe is owned by one stripe, the stripe lock will
2388          * protect it.
2389          */
2390         spin_lock_irq(&sh->stripe_lock);
2391         if (forwrite) {
2392                 bip = &sh->dev[dd_idx].towrite;
2393                 if (*bip == NULL)
2394                         firstwrite = 1;
2395         } else
2396                 bip = &sh->dev[dd_idx].toread;
2397         while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2398                 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2399                         goto overlap;
2400                 bip = & (*bip)->bi_next;
2401         }
2402         if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2403                 goto overlap;
2404
2405         BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2406         if (*bip)
2407                 bi->bi_next = *bip;
2408         *bip = bi;
2409         raid5_inc_bi_active_stripes(bi);
2410
2411         if (forwrite) {
2412                 /* check if page is covered */
2413                 sector_t sector = sh->dev[dd_idx].sector;
2414                 for (bi=sh->dev[dd_idx].towrite;
2415                      sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2416                              bi && bi->bi_sector <= sector;
2417                      bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2418                         if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2419                                 sector = bi->bi_sector + (bi->bi_size>>9);
2420                 }
2421                 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2422                         set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2423         }
2424
2425         pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2426                 (unsigned long long)(*bip)->bi_sector,
2427                 (unsigned long long)sh->sector, dd_idx);
2428         spin_unlock_irq(&sh->stripe_lock);
2429
2430         if (conf->mddev->bitmap && firstwrite) {
2431                 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2432                                   STRIPE_SECTORS, 0);
2433                 sh->bm_seq = conf->seq_flush+1;
2434                 set_bit(STRIPE_BIT_DELAY, &sh->state);
2435         }
2436         return 1;
2437
2438  overlap:
2439         set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2440         spin_unlock_irq(&sh->stripe_lock);
2441         return 0;
2442 }
2443
2444 static void end_reshape(struct r5conf *conf);
2445
2446 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2447                             struct stripe_head *sh)
2448 {
2449         int sectors_per_chunk =
2450                 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2451         int dd_idx;
2452         int chunk_offset = sector_div(stripe, sectors_per_chunk);
2453         int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2454
2455         raid5_compute_sector(conf,
2456                              stripe * (disks - conf->max_degraded)
2457                              *sectors_per_chunk + chunk_offset,
2458                              previous,
2459                              &dd_idx, sh);
2460 }
2461
2462 static void
2463 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2464                                 struct stripe_head_state *s, int disks,
2465                                 struct bio **return_bi)
2466 {
2467         int i;
2468         for (i = disks; i--; ) {
2469                 struct bio *bi;
2470                 int bitmap_end = 0;
2471
2472                 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2473                         struct md_rdev *rdev;
2474                         rcu_read_lock();
2475                         rdev = rcu_dereference(conf->disks[i].rdev);
2476                         if (rdev && test_bit(In_sync, &rdev->flags))
2477                                 atomic_inc(&rdev->nr_pending);
2478                         else
2479                                 rdev = NULL;
2480                         rcu_read_unlock();
2481                         if (rdev) {
2482                                 if (!rdev_set_badblocks(
2483                                             rdev,
2484                                             sh->sector,
2485                                             STRIPE_SECTORS, 0))
2486                                         md_error(conf->mddev, rdev);
2487                                 rdev_dec_pending(rdev, conf->mddev);
2488                         }
2489                 }
2490                 spin_lock_irq(&sh->stripe_lock);
2491                 /* fail all writes first */
2492                 bi = sh->dev[i].towrite;
2493                 sh->dev[i].towrite = NULL;
2494                 spin_unlock_irq(&sh->stripe_lock);
2495                 if (bi)
2496                         bitmap_end = 1;
2497
2498                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2499                         wake_up(&conf->wait_for_overlap);
2500
2501                 while (bi && bi->bi_sector <
2502                         sh->dev[i].sector + STRIPE_SECTORS) {
2503                         struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2504                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2505                         if (!raid5_dec_bi_active_stripes(bi)) {
2506                                 md_write_end(conf->mddev);
2507                                 bi->bi_next = *return_bi;
2508                                 *return_bi = bi;
2509                         }
2510                         bi = nextbi;
2511                 }
2512                 if (bitmap_end)
2513                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2514                                 STRIPE_SECTORS, 0, 0);
2515                 bitmap_end = 0;
2516                 /* and fail all 'written' */
2517                 bi = sh->dev[i].written;
2518                 sh->dev[i].written = NULL;
2519                 if (bi) bitmap_end = 1;
2520                 while (bi && bi->bi_sector <
2521                        sh->dev[i].sector + STRIPE_SECTORS) {
2522                         struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2523                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
2524                         if (!raid5_dec_bi_active_stripes(bi)) {
2525                                 md_write_end(conf->mddev);
2526                                 bi->bi_next = *return_bi;
2527                                 *return_bi = bi;
2528                         }
2529                         bi = bi2;
2530                 }
2531
2532                 /* fail any reads if this device is non-operational and
2533                  * the data has not reached the cache yet.
2534                  */
2535                 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2536                     (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2537                       test_bit(R5_ReadError, &sh->dev[i].flags))) {
2538                         spin_lock_irq(&sh->stripe_lock);
2539                         bi = sh->dev[i].toread;
2540                         sh->dev[i].toread = NULL;
2541                         spin_unlock_irq(&sh->stripe_lock);
2542                         if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2543                                 wake_up(&conf->wait_for_overlap);
2544                         while (bi && bi->bi_sector <
2545                                sh->dev[i].sector + STRIPE_SECTORS) {
2546                                 struct bio *nextbi =
2547                                         r5_next_bio(bi, sh->dev[i].sector);
2548                                 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2549                                 if (!raid5_dec_bi_active_stripes(bi)) {
2550                                         bi->bi_next = *return_bi;
2551                                         *return_bi = bi;
2552                                 }
2553                                 bi = nextbi;
2554                         }
2555                 }
2556                 if (bitmap_end)
2557                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2558                                         STRIPE_SECTORS, 0, 0);
2559                 /* If we were in the middle of a write the parity block might
2560                  * still be locked - so just clear all R5_LOCKED flags
2561                  */
2562                 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2563         }
2564
2565         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2566                 if (atomic_dec_and_test(&conf->pending_full_writes))
2567                         md_wakeup_thread(conf->mddev->thread);
2568 }
2569
2570 static void
2571 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2572                    struct stripe_head_state *s)
2573 {
2574         int abort = 0;
2575         int i;
2576
2577         clear_bit(STRIPE_SYNCING, &sh->state);
2578         if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2579                 wake_up(&conf->wait_for_overlap);
2580         s->syncing = 0;
2581         s->replacing = 0;
2582         /* There is nothing more to do for sync/check/repair.
2583          * Don't even need to abort as that is handled elsewhere
2584          * if needed, and not always wanted e.g. if there is a known
2585          * bad block here.
2586          * For recover/replace we need to record a bad block on all
2587          * non-sync devices, or abort the recovery
2588          */
2589         if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2590                 /* During recovery devices cannot be removed, so
2591                  * locking and refcounting of rdevs is not needed
2592                  */
2593                 for (i = 0; i < conf->raid_disks; i++) {
2594                         struct md_rdev *rdev = conf->disks[i].rdev;
2595                         if (rdev
2596                             && !test_bit(Faulty, &rdev->flags)
2597                             && !test_bit(In_sync, &rdev->flags)
2598                             && !rdev_set_badblocks(rdev, sh->sector,
2599                                                    STRIPE_SECTORS, 0))
2600                                 abort = 1;
2601                         rdev = conf->disks[i].replacement;
2602                         if (rdev
2603                             && !test_bit(Faulty, &rdev->flags)
2604                             && !test_bit(In_sync, &rdev->flags)
2605                             && !rdev_set_badblocks(rdev, sh->sector,
2606                                                    STRIPE_SECTORS, 0))
2607                                 abort = 1;
2608                 }
2609                 if (abort)
2610                         conf->recovery_disabled =
2611                                 conf->mddev->recovery_disabled;
2612         }
2613         md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2614 }
2615
2616 static int want_replace(struct stripe_head *sh, int disk_idx)
2617 {
2618         struct md_rdev *rdev;
2619         int rv = 0;
2620         /* Doing recovery so rcu locking not required */
2621         rdev = sh->raid_conf->disks[disk_idx].replacement;
2622         if (rdev
2623             && !test_bit(Faulty, &rdev->flags)
2624             && !test_bit(In_sync, &rdev->flags)
2625             && (rdev->recovery_offset <= sh->sector
2626                 || rdev->mddev->recovery_cp <= sh->sector))
2627                 rv = 1;
2628
2629         return rv;
2630 }
2631
2632 /* fetch_block - checks the given member device to see if its data needs
2633  * to be read or computed to satisfy a request.
2634  *
2635  * Returns 1 when no more member devices need to be checked, otherwise returns
2636  * 0 to tell the loop in handle_stripe_fill to continue
2637  */
2638 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2639                        int disk_idx, int disks)
2640 {
2641         struct r5dev *dev = &sh->dev[disk_idx];
2642         struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2643                                   &sh->dev[s->failed_num[1]] };
2644
2645         /* is the data in this block needed, and can we get it? */
2646         if (!test_bit(R5_LOCKED, &dev->flags) &&
2647             !test_bit(R5_UPTODATE, &dev->flags) &&
2648             (dev->toread ||
2649              (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2650              s->syncing || s->expanding ||
2651              (s->replacing && want_replace(sh, disk_idx)) ||
2652              (s->failed >= 1 && fdev[0]->toread) ||
2653              (s->failed >= 2 && fdev[1]->toread) ||
2654              (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2655               !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2656              (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2657                 /* we would like to get this block, possibly by computing it,
2658                  * otherwise read it if the backing disk is insync
2659                  */
2660                 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2661                 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2662                 if ((s->uptodate == disks - 1) &&
2663                     (s->failed && (disk_idx == s->failed_num[0] ||
2664                                    disk_idx == s->failed_num[1]))) {
2665                         /* have disk failed, and we're requested to fetch it;
2666                          * do compute it
2667                          */
2668                         pr_debug("Computing stripe %llu block %d\n",
2669                                (unsigned long long)sh->sector, disk_idx);
2670                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2671                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2672                         set_bit(R5_Wantcompute, &dev->flags);
2673                         sh->ops.target = disk_idx;
2674                         sh->ops.target2 = -1; /* no 2nd target */
2675                         s->req_compute = 1;
2676                         /* Careful: from this point on 'uptodate' is in the eye
2677                          * of raid_run_ops which services 'compute' operations
2678                          * before writes. R5_Wantcompute flags a block that will
2679                          * be R5_UPTODATE by the time it is needed for a
2680                          * subsequent operation.
2681                          */
2682                         s->uptodate++;
2683                         return 1;
2684                 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2685                         /* Computing 2-failure is *very* expensive; only
2686                          * do it if failed >= 2
2687                          */
2688                         int other;
2689                         for (other = disks; other--; ) {
2690                                 if (other == disk_idx)
2691                                         continue;
2692                                 if (!test_bit(R5_UPTODATE,
2693                                       &sh->dev[other].flags))
2694                                         break;
2695                         }
2696                         BUG_ON(other < 0);
2697                         pr_debug("Computing stripe %llu blocks %d,%d\n",
2698                                (unsigned long long)sh->sector,
2699                                disk_idx, other);
2700                         set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2701                         set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2702                         set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2703                         set_bit(R5_Wantcompute, &sh->dev[other].flags);
2704                         sh->ops.target = disk_idx;
2705                         sh->ops.target2 = other;
2706                         s->uptodate += 2;
2707                         s->req_compute = 1;
2708                         return 1;
2709                 } else if (test_bit(R5_Insync, &dev->flags)) {
2710                         set_bit(R5_LOCKED, &dev->flags);
2711                         set_bit(R5_Wantread, &dev->flags);
2712                         s->locked++;
2713                         pr_debug("Reading block %d (sync=%d)\n",
2714                                 disk_idx, s->syncing);
2715                 }
2716         }
2717
2718         return 0;
2719 }
2720
2721 /**
2722  * handle_stripe_fill - read or compute data to satisfy pending requests.
2723  */
2724 static void handle_stripe_fill(struct stripe_head *sh,
2725                                struct stripe_head_state *s,
2726                                int disks)
2727 {
2728         int i;
2729
2730         /* look for blocks to read/compute, skip this if a compute
2731          * is already in flight, or if the stripe contents are in the
2732          * midst of changing due to a write
2733          */
2734         if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2735             !sh->reconstruct_state)
2736                 for (i = disks; i--; )
2737                         if (fetch_block(sh, s, i, disks))
2738                                 break;
2739         set_bit(STRIPE_HANDLE, &sh->state);
2740 }
2741
2742
2743 /* handle_stripe_clean_event
2744  * any written block on an uptodate or failed drive can be returned.
2745  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2746  * never LOCKED, so we don't need to test 'failed' directly.
2747  */
2748 static void handle_stripe_clean_event(struct r5conf *conf,
2749         struct stripe_head *sh, int disks, struct bio **return_bi)
2750 {
2751         int i;
2752         struct r5dev *dev;
2753         int discard_pending = 0;
2754
2755         for (i = disks; i--; )
2756                 if (sh->dev[i].written) {
2757                         dev = &sh->dev[i];
2758                         if (!test_bit(R5_LOCKED, &dev->flags) &&
2759                             (test_bit(R5_UPTODATE, &dev->flags) ||
2760                              test_bit(R5_Discard, &dev->flags))) {
2761                                 /* We can return any write requests */
2762                                 struct bio *wbi, *wbi2;
2763                                 pr_debug("Return write for disc %d\n", i);
2764                                 if (test_and_clear_bit(R5_Discard, &dev->flags))
2765                                         clear_bit(R5_UPTODATE, &dev->flags);
2766                                 wbi = dev->written;
2767                                 dev->written = NULL;
2768                                 while (wbi && wbi->bi_sector <
2769                                         dev->sector + STRIPE_SECTORS) {
2770                                         wbi2 = r5_next_bio(wbi, dev->sector);
2771                                         if (!raid5_dec_bi_active_stripes(wbi)) {
2772                                                 md_write_end(conf->mddev);
2773                                                 wbi->bi_next = *return_bi;
2774                                                 *return_bi = wbi;
2775                                         }
2776                                         wbi = wbi2;
2777                                 }
2778                                 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2779                                                 STRIPE_SECTORS,
2780                                          !test_bit(STRIPE_DEGRADED, &sh->state),
2781                                                 0);
2782                         } else if (test_bit(R5_Discard, &dev->flags))
2783                                 discard_pending = 1;
2784                 }
2785         if (!discard_pending &&
2786             test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2787                 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2788                 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2789                 if (sh->qd_idx >= 0) {
2790                         clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2791                         clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2792                 }
2793                 /* now that discard is done we can proceed with any sync */
2794                 clear_bit(STRIPE_DISCARD, &sh->state);
2795                 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2796                         set_bit(STRIPE_HANDLE, &sh->state);
2797
2798         }
2799
2800         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2801                 if (atomic_dec_and_test(&conf->pending_full_writes))
2802                         md_wakeup_thread(conf->mddev->thread);
2803 }
2804
2805 static void handle_stripe_dirtying(struct r5conf *conf,
2806                                    struct stripe_head *sh,
2807                                    struct stripe_head_state *s,
2808                                    int disks)
2809 {
2810         int rmw = 0, rcw = 0, i;
2811         sector_t recovery_cp = conf->mddev->recovery_cp;
2812
2813         /* RAID6 requires 'rcw' in current implementation.
2814          * Otherwise, check whether resync is now happening or should start.
2815          * If yes, then the array is dirty (after unclean shutdown or
2816          * initial creation), so parity in some stripes might be inconsistent.
2817          * In this case, we need to always do reconstruct-write, to ensure
2818          * that in case of drive failure or read-error correction, we
2819          * generate correct data from the parity.
2820          */
2821         if (conf->max_degraded == 2 ||
2822             (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2823                 /* Calculate the real rcw later - for now make it
2824                  * look like rcw is cheaper
2825                  */
2826                 rcw = 1; rmw = 2;
2827                 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2828                          conf->max_degraded, (unsigned long long)recovery_cp,
2829                          (unsigned long long)sh->sector);
2830         } else for (i = disks; i--; ) {
2831                 /* would I have to read this buffer for read_modify_write */
2832                 struct r5dev *dev = &sh->dev[i];
2833                 if ((dev->towrite || i == sh->pd_idx) &&
2834                     !test_bit(R5_LOCKED, &dev->flags) &&
2835                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2836                       test_bit(R5_Wantcompute, &dev->flags))) {
2837                         if (test_bit(R5_Insync, &dev->flags))
2838                                 rmw++;
2839                         else
2840                                 rmw += 2*disks;  /* cannot read it */
2841                 }
2842                 /* Would I have to read this buffer for reconstruct_write */
2843                 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2844                     !test_bit(R5_LOCKED, &dev->flags) &&
2845                     !(test_bit(R5_UPTODATE, &dev->flags) ||
2846                     test_bit(R5_Wantcompute, &dev->flags))) {
2847                         if (test_bit(R5_Insync, &dev->flags)) rcw++;
2848                         else
2849                                 rcw += 2*disks;
2850                 }
2851         }
2852         pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2853                 (unsigned long long)sh->sector, rmw, rcw);
2854         set_bit(STRIPE_HANDLE, &sh->state);
2855         if (rmw < rcw && rmw > 0) {
2856                 /* prefer read-modify-write, but need to get some data */
2857                 if (conf->mddev->queue)
2858                         blk_add_trace_msg(conf->mddev->queue,
2859                                           "raid5 rmw %llu %d",
2860                                           (unsigned long long)sh->sector, rmw);
2861                 for (i = disks; i--; ) {
2862                         struct r5dev *dev = &sh->dev[i];
2863                         if ((dev->towrite || i == sh->pd_idx) &&
2864                             !test_bit(R5_LOCKED, &dev->flags) &&
2865                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2866                             test_bit(R5_Wantcompute, &dev->flags)) &&
2867                             test_bit(R5_Insync, &dev->flags)) {
2868                                 if (
2869                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2870                                         pr_debug("Read_old block "
2871                                                  "%d for r-m-w\n", i);
2872                                         set_bit(R5_LOCKED, &dev->flags);
2873                                         set_bit(R5_Wantread, &dev->flags);
2874                                         s->locked++;
2875                                 } else {
2876                                         set_bit(STRIPE_DELAYED, &sh->state);
2877                                         set_bit(STRIPE_HANDLE, &sh->state);
2878                                 }
2879                         }
2880                 }
2881         }
2882         if (rcw <= rmw && rcw > 0) {
2883                 /* want reconstruct write, but need to get some data */
2884                 int qread =0;
2885                 rcw = 0;
2886                 for (i = disks; i--; ) {
2887                         struct r5dev *dev = &sh->dev[i];
2888                         if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2889                             i != sh->pd_idx && i != sh->qd_idx &&
2890                             !test_bit(R5_LOCKED, &dev->flags) &&
2891                             !(test_bit(R5_UPTODATE, &dev->flags) ||
2892                               test_bit(R5_Wantcompute, &dev->flags))) {
2893                                 rcw++;
2894                                 if (!test_bit(R5_Insync, &dev->flags))
2895                                         continue; /* it's a failed drive */
2896                                 if (
2897                                   test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2898                                         pr_debug("Read_old block "
2899                                                 "%d for Reconstruct\n", i);
2900                                         set_bit(R5_LOCKED, &dev->flags);
2901                                         set_bit(R5_Wantread, &dev->flags);
2902                                         s->locked++;
2903                                         qread++;
2904                                 } else {
2905                                         set_bit(STRIPE_DELAYED, &sh->state);
2906                                         set_bit(STRIPE_HANDLE, &sh->state);
2907                                 }
2908                         }
2909                 }
2910                 if (rcw && conf->mddev->queue)
2911                         blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2912                                           (unsigned long long)sh->sector,
2913                                           rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2914         }
2915         /* now if nothing is locked, and if we have enough data,
2916          * we can start a write request
2917          */
2918         /* since handle_stripe can be called at any time we need to handle the
2919          * case where a compute block operation has been submitted and then a
2920          * subsequent call wants to start a write request.  raid_run_ops only
2921          * handles the case where compute block and reconstruct are requested
2922          * simultaneously.  If this is not the case then new writes need to be
2923          * held off until the compute completes.
2924          */
2925         if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2926             (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2927             !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2928                 schedule_reconstruction(sh, s, rcw == 0, 0);
2929 }
2930
2931 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2932                                 struct stripe_head_state *s, int disks)
2933 {
2934         struct r5dev *dev = NULL;
2935
2936         set_bit(STRIPE_HANDLE, &sh->state);
2937
2938         switch (sh->check_state) {
2939         case check_state_idle:
2940                 /* start a new check operation if there are no failures */
2941                 if (s->failed == 0) {
2942                         BUG_ON(s->uptodate != disks);
2943                         sh->check_state = check_state_run;
2944                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
2945                         clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2946                         s->uptodate--;
2947                         break;
2948                 }
2949                 dev = &sh->dev[s->failed_num[0]];
2950                 /* fall through */
2951         case check_state_compute_result:
2952                 sh->check_state = check_state_idle;
2953                 if (!dev)
2954                         dev = &sh->dev[sh->pd_idx];
2955
2956                 /* check that a write has not made the stripe insync */
2957                 if (test_bit(STRIPE_INSYNC, &sh->state))
2958                         break;
2959
2960                 /* either failed parity check, or recovery is happening */
2961                 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2962                 BUG_ON(s->uptodate != disks);
2963
2964                 set_bit(R5_LOCKED, &dev->flags);
2965                 s->locked++;
2966                 set_bit(R5_Wantwrite, &dev->flags);
2967
2968                 clear_bit(STRIPE_DEGRADED, &sh->state);
2969                 set_bit(STRIPE_INSYNC, &sh->state);
2970                 break;
2971         case check_state_run:
2972                 break; /* we will be called again upon completion */
2973         case check_state_check_result:
2974                 sh->check_state = check_state_idle;
2975
2976                 /* if a failure occurred during the check operation, leave
2977                  * STRIPE_INSYNC not set and let the stripe be handled again
2978                  */
2979                 if (s->failed)
2980                         break;
2981
2982                 /* handle a successful check operation, if parity is correct
2983                  * we are done.  Otherwise update the mismatch count and repair
2984                  * parity if !MD_RECOVERY_CHECK
2985                  */
2986                 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2987                         /* parity is correct (on disc,
2988                          * not in buffer any more)
2989                          */
2990                         set_bit(STRIPE_INSYNC, &sh->state);
2991                 else {
2992                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
2993                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2994                                 /* don't try to repair!! */
2995                                 set_bit(STRIPE_INSYNC, &sh->state);
2996                         else {
2997                                 sh->check_state = check_state_compute_run;
2998                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2999                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3000                                 set_bit(R5_Wantcompute,
3001                                         &sh->dev[sh->pd_idx].flags);
3002                                 sh->ops.target = sh->pd_idx;
3003                                 sh->ops.target2 = -1;
3004                                 s->uptodate++;
3005                         }
3006                 }
3007                 break;
3008         case check_state_compute_run:
3009                 break;
3010         default:
3011                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3012                        __func__, sh->check_state,
3013                        (unsigned long long) sh->sector);
3014                 BUG();
3015         }
3016 }
3017
3018
3019 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3020                                   struct stripe_head_state *s,
3021                                   int disks)
3022 {
3023         int pd_idx = sh->pd_idx;
3024         int qd_idx = sh->qd_idx;
3025         struct r5dev *dev;
3026
3027         set_bit(STRIPE_HANDLE, &sh->state);
3028
3029         BUG_ON(s->failed > 2);
3030
3031         /* Want to check and possibly repair P and Q.
3032          * However there could be one 'failed' device, in which
3033          * case we can only check one of them, possibly using the
3034          * other to generate missing data
3035          */
3036
3037         switch (sh->check_state) {
3038         case check_state_idle:
3039                 /* start a new check operation if there are < 2 failures */
3040                 if (s->failed == s->q_failed) {
3041                         /* The only possible failed device holds Q, so it
3042                          * makes sense to check P (If anything else were failed,
3043                          * we would have used P to recreate it).
3044                          */
3045                         sh->check_state = check_state_run;
3046                 }
3047                 if (!s->q_failed && s->failed < 2) {
3048                         /* Q is not failed, and we didn't use it to generate
3049                          * anything, so it makes sense to check it
3050                          */
3051                         if (sh->check_state == check_state_run)
3052                                 sh->check_state = check_state_run_pq;
3053                         else
3054                                 sh->check_state = check_state_run_q;
3055                 }
3056
3057                 /* discard potentially stale zero_sum_result */
3058                 sh->ops.zero_sum_result = 0;
3059
3060                 if (sh->check_state == check_state_run) {
3061                         /* async_xor_zero_sum destroys the contents of P */
3062                         clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3063                         s->uptodate--;
3064                 }
3065                 if (sh->check_state >= check_state_run &&
3066                     sh->check_state <= check_state_run_pq) {
3067                         /* async_syndrome_zero_sum preserves P and Q, so
3068                          * no need to mark them !uptodate here
3069                          */
3070                         set_bit(STRIPE_OP_CHECK, &s->ops_request);
3071                         break;
3072                 }
3073
3074                 /* we have 2-disk failure */
3075                 BUG_ON(s->failed != 2);
3076                 /* fall through */
3077         case check_state_compute_result:
3078                 sh->check_state = check_state_idle;
3079
3080                 /* check that a write has not made the stripe insync */
3081                 if (test_bit(STRIPE_INSYNC, &sh->state))
3082                         break;
3083
3084                 /* now write out any block on a failed drive,
3085                  * or P or Q if they were recomputed
3086                  */
3087                 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3088                 if (s->failed == 2) {
3089                         dev = &sh->dev[s->failed_num[1]];
3090                         s->locked++;
3091                         set_bit(R5_LOCKED, &dev->flags);
3092                         set_bit(R5_Wantwrite, &dev->flags);
3093                 }
3094                 if (s->failed >= 1) {
3095                         dev = &sh->dev[s->failed_num[0]];
3096                         s->locked++;
3097                         set_bit(R5_LOCKED, &dev->flags);
3098                         set_bit(R5_Wantwrite, &dev->flags);
3099                 }
3100                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3101                         dev = &sh->dev[pd_idx];
3102                         s->locked++;
3103                         set_bit(R5_LOCKED, &dev->flags);
3104                         set_bit(R5_Wantwrite, &dev->flags);
3105                 }
3106                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3107                         dev = &sh->dev[qd_idx];
3108                         s->locked++;
3109                         set_bit(R5_LOCKED, &dev->flags);
3110                         set_bit(R5_Wantwrite, &dev->flags);
3111                 }
3112                 clear_bit(STRIPE_DEGRADED, &sh->state);
3113
3114                 set_bit(STRIPE_INSYNC, &sh->state);
3115                 break;
3116         case check_state_run:
3117         case check_state_run_q:
3118         case check_state_run_pq:
3119                 break; /* we will be called again upon completion */
3120         case check_state_check_result:
3121                 sh->check_state = check_state_idle;
3122
3123                 /* handle a successful check operation, if parity is correct
3124                  * we are done.  Otherwise update the mismatch count and repair
3125                  * parity if !MD_RECOVERY_CHECK
3126                  */
3127                 if (sh->ops.zero_sum_result == 0) {
3128                         /* both parities are correct */
3129                         if (!s->failed)
3130                                 set_bit(STRIPE_INSYNC, &sh->state);
3131                         else {
3132                                 /* in contrast to the raid5 case we can validate
3133                                  * parity, but still have a failure to write
3134                                  * back
3135                                  */
3136                                 sh->check_state = check_state_compute_result;
3137                                 /* Returning at this point means that we may go
3138                                  * off and bring p and/or q uptodate again so
3139                                  * we make sure to check zero_sum_result again
3140                                  * to verify if p or q need writeback
3141                                  */
3142                         }
3143                 } else {
3144                         atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3145                         if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3146                                 /* don't try to repair!! */
3147                                 set_bit(STRIPE_INSYNC, &sh->state);
3148                         else {
3149                                 int *target = &sh->ops.target;
3150
3151                                 sh->ops.target = -1;
3152                                 sh->ops.target2 = -1;
3153                                 sh->check_state = check_state_compute_run;
3154                                 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3155                                 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3156                                 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3157                                         set_bit(R5_Wantcompute,
3158                                                 &sh->dev[pd_idx].flags);
3159                                         *target = pd_idx;
3160                                         target = &sh->ops.target2;
3161                                         s->uptodate++;
3162                                 }
3163                                 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3164                                         set_bit(R5_Wantcompute,
3165                                                 &sh->dev[qd_idx].flags);
3166                                         *target = qd_idx;
3167                                         s->uptodate++;
3168                                 }
3169                         }
3170                 }
3171                 break;
3172         case check_state_compute_run:
3173                 break;
3174         default:
3175                 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3176                        __func__, sh->check_state,
3177                        (unsigned long long) sh->sector);
3178                 BUG();
3179         }
3180 }
3181
3182 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3183 {
3184         int i;
3185
3186         /* We have read all the blocks in this stripe and now we need to
3187          * copy some of them into a target stripe for expand.
3188          */
3189         struct dma_async_tx_descriptor *tx = NULL;
3190         clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3191         for (i = 0; i < sh->disks; i++)
3192                 if (i != sh->pd_idx && i != sh->qd_idx) {
3193                         int dd_idx, j;
3194                         struct stripe_head *sh2;
3195                         struct async_submit_ctl submit;
3196
3197                         sector_t bn = compute_blocknr(sh, i, 1);
3198                         sector_t s = raid5_compute_sector(conf, bn, 0,
3199                                                           &dd_idx, NULL);
3200                         sh2 = get_active_stripe(conf, s, 0, 1, 1);
3201                         if (sh2 == NULL)
3202                                 /* so far only the early blocks of this stripe
3203                                  * have been requested.  When later blocks
3204                                  * get requested, we will try again
3205                                  */
3206                                 continue;
3207                         if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3208                            test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3209                                 /* must have already done this block */
3210                                 release_stripe(sh2);
3211                                 continue;
3212                         }
3213
3214                         /* place all the copies on one channel */
3215                         init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3216                         tx = async_memcpy(sh2->dev[dd_idx].page,
3217                                           sh->dev[i].page, 0, 0, STRIPE_SIZE,
3218                                           &submit);
3219
3220                         set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3221                         set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3222                         for (j = 0; j < conf->raid_disks; j++)
3223                                 if (j != sh2->pd_idx &&
3224                                     j != sh2->qd_idx &&
3225                                     !test_bit(R5_Expanded, &sh2->dev[j].flags))
3226                                         break;
3227                         if (j == conf->raid_disks) {
3228                                 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3229                                 set_bit(STRIPE_HANDLE, &sh2->state);
3230                         }
3231                         release_stripe(sh2);
3232
3233                 }
3234         /* done submitting copies, wait for them to complete */
3235         async_tx_quiesce(&tx);
3236 }
3237
3238 /*
3239  * handle_stripe - do things to a stripe.
3240  *
3241  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3242  * state of various bits to see what needs to be done.
3243  * Possible results:
3244  *    return some read requests which now have data
3245  *    return some write requests which are safely on storage
3246  *    schedule a read on some buffers
3247  *    schedule a write of some buffers
3248  *    return confirmation of parity correctness
3249  *
3250  */
3251
3252 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3253 {
3254         struct r5conf *conf = sh->raid_conf;
3255         int disks = sh->disks;
3256         struct r5dev *dev;
3257         int i;
3258         int do_recovery = 0;
3259
3260         memset(s, 0, sizeof(*s));
3261
3262         s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3263         s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3264         s->failed_num[0] = -1;
3265         s->failed_num[1] = -1;
3266
3267         /* Now to look around and see what can be done */
3268         rcu_read_lock();
3269         for (i=disks; i--; ) {
3270                 struct md_rdev *rdev;
3271                 sector_t first_bad;
3272                 int bad_sectors;
3273                 int is_bad = 0;
3274
3275                 dev = &sh->dev[i];
3276
3277                 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3278                          i, dev->flags,
3279                          dev->toread, dev->towrite, dev->written);
3280                 /* maybe we can reply to a read
3281                  *
3282                  * new wantfill requests are only permitted while
3283                  * ops_complete_biofill is guaranteed to be inactive
3284                  */
3285                 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3286                     !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3287                         set_bit(R5_Wantfill, &dev->flags);
3288
3289                 /* now count some things */
3290                 if (test_bit(R5_LOCKED, &dev->flags))
3291                         s->locked++;
3292                 if (test_bit(R5_UPTODATE, &dev->flags))
3293                         s->uptodate++;
3294                 if (test_bit(R5_Wantcompute, &dev->flags)) {
3295                         s->compute++;
3296                         BUG_ON(s->compute > 2);
3297                 }
3298
3299                 if (test_bit(R5_Wantfill, &dev->flags))
3300                         s->to_fill++;
3301                 else if (dev->toread)
3302                         s->to_read++;
3303                 if (dev->towrite) {
3304                         s->to_write++;
3305                         if (!test_bit(R5_OVERWRITE, &dev->flags))
3306                                 s->non_overwrite++;
3307                 }
3308                 if (dev->written)
3309                         s->written++;
3310                 /* Prefer to use the replacement for reads, but only
3311                  * if it is recovered enough and has no bad blocks.
3312                  */
3313                 rdev = rcu_dereference(conf->disks[i].replacement);
3314                 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3315                     rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3316                     !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3317                                  &first_bad, &bad_sectors))
3318                         set_bit(R5_ReadRepl, &dev->flags);
3319                 else {
3320                         if (rdev)
3321                                 set_bit(R5_NeedReplace, &dev->flags);
3322                         rdev = rcu_dereference(conf->disks[i].rdev);
3323                         clear_bit(R5_ReadRepl, &dev->flags);
3324                 }
3325                 if (rdev && test_bit(Faulty, &rdev->flags))
3326                         rdev = NULL;
3327                 if (rdev) {
3328                         is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3329                                              &first_bad, &bad_sectors);
3330                         if (s->blocked_rdev == NULL
3331                             && (test_bit(Blocked, &rdev->flags)
3332                                 || is_bad < 0)) {
3333                                 if (is_bad < 0)
3334                                         set_bit(BlockedBadBlocks,
3335                                                 &rdev->flags);
3336                                 s->blocked_rdev = rdev;
3337                                 atomic_inc(&rdev->nr_pending);
3338                         }
3339                 }
3340                 clear_bit(R5_Insync, &dev->flags);
3341                 if (!rdev)
3342                         /* Not in-sync */;
3343                 else if (is_bad) {
3344                         /* also not in-sync */
3345                         if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3346                             test_bit(R5_UPTODATE, &dev->flags)) {
3347                                 /* treat as in-sync, but with a read error
3348                                  * which we can now try to correct
3349                                  */
3350                                 set_bit(R5_Insync, &dev->flags);
3351                                 set_bit(R5_ReadError, &dev->flags);
3352                         }
3353                 } else if (test_bit(In_sync, &rdev->flags))
3354                         set_bit(R5_Insync, &dev->flags);
3355                 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3356                         /* in sync if before recovery_offset */
3357                         set_bit(R5_Insync, &dev->flags);
3358                 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3359                          test_bit(R5_Expanded, &dev->flags))
3360                         /* If we've reshaped into here, we assume it is Insync.
3361                          * We will shortly update recovery_offset to make
3362                          * it official.
3363                          */
3364                         set_bit(R5_Insync, &dev->flags);
3365
3366                 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3367                         /* This flag does not apply to '.replacement'
3368                          * only to .rdev, so make sure to check that*/
3369                         struct md_rdev *rdev2 = rcu_dereference(
3370                                 conf->disks[i].rdev);
3371                         if (rdev2 == rdev)
3372                                 clear_bit(R5_Insync, &dev->flags);
3373                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3374                                 s->handle_bad_blocks = 1;
3375                                 atomic_inc(&rdev2->nr_pending);
3376                         } else
3377                                 clear_bit(R5_WriteError, &dev->flags);
3378                 }
3379                 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3380                         /* This flag does not apply to '.replacement'
3381                          * only to .rdev, so make sure to check that*/
3382                         struct md_rdev *rdev2 = rcu_dereference(
3383                                 conf->disks[i].rdev);
3384                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3385                                 s->handle_bad_blocks = 1;
3386                                 atomic_inc(&rdev2->nr_pending);
3387                         } else
3388                                 clear_bit(R5_MadeGood, &dev->flags);
3389                 }
3390                 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3391                         struct md_rdev *rdev2 = rcu_dereference(
3392                                 conf->disks[i].replacement);
3393                         if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3394                                 s->handle_bad_blocks = 1;
3395                                 atomic_inc(&rdev2->nr_pending);
3396                         } else
3397                                 clear_bit(R5_MadeGoodRepl, &dev->flags);
3398                 }
3399                 if (!test_bit(R5_Insync, &dev->flags)) {
3400                         /* The ReadError flag will just be confusing now */
3401                         clear_bit(R5_ReadError, &dev->flags);
3402                         clear_bit(R5_ReWrite, &dev->flags);
3403                 }
3404                 if (test_bit(R5_ReadError, &dev->flags))
3405                         clear_bit(R5_Insync, &dev->flags);
3406                 if (!test_bit(R5_Insync, &dev->flags)) {
3407                         if (s->failed < 2)
3408                                 s->failed_num[s->failed] = i;
3409                         s->failed++;
3410                         if (rdev && !test_bit(Faulty, &rdev->flags))
3411                                 do_recovery = 1;
3412                 }
3413         }
3414         if (test_bit(STRIPE_SYNCING, &sh->state)) {
3415                 /* If there is a failed device being replaced,
3416                  *     we must be recovering.
3417                  * else if we are after recovery_cp, we must be syncing
3418                  * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3419                  * else we can only be replacing
3420                  * sync and recovery both need to read all devices, and so
3421                  * use the same flag.
3422                  */
3423                 if (do_recovery ||
3424                     sh->sector >= conf->mddev->recovery_cp ||
3425                     test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3426                         s->syncing = 1;
3427                 else
3428                         s->replacing = 1;
3429         }
3430         rcu_read_unlock();
3431 }
3432
3433 static void handle_stripe(struct stripe_head *sh)
3434 {
3435         struct stripe_head_state s;
3436         struct r5conf *conf = sh->raid_conf;
3437         int i;
3438         int prexor;
3439         int disks = sh->disks;
3440         struct r5dev *pdev, *qdev;
3441
3442         clear_bit(STRIPE_HANDLE, &sh->state);
3443         if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3444                 /* already being handled, ensure it gets handled
3445                  * again when current action finishes */
3446                 set_bit(STRIPE_HANDLE, &sh->state);
3447                 return;
3448         }
3449
3450         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3451                 spin_lock(&sh->stripe_lock);
3452                 /* Cannot process 'sync' concurrently with 'discard' */
3453                 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3454                     test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3455                         set_bit(STRIPE_SYNCING, &sh->state);
3456                         clear_bit(STRIPE_INSYNC, &sh->state);
3457                 }
3458                 spin_unlock(&sh->stripe_lock);
3459         }
3460         clear_bit(STRIPE_DELAYED, &sh->state);
3461
3462         pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3463                 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3464                (unsigned long long)sh->sector, sh->state,
3465                atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3466                sh->check_state, sh->reconstruct_state);
3467
3468         analyse_stripe(sh, &s);
3469
3470         if (s.handle_bad_blocks) {
3471                 set_bit(STRIPE_HANDLE, &sh->state);
3472                 goto finish;
3473         }
3474
3475         if (unlikely(s.blocked_rdev)) {
3476                 if (s.syncing || s.expanding || s.expanded ||
3477                     s.replacing || s.to_write || s.written) {
3478                         set_bit(STRIPE_HANDLE, &sh->state);
3479                         goto finish;
3480                 }
3481                 /* There is nothing for the blocked_rdev to block */
3482                 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3483                 s.blocked_rdev = NULL;
3484         }
3485
3486         if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3487                 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3488                 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3489         }
3490
3491         pr_debug("locked=%d uptodate=%d to_read=%d"
3492                " to_write=%d failed=%d failed_num=%d,%d\n",
3493                s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3494                s.failed_num[0], s.failed_num[1]);
3495         /* check if the array has lost more than max_degraded devices and,
3496          * if so, some requests might need to be failed.
3497          */
3498         if (s.failed > conf->max_degraded) {
3499                 sh->check_state = 0;
3500                 sh->reconstruct_state = 0;
3501                 if (s.to_read+s.to_write+s.written)
3502                         handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3503                 if (s.syncing + s.replacing)
3504                         handle_failed_sync(conf, sh, &s);
3505         }
3506
3507         /* Now we check to see if any write operations have recently
3508          * completed
3509          */
3510         prexor = 0;
3511         if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3512                 prexor = 1;
3513         if (sh->reconstruct_state == reconstruct_state_drain_result ||
3514             sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3515                 sh->reconstruct_state = reconstruct_state_idle;
3516
3517                 /* All the 'written' buffers and the parity block are ready to
3518                  * be written back to disk
3519                  */
3520                 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3521                        !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3522                 BUG_ON(sh->qd_idx >= 0 &&
3523                        !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3524                        !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3525                 for (i = disks; i--; ) {
3526                         struct r5dev *dev = &sh->dev[i];
3527                         if (test_bit(R5_LOCKED, &dev->flags) &&
3528                                 (i == sh->pd_idx || i == sh->qd_idx ||
3529                                  dev->written)) {
3530                                 pr_debug("Writing block %d\n", i);
3531                                 set_bit(R5_Wantwrite, &dev->flags);
3532                                 if (prexor)
3533                                         continue;
3534                                 if (!test_bit(R5_Insync, &dev->flags) ||
3535                                     ((i == sh->pd_idx || i == sh->qd_idx)  &&
3536                                      s.failed == 0))
3537                                         set_bit(STRIPE_INSYNC, &sh->state);
3538                         }
3539                 }
3540                 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3541                         s.dec_preread_active = 1;
3542         }
3543
3544         /*
3545          * might be able to return some write requests if the parity blocks
3546          * are safe, or on a failed drive
3547          */
3548         pdev = &sh->dev[sh->pd_idx];
3549         s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3550                 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3551         qdev = &sh->dev[sh->qd_idx];
3552         s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3553                 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3554                 || conf->level < 6;
3555
3556         if (s.written &&
3557             (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3558                              && !test_bit(R5_LOCKED, &pdev->flags)
3559                              && (test_bit(R5_UPTODATE, &pdev->flags) ||
3560                                  test_bit(R5_Discard, &pdev->flags))))) &&
3561             (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3562                              && !test_bit(R5_LOCKED, &qdev->flags)
3563                              && (test_bit(R5_UPTODATE, &qdev->flags) ||
3564                                  test_bit(R5_Discard, &qdev->flags))))))
3565                 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3566
3567         /* Now we might consider reading some blocks, either to check/generate
3568          * parity, or to satisfy requests
3569          * or to load a block that is being partially written.
3570          */
3571         if (s.to_read || s.non_overwrite
3572             || (conf->level == 6 && s.to_write && s.failed)
3573             || (s.syncing && (s.uptodate + s.compute < disks))
3574             || s.replacing
3575             || s.expanding)
3576                 handle_stripe_fill(sh, &s, disks);
3577
3578         /* Now to consider new write requests and what else, if anything
3579          * should be read.  We do not handle new writes when:
3580          * 1/ A 'write' operation (copy+xor) is already in flight.
3581          * 2/ A 'check' operation is in flight, as it may clobber the parity
3582          *    block.
3583          */
3584         if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3585                 handle_stripe_dirtying(conf, sh, &s, disks);
3586
3587         /* maybe we need to check and possibly fix the parity for this stripe
3588          * Any reads will already have been scheduled, so we just see if enough
3589          * data is available.  The parity check is held off while parity
3590          * dependent operations are in flight.
3591          */
3592         if (sh->check_state ||
3593             (s.syncing && s.locked == 0 &&
3594              !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3595              !test_bit(STRIPE_INSYNC, &sh->state))) {
3596                 if (conf->level == 6)
3597                         handle_parity_checks6(conf, sh, &s, disks);
3598                 else
3599                         handle_parity_checks5(conf, sh, &s, disks);
3600         }
3601
3602         if (s.replacing && s.locked == 0
3603             && !test_bit(STRIPE_INSYNC, &sh->state)) {
3604                 /* Write out to replacement devices where possible */
3605                 for (i = 0; i < conf->raid_disks; i++)
3606                         if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3607                             test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3608                                 set_bit(R5_WantReplace, &sh->dev[i].flags);
3609                                 set_bit(R5_LOCKED, &sh->dev[i].flags);
3610                                 s.locked++;
3611                         }
3612                 set_bit(STRIPE_INSYNC, &sh->state);
3613         }
3614         if ((s.syncing || s.replacing) && s.locked == 0 &&
3615             test_bit(STRIPE_INSYNC, &sh->state)) {
3616                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3617                 clear_bit(STRIPE_SYNCING, &sh->state);
3618                 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3619                         wake_up(&conf->wait_for_overlap);
3620         }
3621
3622         /* If the failed drives are just a ReadError, then we might need
3623          * to progress the repair/check process
3624          */
3625         if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3626                 for (i = 0; i < s.failed; i++) {
3627                         struct r5dev *dev = &sh->dev[s.failed_num[i]];
3628                         if (test_bit(R5_ReadError, &dev->flags)
3629                             && !test_bit(R5_LOCKED, &dev->flags)
3630                             && test_bit(R5_UPTODATE, &dev->flags)
3631                                 ) {
3632                                 if (!test_bit(R5_ReWrite, &dev->flags)) {
3633                                         set_bit(R5_Wantwrite, &dev->flags);
3634                                         set_bit(R5_ReWrite, &dev->flags);
3635                                         set_bit(R5_LOCKED, &dev->flags);
3636                                         s.locked++;
3637                                 } else {
3638                                         /* let's read it back */
3639                                         set_bit(R5_Wantread, &dev->flags);
3640                                         set_bit(R5_LOCKED, &dev->flags);
3641                                         s.locked++;
3642                                 }
3643                         }
3644                 }
3645
3646
3647         /* Finish reconstruct operations initiated by the expansion process */
3648         if (sh->reconstruct_state == reconstruct_state_result) {
3649                 struct stripe_head *sh_src
3650                         = get_active_stripe(conf, sh->sector, 1, 1, 1);
3651                 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3652                         /* sh cannot be written until sh_src has been read.
3653                          * so arrange for sh to be delayed a little
3654                          */
3655                         set_bit(STRIPE_DELAYED, &sh->state);
3656                         set_bit(STRIPE_HANDLE, &sh->state);
3657                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3658                                               &sh_src->state))
3659                                 atomic_inc(&conf->preread_active_stripes);
3660                         release_stripe(sh_src);
3661                         goto finish;
3662                 }
3663                 if (sh_src)
3664                         release_stripe(sh_src);
3665
3666                 sh->reconstruct_state = reconstruct_state_idle;
3667                 clear_bit(STRIPE_EXPANDING, &sh->state);
3668                 for (i = conf->raid_disks; i--; ) {
3669                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
3670                         set_bit(R5_LOCKED, &sh->dev[i].flags);
3671                         s.locked++;
3672                 }
3673         }
3674
3675         if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3676             !sh->reconstruct_state) {
3677                 /* Need to write out all blocks after computing parity */
3678                 sh->disks = conf->raid_disks;
3679                 stripe_set_idx(sh->sector, conf, 0, sh);
3680                 schedule_reconstruction(sh, &s, 1, 1);
3681         } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3682                 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3683                 atomic_dec(&conf->reshape_stripes);
3684                 wake_up(&conf->wait_for_overlap);
3685                 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3686         }
3687
3688         if (s.expanding && s.locked == 0 &&
3689             !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3690                 handle_stripe_expansion(conf, sh);
3691
3692 finish:
3693         /* wait for this device to become unblocked */
3694         if (unlikely(s.blocked_rdev)) {
3695                 if (conf->mddev->external)
3696                         md_wait_for_blocked_rdev(s.blocked_rdev,
3697                                                  conf->mddev);
3698                 else
3699                         /* Internal metadata will immediately
3700                          * be written by raid5d, so we don't
3701                          * need to wait here.
3702                          */
3703                         rdev_dec_pending(s.blocked_rdev,
3704                                          conf->mddev);
3705         }
3706
3707         if (s.handle_bad_blocks)
3708                 for (i = disks; i--; ) {
3709                         struct md_rdev *rdev;
3710                         struct r5dev *dev = &sh->dev[i];
3711                         if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3712                                 /* We own a safe reference to the rdev */
3713                                 rdev = conf->disks[i].rdev;
3714                                 if (!rdev_set_badblocks(rdev, sh->sector,
3715                                                         STRIPE_SECTORS, 0))
3716                                         md_error(conf->mddev, rdev);
3717                                 rdev_dec_pending(rdev, conf->mddev);
3718                         }
3719                         if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3720                                 rdev = conf->disks[i].rdev;
3721                                 rdev_clear_badblocks(rdev, sh->sector,
3722                                                      STRIPE_SECTORS, 0);
3723                                 rdev_dec_pending(rdev, conf->mddev);
3724                         }
3725                         if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3726                                 rdev = conf->disks[i].replacement;
3727                                 if (!rdev)
3728                                         /* rdev have been moved down */
3729                                         rdev = conf->disks[i].rdev;
3730                                 rdev_clear_badblocks(rdev, sh->sector,
3731                                                      STRIPE_SECTORS, 0);
3732                                 rdev_dec_pending(rdev, conf->mddev);
3733                         }
3734                 }
3735
3736         if (s.ops_request)
3737                 raid_run_ops(sh, s.ops_request);
3738
3739         ops_run_io(sh, &s);
3740
3741         if (s.dec_preread_active) {
3742                 /* We delay this until after ops_run_io so that if make_request
3743                  * is waiting on a flush, it won't continue until the writes
3744                  * have actually been submitted.
3745                  */
3746                 atomic_dec(&conf->preread_active_stripes);
3747                 if (atomic_read(&conf->preread_active_stripes) <
3748                     IO_THRESHOLD)
3749                         md_wakeup_thread(conf->mddev->thread);
3750         }
3751
3752         return_io(s.return_bi);
3753
3754         clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3755 }
3756
3757 static void raid5_activate_delayed(struct r5conf *conf)
3758 {
3759         if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3760                 while (!list_empty(&conf->delayed_list)) {
3761                         struct list_head *l = conf->delayed_list.next;
3762                         struct stripe_head *sh;
3763                         sh = list_entry(l, struct stripe_head, lru);
3764                         list_del_init(l);
3765                         clear_bit(STRIPE_DELAYED, &sh->state);
3766                         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3767                                 atomic_inc(&conf->preread_active_stripes);
3768                         list_add_tail(&sh->lru, &conf->hold_list);
3769                 }
3770         }
3771 }
3772
3773 static void activate_bit_delay(struct r5conf *conf)
3774 {
3775         /* device_lock is held */
3776         struct list_head head;
3777         list_add(&head, &conf->bitmap_list);
3778         list_del_init(&conf->bitmap_list);
3779         while (!list_empty(&head)) {
3780                 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3781                 list_del_init(&sh->lru);
3782                 atomic_inc(&sh->count);
3783                 __release_stripe(conf, sh);
3784         }
3785 }
3786
3787 int md_raid5_congested(struct mddev *mddev, int bits)
3788 {
3789         struct r5conf *conf = mddev->private;
3790
3791         /* No difference between reads and writes.  Just check
3792          * how busy the stripe_cache is
3793          */
3794
3795         if (conf->inactive_blocked)
3796                 return 1;
3797         if (conf->quiesce)
3798                 return 1;
3799         if (list_empty_careful(&conf->inactive_list))
3800                 return 1;
3801
3802         return 0;
3803 }
3804 EXPORT_SYMBOL_GPL(md_raid5_congested);
3805
3806 static int raid5_congested(void *data, int bits)
3807 {
3808         struct mddev *mddev = data;
3809
3810         return mddev_congested(mddev, bits) ||
3811                 md_raid5_congested(mddev, bits);
3812 }
3813
3814 /* We want read requests to align with chunks where possible,
3815  * but write requests don't need to.
3816  */
3817 static int raid5_mergeable_bvec(struct request_queue *q,
3818                                 struct bvec_merge_data *bvm,
3819                                 struct bio_vec *biovec)
3820 {
3821         struct mddev *mddev = q->queuedata;
3822         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3823         int max;
3824         unsigned int chunk_sectors = mddev->chunk_sectors;
3825         unsigned int bio_sectors = bvm->bi_size >> 9;
3826
3827         if ((bvm->bi_rw & 1) == WRITE)
3828                 return biovec->bv_len; /* always allow writes to be mergeable */
3829
3830         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3831                 chunk_sectors = mddev->new_chunk_sectors;
3832         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3833         if (max < 0) max = 0;
3834         if (max <= biovec->bv_len && bio_sectors == 0)
3835                 return biovec->bv_len;
3836         else
3837                 return max;
3838 }
3839
3840
3841 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3842 {
3843         sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3844         unsigned int chunk_sectors = mddev->chunk_sectors;
3845         unsigned int bio_sectors = bio->bi_size >> 9;
3846
3847         if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3848                 chunk_sectors = mddev->new_chunk_sectors;
3849         return  chunk_sectors >=
3850                 ((sector & (chunk_sectors - 1)) + bio_sectors);
3851 }
3852
3853 /*
3854  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3855  *  later sampled by raid5d.
3856  */
3857 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3858 {
3859         unsigned long flags;
3860
3861         spin_lock_irqsave(&conf->device_lock, flags);
3862
3863         bi->bi_next = conf->retry_read_aligned_list;
3864         conf->retry_read_aligned_list = bi;
3865
3866         spin_unlock_irqrestore(&conf->device_lock, flags);
3867         md_wakeup_thread(conf->mddev->thread);
3868 }
3869
3870
3871 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3872 {
3873         struct bio *bi;
3874
3875         bi = conf->retry_read_aligned;
3876         if (bi) {
3877                 conf->retry_read_aligned = NULL;
3878                 return bi;
3879         }
3880         bi = conf->retry_read_aligned_list;
3881         if(bi) {
3882                 conf->retry_read_aligned_list = bi->bi_next;
3883                 bi->bi_next = NULL;
3884                 /*
3885                  * this sets the active strip count to 1 and the processed
3886                  * strip count to zero (upper 8 bits)
3887                  */
3888                 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3889         }
3890
3891         return bi;
3892 }
3893
3894
3895 /*
3896  *  The "raid5_align_endio" should check if the read succeeded and if it
3897  *  did, call bio_endio on the original bio (having bio_put the new bio
3898  *  first).
3899  *  If the read failed..
3900  */
3901 static void raid5_align_endio(struct bio *bi, int error)
3902 {
3903         struct bio* raid_bi  = bi->bi_private;
3904         struct mddev *mddev;
3905         struct r5conf *conf;
3906         int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3907         struct md_rdev *rdev;
3908
3909         bio_put(bi);
3910
3911         rdev = (void*)raid_bi->bi_next;
3912         raid_bi->bi_next = NULL;
3913         mddev = rdev->mddev;
3914         conf = mddev->private;
3915
3916         rdev_dec_pending(rdev, conf->mddev);
3917
3918         if (!error && uptodate) {
3919                 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3920                                          raid_bi, 0);
3921                 bio_endio(raid_bi, 0);
3922                 if (atomic_dec_and_test(&conf->active_aligned_reads))
3923                         wake_up(&conf->wait_for_stripe);
3924                 return;
3925         }
3926
3927
3928         pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3929
3930         add_bio_to_retry(raid_bi, conf);
3931 }
3932
3933 static int bio_fits_rdev(struct bio *bi)
3934 {
3935         struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3936
3937         if ((bi->bi_size>>9) > queue_max_sectors(q))
3938                 return 0;
3939         blk_recount_segments(q, bi);
3940         if (bi->bi_phys_segments > queue_max_segments(q))
3941                 return 0;
3942
3943         if (q->merge_bvec_fn)
3944                 /* it's too hard to apply the merge_bvec_fn at this stage,
3945                  * just just give up
3946                  */
3947                 return 0;
3948
3949         return 1;
3950 }
3951
3952
3953 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3954 {
3955         struct r5conf *conf = mddev->private;
3956         int dd_idx;
3957         struct bio* align_bi;
3958         struct md_rdev *rdev;
3959         sector_t end_sector;
3960
3961         if (!in_chunk_boundary(mddev, raid_bio)) {
3962                 pr_debug("chunk_aligned_read : non aligned\n");
3963                 return 0;
3964         }
3965         /*
3966          * use bio_clone_mddev to make a copy of the bio
3967          */
3968         align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3969         if (!align_bi)
3970                 return 0;
3971         /*
3972          *   set bi_end_io to a new function, and set bi_private to the
3973          *     original bio.
3974          */
3975         align_bi->bi_end_io  = raid5_align_endio;
3976         align_bi->bi_private = raid_bio;
3977         /*
3978          *      compute position
3979          */
3980         align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3981                                                     0,
3982                                                     &dd_idx, NULL);
3983
3984         end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3985         rcu_read_lock();
3986         rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3987         if (!rdev || test_bit(Faulty, &rdev->flags) ||
3988             rdev->recovery_offset < end_sector) {
3989                 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3990                 if (rdev &&
3991                     (test_bit(Faulty, &rdev->flags) ||
3992                     !(test_bit(In_sync, &rdev->flags) ||
3993                       rdev->recovery_offset >= end_sector)))
3994                         rdev = NULL;
3995         }
3996         if (rdev) {
3997                 sector_t first_bad;
3998                 int bad_sectors;
3999
4000                 atomic_inc(&rdev->nr_pending);
4001                 rcu_read_unlock();
4002                 raid_bio->bi_next = (void*)rdev;
4003                 align_bi->bi_bdev =  rdev->bdev;
4004                 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4005
4006                 if (!bio_fits_rdev(align_bi) ||
4007                     is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
4008                                 &first_bad, &bad_sectors)) {
4009                         /* too big in some way, or has a known bad block */
4010                         bio_put(align_bi);
4011                         rdev_dec_pending(rdev, mddev);
4012                         return 0;
4013                 }
4014
4015                 /* No reshape active, so we can trust rdev->data_offset */
4016                 align_bi->bi_sector += rdev->data_offset;
4017
4018                 spin_lock_irq(&conf->device_lock);
4019                 wait_event_lock_irq(conf->wait_for_stripe,
4020                                     conf->quiesce == 0,
4021                                     conf->device_lock);
4022                 atomic_inc(&conf->active_aligned_reads);
4023                 spin_unlock_irq(&conf->device_lock);
4024
4025                 if (mddev->gendisk)
4026                         trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4027                                               align_bi, disk_devt(mddev->gendisk),
4028                                               raid_bio->bi_sector);
4029                 generic_make_request(align_bi);
4030                 return 1;
4031         } else {
4032                 rcu_read_unlock();
4033                 bio_put(align_bi);
4034                 return 0;
4035         }
4036 }
4037
4038 /* __get_priority_stripe - get the next stripe to process
4039  *
4040  * Full stripe writes are allowed to pass preread active stripes up until
4041  * the bypass_threshold is exceeded.  In general the bypass_count
4042  * increments when the handle_list is handled before the hold_list; however, it
4043  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4044  * stripe with in flight i/o.  The bypass_count will be reset when the
4045  * head of the hold_list has changed, i.e. the head was promoted to the
4046  * handle_list.
4047  */
4048 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4049 {
4050         struct stripe_head *sh;
4051
4052         pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4053                   __func__,
4054                   list_empty(&conf->handle_list) ? "empty" : "busy",
4055                   list_empty(&conf->hold_list) ? "empty" : "busy",
4056                   atomic_read(&conf->pending_full_writes), conf->bypass_count);
4057
4058         if (!list_empty(&conf->handle_list)) {
4059                 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4060
4061                 if (list_empty(&conf->hold_list))
4062                         conf->bypass_count = 0;
4063                 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4064                         if (conf->hold_list.next == conf->last_hold)
4065                                 conf->bypass_count++;
4066                         else {
4067                                 conf->last_hold = conf->hold_list.next;
4068                                 conf->bypass_count -= conf->bypass_threshold;
4069                                 if (conf->bypass_count < 0)
4070                                         conf->bypass_count = 0;
4071                         }
4072                 }
4073         } else if (!list_empty(&conf->hold_list) &&
4074                    ((conf->bypass_threshold &&
4075                      conf->bypass_count > conf->bypass_threshold) ||
4076                     atomic_read(&conf->pending_full_writes) == 0)) {
4077                 sh = list_entry(conf->hold_list.next,
4078                                 typeof(*sh), lru);
4079                 conf->bypass_count -= conf->bypass_threshold;
4080                 if (conf->bypass_count < 0)
4081                         conf->bypass_count = 0;
4082         } else
4083                 return NULL;
4084
4085         list_del_init(&sh->lru);
4086         atomic_inc(&sh->count);
4087         BUG_ON(atomic_read(&sh->count) != 1);
4088         return sh;
4089 }
4090
4091 struct raid5_plug_cb {
4092         struct blk_plug_cb      cb;
4093         struct list_head        list;
4094 };
4095
4096 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4097 {
4098         struct raid5_plug_cb *cb = container_of(
4099                 blk_cb, struct raid5_plug_cb, cb);
4100         struct stripe_head *sh;
4101         struct mddev *mddev = cb->cb.data;
4102         struct r5conf *conf = mddev->private;
4103         int cnt = 0;
4104
4105         if (cb->list.next && !list_empty(&cb->list)) {
4106                 spin_lock_irq(&conf->device_lock);
4107                 while (!list_empty(&cb->list)) {
4108                         sh = list_first_entry(&cb->list, struct stripe_head, lru);
4109                         list_del_init(&sh->lru);
4110                         /*
4111                          * avoid race release_stripe_plug() sees
4112                          * STRIPE_ON_UNPLUG_LIST clear but the stripe
4113                          * is still in our list
4114                          */
4115                         smp_mb__before_clear_bit();
4116                         clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4117                         __release_stripe(conf, sh);
4118                         cnt++;
4119                 }
4120                 spin_unlock_irq(&conf->device_lock);
4121         }
4122         if (mddev->queue)
4123                 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4124         kfree(cb);
4125 }
4126
4127 static void release_stripe_plug(struct mddev *mddev,
4128                                 struct stripe_head *sh)
4129 {
4130         struct blk_plug_cb *blk_cb = blk_check_plugged(
4131                 raid5_unplug, mddev,
4132                 sizeof(struct raid5_plug_cb));
4133         struct raid5_plug_cb *cb;
4134
4135         if (!blk_cb) {
4136                 release_stripe(sh);
4137                 return;
4138         }
4139
4140         cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4141
4142         if (cb->list.next == NULL)
4143                 INIT_LIST_HEAD(&cb->list);
4144
4145         if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4146                 list_add_tail(&sh->lru, &cb->list);
4147         else
4148                 release_stripe(sh);
4149 }
4150
4151 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4152 {
4153         struct r5conf *conf = mddev->private;
4154         sector_t logical_sector, last_sector;
4155         struct stripe_head *sh;
4156         int remaining;
4157         int stripe_sectors;
4158
4159         if (mddev->reshape_position != MaxSector)
4160                 /* Skip discard while reshape is happening */
4161                 return;
4162
4163         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4164         last_sector = bi->bi_sector + (bi->bi_size>>9);
4165
4166         bi->bi_next = NULL;
4167         bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4168
4169         stripe_sectors = conf->chunk_sectors *
4170                 (conf->raid_disks - conf->max_degraded);
4171         logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4172                                                stripe_sectors);
4173         sector_div(last_sector, stripe_sectors);
4174
4175         logical_sector *= conf->chunk_sectors;
4176         last_sector *= conf->chunk_sectors;
4177
4178         for (; logical_sector < last_sector;
4179              logical_sector += STRIPE_SECTORS) {
4180                 DEFINE_WAIT(w);
4181                 int d;
4182         again:
4183                 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4184                 prepare_to_wait(&conf->wait_for_overlap, &w,
4185                                 TASK_UNINTERRUPTIBLE);
4186                 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4187                 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4188                         release_stripe(sh);
4189                         schedule();
4190                         goto again;
4191                 }
4192                 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4193                 spin_lock_irq(&sh->stripe_lock);
4194                 for (d = 0; d < conf->raid_disks; d++) {
4195                         if (d == sh->pd_idx || d == sh->qd_idx)
4196                                 continue;
4197                         if (sh->dev[d].towrite || sh->dev[d].toread) {
4198                                 set_bit(R5_Overlap, &sh->dev[d].flags);
4199                                 spin_unlock_irq(&sh->stripe_lock);
4200                                 release_stripe(sh);
4201                                 schedule();
4202                                 goto again;
4203                         }
4204                 }
4205                 set_bit(STRIPE_DISCARD, &sh->state);
4206                 finish_wait(&conf->wait_for_overlap, &w);
4207                 for (d = 0; d < conf->raid_disks; d++) {
4208                         if (d == sh->pd_idx || d == sh->qd_idx)
4209                                 continue;
4210                         sh->dev[d].towrite = bi;
4211                         set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4212                         raid5_inc_bi_active_stripes(bi);
4213                 }
4214                 spin_unlock_irq(&sh->stripe_lock);
4215                 if (conf->mddev->bitmap) {
4216                         for (d = 0;
4217                              d < conf->raid_disks - conf->max_degraded;
4218                              d++)
4219                                 bitmap_startwrite(mddev->bitmap,
4220                                                   sh->sector,
4221                                                   STRIPE_SECTORS,
4222                                                   0);
4223                         sh->bm_seq = conf->seq_flush + 1;
4224                         set_bit(STRIPE_BIT_DELAY, &sh->state);
4225                 }
4226
4227                 set_bit(STRIPE_HANDLE, &sh->state);
4228                 clear_bit(STRIPE_DELAYED, &sh->state);
4229                 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4230                         atomic_inc(&conf->preread_active_stripes);
4231                 release_stripe_plug(mddev, sh);
4232         }
4233
4234         remaining = raid5_dec_bi_active_stripes(bi);
4235         if (remaining == 0) {
4236                 md_write_end(mddev);
4237                 bio_endio(bi, 0);
4238         }
4239 }
4240
4241 static void make_request(struct mddev *mddev, struct bio * bi)
4242 {
4243         struct r5conf *conf = mddev->private;
4244         int dd_idx;
4245         sector_t new_sector;
4246         sector_t logical_sector, last_sector;
4247         struct stripe_head *sh;
4248         const int rw = bio_data_dir(bi);
4249         int remaining;
4250
4251         if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4252                 md_flush_request(mddev, bi);
4253                 return;
4254         }
4255
4256         md_write_start(mddev, bi);
4257
4258         if (rw == READ &&
4259              mddev->reshape_position == MaxSector &&
4260              chunk_aligned_read(mddev,bi))
4261                 return;
4262
4263         if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4264                 make_discard_request(mddev, bi);
4265                 return;
4266         }
4267
4268         logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4269         last_sector = bi->bi_sector + (bi->bi_size>>9);
4270         bi->bi_next = NULL;
4271         bi->bi_phys_segments = 1;       /* over-loaded to count active stripes */
4272
4273         for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4274                 DEFINE_WAIT(w);
4275                 int previous;
4276
4277         retry:
4278                 previous = 0;
4279                 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4280                 if (unlikely(conf->reshape_progress != MaxSector)) {
4281                         /* spinlock is needed as reshape_progress may be
4282                          * 64bit on a 32bit platform, and so it might be
4283                          * possible to see a half-updated value
4284                          * Of course reshape_progress could change after
4285                          * the lock is dropped, so once we get a reference
4286                          * to the stripe that we think it is, we will have
4287                          * to check again.
4288                          */
4289                         spin_lock_irq(&conf->device_lock);
4290                         if (mddev->reshape_backwards
4291                             ? logical_sector < conf->reshape_progress
4292                             : logical_sector >= conf->reshape_progress) {
4293                                 previous = 1;
4294                         } else {
4295                                 if (mddev->reshape_backwards
4296                                     ? logical_sector < conf->reshape_safe
4297                                     : logical_sector >= conf->reshape_safe) {
4298                                         spin_unlock_irq(&conf->device_lock);
4299                                         schedule();
4300                                         goto retry;
4301                                 }
4302                         }
4303                         spin_unlock_irq(&conf->device_lock);
4304                 }
4305
4306                 new_sector = raid5_compute_sector(conf, logical_sector,
4307                                                   previous,
4308                                                   &dd_idx, NULL);
4309                 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4310                         (unsigned long long)new_sector,
4311                         (unsigned long long)logical_sector);
4312
4313                 sh = get_active_stripe(conf, new_sector, previous,
4314                                        (bi->bi_rw&RWA_MASK), 0);
4315                 if (sh) {
4316                         if (unlikely(previous)) {
4317                                 /* expansion might have moved on while waiting for a
4318                                  * stripe, so we must do the range check again.
4319                                  * Expansion could still move past after this
4320                                  * test, but as we are holding a reference to
4321                                  * 'sh', we know that if that happens,
4322                                  *  STRIPE_EXPANDING will get set and the expansion
4323                                  * won't proceed until we finish with the stripe.
4324                                  */
4325                                 int must_retry = 0;
4326                                 spin_lock_irq(&conf->device_lock);
4327                                 if (mddev->reshape_backwards
4328                                     ? logical_sector >= conf->reshape_progress
4329                                     : logical_sector < conf->reshape_progress)
4330                                         /* mismatch, need to try again */
4331                                         must_retry = 1;
4332                                 spin_unlock_irq(&conf->device_lock);
4333                                 if (must_retry) {
4334                                         release_stripe(sh);
4335                                         schedule();
4336                                         goto retry;
4337                                 }
4338                         }
4339
4340                         if (rw == WRITE &&
4341                             logical_sector >= mddev->suspend_lo &&
4342                             logical_sector < mddev->suspend_hi) {
4343                                 release_stripe(sh);
4344                                 /* As the suspend_* range is controlled by
4345                                  * userspace, we want an interruptible
4346                                  * wait.
4347                                  */
4348                                 flush_signals(current);
4349                                 prepare_to_wait(&conf->wait_for_overlap,
4350                                                 &w, TASK_INTERRUPTIBLE);
4351                                 if (logical_sector >= mddev->suspend_lo &&
4352                                     logical_sector < mddev->suspend_hi)
4353                                         schedule();
4354                                 goto retry;
4355                         }
4356
4357                         if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4358                             !add_stripe_bio(sh, bi, dd_idx, rw)) {
4359                                 /* Stripe is busy expanding or
4360                                  * add failed due to overlap.  Flush everything
4361                                  * and wait a while
4362                                  */
4363                                 md_wakeup_thread(mddev->thread);
4364                                 release_stripe(sh);
4365                                 schedule();
4366                                 goto retry;
4367                         }
4368                         finish_wait(&conf->wait_for_overlap, &w);
4369                         set_bit(STRIPE_HANDLE, &sh->state);
4370                         clear_bit(STRIPE_DELAYED, &sh->state);
4371                         if ((bi->bi_rw & REQ_SYNC) &&
4372                             !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4373                                 atomic_inc(&conf->preread_active_stripes);
4374                         release_stripe_plug(mddev, sh);
4375                 } else {
4376                         /* cannot get stripe for read-ahead, just give-up */
4377                         clear_bit(BIO_UPTODATE, &bi->bi_flags);
4378                         finish_wait(&conf->wait_for_overlap, &w);
4379                         break;
4380                 }
4381         }
4382
4383         remaining = raid5_dec_bi_active_stripes(bi);
4384         if (remaining == 0) {
4385
4386                 if ( rw == WRITE )
4387                         md_write_end(mddev);
4388
4389                 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4390                                          bi, 0);
4391                 bio_endio(bi, 0);
4392         }
4393 }
4394
4395 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4396
4397 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4398 {
4399         /* reshaping is quite different to recovery/resync so it is
4400          * handled quite separately ... here.
4401          *
4402          * On each call to sync_request, we gather one chunk worth of
4403          * destination stripes and flag them as expanding.
4404          * Then we find all the source stripes and request reads.
4405          * As the reads complete, handle_stripe will copy the data
4406          * into the destination stripe and release that stripe.
4407          */
4408         struct r5conf *conf = mddev->private;
4409         struct stripe_head *sh;
4410         sector_t first_sector, last_sector;
4411         int raid_disks = conf->previous_raid_disks;
4412         int data_disks = raid_disks - conf->max_degraded;
4413         int new_data_disks = conf->raid_disks - conf->max_degraded;
4414         int i;
4415         int dd_idx;
4416         sector_t writepos, readpos, safepos;
4417         sector_t stripe_addr;
4418         int reshape_sectors;
4419         struct list_head stripes;
4420
4421         if (sector_nr == 0) {
4422                 /* If restarting in the middle, skip the initial sectors */
4423                 if (mddev->reshape_backwards &&
4424                     conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4425                         sector_nr = raid5_size(mddev, 0, 0)
4426                                 - conf->reshape_progress;
4427                 } else if (!mddev->reshape_backwards &&
4428                            conf->reshape_progress > 0)
4429                         sector_nr = conf->reshape_progress;
4430                 sector_div(sector_nr, new_data_disks);
4431                 if (sector_nr) {
4432                         mddev->curr_resync_completed = sector_nr;
4433                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4434                         *skipped = 1;
4435                         return sector_nr;
4436                 }
4437         }
4438
4439         /* We need to process a full chunk at a time.
4440          * If old and new chunk sizes differ, we need to process the
4441          * largest of these
4442          */
4443         if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4444                 reshape_sectors = mddev->new_chunk_sectors;
4445         else
4446                 reshape_sectors = mddev->chunk_sectors;
4447
4448         /* We update the metadata at least every 10 seconds, or when
4449          * the data about to be copied would over-write the source of
4450          * the data at the front of the range.  i.e. one new_stripe
4451          * along from reshape_progress new_maps to after where
4452          * reshape_safe old_maps to
4453          */
4454         writepos = conf->reshape_progress;
4455         sector_div(writepos, new_data_disks);
4456         readpos = conf->reshape_progress;
4457         sector_div(readpos, data_disks);
4458         safepos = conf->reshape_safe;
4459         sector_div(safepos, data_disks);
4460         if (mddev->reshape_backwards) {
4461                 writepos -= min_t(sector_t, reshape_sectors, writepos);
4462                 readpos += reshape_sectors;
4463                 safepos += reshape_sectors;
4464         } else {
4465                 writepos += reshape_sectors;
4466                 readpos -= min_t(sector_t, reshape_sectors, readpos);
4467                 safepos -= min_t(sector_t, reshape_sectors, safepos);
4468         }
4469
4470         /* Having calculated the 'writepos' possibly use it
4471          * to set 'stripe_addr' which is where we will write to.
4472          */
4473         if (mddev->reshape_backwards) {
4474                 BUG_ON(conf->reshape_progress == 0);
4475                 stripe_addr = writepos;
4476                 BUG_ON((mddev->dev_sectors &
4477                         ~((sector_t)reshape_sectors - 1))
4478                        - reshape_sectors - stripe_addr
4479                        != sector_nr);
4480         } else {
4481                 BUG_ON(writepos != sector_nr + reshape_sectors);
4482                 stripe_addr = sector_nr;
4483         }
4484
4485         /* 'writepos' is the most advanced device address we might write.
4486          * 'readpos' is the least advanced device address we might read.
4487          * 'safepos' is the least address recorded in the metadata as having
4488          *     been reshaped.
4489          * If there is a min_offset_diff, these are adjusted either by
4490          * increasing the safepos/readpos if diff is negative, or
4491          * increasing writepos if diff is positive.
4492          * If 'readpos' is then behind 'writepos', there is no way that we can
4493          * ensure safety in the face of a crash - that must be done by userspace
4494          * making a backup of the data.  So in that case there is no particular
4495          * rush to update metadata.
4496          * Otherwise if 'safepos' is behind 'writepos', then we really need to
4497          * update the metadata to advance 'safepos' to match 'readpos' so that
4498          * we can be safe in the event of a crash.
4499          * So we insist on updating metadata if safepos is behind writepos and
4500          * readpos is beyond writepos.
4501          * In any case, update the metadata every 10 seconds.
4502          * Maybe that number should be configurable, but I'm not sure it is
4503          * worth it.... maybe it could be a multiple of safemode_delay???
4504          */
4505         if (conf->min_offset_diff < 0) {
4506                 safepos += -conf->min_offset_diff;
4507                 readpos += -conf->min_offset_diff;
4508         } else
4509                 writepos += conf->min_offset_diff;
4510
4511         if ((mddev->reshape_backwards
4512              ? (safepos > writepos && readpos < writepos)
4513              : (safepos < writepos && readpos > writepos)) ||
4514             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4515                 /* Cannot proceed until we've updated the superblock... */
4516                 wait_event(conf->wait_for_overlap,
4517                            atomic_read(&conf->reshape_stripes)==0);
4518                 mddev->reshape_position = conf->reshape_progress;
4519                 mddev->curr_resync_completed = sector_nr;
4520                 conf->reshape_checkpoint = jiffies;
4521                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4522                 md_wakeup_thread(mddev->thread);
4523                 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4524                            kthread_should_stop());
4525                 spin_lock_irq(&conf->device_lock);
4526                 conf->reshape_safe = mddev->reshape_position;
4527                 spin_unlock_irq(&conf->device_lock);
4528                 wake_up(&conf->wait_for_overlap);
4529                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4530         }
4531
4532         INIT_LIST_HEAD(&stripes);
4533         for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4534                 int j;
4535                 int skipped_disk = 0;
4536                 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4537                 set_bit(STRIPE_EXPANDING, &sh->state);
4538                 atomic_inc(&conf->reshape_stripes);
4539                 /* If any of this stripe is beyond the end of the old
4540                  * array, then we need to zero those blocks
4541                  */
4542                 for (j=sh->disks; j--;) {
4543                         sector_t s;
4544                         if (j == sh->pd_idx)
4545                                 continue;
4546                         if (conf->level == 6 &&
4547                             j == sh->qd_idx)
4548                                 continue;
4549                         s = compute_blocknr(sh, j, 0);
4550                         if (s < raid5_size(mddev, 0, 0)) {
4551                                 skipped_disk = 1;
4552                                 continue;
4553                         }
4554                         memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4555                         set_bit(R5_Expanded, &sh->dev[j].flags);
4556                         set_bit(R5_UPTODATE, &sh->dev[j].flags);
4557                 }
4558                 if (!skipped_disk) {
4559                         set_bit(STRIPE_EXPAND_READY, &sh->state);
4560                         set_bit(STRIPE_HANDLE, &sh->state);
4561                 }
4562                 list_add(&sh->lru, &stripes);
4563         }
4564         spin_lock_irq(&conf->device_lock);
4565         if (mddev->reshape_backwards)
4566                 conf->reshape_progress -= reshape_sectors * new_data_disks;
4567         else
4568                 conf->reshape_progress += reshape_sectors * new_data_disks;
4569         spin_unlock_irq(&conf->device_lock);
4570         /* Ok, those stripe are ready. We can start scheduling
4571          * reads on the source stripes.
4572          * The source stripes are determined by mapping the first and last
4573          * block on the destination stripes.
4574          */
4575         first_sector =
4576                 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4577                                      1, &dd_idx, NULL);
4578         last_sector =
4579                 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4580                                             * new_data_disks - 1),
4581                                      1, &dd_idx, NULL);
4582         if (last_sector >= mddev->dev_sectors)
4583                 last_sector = mddev->dev_sectors - 1;
4584         while (first_sector <= last_sector) {
4585                 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4586                 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4587                 set_bit(STRIPE_HANDLE, &sh->state);
4588                 release_stripe(sh);
4589                 first_sector += STRIPE_SECTORS;
4590         }
4591         /* Now that the sources are clearly marked, we can release
4592          * the destination stripes
4593          */
4594         while (!list_empty(&stripes)) {
4595                 sh = list_entry(stripes.next, struct stripe_head, lru);
4596                 list_del_init(&sh->lru);
4597                 release_stripe(sh);
4598         }
4599         /* If this takes us to the resync_max point where we have to pause,
4600          * then we need to write out the superblock.
4601          */
4602         sector_nr += reshape_sectors;
4603         if ((sector_nr - mddev->curr_resync_completed) * 2
4604             >= mddev->resync_max - mddev->curr_resync_completed) {
4605                 /* Cannot proceed until we've updated the superblock... */
4606                 wait_event(conf->wait_for_overlap,
4607                            atomic_read(&conf->reshape_stripes) == 0);
4608                 mddev->reshape_position = conf->reshape_progress;
4609                 mddev->curr_resync_completed = sector_nr;
4610                 conf->reshape_checkpoint = jiffies;
4611                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4612                 md_wakeup_thread(mddev->thread);
4613                 wait_event(mddev->sb_wait,
4614                            !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4615                            || kthread_should_stop());
4616                 spin_lock_irq(&conf->device_lock);
4617                 conf->reshape_safe = mddev->reshape_position;
4618                 spin_unlock_irq(&conf->device_lock);
4619                 wake_up(&conf->wait_for_overlap);
4620                 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4621         }
4622         return reshape_sectors;
4623 }
4624
4625 /* FIXME go_faster isn't used */
4626 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4627 {
4628         struct r5conf *conf = mddev->private;
4629         struct stripe_head *sh;
4630         sector_t max_sector = mddev->dev_sectors;
4631         sector_t sync_blocks;
4632         int still_degraded = 0;
4633         int i;
4634
4635         if (sector_nr >= max_sector) {
4636                 /* just being told to finish up .. nothing much to do */
4637
4638                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4639                         end_reshape(conf);
4640                         return 0;
4641                 }
4642
4643                 if (mddev->curr_resync < max_sector) /* aborted */
4644                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4645                                         &sync_blocks, 1);
4646                 else /* completed sync */
4647                         conf->fullsync = 0;
4648                 bitmap_close_sync(mddev->bitmap);
4649
4650                 return 0;
4651         }
4652
4653         /* Allow raid5_quiesce to complete */
4654         wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4655
4656         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4657                 return reshape_request(mddev, sector_nr, skipped);
4658
4659         /* No need to check resync_max as we never do more than one
4660          * stripe, and as resync_max will always be on a chunk boundary,
4661          * if the check in md_do_sync didn't fire, there is no chance
4662          * of overstepping resync_max here
4663          */
4664
4665         /* if there is too many failed drives and we are trying
4666          * to resync, then assert that we are finished, because there is
4667          * nothing we can do.
4668          */
4669         if (mddev->degraded >= conf->max_degraded &&
4670             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4671                 sector_t rv = mddev->dev_sectors - sector_nr;
4672                 *skipped = 1;
4673                 return rv;
4674         }
4675         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4676             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4677             !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4678                 /* we can skip this block, and probably more */
4679                 sync_blocks /= STRIPE_SECTORS;
4680                 *skipped = 1;
4681                 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4682         }
4683
4684         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4685
4686         sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4687         if (sh == NULL) {
4688                 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4689                 /* make sure we don't swamp the stripe cache if someone else
4690                  * is trying to get access
4691                  */
4692                 schedule_timeout_uninterruptible(1);
4693         }
4694         /* Need to check if array will still be degraded after recovery/resync
4695          * We don't need to check the 'failed' flag as when that gets set,
4696          * recovery aborts.
4697          */
4698         for (i = 0; i < conf->raid_disks; i++)
4699                 if (conf->disks[i].rdev == NULL)
4700                         still_degraded = 1;
4701
4702         bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4703
4704         set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4705
4706         handle_stripe(sh);
4707         release_stripe(sh);
4708
4709         return STRIPE_SECTORS;
4710 }
4711
4712 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4713 {
4714         /* We may not be able to submit a whole bio at once as there
4715          * may not be enough stripe_heads available.
4716          * We cannot pre-allocate enough stripe_heads as we may need
4717          * more than exist in the cache (if we allow ever large chunks).
4718          * So we do one stripe head at a time and record in
4719          * ->bi_hw_segments how many have been done.
4720          *
4721          * We *know* that this entire raid_bio is in one chunk, so
4722          * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4723          */
4724         struct stripe_head *sh;
4725         int dd_idx;
4726         sector_t sector, logical_sector, last_sector;
4727         int scnt = 0;
4728         int remaining;
4729         int handled = 0;
4730
4731         logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4732         sector = raid5_compute_sector(conf, logical_sector,
4733                                       0, &dd_idx, NULL);
4734         last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4735
4736         for (; logical_sector < last_sector;
4737              logical_sector += STRIPE_SECTORS,
4738                      sector += STRIPE_SECTORS,
4739                      scnt++) {
4740
4741                 if (scnt < raid5_bi_processed_stripes(raid_bio))
4742                         /* already done this stripe */
4743                         continue;
4744
4745                 sh = get_active_stripe(conf, sector, 0, 1, 0);
4746
4747                 if (!sh) {
4748                         /* failed to get a stripe - must wait */
4749                         raid5_set_bi_processed_stripes(raid_bio, scnt);
4750                         conf->retry_read_aligned = raid_bio;
4751                         return handled;
4752                 }
4753
4754                 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4755                         release_stripe(sh);
4756                         raid5_set_bi_processed_stripes(raid_bio, scnt);
4757                         conf->retry_read_aligned = raid_bio;
4758                         return handled;
4759                 }
4760
4761                 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4762                 handle_stripe(sh);
4763                 release_stripe(sh);
4764                 handled++;
4765         }
4766         remaining = raid5_dec_bi_active_stripes(raid_bio);
4767         if (remaining == 0) {
4768                 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4769                                          raid_bio, 0);
4770                 bio_endio(raid_bio, 0);
4771         }
4772         if (atomic_dec_and_test(&conf->active_aligned_reads))
4773                 wake_up(&conf->wait_for_stripe);
4774         return handled;
4775 }
4776
4777 #define MAX_STRIPE_BATCH 8
4778 static int handle_active_stripes(struct r5conf *conf)
4779 {
4780         struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4781         int i, batch_size = 0;
4782
4783         while (batch_size < MAX_STRIPE_BATCH &&
4784                         (sh = __get_priority_stripe(conf)) != NULL)
4785                 batch[batch_size++] = sh;
4786
4787         if (batch_size == 0)
4788                 return batch_size;
4789         spin_unlock_irq(&conf->device_lock);
4790
4791         for (i = 0; i < batch_size; i++)
4792                 handle_stripe(batch[i]);
4793
4794         cond_resched();
4795
4796         spin_lock_irq(&conf->device_lock);
4797         for (i = 0; i < batch_size; i++)
4798                 __release_stripe(conf, batch[i]);
4799         return batch_size;
4800 }
4801
4802 /*
4803  * This is our raid5 kernel thread.
4804  *
4805  * We scan the hash table for stripes which can be handled now.
4806  * During the scan, completed stripes are saved for us by the interrupt
4807  * handler, so that they will not have to wait for our next wakeup.
4808  */
4809 static void raid5d(struct md_thread *thread)
4810 {
4811         struct mddev *mddev = thread->mddev;
4812         struct r5conf *conf = mddev->private;
4813         int handled;
4814         struct blk_plug plug;
4815
4816         pr_debug("+++ raid5d active\n");
4817
4818         md_check_recovery(mddev);
4819
4820         blk_start_plug(&plug);
4821         handled = 0;
4822         spin_lock_irq(&conf->device_lock);
4823         while (1) {
4824                 struct bio *bio;
4825                 int batch_size;
4826
4827                 if (
4828                     !list_empty(&conf->bitmap_list)) {
4829                         /* Now is a good time to flush some bitmap updates */
4830                         conf->seq_flush++;
4831                         spin_unlock_irq(&conf->device_lock);
4832                         bitmap_unplug(mddev->bitmap);
4833                         spin_lock_irq(&conf->device_lock);
4834                         conf->seq_write = conf->seq_flush;
4835                         activate_bit_delay(conf);
4836                 }
4837                 raid5_activate_delayed(conf);
4838
4839                 while ((bio = remove_bio_from_retry(conf))) {
4840                         int ok;
4841                         spin_unlock_irq(&conf->device_lock);
4842                         ok = retry_aligned_read(conf, bio);
4843                         spin_lock_irq(&conf->device_lock);
4844                         if (!ok)
4845                                 break;
4846                         handled++;
4847                 }
4848
4849                 batch_size = handle_active_stripes(conf);
4850                 if (!batch_size)
4851                         break;
4852                 handled += batch_size;
4853
4854                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4855                         spin_unlock_irq(&conf->device_lock);
4856                         md_check_recovery(mddev);
4857                         spin_lock_irq(&conf->device_lock);
4858                 }
4859         }
4860         pr_debug("%d stripes handled\n", handled);
4861
4862         spin_unlock_irq(&conf->device_lock);
4863
4864         async_tx_issue_pending_all();
4865         blk_finish_plug(&plug);
4866
4867         pr_debug("--- raid5d inactive\n");
4868 }
4869
4870 static ssize_t
4871 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4872 {
4873         struct r5conf *conf = mddev->private;
4874         if (conf)
4875                 return sprintf(page, "%d\n", conf->max_nr_stripes);
4876         else
4877                 return 0;
4878 }
4879
4880 int
4881 raid5_set_cache_size(struct mddev *mddev, int size)
4882 {
4883         struct r5conf *conf = mddev->private;
4884         int err;
4885
4886         if (size <= 16 || size > 32768)
4887                 return -EINVAL;
4888         while (size < conf->max_nr_stripes) {
4889                 if (drop_one_stripe(conf))
4890                         conf->max_nr_stripes--;
4891                 else
4892                         break;
4893         }
4894         err = md_allow_write(mddev);
4895         if (err)
4896                 return err;
4897         while (size > conf->max_nr_stripes) {
4898                 if (grow_one_stripe(conf))
4899                         conf->max_nr_stripes++;
4900                 else break;
4901         }
4902         return 0;
4903 }
4904 EXPORT_SYMBOL(raid5_set_cache_size);
4905
4906 static ssize_t
4907 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4908 {
4909         struct r5conf *conf = mddev->private;
4910         unsigned long new;
4911         int err;
4912
4913         if (len >= PAGE_SIZE)
4914                 return -EINVAL;
4915         if (!conf)
4916                 return -ENODEV;
4917
4918         if (strict_strtoul(page, 10, &new))
4919                 return -EINVAL;
4920         err = raid5_set_cache_size(mddev, new);
4921         if (err)
4922                 return err;
4923         return len;
4924 }
4925
4926 static struct md_sysfs_entry
4927 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4928                                 raid5_show_stripe_cache_size,
4929                                 raid5_store_stripe_cache_size);
4930
4931 static ssize_t
4932 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4933 {
4934         struct r5conf *conf = mddev->private;
4935         if (conf)
4936                 return sprintf(page, "%d\n", conf->bypass_threshold);
4937         else
4938                 return 0;
4939 }
4940
4941 static ssize_t
4942 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4943 {
4944         struct r5conf *conf = mddev->private;
4945         unsigned long new;
4946         if (len >= PAGE_SIZE)
4947                 return -EINVAL;
4948         if (!conf)
4949                 return -ENODEV;
4950
4951         if (strict_strtoul(page, 10, &new))
4952                 return -EINVAL;
4953         if (new > conf->max_nr_stripes)
4954                 return -EINVAL;
4955         conf->bypass_threshold = new;
4956         return len;
4957 }
4958
4959 static struct md_sysfs_entry
4960 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4961                                         S_IRUGO | S_IWUSR,
4962                                         raid5_show_preread_threshold,
4963                                         raid5_store_preread_threshold);
4964
4965 static ssize_t
4966 stripe_cache_active_show(struct mddev *mddev, char *page)
4967 {
4968         struct r5conf *conf = mddev->private;
4969         if (conf)
4970                 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4971         else
4972                 return 0;
4973 }
4974
4975 static struct md_sysfs_entry
4976 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4977
4978 static struct attribute *raid5_attrs[] =  {
4979         &raid5_stripecache_size.attr,
4980         &raid5_stripecache_active.attr,
4981         &raid5_preread_bypass_threshold.attr,
4982         NULL,
4983 };
4984 static struct attribute_group raid5_attrs_group = {
4985         .name = NULL,
4986         .attrs = raid5_attrs,
4987 };
4988
4989 static sector_t
4990 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4991 {
4992         struct r5conf *conf = mddev->private;
4993
4994         if (!sectors)
4995                 sectors = mddev->dev_sectors;
4996         if (!raid_disks)
4997                 /* size is defined by the smallest of previous and new size */
4998                 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4999
5000         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5001         sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5002         return sectors * (raid_disks - conf->max_degraded);
5003 }
5004
5005 static void raid5_free_percpu(struct r5conf *conf)
5006 {
5007         struct raid5_percpu *percpu;
5008         unsigned long cpu;
5009
5010         if (!conf->percpu)
5011                 return;
5012
5013         get_online_cpus();
5014         for_each_possible_cpu(cpu) {
5015                 percpu = per_cpu_ptr(conf->percpu, cpu);
5016                 safe_put_page(percpu->spare_page);
5017                 kfree(percpu->scribble);
5018         }
5019 #ifdef CONFIG_HOTPLUG_CPU
5020         unregister_cpu_notifier(&conf->cpu_notify);
5021 #endif
5022         put_online_cpus();
5023
5024         free_percpu(conf->percpu);
5025 }
5026
5027 static void free_conf(struct r5conf *conf)
5028 {
5029         shrink_stripes(conf);
5030         raid5_free_percpu(conf);
5031         kfree(conf->disks);
5032         kfree(conf->stripe_hashtbl);
5033         kfree(conf);
5034 }
5035
5036 #ifdef CONFIG_HOTPLUG_CPU
5037 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5038                               void *hcpu)
5039 {
5040         struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5041         long cpu = (long)hcpu;
5042         struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5043
5044         switch (action) {
5045         case CPU_UP_PREPARE:
5046         case CPU_UP_PREPARE_FROZEN:
5047                 if (conf->level == 6 && !percpu->spare_page)
5048                         percpu->spare_page = alloc_page(GFP_KERNEL);
5049                 if (!percpu->scribble)
5050                         percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5051
5052                 if (!percpu->scribble ||
5053                     (conf->level == 6 && !percpu->spare_page)) {
5054                         safe_put_page(percpu->spare_page);
5055                         kfree(percpu->scribble);
5056                         pr_err("%s: failed memory allocation for cpu%ld\n",
5057                                __func__, cpu);
5058                         return notifier_from_errno(-ENOMEM);
5059                 }
5060                 break;
5061         case CPU_DEAD:
5062         case CPU_DEAD_FROZEN:
5063                 safe_put_page(percpu->spare_page);
5064                 kfree(percpu->scribble);
5065                 percpu->spare_page = NULL;
5066                 percpu->scribble = NULL;
5067                 break;
5068         default:
5069                 break;
5070         }
5071         return NOTIFY_OK;
5072 }
5073 #endif
5074
5075 static int raid5_alloc_percpu(struct r5conf *conf)
5076 {
5077         unsigned long cpu;
5078         struct page *spare_page;
5079         struct raid5_percpu __percpu *allcpus;
5080         void *scribble;
5081         int err;
5082
5083         allcpus = alloc_percpu(struct raid5_percpu);
5084         if (!allcpus)
5085                 return -ENOMEM;
5086         conf->percpu = allcpus;
5087
5088         get_online_cpus();
5089         err = 0;
5090         for_each_present_cpu(cpu) {
5091                 if (conf->level == 6) {
5092                         spare_page = alloc_page(GFP_KERNEL);
5093                         if (!spare_page) {
5094                                 err = -ENOMEM;
5095                                 break;
5096                         }
5097                         per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5098                 }
5099                 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5100                 if (!scribble) {
5101                         err = -ENOMEM;
5102                         break;
5103                 }
5104                 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5105         }
5106 #ifdef CONFIG_HOTPLUG_CPU
5107         conf->cpu_notify.notifier_call = raid456_cpu_notify;
5108         conf->cpu_notify.priority = 0;
5109         if (err == 0)
5110                 err = register_cpu_notifier(&conf->cpu_notify);
5111 #endif
5112         put_online_cpus();
5113
5114         return err;
5115 }
5116
5117 static struct r5conf *setup_conf(struct mddev *mddev)
5118 {
5119         struct r5conf *conf;
5120         int raid_disk, memory, max_disks;
5121         struct md_rdev *rdev;
5122         struct disk_info *disk;
5123         char pers_name[6];
5124
5125         if (mddev->new_level != 5
5126             && mddev->new_level != 4
5127             && mddev->new_level != 6) {
5128                 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5129                        mdname(mddev), mddev->new_level);
5130                 return ERR_PTR(-EIO);
5131         }
5132         if ((mddev->new_level == 5
5133              && !algorithm_valid_raid5(mddev->new_layout)) ||
5134             (mddev->new_level == 6
5135              && !algorithm_valid_raid6(mddev->new_layout))) {
5136                 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5137                        mdname(mddev), mddev->new_layout);
5138                 return ERR_PTR(-EIO);
5139         }
5140         if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5141                 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5142                        mdname(mddev), mddev->raid_disks);
5143                 return ERR_PTR(-EINVAL);
5144         }
5145
5146         if (!mddev->new_chunk_sectors ||
5147             (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5148             !is_power_of_2(mddev->new_chunk_sectors)) {
5149                 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5150                        mdname(mddev), mddev->new_chunk_sectors << 9);
5151                 return ERR_PTR(-EINVAL);
5152         }
5153
5154         conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5155         if (conf == NULL)
5156                 goto abort;
5157         spin_lock_init(&conf->device_lock);
5158         init_waitqueue_head(&conf->wait_for_stripe);
5159         init_waitqueue_head(&conf->wait_for_overlap);
5160         INIT_LIST_HEAD(&conf->handle_list);
5161         INIT_LIST_HEAD(&conf->hold_list);
5162         INIT_LIST_HEAD(&conf->delayed_list);
5163         INIT_LIST_HEAD(&conf->bitmap_list);
5164         INIT_LIST_HEAD(&conf->inactive_list);
5165         atomic_set(&conf->active_stripes, 0);
5166         atomic_set(&conf->preread_active_stripes, 0);
5167         atomic_set(&conf->active_aligned_reads, 0);
5168         conf->bypass_threshold = BYPASS_THRESHOLD;
5169         conf->recovery_disabled = mddev->recovery_disabled - 1;
5170
5171         conf->raid_disks = mddev->raid_disks;
5172         if (mddev->reshape_position == MaxSector)
5173                 conf->previous_raid_disks = mddev->raid_disks;
5174         else
5175                 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5176         max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5177         conf->scribble_len = scribble_len(max_disks);
5178
5179         conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5180                               GFP_KERNEL);
5181         if (!conf->disks)
5182                 goto abort;
5183
5184         conf->mddev = mddev;
5185
5186         if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5187                 goto abort;
5188
5189         conf->level = mddev->new_level;
5190         if (raid5_alloc_percpu(conf) != 0)
5191                 goto abort;
5192
5193         pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5194
5195         rdev_for_each(rdev, mddev) {
5196                 raid_disk = rdev->raid_disk;
5197                 if (raid_disk >= max_disks
5198                     || raid_disk < 0)
5199                         continue;
5200                 disk = conf->disks + raid_disk;
5201
5202                 if (test_bit(Replacement, &rdev->flags)) {
5203                         if (disk->replacement)
5204                                 goto abort;
5205                         disk->replacement = rdev;
5206                 } else {
5207                         if (disk->rdev)
5208                                 goto abort;
5209                         disk->rdev = rdev;
5210                 }
5211
5212                 if (test_bit(In_sync, &rdev->flags)) {
5213                         char b[BDEVNAME_SIZE];
5214                         printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5215                                " disk %d\n",
5216                                mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5217                 } else if (rdev->saved_raid_disk != raid_disk)
5218                         /* Cannot rely on bitmap to complete recovery */
5219                         conf->fullsync = 1;
5220         }
5221
5222         conf->chunk_sectors = mddev->new_chunk_sectors;
5223         conf->level = mddev->new_level;
5224         if (conf->level == 6)
5225                 conf->max_degraded = 2;
5226         else
5227                 conf->max_degraded = 1;
5228         conf->algorithm = mddev->new_layout;
5229         conf->max_nr_stripes = NR_STRIPES;
5230         conf->reshape_progress = mddev->reshape_position;
5231         if (conf->reshape_progress != MaxSector) {
5232                 conf->prev_chunk_sectors = mddev->chunk_sectors;
5233                 conf->prev_algo = mddev->layout;
5234         }
5235
5236         memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5237                  max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5238         if (grow_stripes(conf, conf->max_nr_stripes)) {
5239                 printk(KERN_ERR
5240                        "md/raid:%s: couldn't allocate %dkB for buffers\n",
5241                        mdname(mddev), memory);
5242                 goto abort;
5243         } else
5244                 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5245                        mdname(mddev), memory);
5246
5247         sprintf(pers_name, "raid%d", mddev->new_level);
5248         conf->thread = md_register_thread(raid5d, mddev, pers_name);
5249         if (!conf->thread) {
5250                 printk(KERN_ERR
5251                        "md/raid:%s: couldn't allocate thread.\n",
5252                        mdname(mddev));
5253                 goto abort;
5254         }
5255
5256         return conf;
5257
5258  abort:
5259         if (conf) {
5260                 free_conf(conf);
5261                 return ERR_PTR(-EIO);
5262         } else
5263                 return ERR_PTR(-ENOMEM);
5264 }
5265
5266
5267 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5268 {
5269         switch (algo) {
5270         case ALGORITHM_PARITY_0:
5271                 if (raid_disk < max_degraded)
5272                         return 1;
5273                 break;
5274         case ALGORITHM_PARITY_N:
5275                 if (raid_disk >= raid_disks - max_degraded)
5276                         return 1;
5277                 break;
5278         case ALGORITHM_PARITY_0_6:
5279                 if (raid_disk == 0 ||
5280                     raid_disk == raid_disks - 1)
5281                         return 1;
5282                 break;
5283         case ALGORITHM_LEFT_ASYMMETRIC_6:
5284         case ALGORITHM_RIGHT_ASYMMETRIC_6:
5285         case ALGORITHM_LEFT_SYMMETRIC_6:
5286         case ALGORITHM_RIGHT_SYMMETRIC_6:
5287                 if (raid_disk == raid_disks - 1)
5288                         return 1;
5289         }
5290         return 0;
5291 }
5292
5293 static int run(struct mddev *mddev)
5294 {
5295         struct r5conf *conf;
5296         int working_disks = 0;
5297         int dirty_parity_disks = 0;
5298         struct md_rdev *rdev;
5299         sector_t reshape_offset = 0;
5300         int i;
5301         long long min_offset_diff = 0;
5302         int first = 1;
5303
5304         if (mddev->recovery_cp != MaxSector)
5305                 printk(KERN_NOTICE "md/raid:%s: not clean"
5306                        " -- starting background reconstruction\n",
5307                        mdname(mddev));
5308
5309         rdev_for_each(rdev, mddev) {
5310                 long long diff;
5311                 if (rdev->raid_disk < 0)
5312                         continue;
5313                 diff = (rdev->new_data_offset - rdev->data_offset);
5314                 if (first) {
5315                         min_offset_diff = diff;
5316                         first = 0;
5317                 } else if (mddev->reshape_backwards &&
5318                          diff < min_offset_diff)
5319                         min_offset_diff = diff;
5320                 else if (!mddev->reshape_backwards &&
5321                          diff > min_offset_diff)
5322                         min_offset_diff = diff;
5323         }
5324
5325         if (mddev->reshape_position != MaxSector) {
5326                 /* Check that we can continue the reshape.
5327                  * Difficulties arise if the stripe we would write to
5328                  * next is at or after the stripe we would read from next.
5329                  * For a reshape that changes the number of devices, this
5330                  * is only possible for a very short time, and mdadm makes
5331                  * sure that time appears to have past before assembling
5332                  * the array.  So we fail if that time hasn't passed.
5333                  * For a reshape that keeps the number of devices the same
5334                  * mdadm must be monitoring the reshape can keeping the
5335                  * critical areas read-only and backed up.  It will start
5336                  * the array in read-only mode, so we check for that.
5337                  */
5338                 sector_t here_new, here_old;
5339                 int old_disks;
5340                 int max_degraded = (mddev->level == 6 ? 2 : 1);
5341
5342                 if (mddev->new_level != mddev->level) {
5343                         printk(KERN_ERR "md/raid:%s: unsupported reshape "
5344                                "required - aborting.\n",
5345                                mdname(mddev));
5346                         return -EINVAL;
5347                 }
5348                 old_disks = mddev->raid_disks - mddev->delta_disks;
5349                 /* reshape_position must be on a new-stripe boundary, and one
5350                  * further up in new geometry must map after here in old
5351                  * geometry.
5352                  */
5353                 here_new = mddev->reshape_position;
5354                 if (sector_div(here_new, mddev->new_chunk_sectors *
5355                                (mddev->raid_disks - max_degraded))) {
5356                         printk(KERN_ERR "md/raid:%s: reshape_position not "
5357                                "on a stripe boundary\n", mdname(mddev));
5358                         return -EINVAL;
5359                 }
5360                 reshape_offset = here_new * mddev->new_chunk_sectors;
5361                 /* here_new is the stripe we will write to */
5362                 here_old = mddev->reshape_position;
5363                 sector_div(here_old, mddev->chunk_sectors *
5364                            (old_disks-max_degraded));
5365                 /* here_old is the first stripe that we might need to read
5366                  * from */
5367                 if (mddev->delta_disks == 0) {
5368                         if ((here_new * mddev->new_chunk_sectors !=
5369                              here_old * mddev->chunk_sectors)) {
5370                                 printk(KERN_ERR "md/raid:%s: reshape position is"
5371                                        " confused - aborting\n", mdname(mddev));
5372                                 return -EINVAL;
5373                         }
5374                         /* We cannot be sure it is safe to start an in-place
5375                          * reshape.  It is only safe if user-space is monitoring
5376                          * and taking constant backups.
5377                          * mdadm always starts a situation like this in
5378                          * readonly mode so it can take control before
5379                          * allowing any writes.  So just check for that.
5380                          */
5381                         if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5382                             abs(min_offset_diff) >= mddev->new_chunk_sectors)
5383                                 /* not really in-place - so OK */;
5384                         else if (mddev->ro == 0) {
5385                                 printk(KERN_ERR "md/raid:%s: in-place reshape "
5386                                        "must be started in read-only mode "
5387                                        "- aborting\n",
5388                                        mdname(mddev));
5389                                 return -EINVAL;
5390                         }
5391                 } else if (mddev->reshape_backwards
5392                     ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5393                        here_old * mddev->chunk_sectors)
5394                     : (here_new * mddev->new_chunk_sectors >=
5395                        here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5396                         /* Reading from the same stripe as writing to - bad */
5397                         printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5398                                "auto-recovery - aborting.\n",
5399                                mdname(mddev));
5400                         return -EINVAL;
5401                 }
5402                 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5403                        mdname(mddev));
5404                 /* OK, we should be able to continue; */
5405         } else {
5406                 BUG_ON(mddev->level != mddev->new_level);
5407                 BUG_ON(mddev->layout != mddev->new_layout);
5408                 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5409                 BUG_ON(mddev->delta_disks != 0);
5410         }
5411
5412         if (mddev->private == NULL)
5413                 conf = setup_conf(mddev);
5414         else
5415                 conf = mddev->private;
5416
5417         if (IS_ERR(conf))
5418                 return PTR_ERR(conf);
5419
5420         conf->min_offset_diff = min_offset_diff;
5421         mddev->thread = conf->thread;
5422         conf->thread = NULL;
5423         mddev->private = conf;
5424
5425         for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5426              i++) {
5427                 rdev = conf->disks[i].rdev;
5428                 if (!rdev && conf->disks[i].replacement) {
5429                         /* The replacement is all we have yet */
5430                         rdev = conf->disks[i].replacement;
5431                         conf->disks[i].replacement = NULL;
5432                         clear_bit(Replacement, &rdev->flags);
5433                         conf->disks[i].rdev = rdev;
5434                 }
5435                 if (!rdev)
5436                         continue;
5437                 if (conf->disks[i].replacement &&
5438                     conf->reshape_progress != MaxSector) {
5439                         /* replacements and reshape simply do not mix. */
5440                         printk(KERN_ERR "md: cannot handle concurrent "
5441                                "replacement and reshape.\n");
5442                         goto abort;
5443                 }
5444                 if (test_bit(In_sync, &rdev->flags)) {
5445                         working_disks++;
5446                         continue;
5447                 }
5448                 /* This disc is not fully in-sync.  However if it
5449                  * just stored parity (beyond the recovery_offset),
5450                  * when we don't need to be concerned about the
5451                  * array being dirty.
5452                  * When reshape goes 'backwards', we never have
5453                  * partially completed devices, so we only need
5454                  * to worry about reshape going forwards.
5455                  */
5456                 /* Hack because v0.91 doesn't store recovery_offset properly. */
5457                 if (mddev->major_version == 0 &&
5458                     mddev->minor_version > 90)
5459                         rdev->recovery_offset = reshape_offset;
5460
5461                 if (rdev->recovery_offset < reshape_offset) {
5462                         /* We need to check old and new layout */
5463                         if (!only_parity(rdev->raid_disk,
5464                                          conf->algorithm,
5465                                          conf->raid_disks,
5466                                          conf->max_degraded))
5467                                 continue;
5468                 }
5469                 if (!only_parity(rdev->raid_disk,
5470                                  conf->prev_algo,
5471                                  conf->previous_raid_disks,
5472                                  conf->max_degraded))
5473                         continue;
5474                 dirty_parity_disks++;
5475         }
5476
5477         /*
5478          * 0 for a fully functional array, 1 or 2 for a degraded array.
5479          */
5480         mddev->degraded = calc_degraded(conf);
5481
5482         if (has_failed(conf)) {
5483                 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5484                         " (%d/%d failed)\n",
5485                         mdname(mddev), mddev->degraded, conf->raid_disks);
5486                 goto abort;
5487         }
5488
5489         /* device size must be a multiple of chunk size */
5490         mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5491         mddev->resync_max_sectors = mddev->dev_sectors;
5492
5493         if (mddev->degraded > dirty_parity_disks &&
5494             mddev->recovery_cp != MaxSector) {
5495                 if (mddev->ok_start_degraded)
5496                         printk(KERN_WARNING
5497                                "md/raid:%s: starting dirty degraded array"
5498                                " - data corruption possible.\n",
5499                                mdname(mddev));
5500                 else {
5501                         printk(KERN_ERR
5502                                "md/raid:%s: cannot start dirty degraded array.\n",
5503                                mdname(mddev));
5504                         goto abort;
5505                 }
5506         }
5507
5508         if (mddev->degraded == 0)
5509                 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5510                        " devices, algorithm %d\n", mdname(mddev), conf->level,
5511                        mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5512                        mddev->new_layout);
5513         else
5514                 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5515                        " out of %d devices, algorithm %d\n",
5516                        mdname(mddev), conf->level,
5517                        mddev->raid_disks - mddev->degraded,
5518                        mddev->raid_disks, mddev->new_layout);
5519
5520         print_raid5_conf(conf);
5521
5522         if (conf->reshape_progress != MaxSector) {
5523                 conf->reshape_safe = conf->reshape_progress;
5524                 atomic_set(&conf->reshape_stripes, 0);
5525                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5526                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5527                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5528                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5529                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5530                                                         "reshape");
5531         }
5532
5533
5534         /* Ok, everything is just fine now */
5535         if (mddev->to_remove == &raid5_attrs_group)
5536                 mddev->to_remove = NULL;
5537         else if (mddev->kobj.sd &&
5538             sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5539                 printk(KERN_WARNING
5540                        "raid5: failed to create sysfs attributes for %s\n",
5541                        mdname(mddev));
5542         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5543
5544         if (mddev->queue) {
5545                 int chunk_size;
5546                 bool discard_supported = true;
5547                 /* read-ahead size must cover two whole stripes, which
5548                  * is 2 * (datadisks) * chunksize where 'n' is the
5549                  * number of raid devices
5550                  */
5551                 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5552                 int stripe = data_disks *
5553                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5554                 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5555                         mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5556
5557                 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5558
5559                 mddev->queue->backing_dev_info.congested_data = mddev;
5560                 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5561
5562                 chunk_size = mddev->chunk_sectors << 9;
5563                 blk_queue_io_min(mddev->queue, chunk_size);
5564                 blk_queue_io_opt(mddev->queue, chunk_size *
5565                                  (conf->raid_disks - conf->max_degraded));
5566                 /*
5567                  * We can only discard a whole stripe. It doesn't make sense to
5568                  * discard data disk but write parity disk
5569                  */
5570                 stripe = stripe * PAGE_SIZE;
5571                 /* Round up to power of 2, as discard handling
5572                  * currently assumes that */
5573                 while ((stripe-1) & stripe)
5574                         stripe = (stripe | (stripe-1)) + 1;
5575                 mddev->queue->limits.discard_alignment = stripe;
5576                 mddev->queue->limits.discard_granularity = stripe;
5577                 /*
5578                  * unaligned part of discard request will be ignored, so can't
5579                  * guarantee discard_zerors_data
5580                  */
5581                 mddev->queue->limits.discard_zeroes_data = 0;
5582
5583                 rdev_for_each(rdev, mddev) {
5584                         disk_stack_limits(mddev->gendisk, rdev->bdev,
5585                                           rdev->data_offset << 9);
5586                         disk_stack_limits(mddev->gendisk, rdev->bdev,
5587                                           rdev->new_data_offset << 9);
5588                         /*
5589                          * discard_zeroes_data is required, otherwise data
5590                          * could be lost. Consider a scenario: discard a stripe
5591                          * (the stripe could be inconsistent if
5592                          * discard_zeroes_data is 0); write one disk of the
5593                          * stripe (the stripe could be inconsistent again
5594                          * depending on which disks are used to calculate
5595                          * parity); the disk is broken; The stripe data of this
5596                          * disk is lost.
5597                          */
5598                         if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5599                             !bdev_get_queue(rdev->bdev)->
5600                                                 limits.discard_zeroes_data)
5601                                 discard_supported = false;
5602                 }
5603
5604                 if (discard_supported &&
5605                    mddev->queue->limits.max_discard_sectors >= stripe &&
5606                    mddev->queue->limits.discard_granularity >= stripe)
5607                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5608                                                 mddev->queue);
5609                 else
5610                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5611                                                 mddev->queue);
5612         }
5613
5614         return 0;
5615 abort:
5616         md_unregister_thread(&mddev->thread);
5617         print_raid5_conf(conf);
5618         free_conf(conf);
5619         mddev->private = NULL;
5620         printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5621         return -EIO;
5622 }
5623
5624 static int stop(struct mddev *mddev)
5625 {
5626         struct r5conf *conf = mddev->private;
5627
5628         md_unregister_thread(&mddev->thread);
5629         if (mddev->queue)
5630                 mddev->queue->backing_dev_info.congested_fn = NULL;
5631         free_conf(conf);
5632         mddev->private = NULL;
5633         mddev->to_remove = &raid5_attrs_group;
5634         return 0;
5635 }
5636
5637 static void status(struct seq_file *seq, struct mddev *mddev)
5638 {
5639         struct r5conf *conf = mddev->private;
5640         int i;
5641
5642         seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5643                 mddev->chunk_sectors / 2, mddev->layout);
5644         seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5645         for (i = 0; i < conf->raid_disks; i++)
5646                 seq_printf (seq, "%s",
5647                                conf->disks[i].rdev &&
5648                                test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5649         seq_printf (seq, "]");
5650 }
5651
5652 static void print_raid5_conf (struct r5conf *conf)
5653 {
5654         int i;
5655         struct disk_info *tmp;
5656
5657         printk(KERN_DEBUG "RAID conf printout:\n");
5658         if (!conf) {
5659                 printk("(conf==NULL)\n");
5660                 return;
5661         }
5662         printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5663                conf->raid_disks,
5664                conf->raid_disks - conf->mddev->degraded);
5665
5666         for (i = 0; i < conf->raid_disks; i++) {
5667                 char b[BDEVNAME_SIZE];
5668                 tmp = conf->disks + i;
5669                 if (tmp->rdev)
5670                         printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5671                                i, !test_bit(Faulty, &tmp->rdev->flags),
5672                                bdevname(tmp->rdev->bdev, b));
5673         }
5674 }
5675
5676 static int raid5_spare_active(struct mddev *mddev)
5677 {
5678         int i;
5679         struct r5conf *conf = mddev->private;
5680         struct disk_info *tmp;
5681         int count = 0;
5682         unsigned long flags;
5683
5684         for (i = 0; i < conf->raid_disks; i++) {
5685                 tmp = conf->disks + i;
5686                 if (tmp->replacement
5687                     && tmp->replacement->recovery_offset == MaxSector
5688                     && !test_bit(Faulty, &tmp->replacement->flags)
5689                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5690                         /* Replacement has just become active. */
5691                         if (!tmp->rdev
5692                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5693                                 count++;
5694                         if (tmp->rdev) {
5695                                 /* Replaced device not technically faulty,
5696                                  * but we need to be sure it gets removed
5697                                  * and never re-added.
5698                                  */
5699                                 set_bit(Faulty, &tmp->rdev->flags);
5700                                 sysfs_notify_dirent_safe(
5701                                         tmp->rdev->sysfs_state);
5702                         }
5703                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5704                 } else if (tmp->rdev
5705                     && tmp->rdev->recovery_offset == MaxSector
5706                     && !test_bit(Faulty, &tmp->rdev->flags)
5707                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5708                         count++;
5709                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5710                 }
5711         }
5712         spin_lock_irqsave(&conf->device_lock, flags);
5713         mddev->degraded = calc_degraded(conf);
5714         spin_unlock_irqrestore(&conf->device_lock, flags);
5715         print_raid5_conf(conf);
5716         return count;
5717 }
5718
5719 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5720 {
5721         struct r5conf *conf = mddev->private;
5722         int err = 0;
5723         int number = rdev->raid_disk;
5724         struct md_rdev **rdevp;
5725         struct disk_info *p = conf->disks + number;
5726
5727         print_raid5_conf(conf);
5728         if (rdev == p->rdev)
5729                 rdevp = &p->rdev;
5730         else if (rdev == p->replacement)
5731                 rdevp = &p->replacement;
5732         else
5733                 return 0;
5734
5735         if (number >= conf->raid_disks &&
5736             conf->reshape_progress == MaxSector)
5737                 clear_bit(In_sync, &rdev->flags);
5738
5739         if (test_bit(In_sync, &rdev->flags) ||
5740             atomic_read(&rdev->nr_pending)) {
5741                 err = -EBUSY;
5742                 goto abort;
5743         }
5744         /* Only remove non-faulty devices if recovery
5745          * isn't possible.
5746          */
5747         if (!test_bit(Faulty, &rdev->flags) &&
5748             mddev->recovery_disabled != conf->recovery_disabled &&
5749             !has_failed(conf) &&
5750             (!p->replacement || p->replacement == rdev) &&
5751             number < conf->raid_disks) {
5752                 err = -EBUSY;
5753                 goto abort;
5754         }
5755         *rdevp = NULL;
5756         synchronize_rcu();
5757         if (atomic_read(&rdev->nr_pending)) {
5758                 /* lost the race, try later */
5759                 err = -EBUSY;
5760                 *rdevp = rdev;
5761         } else if (p->replacement) {
5762                 /* We must have just cleared 'rdev' */
5763                 p->rdev = p->replacement;
5764                 clear_bit(Replacement, &p->replacement->flags);
5765                 smp_mb(); /* Make sure other CPUs may see both as identical
5766                            * but will never see neither - if they are careful
5767                            */
5768                 p->replacement = NULL;
5769                 clear_bit(WantReplacement, &rdev->flags);
5770         } else
5771                 /* We might have just removed the Replacement as faulty-
5772                  * clear the bit just in case
5773                  */
5774                 clear_bit(WantReplacement, &rdev->flags);
5775 abort:
5776
5777         print_raid5_conf(conf);
5778         return err;
5779 }
5780
5781 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5782 {
5783         struct r5conf *conf = mddev->private;
5784         int err = -EEXIST;
5785         int disk;
5786         struct disk_info *p;
5787         int first = 0;
5788         int last = conf->raid_disks - 1;
5789
5790         if (mddev->recovery_disabled == conf->recovery_disabled)
5791                 return -EBUSY;
5792
5793         if (rdev->saved_raid_disk < 0 && has_failed(conf))
5794                 /* no point adding a device */
5795                 return -EINVAL;
5796
5797         if (rdev->raid_disk >= 0)
5798                 first = last = rdev->raid_disk;
5799
5800         /*
5801          * find the disk ... but prefer rdev->saved_raid_disk
5802          * if possible.
5803          */
5804         if (rdev->saved_raid_disk >= 0 &&
5805             rdev->saved_raid_disk >= first &&
5806             conf->disks[rdev->saved_raid_disk].rdev == NULL)
5807                 first = rdev->saved_raid_disk;
5808
5809         for (disk = first; disk <= last; disk++) {
5810                 p = conf->disks + disk;
5811                 if (p->rdev == NULL) {
5812                         clear_bit(In_sync, &rdev->flags);
5813                         rdev->raid_disk = disk;
5814                         err = 0;
5815                         if (rdev->saved_raid_disk != disk)
5816                                 conf->fullsync = 1;
5817                         rcu_assign_pointer(p->rdev, rdev);
5818                         goto out;
5819                 }
5820         }
5821         for (disk = first; disk <= last; disk++) {
5822                 p = conf->disks + disk;
5823                 if (test_bit(WantReplacement, &p->rdev->flags) &&
5824                     p->replacement == NULL) {
5825                         clear_bit(In_sync, &rdev->flags);
5826                         set_bit(Replacement, &rdev->flags);
5827                         rdev->raid_disk = disk;
5828                         err = 0;
5829                         conf->fullsync = 1;
5830                         rcu_assign_pointer(p->replacement, rdev);
5831                         break;
5832                 }
5833         }
5834 out:
5835         print_raid5_conf(conf);
5836         return err;
5837 }
5838
5839 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5840 {
5841         /* no resync is happening, and there is enough space
5842          * on all devices, so we can resize.
5843          * We need to make sure resync covers any new space.
5844          * If the array is shrinking we should possibly wait until
5845          * any io in the removed space completes, but it hardly seems
5846          * worth it.
5847          */
5848         sector_t newsize;
5849         sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5850         newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5851         if (mddev->external_size &&
5852             mddev->array_sectors > newsize)
5853                 return -EINVAL;
5854         if (mddev->bitmap) {
5855                 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5856                 if (ret)
5857                         return ret;
5858         }
5859         md_set_array_sectors(mddev, newsize);
5860         set_capacity(mddev->gendisk, mddev->array_sectors);
5861         revalidate_disk(mddev->gendisk);
5862         if (sectors > mddev->dev_sectors &&
5863             mddev->recovery_cp > mddev->dev_sectors) {
5864                 mddev->recovery_cp = mddev->dev_sectors;
5865                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5866         }
5867         mddev->dev_sectors = sectors;
5868         mddev->resync_max_sectors = sectors;
5869         return 0;
5870 }
5871
5872 static int check_stripe_cache(struct mddev *mddev)
5873 {
5874         /* Can only proceed if there are plenty of stripe_heads.
5875          * We need a minimum of one full stripe,, and for sensible progress
5876          * it is best to have about 4 times that.
5877          * If we require 4 times, then the default 256 4K stripe_heads will
5878          * allow for chunk sizes up to 256K, which is probably OK.
5879          * If the chunk size is greater, user-space should request more
5880          * stripe_heads first.
5881          */
5882         struct r5conf *conf = mddev->private;
5883         if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5884             > conf->max_nr_stripes ||
5885             ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5886             > conf->max_nr_stripes) {
5887                 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
5888                        mdname(mddev),
5889                        ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5890                         / STRIPE_SIZE)*4);
5891                 return 0;
5892         }
5893         return 1;
5894 }
5895
5896 static int check_reshape(struct mddev *mddev)
5897 {
5898         struct r5conf *conf = mddev->private;
5899
5900         if (mddev->delta_disks == 0 &&
5901             mddev->new_layout == mddev->layout &&
5902             mddev->new_chunk_sectors == mddev->chunk_sectors)
5903                 return 0; /* nothing to do */
5904         if (has_failed(conf))
5905                 return -EINVAL;
5906         if (mddev->delta_disks < 0) {
5907                 /* We might be able to shrink, but the devices must
5908                  * be made bigger first.
5909                  * For raid6, 4 is the minimum size.
5910                  * Otherwise 2 is the minimum
5911                  */
5912                 int min = 2;
5913                 if (mddev->level == 6)
5914                         min = 4;
5915                 if (mddev->raid_disks + mddev->delta_disks < min)
5916                         return -EINVAL;
5917         }
5918
5919         if (!check_stripe_cache(mddev))
5920                 return -ENOSPC;
5921
5922         return resize_stripes(conf, (conf->previous_raid_disks
5923                                      + mddev->delta_disks));
5924 }
5925
5926 static int raid5_start_reshape(struct mddev *mddev)
5927 {
5928         struct r5conf *conf = mddev->private;
5929         struct md_rdev *rdev;
5930         int spares = 0;
5931         unsigned long flags;
5932
5933         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5934                 return -EBUSY;
5935
5936         if (!check_stripe_cache(mddev))
5937                 return -ENOSPC;
5938
5939         if (has_failed(conf))
5940                 return -EINVAL;
5941
5942         rdev_for_each(rdev, mddev) {
5943                 if (!test_bit(In_sync, &rdev->flags)
5944                     && !test_bit(Faulty, &rdev->flags))
5945                         spares++;
5946         }
5947
5948         if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5949                 /* Not enough devices even to make a degraded array
5950                  * of that size
5951                  */
5952                 return -EINVAL;
5953
5954         /* Refuse to reduce size of the array.  Any reductions in
5955          * array size must be through explicit setting of array_size
5956          * attribute.
5957          */
5958         if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5959             < mddev->array_sectors) {
5960                 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5961                        "before number of disks\n", mdname(mddev));
5962                 return -EINVAL;
5963         }
5964
5965         atomic_set(&conf->reshape_stripes, 0);
5966         spin_lock_irq(&conf->device_lock);
5967         conf->previous_raid_disks = conf->raid_disks;
5968         conf->raid_disks += mddev->delta_disks;
5969         conf->prev_chunk_sectors = conf->chunk_sectors;
5970         conf->chunk_sectors = mddev->new_chunk_sectors;
5971         conf->prev_algo = conf->algorithm;
5972         conf->algorithm = mddev->new_layout;
5973         conf->generation++;
5974         /* Code that selects data_offset needs to see the generation update
5975          * if reshape_progress has been set - so a memory barrier needed.
5976          */
5977         smp_mb();
5978         if (mddev->reshape_backwards)
5979                 conf->reshape_progress = raid5_size(mddev, 0, 0);
5980         else
5981                 conf->reshape_progress = 0;
5982         conf->reshape_safe = conf->reshape_progress;
5983         spin_unlock_irq(&conf->device_lock);
5984
5985         /* Add some new drives, as many as will fit.
5986          * We know there are enough to make the newly sized array work.
5987          * Don't add devices if we are reducing the number of
5988          * devices in the array.  This is because it is not possible
5989          * to correctly record the "partially reconstructed" state of
5990          * such devices during the reshape and confusion could result.
5991          */
5992         if (mddev->delta_disks >= 0) {
5993                 rdev_for_each(rdev, mddev)
5994                         if (rdev->raid_disk < 0 &&
5995                             !test_bit(Faulty, &rdev->flags)) {
5996                                 if (raid5_add_disk(mddev, rdev) == 0) {
5997                                         if (rdev->raid_disk
5998                                             >= conf->previous_raid_disks)
5999                                                 set_bit(In_sync, &rdev->flags);
6000                                         else
6001                                                 rdev->recovery_offset = 0;
6002
6003                                         if (sysfs_link_rdev(mddev, rdev))
6004                                                 /* Failure here is OK */;
6005                                 }
6006                         } else if (rdev->raid_disk >= conf->previous_raid_disks
6007                                    && !test_bit(Faulty, &rdev->flags)) {
6008                                 /* This is a spare that was manually added */
6009                                 set_bit(In_sync, &rdev->flags);
6010                         }
6011
6012                 /* When a reshape changes the number of devices,
6013                  * ->degraded is measured against the larger of the
6014                  * pre and post number of devices.
6015                  */
6016                 spin_lock_irqsave(&conf->device_lock, flags);
6017                 mddev->degraded = calc_degraded(conf);
6018                 spin_unlock_irqrestore(&conf->device_lock, flags);
6019         }
6020         mddev->raid_disks = conf->raid_disks;
6021         mddev->reshape_position = conf->reshape_progress;
6022         set_bit(MD_CHANGE_DEVS, &mddev->flags);
6023
6024         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6025         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6026         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6027         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6028         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6029                                                 "reshape");
6030         if (!mddev->sync_thread) {
6031                 mddev->recovery = 0;
6032                 spin_lock_irq(&conf->device_lock);
6033                 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6034                 rdev_for_each(rdev, mddev)
6035                         rdev->new_data_offset = rdev->data_offset;
6036                 smp_wmb();
6037                 conf->reshape_progress = MaxSector;
6038                 mddev->reshape_position = MaxSector;
6039                 spin_unlock_irq(&conf->device_lock);
6040                 return -EAGAIN;
6041         }
6042         conf->reshape_checkpoint = jiffies;
6043         md_wakeup_thread(mddev->sync_thread);
6044         md_new_event(mddev);
6045         return 0;
6046 }
6047
6048 /* This is called from the reshape thread and should make any
6049  * changes needed in 'conf'
6050  */
6051 static void end_reshape(struct r5conf *conf)
6052 {
6053
6054         if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6055                 struct md_rdev *rdev;
6056
6057                 spin_lock_irq(&conf->device_lock);
6058                 conf->previous_raid_disks = conf->raid_disks;
6059                 rdev_for_each(rdev, conf->mddev)
6060                         rdev->data_offset = rdev->new_data_offset;
6061                 smp_wmb();
6062                 conf->reshape_progress = MaxSector;
6063                 spin_unlock_irq(&conf->device_lock);
6064                 wake_up(&conf->wait_for_overlap);
6065
6066                 /* read-ahead size must cover two whole stripes, which is
6067                  * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6068                  */
6069                 if (conf->mddev->queue) {
6070                         int data_disks = conf->raid_disks - conf->max_degraded;
6071                         int stripe = data_disks * ((conf->chunk_sectors << 9)
6072                                                    / PAGE_SIZE);
6073                         if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6074                                 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6075                 }
6076         }
6077 }
6078
6079 /* This is called from the raid5d thread with mddev_lock held.
6080  * It makes config changes to the device.
6081  */
6082 static void raid5_finish_reshape(struct mddev *mddev)
6083 {
6084         struct r5conf *conf = mddev->private;
6085
6086         if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6087
6088                 if (mddev->delta_disks > 0) {
6089                         md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6090                         set_capacity(mddev->gendisk, mddev->array_sectors);
6091                         revalidate_disk(mddev->gendisk);
6092                 } else {
6093                         int d;
6094                         spin_lock_irq(&conf->device_lock);
6095                         mddev->degraded = calc_degraded(conf);
6096                         spin_unlock_irq(&conf->device_lock);
6097                         for (d = conf->raid_disks ;
6098                              d < conf->raid_disks - mddev->delta_disks;
6099                              d++) {
6100                                 struct md_rdev *rdev = conf->disks[d].rdev;
6101                                 if (rdev)
6102                                         clear_bit(In_sync, &rdev->flags);
6103                                 rdev = conf->disks[d].replacement;
6104                                 if (rdev)
6105                                         clear_bit(In_sync, &rdev->flags);
6106                         }
6107                 }
6108                 mddev->layout = conf->algorithm;
6109                 mddev->chunk_sectors = conf->chunk_sectors;
6110                 mddev->reshape_position = MaxSector;
6111                 mddev->delta_disks = 0;
6112                 mddev->reshape_backwards = 0;
6113         }
6114 }
6115
6116 static void raid5_quiesce(struct mddev *mddev, int state)
6117 {
6118         struct r5conf *conf = mddev->private;
6119
6120         switch(state) {
6121         case 2: /* resume for a suspend */
6122                 wake_up(&conf->wait_for_overlap);
6123                 break;
6124
6125         case 1: /* stop all writes */
6126                 spin_lock_irq(&conf->device_lock);
6127                 /* '2' tells resync/reshape to pause so that all
6128                  * active stripes can drain
6129                  */
6130                 conf->quiesce = 2;
6131                 wait_event_lock_irq(conf->wait_for_stripe,
6132                                     atomic_read(&conf->active_stripes) == 0 &&
6133                                     atomic_read(&conf->active_aligned_reads) == 0,
6134                                     conf->device_lock);
6135                 conf->quiesce = 1;
6136                 spin_unlock_irq(&conf->device_lock);
6137                 /* allow reshape to continue */
6138                 wake_up(&conf->wait_for_overlap);
6139                 break;
6140
6141         case 0: /* re-enable writes */
6142                 spin_lock_irq(&conf->device_lock);
6143                 conf->quiesce = 0;
6144                 wake_up(&conf->wait_for_stripe);
6145                 wake_up(&conf->wait_for_overlap);
6146                 spin_unlock_irq(&conf->device_lock);
6147                 break;
6148         }
6149 }
6150
6151
6152 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6153 {
6154         struct r0conf *raid0_conf = mddev->private;
6155         sector_t sectors;
6156
6157         /* for raid0 takeover only one zone is supported */
6158         if (raid0_conf->nr_strip_zones > 1) {
6159                 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6160                        mdname(mddev));
6161                 return ERR_PTR(-EINVAL);
6162         }
6163
6164         sectors = raid0_conf->strip_zone[0].zone_end;
6165         sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6166         mddev->dev_sectors = sectors;
6167         mddev->new_level = level;
6168         mddev->new_layout = ALGORITHM_PARITY_N;
6169         mddev->new_chunk_sectors = mddev->chunk_sectors;
6170         mddev->raid_disks += 1;
6171         mddev->delta_disks = 1;
6172         /* make sure it will be not marked as dirty */
6173         mddev->recovery_cp = MaxSector;
6174
6175         return setup_conf(mddev);
6176 }
6177
6178
6179 static void *raid5_takeover_raid1(struct mddev *mddev)
6180 {
6181         int chunksect;
6182
6183         if (mddev->raid_disks != 2 ||
6184             mddev->degraded > 1)
6185                 return ERR_PTR(-EINVAL);
6186
6187         /* Should check if there are write-behind devices? */
6188
6189         chunksect = 64*2; /* 64K by default */
6190
6191         /* The array must be an exact multiple of chunksize */
6192         while (chunksect && (mddev->array_sectors & (chunksect-1)))
6193                 chunksect >>= 1;
6194
6195         if ((chunksect<<9) < STRIPE_SIZE)
6196                 /* array size does not allow a suitable chunk size */
6197                 return ERR_PTR(-EINVAL);
6198
6199         mddev->new_level = 5;
6200         mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6201         mddev->new_chunk_sectors = chunksect;
6202
6203         return setup_conf(mddev);
6204 }
6205
6206 static void *raid5_takeover_raid6(struct mddev *mddev)
6207 {
6208         int new_layout;
6209
6210         switch (mddev->layout) {
6211         case ALGORITHM_LEFT_ASYMMETRIC_6:
6212                 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6213                 break;
6214         case ALGORITHM_RIGHT_ASYMMETRIC_6:
6215                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6216                 break;
6217         case ALGORITHM_LEFT_SYMMETRIC_6:
6218                 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6219                 break;
6220         case ALGORITHM_RIGHT_SYMMETRIC_6:
6221                 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6222                 break;
6223         case ALGORITHM_PARITY_0_6:
6224                 new_layout = ALGORITHM_PARITY_0;
6225                 break;
6226         case ALGORITHM_PARITY_N:
6227                 new_layout = ALGORITHM_PARITY_N;
6228                 break;
6229         default:
6230                 return ERR_PTR(-EINVAL);
6231         }
6232         mddev->new_level = 5;
6233         mddev->new_layout = new_layout;
6234         mddev->delta_disks = -1;
6235         mddev->raid_disks -= 1;
6236         return setup_conf(mddev);
6237 }
6238
6239
6240 static int raid5_check_reshape(struct mddev *mddev)
6241 {
6242         /* For a 2-drive array, the layout and chunk size can be changed
6243          * immediately as not restriping is needed.
6244          * For larger arrays we record the new value - after validation
6245          * to be used by a reshape pass.
6246          */
6247         struct r5conf *conf = mddev->private;
6248         int new_chunk = mddev->new_chunk_sectors;
6249
6250         if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6251                 return -EINVAL;
6252         if (new_chunk > 0) {
6253                 if (!is_power_of_2(new_chunk))
6254                         return -EINVAL;
6255                 if (new_chunk < (PAGE_SIZE>>9))
6256                         return -EINVAL;
6257                 if (mddev->array_sectors & (new_chunk-1))
6258                         /* not factor of array size */
6259                         return -EINVAL;
6260         }
6261
6262         /* They look valid */
6263
6264         if (mddev->raid_disks == 2) {
6265                 /* can make the change immediately */
6266                 if (mddev->new_layout >= 0) {
6267                         conf->algorithm = mddev->new_layout;
6268                         mddev->layout = mddev->new_layout;
6269                 }
6270                 if (new_chunk > 0) {
6271                         conf->chunk_sectors = new_chunk ;
6272                         mddev->chunk_sectors = new_chunk;
6273                 }
6274                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6275                 md_wakeup_thread(mddev->thread);
6276         }
6277         return check_reshape(mddev);
6278 }
6279
6280 static int raid6_check_reshape(struct mddev *mddev)
6281 {
6282         int new_chunk = mddev->new_chunk_sectors;
6283
6284         if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6285                 return -EINVAL;
6286         if (new_chunk > 0) {
6287                 if (!is_power_of_2(new_chunk))
6288                         return -EINVAL;
6289                 if (new_chunk < (PAGE_SIZE >> 9))
6290                         return -EINVAL;
6291                 if (mddev->array_sectors & (new_chunk-1))
6292                         /* not factor of array size */
6293                         return -EINVAL;
6294         }
6295
6296         /* They look valid */
6297         return check_reshape(mddev);
6298 }
6299
6300 static void *raid5_takeover(struct mddev *mddev)
6301 {
6302         /* raid5 can take over:
6303          *  raid0 - if there is only one strip zone - make it a raid4 layout
6304          *  raid1 - if there are two drives.  We need to know the chunk size
6305          *  raid4 - trivial - just use a raid4 layout.
6306          *  raid6 - Providing it is a *_6 layout
6307          */
6308         if (mddev->level == 0)
6309                 return raid45_takeover_raid0(mddev, 5);
6310         if (mddev->level == 1)
6311                 return raid5_takeover_raid1(mddev);
6312         if (mddev->level == 4) {
6313                 mddev->new_layout = ALGORITHM_PARITY_N;
6314                 mddev->new_level = 5;
6315                 return setup_conf(mddev);
6316         }
6317         if (mddev->level == 6)
6318                 return raid5_takeover_raid6(mddev);
6319
6320         return ERR_PTR(-EINVAL);
6321 }
6322
6323 static void *raid4_takeover(struct mddev *mddev)
6324 {
6325         /* raid4 can take over:
6326          *  raid0 - if there is only one strip zone
6327          *  raid5 - if layout is right
6328          */
6329         if (mddev->level == 0)
6330                 return raid45_takeover_raid0(mddev, 4);
6331         if (mddev->level == 5 &&
6332             mddev->layout == ALGORITHM_PARITY_N) {
6333                 mddev->new_layout = 0;
6334                 mddev->new_level = 4;
6335                 return setup_conf(mddev);
6336         }
6337         return ERR_PTR(-EINVAL);
6338 }
6339
6340 static struct md_personality raid5_personality;
6341
6342 static void *raid6_takeover(struct mddev *mddev)
6343 {
6344         /* Currently can only take over a raid5.  We map the
6345          * personality to an equivalent raid6 personality
6346          * with the Q block at the end.
6347          */
6348         int new_layout;
6349
6350         if (mddev->pers != &raid5_personality)
6351                 return ERR_PTR(-EINVAL);
6352         if (mddev->degraded > 1)
6353                 return ERR_PTR(-EINVAL);
6354         if (mddev->raid_disks > 253)
6355                 return ERR_PTR(-EINVAL);
6356         if (mddev->raid_disks < 3)
6357                 return ERR_PTR(-EINVAL);
6358
6359         switch (mddev->layout) {
6360         case ALGORITHM_LEFT_ASYMMETRIC:
6361                 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6362                 break;
6363         case ALGORITHM_RIGHT_ASYMMETRIC:
6364                 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6365                 break;
6366         case ALGORITHM_LEFT_SYMMETRIC:
6367                 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6368                 break;
6369         case ALGORITHM_RIGHT_SYMMETRIC:
6370                 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6371                 break;
6372         case ALGORITHM_PARITY_0:
6373                 new_layout = ALGORITHM_PARITY_0_6;
6374                 break;
6375         case ALGORITHM_PARITY_N:
6376                 new_layout = ALGORITHM_PARITY_N;
6377                 break;
6378         default:
6379                 return ERR_PTR(-EINVAL);
6380         }
6381         mddev->new_level = 6;
6382         mddev->new_layout = new_layout;
6383         mddev->delta_disks = 1;
6384         mddev->raid_disks += 1;
6385         return setup_conf(mddev);
6386 }
6387
6388
6389 static struct md_personality raid6_personality =
6390 {
6391         .name           = "raid6",
6392         .level          = 6,
6393         .owner          = THIS_MODULE,
6394         .make_request   = make_request,
6395         .run            = run,
6396         .stop           = stop,
6397         .status         = status,
6398         .error_handler  = error,
6399         .hot_add_disk   = raid5_add_disk,
6400         .hot_remove_disk= raid5_remove_disk,
6401         .spare_active   = raid5_spare_active,
6402         .sync_request   = sync_request,
6403         .resize         = raid5_resize,
6404         .size           = raid5_size,
6405         .check_reshape  = raid6_check_reshape,
6406         .start_reshape  = raid5_start_reshape,
6407         .finish_reshape = raid5_finish_reshape,
6408         .quiesce        = raid5_quiesce,
6409         .takeover       = raid6_takeover,
6410 };
6411 static struct md_personality raid5_personality =
6412 {
6413         .name           = "raid5",
6414         .level          = 5,
6415         .owner          = THIS_MODULE,
6416         .make_request   = make_request,
6417         .run            = run,
6418         .stop           = stop,
6419         .status         = status,
6420         .error_handler  = error,
6421         .hot_add_disk   = raid5_add_disk,
6422         .hot_remove_disk= raid5_remove_disk,
6423         .spare_active   = raid5_spare_active,
6424         .sync_request   = sync_request,
6425         .resize         = raid5_resize,
6426         .size           = raid5_size,
6427         .check_reshape  = raid5_check_reshape,
6428         .start_reshape  = raid5_start_reshape,
6429         .finish_reshape = raid5_finish_reshape,
6430         .quiesce        = raid5_quiesce,
6431         .takeover       = raid5_takeover,
6432 };
6433
6434 static struct md_personality raid4_personality =
6435 {
6436         .name           = "raid4",
6437         .level          = 4,
6438         .owner          = THIS_MODULE,
6439         .make_request   = make_request,
6440         .run            = run,
6441         .stop           = stop,
6442         .status         = status,
6443         .error_handler  = error,
6444         .hot_add_disk   = raid5_add_disk,
6445         .hot_remove_disk= raid5_remove_disk,
6446         .spare_active   = raid5_spare_active,
6447         .sync_request   = sync_request,
6448         .resize         = raid5_resize,
6449         .size           = raid5_size,
6450         .check_reshape  = raid5_check_reshape,
6451         .start_reshape  = raid5_start_reshape,
6452         .finish_reshape = raid5_finish_reshape,
6453         .quiesce        = raid5_quiesce,
6454         .takeover       = raid4_takeover,
6455 };
6456
6457 static int __init raid5_init(void)
6458 {
6459         register_md_personality(&raid6_personality);
6460         register_md_personality(&raid5_personality);
6461         register_md_personality(&raid4_personality);
6462         return 0;
6463 }
6464
6465 static void raid5_exit(void)
6466 {
6467         unregister_md_personality(&raid6_personality);
6468         unregister_md_personality(&raid5_personality);
6469         unregister_md_personality(&raid4_personality);
6470 }
6471
6472 module_init(raid5_init);
6473 module_exit(raid5_exit);
6474 MODULE_LICENSE("GPL");
6475 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6476 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6477 MODULE_ALIAS("md-raid5");
6478 MODULE_ALIAS("md-raid4");
6479 MODULE_ALIAS("md-level-5");
6480 MODULE_ALIAS("md-level-4");
6481 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6482 MODULE_ALIAS("md-raid6");
6483 MODULE_ALIAS("md-level-6");
6484
6485 /* This used to be two separate modules, they were: */
6486 MODULE_ALIAS("raid5");
6487 MODULE_ALIAS("raid6");