drivers/md/raid5-ppl.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Partial Parity Log for closing the RAID5 write hole
   4  * Copyright (c) 2017, Intel Corporation.
   5  */
   6
   7 #include <linux/kernel.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/slab.h>
  10 #include <linux/crc32c.h>
  11 #include <linux/async_tx.h>
  12 #include <linux/raid/md_p.h>
  13 #include "md.h"
  14 #include "raid5.h"
  15 #include "raid5-log.h"
  16
  17 /*
  18  * PPL consists of a 4KB header (struct ppl_header) and at least 128KB for
  19  * partial parity data. The header contains an array of entries
  20  * (struct ppl_header_entry) which describe the logged write requests.
  21  * Partial parity for the entries comes after the header, written in the same
  22  * sequence as the entries:
  23  *
  24  * Header
  25  *   entry0
  26  *   ...
  27  *   entryN
  28  * PP data
  29  *   PP for entry0
  30  *   ...
  31  *   PP for entryN
  32  *
  33  * An entry describes one or more consecutive stripe_heads, up to a full
  34  * stripe. The modifed raid data chunks form an m-by-n matrix, where m is the
  35  * number of stripe_heads in the entry and n is the number of modified data
  36  * disks. Every stripe_head in the entry must write to the same data disks.
  37  * An example of a valid case described by a single entry (writes to the first
  38  * stripe of a 4 disk array, 16k chunk size):
  39  *
  40  * sh->sector   dd0   dd1   dd2    ppl
  41  *            +-----+-----+-----+
  42  * 0          | --- | --- | --- | +----+
  43  * 8          | -W- | -W- | --- | | pp |   data_sector = 8
  44  * 16         | -W- | -W- | --- | | pp |   data_size = 3 * 2 * 4k
  45  * 24         | -W- | -W- | --- | | pp |   pp_size = 3 * 4k
  46  *            +-----+-----+-----+ +----+
  47  *
  48  * data_sector is the first raid sector of the modified data, data_size is the
  49  * total size of modified data and pp_size is the size of partial parity for
  50  * this entry. Entries for full stripe writes contain no partial parity
  51  * (pp_size = 0), they only mark the stripes for which parity should be
  52  * recalculated after an unclean shutdown. Every entry holds a checksum of its
  53  * partial parity, the header also has a checksum of the header itself.
  54  *
  55  * A write request is always logged to the PPL instance stored on the parity
  56  * disk of the corresponding stripe. For each member disk there is one ppl_log
  57  * used to handle logging for this disk, independently from others. They are
  58  * grouped in child_logs array in struct ppl_conf, which is assigned to
  59  * r5conf->log_private.
  60  *
  61  * ppl_io_unit represents a full PPL write, header_page contains the ppl_header.
  62  * PPL entries for logged stripes are added in ppl_log_stripe(). A stripe_head
  63  * can be appended to the last entry if it meets the conditions for a valid
  64  * entry described above, otherwise a new entry is added. Checksums of entries
  65  * are calculated incrementally as stripes containing partial parity are being
  66  * added. ppl_submit_iounit() calculates the checksum of the header and submits
  67  * a bio containing the header page and partial parity pages (sh->ppl_page) for
  68  * all stripes of the io_unit. When the PPL write completes, the stripes
  69  * associated with the io_unit are released and raid5d starts writing their data
  70  * and parity. When all stripes are written, the io_unit is freed and the next
  71  * can be submitted.
  72  *
  73  * An io_unit is used to gather stripes until it is submitted or becomes full
  74  * (if the maximum number of entries or size of PPL is reached). Another io_unit
  75  * can't be submitted until the previous has completed (PPL and stripe
  76  * data+parity is written). The log->io_list tracks all io_units of a log
  77  * (for a single member disk). New io_units are added to the end of the list
  78  * and the first io_unit is submitted, if it is not submitted already.
  79  * The current io_unit accepting new stripes is always at the end of the list.
  80  *
  81  * If write-back cache is enabled for any of the disks in the array, its data
  82  * must be flushed before next io_unit is submitted.
  83  */
  84
  85 #define PPL_SPACE_SIZE (128 * 1024)
  86
  87 struct ppl_conf {
  88         struct mddev *mddev;
  89
  90         /* array of child logs, one for each raid disk */
  91         struct ppl_log *child_logs;
  92         int count;
  93
  94         int block_size;         /* the logical block size used for data_sector
  95                                  * in ppl_header_entry */
  96         u32 signature;          /* raid array identifier */
  97         atomic64_t seq;         /* current log write sequence number */
  98
  99         struct kmem_cache *io_kc;
 100         mempool_t io_pool;
 101         struct bio_set bs;
 102         struct bio_set flush_bs;
 103
 104         /* used only for recovery */
 105         int recovered_entries;
 106         int mismatch_count;
 107
 108         /* stripes to retry if failed to allocate io_unit */
 109         struct list_head no_mem_stripes;
 110         spinlock_t no_mem_stripes_lock;
 111
 112         unsigned short write_hint;
 113 };
 114
 115 struct ppl_log {
 116         struct ppl_conf *ppl_conf;      /* shared between all log instances */
 117
 118         struct md_rdev *rdev;           /* array member disk associated with
 119                                          * this log instance */
 120         struct mutex io_mutex;
 121         struct ppl_io_unit *current_io; /* current io_unit accepting new data
 122                                          * always at the end of io_list */
 123         spinlock_t io_list_lock;
 124         struct list_head io_list;       /* all io_units of this log */
 125
 126         sector_t next_io_sector;
 127         unsigned int entry_space;
 128         bool use_multippl;
 129         bool wb_cache_on;
 130         unsigned long disk_flush_bitmap;
 131 };
 132
 133 #define PPL_IO_INLINE_BVECS 32
 134
 135 struct ppl_io_unit {
 136         struct ppl_log *log;
 137
 138         struct page *header_page;       /* for ppl_header */
 139
 140         unsigned int entries_count;     /* number of entries in ppl_header */
 141         unsigned int pp_size;           /* total size current of partial parity */
 142
 143         u64 seq;                        /* sequence number of this log write */
 144         struct list_head log_sibling;   /* log->io_list */
 145
 146         struct list_head stripe_list;   /* stripes added to the io_unit */
 147         atomic_t pending_stripes;       /* how many stripes not written to raid */
 148         atomic_t pending_flushes;       /* how many disk flushes are in progress */
 149
 150         bool submitted;                 /* true if write to log started */
 151
 152         /* inline bio and its biovec for submitting the iounit */
 153         struct bio bio;
 154         struct bio_vec biovec[PPL_IO_INLINE_BVECS];
 155 };
 156
 157 struct dma_async_tx_descriptor *
 158 ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu,
 159                        struct dma_async_tx_descriptor *tx)
 160 {
 161         int disks = sh->disks;
 162         struct page **srcs = percpu->scribble;
 163         int count = 0, pd_idx = sh->pd_idx, i;
 164         struct async_submit_ctl submit;
 165
 166         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
 167
 168         /*
 169          * Partial parity is the XOR of stripe data chunks that are not changed
 170          * during the write request. Depending on available data
 171          * (read-modify-write vs. reconstruct-write case) we calculate it
 172          * differently.
 173          */
 174         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
 175                 /*
 176                  * rmw: xor old data and parity from updated disks
 177                  * This is calculated earlier by ops_run_prexor5() so just copy
 178                  * the parity dev page.
 179                  */
 180                 srcs[count++] = sh->dev[pd_idx].page;
 181         } else if (sh->reconstruct_state == reconstruct_state_drain_run) {
 182                 /* rcw: xor data from all not updated disks */
 183                 for (i = disks; i--;) {
 184                         struct r5dev *dev = &sh->dev[i];
 185                         if (test_bit(R5_UPTODATE, &dev->flags))
 186                                 srcs[count++] = dev->page;
 187                 }
 188         } else {
 189                 return tx;
 190         }
 191
 192         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, tx,
 193                           NULL, sh, (void *) (srcs + sh->disks + 2));
 194
 195         if (count == 1)
 196                 tx = async_memcpy(sh->ppl_page, srcs[0], 0, 0, PAGE_SIZE,
 197                                   &submit);
 198         else
 199                 tx = async_xor(sh->ppl_page, srcs, 0, count, PAGE_SIZE,
 200                                &submit);
 201
 202         return tx;
 203 }
 204
 205 static void *ppl_io_pool_alloc(gfp_t gfp_mask, void *pool_data)
 206 {
 207         struct kmem_cache *kc = pool_data;
 208         struct ppl_io_unit *io;
 209
 210         io = kmem_cache_alloc(kc, gfp_mask);
 211         if (!io)
 212                 return NULL;
 213
 214         io->header_page = alloc_page(gfp_mask);
 215         if (!io->header_page) {
 216                 kmem_cache_free(kc, io);
 217                 return NULL;
 218         }
 219
 220         return io;
 221 }
 222
 223 static void ppl_io_pool_free(void *element, void *pool_data)
 224 {
 225         struct kmem_cache *kc = pool_data;
 226         struct ppl_io_unit *io = element;
 227
 228         __free_page(io->header_page);
 229         kmem_cache_free(kc, io);
 230 }
 231
 232 static struct ppl_io_unit *ppl_new_iounit(struct ppl_log *log,
 233                                           struct stripe_head *sh)
 234 {
 235         struct ppl_conf *ppl_conf = log->ppl_conf;
 236         struct ppl_io_unit *io;
 237         struct ppl_header *pplhdr;
 238         struct page *header_page;
 239
 240         io = mempool_alloc(&ppl_conf->io_pool, GFP_NOWAIT);
 241         if (!io)
 242                 return NULL;
 243
 244         header_page = io->header_page;
 245         memset(io, 0, sizeof(*io));
 246         io->header_page = header_page;
 247
 248         io->log = log;
 249         INIT_LIST_HEAD(&io->log_sibling);
 250         INIT_LIST_HEAD(&io->stripe_list);
 251         atomic_set(&io->pending_stripes, 0);
 252         atomic_set(&io->pending_flushes, 0);
 253         bio_init(&io->bio, log->rdev->bdev, io->biovec, PPL_IO_INLINE_BVECS,
 254                  REQ_OP_WRITE | REQ_FUA);
 255
 256         pplhdr = page_address(io->header_page);
 257         clear_page(pplhdr);
 258         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
 259         pplhdr->signature = cpu_to_le32(ppl_conf->signature);
 260
 261         io->seq = atomic64_add_return(1, &ppl_conf->seq);
 262         pplhdr->generation = cpu_to_le64(io->seq);
 263
 264         return io;
 265 }
 266
 267 static int ppl_log_stripe(struct ppl_log *log, struct stripe_head *sh)
 268 {
 269         struct ppl_io_unit *io = log->current_io;
 270         struct ppl_header_entry *e = NULL;
 271         struct ppl_header *pplhdr;
 272         int i;
 273         sector_t data_sector = 0;
 274         int data_disks = 0;
 275         struct r5conf *conf = sh->raid_conf;
 276
 277         pr_debug("%s: stripe: %llu\n", __func__, (unsigned long long)sh->sector);
 278
 279         /* check if current io_unit is full */
 280         if (io && (io->pp_size == log->entry_space ||
 281                    io->entries_count == PPL_HDR_MAX_ENTRIES)) {
 282                 pr_debug("%s: add io_unit blocked by seq: %llu\n",
 283                          __func__, io->seq);
 284                 io = NULL;
 285         }
 286
 287         /* add a new unit if there is none or the current is full */
 288         if (!io) {
 289                 io = ppl_new_iounit(log, sh);
 290                 if (!io)
 291                         return -ENOMEM;
 292                 spin_lock_irq(&log->io_list_lock);
 293                 list_add_tail(&io->log_sibling, &log->io_list);
 294                 spin_unlock_irq(&log->io_list_lock);
 295
 296                 log->current_io = io;
 297         }
 298
 299         for (i = 0; i < sh->disks; i++) {
 300                 struct r5dev *dev = &sh->dev[i];
 301
 302                 if (i != sh->pd_idx && test_bit(R5_Wantwrite, &dev->flags)) {
 303                         if (!data_disks || dev->sector < data_sector)
 304                                 data_sector = dev->sector;
 305                         data_disks++;
 306                 }
 307         }
 308         BUG_ON(!data_disks);
 309
 310         pr_debug("%s: seq: %llu data_sector: %llu data_disks: %d\n", __func__,
 311                  io->seq, (unsigned long long)data_sector, data_disks);
 312
 313         pplhdr = page_address(io->header_page);
 314
 315         if (io->entries_count > 0) {
 316                 struct ppl_header_entry *last =
 317                                 &pplhdr->entries[io->entries_count - 1];
 318                 struct stripe_head *sh_last = list_last_entry(
 319                                 &io->stripe_list, struct stripe_head, log_list);
 320                 u64 data_sector_last = le64_to_cpu(last->data_sector);
 321                 u32 data_size_last = le32_to_cpu(last->data_size);
 322
 323                 /*
 324                  * Check if we can append the stripe to the last entry. It must
 325                  * be just after the last logged stripe and write to the same
 326                  * disks. Use bit shift and logarithm to avoid 64-bit division.
 327                  */
 328                 if ((sh->sector == sh_last->sector + RAID5_STRIPE_SECTORS(conf)) &&
 329                     (data_sector >> ilog2(conf->chunk_sectors) ==
 330                      data_sector_last >> ilog2(conf->chunk_sectors)) &&
 331                     ((data_sector - data_sector_last) * data_disks ==
 332                      data_size_last >> 9))
 333                         e = last;
 334         }
 335
 336         if (!e) {
 337                 e = &pplhdr->entries[io->entries_count++];
 338                 e->data_sector = cpu_to_le64(data_sector);
 339                 e->parity_disk = cpu_to_le32(sh->pd_idx);
 340                 e->checksum = cpu_to_le32(~0);
 341         }
 342
 343         le32_add_cpu(&e->data_size, data_disks << PAGE_SHIFT);
 344
 345         /* don't write any PP if full stripe write */
 346         if (!test_bit(STRIPE_FULL_WRITE, &sh->state)) {
 347                 le32_add_cpu(&e->pp_size, PAGE_SIZE);
 348                 io->pp_size += PAGE_SIZE;
 349                 e->checksum = cpu_to_le32(crc32c_le(le32_to_cpu(e->checksum),
 350                                                     page_address(sh->ppl_page),
 351                                                     PAGE_SIZE));
 352         }
 353
 354         list_add_tail(&sh->log_list, &io->stripe_list);
 355         atomic_inc(&io->pending_stripes);
 356         sh->ppl_io = io;
 357
 358         return 0;
 359 }
 360
 361 int ppl_write_stripe(struct r5conf *conf, struct stripe_head *sh)
 362 {
 363         struct ppl_conf *ppl_conf = conf->log_private;
 364         struct ppl_io_unit *io = sh->ppl_io;
 365         struct ppl_log *log;
 366
 367         if (io || test_bit(STRIPE_SYNCING, &sh->state) || !sh->ppl_page ||
 368             !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 369             !test_bit(R5_Insync, &sh->dev[sh->pd_idx].flags)) {
 370                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 371                 return -EAGAIN;
 372         }
 373
 374         log = &ppl_conf->child_logs[sh->pd_idx];
 375
 376         mutex_lock(&log->io_mutex);
 377
 378         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 379                 mutex_unlock(&log->io_mutex);
 380                 return -EAGAIN;
 381         }
 382
 383         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 384         clear_bit(STRIPE_DELAYED, &sh->state);
 385         atomic_inc(&sh->count);
 386
 387         if (ppl_log_stripe(log, sh)) {
 388                 spin_lock_irq(&ppl_conf->no_mem_stripes_lock);
 389                 list_add_tail(&sh->log_list, &ppl_conf->no_mem_stripes);
 390                 spin_unlock_irq(&ppl_conf->no_mem_stripes_lock);
 391         }
 392
 393         mutex_unlock(&log->io_mutex);
 394
 395         return 0;
 396 }
 397
 398 static void ppl_log_endio(struct bio *bio)
 399 {
 400         struct ppl_io_unit *io = bio->bi_private;
 401         struct ppl_log *log = io->log;
 402         struct ppl_conf *ppl_conf = log->ppl_conf;
 403         struct stripe_head *sh, *next;
 404
 405         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 406
 407         if (bio->bi_status)
 408                 md_error(ppl_conf->mddev, log->rdev);
 409
 410         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 411                 list_del_init(&sh->log_list);
 412
 413                 set_bit(STRIPE_HANDLE, &sh->state);
 414                 raid5_release_stripe(sh);
 415         }
 416 }
 417
 418 static void ppl_submit_iounit_bio(struct ppl_io_unit *io, struct bio *bio)
 419 {
 420         pr_debug("%s: seq: %llu size: %u sector: %llu dev: %pg\n",
 421                  __func__, io->seq, bio->bi_iter.bi_size,
 422                  (unsigned long long)bio->bi_iter.bi_sector,
 423                  bio->bi_bdev);
 424
 425         submit_bio(bio);
 426 }
 427
 428 static void ppl_submit_iounit(struct ppl_io_unit *io)
 429 {
 430         struct ppl_log *log = io->log;
 431         struct ppl_conf *ppl_conf = log->ppl_conf;
 432         struct ppl_header *pplhdr = page_address(io->header_page);
 433         struct bio *bio = &io->bio;
 434         struct stripe_head *sh;
 435         int i;
 436
 437         bio->bi_private = io;
 438
 439         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 440                 ppl_log_endio(bio);
 441                 return;
 442         }
 443
 444         for (i = 0; i < io->entries_count; i++) {
 445                 struct ppl_header_entry *e = &pplhdr->entries[i];
 446
 447                 pr_debug("%s: seq: %llu entry: %d data_sector: %llu pp_size: %u data_size: %u\n",
 448                          __func__, io->seq, i, le64_to_cpu(e->data_sector),
 449                          le32_to_cpu(e->pp_size), le32_to_cpu(e->data_size));
 450
 451                 e->data_sector = cpu_to_le64(le64_to_cpu(e->data_sector) >>
 452                                              ilog2(ppl_conf->block_size >> 9));
 453                 e->checksum = cpu_to_le32(~le32_to_cpu(e->checksum));
 454         }
 455
 456         pplhdr->entries_count = cpu_to_le32(io->entries_count);
 457         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PPL_HEADER_SIZE));
 458
 459         /* Rewind the buffer if current PPL is larger then remaining space */
 460         if (log->use_multippl &&
 461             log->rdev->ppl.sector + log->rdev->ppl.size - log->next_io_sector <
 462             (PPL_HEADER_SIZE + io->pp_size) >> 9)
 463                 log->next_io_sector = log->rdev->ppl.sector;
 464
 465
 466         bio->bi_end_io = ppl_log_endio;
 467         bio->bi_iter.bi_sector = log->next_io_sector;
 468         bio_add_page(bio, io->header_page, PAGE_SIZE, 0);
 469
 470         pr_debug("%s: log->current_io_sector: %llu\n", __func__,
 471             (unsigned long long)log->next_io_sector);
 472
 473         if (log->use_multippl)
 474                 log->next_io_sector += (PPL_HEADER_SIZE + io->pp_size) >> 9;
 475
 476         WARN_ON(log->disk_flush_bitmap != 0);
 477
 478         list_for_each_entry(sh, &io->stripe_list, log_list) {
 479                 for (i = 0; i < sh->disks; i++) {
 480                         struct r5dev *dev = &sh->dev[i];
 481
 482                         if ((ppl_conf->child_logs[i].wb_cache_on) &&
 483                             (test_bit(R5_Wantwrite, &dev->flags))) {
 484                                 set_bit(i, &log->disk_flush_bitmap);
 485                         }
 486                 }
 487
 488                 /* entries for full stripe writes have no partial parity */
 489                 if (test_bit(STRIPE_FULL_WRITE, &sh->state))
 490                         continue;
 491
 492                 if (!bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0)) {
 493                         struct bio *prev = bio;
 494
 495                         bio = bio_alloc_bioset(prev->bi_bdev, BIO_MAX_VECS,
 496                                                prev->bi_opf, GFP_NOIO,
 497                                                &ppl_conf->bs);
 498                         bio->bi_iter.bi_sector = bio_end_sector(prev);
 499                         bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0);
 500
 501                         bio_chain(bio, prev);
 502                         ppl_submit_iounit_bio(io, prev);
 503                 }
 504         }
 505
 506         ppl_submit_iounit_bio(io, bio);
 507 }
 508
 509 static void ppl_submit_current_io(struct ppl_log *log)
 510 {
 511         struct ppl_io_unit *io;
 512
 513         spin_lock_irq(&log->io_list_lock);
 514
 515         io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
 516                                       log_sibling);
 517         if (io && io->submitted)
 518                 io = NULL;
 519
 520         spin_unlock_irq(&log->io_list_lock);
 521
 522         if (io) {
 523                 io->submitted = true;
 524
 525                 if (io == log->current_io)
 526                         log->current_io = NULL;
 527
 528                 ppl_submit_iounit(io);
 529         }
 530 }
 531
 532 void ppl_write_stripe_run(struct r5conf *conf)
 533 {
 534         struct ppl_conf *ppl_conf = conf->log_private;
 535         struct ppl_log *log;
 536         int i;
 537
 538         for (i = 0; i < ppl_conf->count; i++) {
 539                 log = &ppl_conf->child_logs[i];
 540
 541                 mutex_lock(&log->io_mutex);
 542                 ppl_submit_current_io(log);
 543                 mutex_unlock(&log->io_mutex);
 544         }
 545 }
 546
 547 static void ppl_io_unit_finished(struct ppl_io_unit *io)
 548 {
 549         struct ppl_log *log = io->log;
 550         struct ppl_conf *ppl_conf = log->ppl_conf;
 551         struct r5conf *conf = ppl_conf->mddev->private;
 552         unsigned long flags;
 553
 554         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 555
 556         local_irq_save(flags);
 557
 558         spin_lock(&log->io_list_lock);
 559         list_del(&io->log_sibling);
 560         spin_unlock(&log->io_list_lock);
 561
 562         mempool_free(io, &ppl_conf->io_pool);
 563
 564         spin_lock(&ppl_conf->no_mem_stripes_lock);
 565         if (!list_empty(&ppl_conf->no_mem_stripes)) {
 566                 struct stripe_head *sh;
 567
 568                 sh = list_first_entry(&ppl_conf->no_mem_stripes,
 569                                       struct stripe_head, log_list);
 570                 list_del_init(&sh->log_list);
 571                 set_bit(STRIPE_HANDLE, &sh->state);
 572                 raid5_release_stripe(sh);
 573         }
 574         spin_unlock(&ppl_conf->no_mem_stripes_lock);
 575
 576         local_irq_restore(flags);
 577
 578         wake_up(&conf->wait_for_quiescent);
 579 }
 580
 581 static void ppl_flush_endio(struct bio *bio)
 582 {
 583         struct ppl_io_unit *io = bio->bi_private;
 584         struct ppl_log *log = io->log;
 585         struct ppl_conf *ppl_conf = log->ppl_conf;
 586         struct r5conf *conf = ppl_conf->mddev->private;
 587
 588         pr_debug("%s: dev: %pg\n", __func__, bio->bi_bdev);
 589
 590         if (bio->bi_status) {
 591                 struct md_rdev *rdev;
 592
 593                 rcu_read_lock();
 594                 rdev = md_find_rdev_rcu(conf->mddev, bio_dev(bio));
 595                 if (rdev)
 596                         md_error(rdev->mddev, rdev);
 597                 rcu_read_unlock();
 598         }
 599
 600         bio_put(bio);
 601
 602         if (atomic_dec_and_test(&io->pending_flushes)) {
 603                 ppl_io_unit_finished(io);
 604                 md_wakeup_thread(conf->mddev->thread);
 605         }
 606 }
 607
 608 static void ppl_do_flush(struct ppl_io_unit *io)
 609 {
 610         struct ppl_log *log = io->log;
 611         struct ppl_conf *ppl_conf = log->ppl_conf;
 612         struct r5conf *conf = ppl_conf->mddev->private;
 613         int raid_disks = conf->raid_disks;
 614         int flushed_disks = 0;
 615         int i;
 616
 617         atomic_set(&io->pending_flushes, raid_disks);
 618
 619         for_each_set_bit(i, &log->disk_flush_bitmap, raid_disks) {
 620                 struct md_rdev *rdev;
 621                 struct block_device *bdev = NULL;
 622
 623                 rcu_read_lock();
 624                 rdev = rcu_dereference(conf->disks[i].rdev);
 625                 if (rdev && !test_bit(Faulty, &rdev->flags))
 626                         bdev = rdev->bdev;
 627                 rcu_read_unlock();
 628
 629                 if (bdev) {
 630                         struct bio *bio;
 631
 632                         bio = bio_alloc_bioset(bdev, 0,
 633                                                REQ_OP_WRITE | REQ_PREFLUSH,
 634                                                GFP_NOIO, &ppl_conf->flush_bs);
 635                         bio->bi_private = io;
 636                         bio->bi_end_io = ppl_flush_endio;
 637
 638                         pr_debug("%s: dev: %ps\n", __func__, bio->bi_bdev);
 639
 640                         submit_bio(bio);
 641                         flushed_disks++;
 642                 }
 643         }
 644
 645         log->disk_flush_bitmap = 0;
 646
 647         for (i = flushed_disks ; i < raid_disks; i++) {
 648                 if (atomic_dec_and_test(&io->pending_flushes))
 649                         ppl_io_unit_finished(io);
 650         }
 651 }
 652
 653 static inline bool ppl_no_io_unit_submitted(struct r5conf *conf,
 654                                             struct ppl_log *log)
 655 {
 656         struct ppl_io_unit *io;
 657
 658         io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
 659                                       log_sibling);
 660
 661         return !io || !io->submitted;
 662 }
 663
 664 void ppl_quiesce(struct r5conf *conf, int quiesce)
 665 {
 666         struct ppl_conf *ppl_conf = conf->log_private;
 667         int i;
 668
 669         if (quiesce) {
 670                 for (i = 0; i < ppl_conf->count; i++) {
 671                         struct ppl_log *log = &ppl_conf->child_logs[i];
 672
 673                         spin_lock_irq(&log->io_list_lock);
 674                         wait_event_lock_irq(conf->wait_for_quiescent,
 675                                             ppl_no_io_unit_submitted(conf, log),
 676                                             log->io_list_lock);
 677                         spin_unlock_irq(&log->io_list_lock);
 678                 }
 679         }
 680 }
 681
 682 int ppl_handle_flush_request(struct bio *bio)
 683 {
 684         if (bio->bi_iter.bi_size == 0) {
 685                 bio_endio(bio);
 686                 return 0;
 687         }
 688         bio->bi_opf &= ~REQ_PREFLUSH;
 689         return -EAGAIN;
 690 }
 691
 692 void ppl_stripe_write_finished(struct stripe_head *sh)
 693 {
 694         struct ppl_io_unit *io;
 695
 696         io = sh->ppl_io;
 697         sh->ppl_io = NULL;
 698
 699         if (io && atomic_dec_and_test(&io->pending_stripes)) {
 700                 if (io->log->disk_flush_bitmap)
 701                         ppl_do_flush(io);
 702                 else
 703                         ppl_io_unit_finished(io);
 704         }
 705 }
 706
 707 static void ppl_xor(int size, struct page *page1, struct page *page2)
 708 {
 709         struct async_submit_ctl submit;
 710         struct dma_async_tx_descriptor *tx;
 711         struct page *xor_srcs[] = { page1, page2 };
 712
 713         init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST,
 714                           NULL, NULL, NULL, NULL);
 715         tx = async_xor(page1, xor_srcs, 0, 2, size, &submit);
 716
 717         async_tx_quiesce(&tx);
 718 }
 719
 720 /*
 721  * PPL recovery strategy: xor partial parity and data from all modified data
 722  * disks within a stripe and write the result as the new stripe parity. If all
 723  * stripe data disks are modified (full stripe write), no partial parity is
 724  * available, so just xor the data disks.
 725  *
 726  * Recovery of a PPL entry shall occur only if all modified data disks are
 727  * available and read from all of them succeeds.
 728  *
 729  * A PPL entry applies to a stripe, partial parity size for an entry is at most
 730  * the size of the chunk. Examples of possible cases for a single entry:
 731  *
 732  * case 0: single data disk write:
 733  *   data0    data1    data2     ppl        parity
 734  * +--------+--------+--------+           +--------------------+
 735  * | ------ | ------ | ------ | +----+    | (no change)        |
 736  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 737  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 738  * | ------ | ------ | ------ | +----+    | (no change)        |
 739  * +--------+--------+--------+           +--------------------+
 740  * pp_size = data_size
 741  *
 742  * case 1: more than one data disk write:
 743  *   data0    data1    data2     ppl        parity
 744  * +--------+--------+--------+           +--------------------+
 745  * | ------ | ------ | ------ | +----+    | (no change)        |
 746  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 747  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 748  * | ------ | ------ | ------ | +----+    | (no change)        |
 749  * +--------+--------+--------+           +--------------------+
 750  * pp_size = data_size / modified_data_disks
 751  *
 752  * case 2: write to all data disks (also full stripe write):
 753  *   data0    data1    data2                parity
 754  * +--------+--------+--------+           +--------------------+
 755  * | ------ | ------ | ------ |           | (no change)        |
 756  * | -data- | -data- | -data- | --------> | xor all data       |
 757  * | ------ | ------ | ------ | --------> | (no change)        |
 758  * | ------ | ------ | ------ |           | (no change)        |
 759  * +--------+--------+--------+           +--------------------+
 760  * pp_size = 0
 761  *
 762  * The following cases are possible only in other implementations. The recovery
 763  * code can handle them, but they are not generated at runtime because they can
 764  * be reduced to cases 0, 1 and 2:
 765  *
 766  * case 3:
 767  *   data0    data1    data2     ppl        parity
 768  * +--------+--------+--------+ +----+    +--------------------+
 769  * | ------ | -data- | -data- | | pp |    | data1 ^ data2 ^ pp |
 770  * | ------ | -data- | -data- | | pp | -> | data1 ^ data2 ^ pp |
 771  * | -data- | -data- | -data- | | -- | -> | xor all data       |
 772  * | -data- | -data- | ------ | | pp |    | data0 ^ data1 ^ pp |
 773  * +--------+--------+--------+ +----+    +--------------------+
 774  * pp_size = chunk_size
 775  *
 776  * case 4:
 777  *   data0    data1    data2     ppl        parity
 778  * +--------+--------+--------+ +----+    +--------------------+
 779  * | ------ | -data- | ------ | | pp |    | data1 ^ pp         |
 780  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 781  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 782  * | -data- | ------ | ------ | | pp |    | data0 ^ pp         |
 783  * +--------+--------+--------+ +----+    +--------------------+
 784  * pp_size = chunk_size
 785  */
 786 static int ppl_recover_entry(struct ppl_log *log, struct ppl_header_entry *e,
 787                              sector_t ppl_sector)
 788 {
 789         struct ppl_conf *ppl_conf = log->ppl_conf;
 790         struct mddev *mddev = ppl_conf->mddev;
 791         struct r5conf *conf = mddev->private;
 792         int block_size = ppl_conf->block_size;
 793         struct page *page1;
 794         struct page *page2;
 795         sector_t r_sector_first;
 796         sector_t r_sector_last;
 797         int strip_sectors;
 798         int data_disks;
 799         int i;
 800         int ret = 0;
 801         unsigned int pp_size = le32_to_cpu(e->pp_size);
 802         unsigned int data_size = le32_to_cpu(e->data_size);
 803
 804         page1 = alloc_page(GFP_KERNEL);
 805         page2 = alloc_page(GFP_KERNEL);
 806
 807         if (!page1 || !page2) {
 808                 ret = -ENOMEM;
 809                 goto out;
 810         }
 811
 812         r_sector_first = le64_to_cpu(e->data_sector) * (block_size >> 9);
 813
 814         if ((pp_size >> 9) < conf->chunk_sectors) {
 815                 if (pp_size > 0) {
 816                         data_disks = data_size / pp_size;
 817                         strip_sectors = pp_size >> 9;
 818                 } else {
 819                         data_disks = conf->raid_disks - conf->max_degraded;
 820                         strip_sectors = (data_size >> 9) / data_disks;
 821                 }
 822                 r_sector_last = r_sector_first +
 823                                 (data_disks - 1) * conf->chunk_sectors +
 824                                 strip_sectors;
 825         } else {
 826                 data_disks = conf->raid_disks - conf->max_degraded;
 827                 strip_sectors = conf->chunk_sectors;
 828                 r_sector_last = r_sector_first + (data_size >> 9);
 829         }
 830
 831         pr_debug("%s: array sector first: %llu last: %llu\n", __func__,
 832                  (unsigned long long)r_sector_first,
 833                  (unsigned long long)r_sector_last);
 834
 835         /* if start and end is 4k aligned, use a 4k block */
 836         if (block_size == 512 &&
 837             (r_sector_first & (RAID5_STRIPE_SECTORS(conf) - 1)) == 0 &&
 838             (r_sector_last & (RAID5_STRIPE_SECTORS(conf) - 1)) == 0)
 839                 block_size = RAID5_STRIPE_SIZE(conf);
 840
 841         /* iterate through blocks in strip */
 842         for (i = 0; i < strip_sectors; i += (block_size >> 9)) {
 843                 bool update_parity = false;
 844                 sector_t parity_sector;
 845                 struct md_rdev *parity_rdev;
 846                 struct stripe_head sh;
 847                 int disk;
 848                 int indent = 0;
 849
 850                 pr_debug("%s:%*s iter %d start\n", __func__, indent, "", i);
 851                 indent += 2;
 852
 853                 memset(page_address(page1), 0, PAGE_SIZE);
 854
 855                 /* iterate through data member disks */
 856                 for (disk = 0; disk < data_disks; disk++) {
 857                         int dd_idx;
 858                         struct md_rdev *rdev;
 859                         sector_t sector;
 860                         sector_t r_sector = r_sector_first + i +
 861                                             (disk * conf->chunk_sectors);
 862
 863                         pr_debug("%s:%*s data member disk %d start\n",
 864                                  __func__, indent, "", disk);
 865                         indent += 2;
 866
 867                         if (r_sector >= r_sector_last) {
 868                                 pr_debug("%s:%*s array sector %llu doesn't need parity update\n",
 869                                          __func__, indent, "",
 870                                          (unsigned long long)r_sector);
 871                                 indent -= 2;
 872                                 continue;
 873                         }
 874
 875                         update_parity = true;
 876
 877                         /* map raid sector to member disk */
 878                         sector = raid5_compute_sector(conf, r_sector, 0,
 879                                                       &dd_idx, NULL);
 880                         pr_debug("%s:%*s processing array sector %llu => data member disk %d, sector %llu\n",
 881                                  __func__, indent, "",
 882                                  (unsigned long long)r_sector, dd_idx,
 883                                  (unsigned long long)sector);
 884
 885                         /* Array has not started so rcu dereference is safe */
 886                         rdev = rcu_dereference_protected(
 887                                         conf->disks[dd_idx].rdev, 1);
 888                         if (!rdev || (!test_bit(In_sync, &rdev->flags) &&
 889                                       sector >= rdev->recovery_offset)) {
 890                                 pr_debug("%s:%*s data member disk %d missing\n",
 891                                          __func__, indent, "", dd_idx);
 892                                 update_parity = false;
 893                                 break;
 894                         }
 895
 896                         pr_debug("%s:%*s reading data member disk %pg sector %llu\n",
 897                                  __func__, indent, "", rdev->bdev,
 898                                  (unsigned long long)sector);
 899                         if (!sync_page_io(rdev, sector, block_size, page2,
 900                                         REQ_OP_READ, false)) {
 901                                 md_error(mddev, rdev);
 902                                 pr_debug("%s:%*s read failed!\n", __func__,
 903                                          indent, "");
 904                                 ret = -EIO;
 905                                 goto out;
 906                         }
 907
 908                         ppl_xor(block_size, page1, page2);
 909
 910                         indent -= 2;
 911                 }
 912
 913                 if (!update_parity)
 914                         continue;
 915
 916                 if (pp_size > 0) {
 917                         pr_debug("%s:%*s reading pp disk sector %llu\n",
 918                                  __func__, indent, "",
 919                                  (unsigned long long)(ppl_sector + i));
 920                         if (!sync_page_io(log->rdev,
 921                                         ppl_sector - log->rdev->data_offset + i,
 922                                         block_size, page2, REQ_OP_READ,
 923                                         false)) {
 924                                 pr_debug("%s:%*s read failed!\n", __func__,
 925                                          indent, "");
 926                                 md_error(mddev, log->rdev);
 927                                 ret = -EIO;
 928                                 goto out;
 929                         }
 930
 931                         ppl_xor(block_size, page1, page2);
 932                 }
 933
 934                 /* map raid sector to parity disk */
 935                 parity_sector = raid5_compute_sector(conf, r_sector_first + i,
 936                                 0, &disk, &sh);
 937                 BUG_ON(sh.pd_idx != le32_to_cpu(e->parity_disk));
 938
 939                 /* Array has not started so rcu dereference is safe */
 940                 parity_rdev = rcu_dereference_protected(
 941                                         conf->disks[sh.pd_idx].rdev, 1);
 942
 943                 BUG_ON(parity_rdev->bdev->bd_dev != log->rdev->bdev->bd_dev);
 944                 pr_debug("%s:%*s write parity at sector %llu, disk %pg\n",
 945                          __func__, indent, "",
 946                          (unsigned long long)parity_sector,
 947                          parity_rdev->bdev);
 948                 if (!sync_page_io(parity_rdev, parity_sector, block_size,
 949                                   page1, REQ_OP_WRITE, false)) {
 950                         pr_debug("%s:%*s parity write error!\n", __func__,
 951                                  indent, "");
 952                         md_error(mddev, parity_rdev);
 953                         ret = -EIO;
 954                         goto out;
 955                 }
 956         }
 957 out:
 958         if (page1)
 959                 __free_page(page1);
 960         if (page2)
 961                 __free_page(page2);
 962         return ret;
 963 }
 964
 965 static int ppl_recover(struct ppl_log *log, struct ppl_header *pplhdr,
 966                        sector_t offset)
 967 {
 968         struct ppl_conf *ppl_conf = log->ppl_conf;
 969         struct md_rdev *rdev = log->rdev;
 970         struct mddev *mddev = rdev->mddev;
 971         sector_t ppl_sector = rdev->ppl.sector + offset +
 972                               (PPL_HEADER_SIZE >> 9);
 973         struct page *page;
 974         int i;
 975         int ret = 0;
 976
 977         page = alloc_page(GFP_KERNEL);
 978         if (!page)
 979                 return -ENOMEM;
 980
 981         /* iterate through all PPL entries saved */
 982         for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++) {
 983                 struct ppl_header_entry *e = &pplhdr->entries[i];
 984                 u32 pp_size = le32_to_cpu(e->pp_size);
 985                 sector_t sector = ppl_sector;
 986                 int ppl_entry_sectors = pp_size >> 9;
 987                 u32 crc, crc_stored;
 988
 989                 pr_debug("%s: disk: %d entry: %d ppl_sector: %llu pp_size: %u\n",
 990                          __func__, rdev->raid_disk, i,
 991                          (unsigned long long)ppl_sector, pp_size);
 992
 993                 crc = ~0;
 994                 crc_stored = le32_to_cpu(e->checksum);
 995
 996                 /* read parial parity for this entry and calculate its checksum */
 997                 while (pp_size) {
 998                         int s = pp_size > PAGE_SIZE ? PAGE_SIZE : pp_size;
 999
1000                         if (!sync_page_io(rdev, sector - rdev->data_offset,
1001                                         s, page, REQ_OP_READ, false)) {
1002                                 md_error(mddev, rdev);
1003                                 ret = -EIO;
1004                                 goto out;
1005                         }
1006
1007                         crc = crc32c_le(crc, page_address(page), s);
1008
1009                         pp_size -= s;
1010                         sector += s >> 9;
1011                 }
1012
1013                 crc = ~crc;
1014
1015                 if (crc != crc_stored) {
1016                         /*
1017                          * Don't recover this entry if the checksum does not
1018                          * match, but keep going and try to recover other
1019                          * entries.
1020                          */
1021                         pr_debug("%s: ppl entry crc does not match: stored: 0x%x calculated: 0x%x\n",
1022                                  __func__, crc_stored, crc);
1023                         ppl_conf->mismatch_count++;
1024                 } else {
1025                         ret = ppl_recover_entry(log, e, ppl_sector);
1026                         if (ret)
1027                                 goto out;
1028                         ppl_conf->recovered_entries++;
1029                 }
1030
1031                 ppl_sector += ppl_entry_sectors;
1032         }
1033
1034         /* flush the disk cache after recovery if necessary */
1035         ret = blkdev_issue_flush(rdev->bdev);
1036 out:
1037         __free_page(page);
1038         return ret;
1039 }
1040
1041 static int ppl_write_empty_header(struct ppl_log *log)
1042 {
1043         struct page *page;
1044         struct ppl_header *pplhdr;
1045         struct md_rdev *rdev = log->rdev;
1046         int ret = 0;
1047
1048         pr_debug("%s: disk: %d ppl_sector: %llu\n", __func__,
1049                  rdev->raid_disk, (unsigned long long)rdev->ppl.sector);
1050
1051         page = alloc_page(GFP_NOIO | __GFP_ZERO);
1052         if (!page)
1053                 return -ENOMEM;
1054
1055         pplhdr = page_address(page);
1056         /* zero out PPL space to avoid collision with old PPLs */
1057         blkdev_issue_zeroout(rdev->bdev, rdev->ppl.sector,
1058                             log->rdev->ppl.size, GFP_NOIO, 0);
1059         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
1060         pplhdr->signature = cpu_to_le32(log->ppl_conf->signature);
1061         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PAGE_SIZE));
1062
1063         if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset,
1064                           PPL_HEADER_SIZE, page, REQ_OP_WRITE | REQ_SYNC |
1065                           REQ_FUA, false)) {
1066                 md_error(rdev->mddev, rdev);
1067                 ret = -EIO;
1068         }
1069
1070         __free_page(page);
1071         return ret;
1072 }
1073
1074 static int ppl_load_distributed(struct ppl_log *log)
1075 {
1076         struct ppl_conf *ppl_conf = log->ppl_conf;
1077         struct md_rdev *rdev = log->rdev;
1078         struct mddev *mddev = rdev->mddev;
1079         struct page *page, *page2;
1080         struct ppl_header *pplhdr = NULL, *prev_pplhdr = NULL;
1081         u32 crc, crc_stored;
1082         u32 signature;
1083         int ret = 0, i;
1084         sector_t pplhdr_offset = 0, prev_pplhdr_offset = 0;
1085
1086         pr_debug("%s: disk: %d\n", __func__, rdev->raid_disk);
1087         /* read PPL headers, find the recent one */
1088         page = alloc_page(GFP_KERNEL);
1089         if (!page)
1090                 return -ENOMEM;
1091
1092         page2 = alloc_page(GFP_KERNEL);
1093         if (!page2) {
1094                 __free_page(page);
1095                 return -ENOMEM;
1096         }
1097
1098         /* searching ppl area for latest ppl */
1099         while (pplhdr_offset < rdev->ppl.size - (PPL_HEADER_SIZE >> 9)) {
1100                 if (!sync_page_io(rdev,
1101                                   rdev->ppl.sector - rdev->data_offset +
1102                                   pplhdr_offset, PAGE_SIZE, page, REQ_OP_READ,
1103                                   false)) {
1104                         md_error(mddev, rdev);
1105                         ret = -EIO;
1106                         /* if not able to read - don't recover any PPL */
1107                         pplhdr = NULL;
1108                         break;
1109                 }
1110                 pplhdr = page_address(page);
1111
1112                 /* check header validity */
1113                 crc_stored = le32_to_cpu(pplhdr->checksum);
1114                 pplhdr->checksum = 0;
1115                 crc = ~crc32c_le(~0, pplhdr, PAGE_SIZE);
1116
1117                 if (crc_stored != crc) {
1118                         pr_debug("%s: ppl header crc does not match: stored: 0x%x calculated: 0x%x (offset: %llu)\n",
1119                                  __func__, crc_stored, crc,
1120                                  (unsigned long long)pplhdr_offset);
1121                         pplhdr = prev_pplhdr;
1122                         pplhdr_offset = prev_pplhdr_offset;
1123                         break;
1124                 }
1125
1126                 signature = le32_to_cpu(pplhdr->signature);
1127
1128                 if (mddev->external) {
1129                         /*
1130                          * For external metadata the header signature is set and
1131                          * validated in userspace.
1132                          */
1133                         ppl_conf->signature = signature;
1134                 } else if (ppl_conf->signature != signature) {
1135                         pr_debug("%s: ppl header signature does not match: stored: 0x%x configured: 0x%x (offset: %llu)\n",
1136                                  __func__, signature, ppl_conf->signature,
1137                                  (unsigned long long)pplhdr_offset);
1138                         pplhdr = prev_pplhdr;
1139                         pplhdr_offset = prev_pplhdr_offset;
1140                         break;
1141                 }
1142
1143                 if (prev_pplhdr && le64_to_cpu(prev_pplhdr->generation) >
1144                     le64_to_cpu(pplhdr->generation)) {
1145                         /* previous was newest */
1146                         pplhdr = prev_pplhdr;
1147                         pplhdr_offset = prev_pplhdr_offset;
1148                         break;
1149                 }
1150
1151                 prev_pplhdr_offset = pplhdr_offset;
1152                 prev_pplhdr = pplhdr;
1153
1154                 swap(page, page2);
1155
1156                 /* calculate next potential ppl offset */
1157                 for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++)
1158                         pplhdr_offset +=
1159                             le32_to_cpu(pplhdr->entries[i].pp_size) >> 9;
1160                 pplhdr_offset += PPL_HEADER_SIZE >> 9;
1161         }
1162
1163         /* no valid ppl found */
1164         if (!pplhdr)
1165                 ppl_conf->mismatch_count++;
1166         else
1167                 pr_debug("%s: latest PPL found at offset: %llu, with generation: %llu\n",
1168                     __func__, (unsigned long long)pplhdr_offset,
1169                     le64_to_cpu(pplhdr->generation));
1170
1171         /* attempt to recover from log if we are starting a dirty array */
1172         if (pplhdr && !mddev->pers && mddev->recovery_cp != MaxSector)
1173                 ret = ppl_recover(log, pplhdr, pplhdr_offset);
1174
1175         /* write empty header if we are starting the array */
1176         if (!ret && !mddev->pers)
1177                 ret = ppl_write_empty_header(log);
1178
1179         __free_page(page);
1180         __free_page(page2);
1181
1182         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1183                  __func__, ret, ppl_conf->mismatch_count,
1184                  ppl_conf->recovered_entries);
1185         return ret;
1186 }
1187
1188 static int ppl_load(struct ppl_conf *ppl_conf)
1189 {
1190         int ret = 0;
1191         u32 signature = 0;
1192         bool signature_set = false;
1193         int i;
1194
1195         for (i = 0; i < ppl_conf->count; i++) {
1196                 struct ppl_log *log = &ppl_conf->child_logs[i];
1197
1198                 /* skip missing drive */
1199                 if (!log->rdev)
1200                         continue;
1201
1202                 ret = ppl_load_distributed(log);
1203                 if (ret)
1204                         break;
1205
1206                 /*
1207                  * For external metadata we can't check if the signature is
1208                  * correct on a single drive, but we can check if it is the same
1209                  * on all drives.
1210                  */
1211                 if (ppl_conf->mddev->external) {
1212                         if (!signature_set) {
1213                                 signature = ppl_conf->signature;
1214                                 signature_set = true;
1215                         } else if (signature != ppl_conf->signature) {
1216                                 pr_warn("md/raid:%s: PPL header signature does not match on all member drives\n",
1217                                         mdname(ppl_conf->mddev));
1218                                 ret = -EINVAL;
1219                                 break;
1220                         }
1221                 }
1222         }
1223
1224         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1225                  __func__, ret, ppl_conf->mismatch_count,
1226                  ppl_conf->recovered_entries);
1227         return ret;
1228 }
1229
1230 static void __ppl_exit_log(struct ppl_conf *ppl_conf)
1231 {
1232         clear_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1233         clear_bit(MD_HAS_MULTIPLE_PPLS, &ppl_conf->mddev->flags);
1234
1235         kfree(ppl_conf->child_logs);
1236
1237         bioset_exit(&ppl_conf->bs);
1238         bioset_exit(&ppl_conf->flush_bs);
1239         mempool_exit(&ppl_conf->io_pool);
1240         kmem_cache_destroy(ppl_conf->io_kc);
1241
1242         kfree(ppl_conf);
1243 }
1244
1245 void ppl_exit_log(struct r5conf *conf)
1246 {
1247         struct ppl_conf *ppl_conf = conf->log_private;
1248
1249         if (ppl_conf) {
1250                 __ppl_exit_log(ppl_conf);
1251                 conf->log_private = NULL;
1252         }
1253 }
1254
1255 static int ppl_validate_rdev(struct md_rdev *rdev)
1256 {
1257         int ppl_data_sectors;
1258         int ppl_size_new;
1259
1260         /*
1261          * The configured PPL size must be enough to store
1262          * the header and (at the very least) partial parity
1263          * for one stripe. Round it down to ensure the data
1264          * space is cleanly divisible by stripe size.
1265          */
1266         ppl_data_sectors = rdev->ppl.size - (PPL_HEADER_SIZE >> 9);
1267
1268         if (ppl_data_sectors > 0)
1269                 ppl_data_sectors = rounddown(ppl_data_sectors,
1270                                 RAID5_STRIPE_SECTORS((struct r5conf *)rdev->mddev->private));
1271
1272         if (ppl_data_sectors <= 0) {
1273                 pr_warn("md/raid:%s: PPL space too small on %pg\n",
1274                         mdname(rdev->mddev), rdev->bdev);
1275                 return -ENOSPC;
1276         }
1277
1278         ppl_size_new = ppl_data_sectors + (PPL_HEADER_SIZE >> 9);
1279
1280         if ((rdev->ppl.sector < rdev->data_offset &&
1281              rdev->ppl.sector + ppl_size_new > rdev->data_offset) ||
1282             (rdev->ppl.sector >= rdev->data_offset &&
1283              rdev->data_offset + rdev->sectors > rdev->ppl.sector)) {
1284                 pr_warn("md/raid:%s: PPL space overlaps with data on %pg\n",
1285                         mdname(rdev->mddev), rdev->bdev);
1286                 return -EINVAL;
1287         }
1288
1289         if (!rdev->mddev->external &&
1290             ((rdev->ppl.offset > 0 && rdev->ppl.offset < (rdev->sb_size >> 9)) ||
1291              (rdev->ppl.offset <= 0 && rdev->ppl.offset + ppl_size_new > 0))) {
1292                 pr_warn("md/raid:%s: PPL space overlaps with superblock on %pg\n",
1293                         mdname(rdev->mddev), rdev->bdev);
1294                 return -EINVAL;
1295         }
1296
1297         rdev->ppl.size = ppl_size_new;
1298
1299         return 0;
1300 }
1301
1302 static void ppl_init_child_log(struct ppl_log *log, struct md_rdev *rdev)
1303 {
1304         struct request_queue *q;
1305
1306         if ((rdev->ppl.size << 9) >= (PPL_SPACE_SIZE +
1307                                       PPL_HEADER_SIZE) * 2) {
1308                 log->use_multippl = true;
1309                 set_bit(MD_HAS_MULTIPLE_PPLS,
1310                         &log->ppl_conf->mddev->flags);
1311                 log->entry_space = PPL_SPACE_SIZE;
1312         } else {
1313                 log->use_multippl = false;
1314                 log->entry_space = (log->rdev->ppl.size << 9) -
1315                                    PPL_HEADER_SIZE;
1316         }
1317         log->next_io_sector = rdev->ppl.sector;
1318
1319         q = bdev_get_queue(rdev->bdev);
1320         if (test_bit(QUEUE_FLAG_WC, &q->queue_flags))
1321                 log->wb_cache_on = true;
1322 }
1323
1324 int ppl_init_log(struct r5conf *conf)
1325 {
1326         struct ppl_conf *ppl_conf;
1327         struct mddev *mddev = conf->mddev;
1328         int ret = 0;
1329         int max_disks;
1330         int i;
1331
1332         pr_debug("md/raid:%s: enabling distributed Partial Parity Log\n",
1333                  mdname(conf->mddev));
1334
1335         if (PAGE_SIZE != 4096)
1336                 return -EINVAL;
1337
1338         if (mddev->level != 5) {
1339                 pr_warn("md/raid:%s PPL is not compatible with raid level %d\n",
1340                         mdname(mddev), mddev->level);
1341                 return -EINVAL;
1342         }
1343
1344         if (mddev->bitmap_info.file || mddev->bitmap_info.offset) {
1345                 pr_warn("md/raid:%s PPL is not compatible with bitmap\n",
1346                         mdname(mddev));
1347                 return -EINVAL;
1348         }
1349
1350         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
1351                 pr_warn("md/raid:%s PPL is not compatible with journal\n",
1352                         mdname(mddev));
1353                 return -EINVAL;
1354         }
1355
1356         max_disks = sizeof_field(struct ppl_log, disk_flush_bitmap) *
1357                 BITS_PER_BYTE;
1358         if (conf->raid_disks > max_disks) {
1359                 pr_warn("md/raid:%s PPL doesn't support over %d disks in the array\n",
1360                         mdname(mddev), max_disks);
1361                 return -EINVAL;
1362         }
1363
1364         ppl_conf = kzalloc(sizeof(struct ppl_conf), GFP_KERNEL);
1365         if (!ppl_conf)
1366                 return -ENOMEM;
1367
1368         ppl_conf->mddev = mddev;
1369
1370         ppl_conf->io_kc = KMEM_CACHE(ppl_io_unit, 0);
1371         if (!ppl_conf->io_kc) {
1372                 ret = -ENOMEM;
1373                 goto err;
1374         }
1375
1376         ret = mempool_init(&ppl_conf->io_pool, conf->raid_disks, ppl_io_pool_alloc,
1377                            ppl_io_pool_free, ppl_conf->io_kc);
1378         if (ret)
1379                 goto err;
1380
1381         ret = bioset_init(&ppl_conf->bs, conf->raid_disks, 0, BIOSET_NEED_BVECS);
1382         if (ret)
1383                 goto err;
1384
1385         ret = bioset_init(&ppl_conf->flush_bs, conf->raid_disks, 0, 0);
1386         if (ret)
1387                 goto err;
1388
1389         ppl_conf->count = conf->raid_disks;
1390         ppl_conf->child_logs = kcalloc(ppl_conf->count, sizeof(struct ppl_log),
1391                                        GFP_KERNEL);
1392         if (!ppl_conf->child_logs) {
1393                 ret = -ENOMEM;
1394                 goto err;
1395         }
1396
1397         atomic64_set(&ppl_conf->seq, 0);
1398         INIT_LIST_HEAD(&ppl_conf->no_mem_stripes);
1399         spin_lock_init(&ppl_conf->no_mem_stripes_lock);
1400
1401         if (!mddev->external) {
1402                 ppl_conf->signature = ~crc32c_le(~0, mddev->uuid, sizeof(mddev->uuid));
1403                 ppl_conf->block_size = 512;
1404         } else {
1405                 ppl_conf->block_size = queue_logical_block_size(mddev->queue);
1406         }
1407
1408         for (i = 0; i < ppl_conf->count; i++) {
1409                 struct ppl_log *log = &ppl_conf->child_logs[i];
1410                 /* Array has not started so rcu dereference is safe */
1411                 struct md_rdev *rdev =
1412                         rcu_dereference_protected(conf->disks[i].rdev, 1);
1413
1414                 mutex_init(&log->io_mutex);
1415                 spin_lock_init(&log->io_list_lock);
1416                 INIT_LIST_HEAD(&log->io_list);
1417
1418                 log->ppl_conf = ppl_conf;
1419                 log->rdev = rdev;
1420
1421                 if (rdev) {
1422                         ret = ppl_validate_rdev(rdev);
1423                         if (ret)
1424                                 goto err;
1425
1426                         ppl_init_child_log(log, rdev);
1427                 }
1428         }
1429
1430         /* load and possibly recover the logs from the member disks */
1431         ret = ppl_load(ppl_conf);
1432
1433         if (ret) {
1434                 goto err;
1435         } else if (!mddev->pers && mddev->recovery_cp == 0 &&
1436                    ppl_conf->recovered_entries > 0 &&
1437                    ppl_conf->mismatch_count == 0) {
1438                 /*
1439                  * If we are starting a dirty array and the recovery succeeds
1440                  * without any issues, set the array as clean.
1441                  */
1442                 mddev->recovery_cp = MaxSector;
1443                 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
1444         } else if (mddev->pers && ppl_conf->mismatch_count > 0) {
1445                 /* no mismatch allowed when enabling PPL for a running array */
1446                 ret = -EINVAL;
1447                 goto err;
1448         }
1449
1450         conf->log_private = ppl_conf;
1451         set_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1452
1453         return 0;
1454 err:
1455         __ppl_exit_log(ppl_conf);
1456         return ret;
1457 }
1458
1459 int ppl_modify_log(struct r5conf *conf, struct md_rdev *rdev, bool add)
1460 {
1461         struct ppl_conf *ppl_conf = conf->log_private;
1462         struct ppl_log *log;
1463         int ret = 0;
1464
1465         if (!rdev)
1466                 return -EINVAL;
1467
1468         pr_debug("%s: disk: %d operation: %s dev: %pg\n",
1469                  __func__, rdev->raid_disk, add ? "add" : "remove",
1470                  rdev->bdev);
1471
1472         if (rdev->raid_disk < 0)
1473                 return 0;
1474
1475         if (rdev->raid_disk >= ppl_conf->count)
1476                 return -ENODEV;
1477
1478         log = &ppl_conf->child_logs[rdev->raid_disk];
1479
1480         mutex_lock(&log->io_mutex);
1481         if (add) {
1482                 ret = ppl_validate_rdev(rdev);
1483                 if (!ret) {
1484                         log->rdev = rdev;
1485                         ret = ppl_write_empty_header(log);
1486                         ppl_init_child_log(log, rdev);
1487                 }
1488         } else {
1489                 log->rdev = NULL;
1490         }
1491         mutex_unlock(&log->io_mutex);
1492
1493         return ret;
1494 }
1495
1496 static ssize_t
1497 ppl_write_hint_show(struct mddev *mddev, char *buf)
1498 {
1499         return sprintf(buf, "%d\n", 0);
1500 }
1501
1502 static ssize_t
1503 ppl_write_hint_store(struct mddev *mddev, const char *page, size_t len)
1504 {
1505         struct r5conf *conf;
1506         int err = 0;
1507         unsigned short new;
1508
1509         if (len >= PAGE_SIZE)
1510                 return -EINVAL;
1511         if (kstrtou16(page, 10, &new))
1512                 return -EINVAL;
1513
1514         err = mddev_lock(mddev);
1515         if (err)
1516                 return err;
1517
1518         conf = mddev->private;
1519         if (!conf)
1520                 err = -ENODEV;
1521         else if (!raid5_has_ppl(conf) || !conf->log_private)
1522                 err = -EINVAL;
1523
1524         mddev_unlock(mddev);
1525
1526         return err ?: len;
1527 }
1528
1529 struct md_sysfs_entry
1530 ppl_write_hint = __ATTR(ppl_write_hint, S_IRUGO | S_IWUSR,
1531                         ppl_write_hint_show,
1532                         ppl_write_hint_store);