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, io->biovec, PPL_IO_INLINE_BVECS);
 254
 255         pplhdr = page_address(io->header_page);
 256         clear_page(pplhdr);
 257         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
 258         pplhdr->signature = cpu_to_le32(ppl_conf->signature);
 259
 260         io->seq = atomic64_add_return(1, &ppl_conf->seq);
 261         pplhdr->generation = cpu_to_le64(io->seq);
 262
 263         return io;
 264 }
 265
 266 static int ppl_log_stripe(struct ppl_log *log, struct stripe_head *sh)
 267 {
 268         struct ppl_io_unit *io = log->current_io;
 269         struct ppl_header_entry *e = NULL;
 270         struct ppl_header *pplhdr;
 271         int i;
 272         sector_t data_sector = 0;
 273         int data_disks = 0;
 274         struct r5conf *conf = sh->raid_conf;
 275
 276         pr_debug("%s: stripe: %llu\n", __func__, (unsigned long long)sh->sector);
 277
 278         /* check if current io_unit is full */
 279         if (io && (io->pp_size == log->entry_space ||
 280                    io->entries_count == PPL_HDR_MAX_ENTRIES)) {
 281                 pr_debug("%s: add io_unit blocked by seq: %llu\n",
 282                          __func__, io->seq);
 283                 io = NULL;
 284         }
 285
 286         /* add a new unit if there is none or the current is full */
 287         if (!io) {
 288                 io = ppl_new_iounit(log, sh);
 289                 if (!io)
 290                         return -ENOMEM;
 291                 spin_lock_irq(&log->io_list_lock);
 292                 list_add_tail(&io->log_sibling, &log->io_list);
 293                 spin_unlock_irq(&log->io_list_lock);
 294
 295                 log->current_io = io;
 296         }
 297
 298         for (i = 0; i < sh->disks; i++) {
 299                 struct r5dev *dev = &sh->dev[i];
 300
 301                 if (i != sh->pd_idx && test_bit(R5_Wantwrite, &dev->flags)) {
 302                         if (!data_disks || dev->sector < data_sector)
 303                                 data_sector = dev->sector;
 304                         data_disks++;
 305                 }
 306         }
 307         BUG_ON(!data_disks);
 308
 309         pr_debug("%s: seq: %llu data_sector: %llu data_disks: %d\n", __func__,
 310                  io->seq, (unsigned long long)data_sector, data_disks);
 311
 312         pplhdr = page_address(io->header_page);
 313
 314         if (io->entries_count > 0) {
 315                 struct ppl_header_entry *last =
 316                                 &pplhdr->entries[io->entries_count - 1];
 317                 struct stripe_head *sh_last = list_last_entry(
 318                                 &io->stripe_list, struct stripe_head, log_list);
 319                 u64 data_sector_last = le64_to_cpu(last->data_sector);
 320                 u32 data_size_last = le32_to_cpu(last->data_size);
 321
 322                 /*
 323                  * Check if we can append the stripe to the last entry. It must
 324                  * be just after the last logged stripe and write to the same
 325                  * disks. Use bit shift and logarithm to avoid 64-bit division.
 326                  */
 327                 if ((sh->sector == sh_last->sector + RAID5_STRIPE_SECTORS(conf)) &&
 328                     (data_sector >> ilog2(conf->chunk_sectors) ==
 329                      data_sector_last >> ilog2(conf->chunk_sectors)) &&
 330                     ((data_sector - data_sector_last) * data_disks ==
 331                      data_size_last >> 9))
 332                         e = last;
 333         }
 334
 335         if (!e) {
 336                 e = &pplhdr->entries[io->entries_count++];
 337                 e->data_sector = cpu_to_le64(data_sector);
 338                 e->parity_disk = cpu_to_le32(sh->pd_idx);
 339                 e->checksum = cpu_to_le32(~0);
 340         }
 341
 342         le32_add_cpu(&e->data_size, data_disks << PAGE_SHIFT);
 343
 344         /* don't write any PP if full stripe write */
 345         if (!test_bit(STRIPE_FULL_WRITE, &sh->state)) {
 346                 le32_add_cpu(&e->pp_size, PAGE_SIZE);
 347                 io->pp_size += PAGE_SIZE;
 348                 e->checksum = cpu_to_le32(crc32c_le(le32_to_cpu(e->checksum),
 349                                                     page_address(sh->ppl_page),
 350                                                     PAGE_SIZE));
 351         }
 352
 353         list_add_tail(&sh->log_list, &io->stripe_list);
 354         atomic_inc(&io->pending_stripes);
 355         sh->ppl_io = io;
 356
 357         return 0;
 358 }
 359
 360 int ppl_write_stripe(struct r5conf *conf, struct stripe_head *sh)
 361 {
 362         struct ppl_conf *ppl_conf = conf->log_private;
 363         struct ppl_io_unit *io = sh->ppl_io;
 364         struct ppl_log *log;
 365
 366         if (io || test_bit(STRIPE_SYNCING, &sh->state) || !sh->ppl_page ||
 367             !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 368             !test_bit(R5_Insync, &sh->dev[sh->pd_idx].flags)) {
 369                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 370                 return -EAGAIN;
 371         }
 372
 373         log = &ppl_conf->child_logs[sh->pd_idx];
 374
 375         mutex_lock(&log->io_mutex);
 376
 377         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 378                 mutex_unlock(&log->io_mutex);
 379                 return -EAGAIN;
 380         }
 381
 382         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 383         clear_bit(STRIPE_DELAYED, &sh->state);
 384         atomic_inc(&sh->count);
 385
 386         if (ppl_log_stripe(log, sh)) {
 387                 spin_lock_irq(&ppl_conf->no_mem_stripes_lock);
 388                 list_add_tail(&sh->log_list, &ppl_conf->no_mem_stripes);
 389                 spin_unlock_irq(&ppl_conf->no_mem_stripes_lock);
 390         }
 391
 392         mutex_unlock(&log->io_mutex);
 393
 394         return 0;
 395 }
 396
 397 static void ppl_log_endio(struct bio *bio)
 398 {
 399         struct ppl_io_unit *io = bio->bi_private;
 400         struct ppl_log *log = io->log;
 401         struct ppl_conf *ppl_conf = log->ppl_conf;
 402         struct stripe_head *sh, *next;
 403
 404         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 405
 406         if (bio->bi_status)
 407                 md_error(ppl_conf->mddev, log->rdev);
 408
 409         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 410                 list_del_init(&sh->log_list);
 411
 412                 set_bit(STRIPE_HANDLE, &sh->state);
 413                 raid5_release_stripe(sh);
 414         }
 415 }
 416
 417 static void ppl_submit_iounit_bio(struct ppl_io_unit *io, struct bio *bio)
 418 {
 419         char b[BDEVNAME_SIZE];
 420
 421         pr_debug("%s: seq: %llu size: %u sector: %llu dev: %s\n",
 422                  __func__, io->seq, bio->bi_iter.bi_size,
 423                  (unsigned long long)bio->bi_iter.bi_sector,
 424                  bio_devname(bio, b));
 425
 426         submit_bio(bio);
 427 }
 428
 429 static void ppl_submit_iounit(struct ppl_io_unit *io)
 430 {
 431         struct ppl_log *log = io->log;
 432         struct ppl_conf *ppl_conf = log->ppl_conf;
 433         struct ppl_header *pplhdr = page_address(io->header_page);
 434         struct bio *bio = &io->bio;
 435         struct stripe_head *sh;
 436         int i;
 437
 438         bio->bi_private = io;
 439
 440         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 441                 ppl_log_endio(bio);
 442                 return;
 443         }
 444
 445         for (i = 0; i < io->entries_count; i++) {
 446                 struct ppl_header_entry *e = &pplhdr->entries[i];
 447
 448                 pr_debug("%s: seq: %llu entry: %d data_sector: %llu pp_size: %u data_size: %u\n",
 449                          __func__, io->seq, i, le64_to_cpu(e->data_sector),
 450                          le32_to_cpu(e->pp_size), le32_to_cpu(e->data_size));
 451
 452                 e->data_sector = cpu_to_le64(le64_to_cpu(e->data_sector) >>
 453                                              ilog2(ppl_conf->block_size >> 9));
 454                 e->checksum = cpu_to_le32(~le32_to_cpu(e->checksum));
 455         }
 456
 457         pplhdr->entries_count = cpu_to_le32(io->entries_count);
 458         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PPL_HEADER_SIZE));
 459
 460         /* Rewind the buffer if current PPL is larger then remaining space */
 461         if (log->use_multippl &&
 462             log->rdev->ppl.sector + log->rdev->ppl.size - log->next_io_sector <
 463             (PPL_HEADER_SIZE + io->pp_size) >> 9)
 464                 log->next_io_sector = log->rdev->ppl.sector;
 465
 466
 467         bio->bi_end_io = ppl_log_endio;
 468         bio->bi_opf = REQ_OP_WRITE | REQ_FUA;
 469         bio_set_dev(bio, log->rdev->bdev);
 470         bio->bi_iter.bi_sector = log->next_io_sector;
 471         bio_add_page(bio, io->header_page, PAGE_SIZE, 0);
 472         bio->bi_write_hint = ppl_conf->write_hint;
 473
 474         pr_debug("%s: log->current_io_sector: %llu\n", __func__,
 475             (unsigned long long)log->next_io_sector);
 476
 477         if (log->use_multippl)
 478                 log->next_io_sector += (PPL_HEADER_SIZE + io->pp_size) >> 9;
 479
 480         WARN_ON(log->disk_flush_bitmap != 0);
 481
 482         list_for_each_entry(sh, &io->stripe_list, log_list) {
 483                 for (i = 0; i < sh->disks; i++) {
 484                         struct r5dev *dev = &sh->dev[i];
 485
 486                         if ((ppl_conf->child_logs[i].wb_cache_on) &&
 487                             (test_bit(R5_Wantwrite, &dev->flags))) {
 488                                 set_bit(i, &log->disk_flush_bitmap);
 489                         }
 490                 }
 491
 492                 /* entries for full stripe writes have no partial parity */
 493                 if (test_bit(STRIPE_FULL_WRITE, &sh->state))
 494                         continue;
 495
 496                 if (!bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0)) {
 497                         struct bio *prev = bio;
 498
 499                         bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES,
 500                                                &ppl_conf->bs);
 501                         bio->bi_opf = prev->bi_opf;
 502                         bio->bi_write_hint = prev->bi_write_hint;
 503                         bio_copy_dev(bio, prev);
 504                         bio->bi_iter.bi_sector = bio_end_sector(prev);
 505                         bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0);
 506
 507                         bio_chain(bio, prev);
 508                         ppl_submit_iounit_bio(io, prev);
 509                 }
 510         }
 511
 512         ppl_submit_iounit_bio(io, bio);
 513 }
 514
 515 static void ppl_submit_current_io(struct ppl_log *log)
 516 {
 517         struct ppl_io_unit *io;
 518
 519         spin_lock_irq(&log->io_list_lock);
 520
 521         io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
 522                                       log_sibling);
 523         if (io && io->submitted)
 524                 io = NULL;
 525
 526         spin_unlock_irq(&log->io_list_lock);
 527
 528         if (io) {
 529                 io->submitted = true;
 530
 531                 if (io == log->current_io)
 532                         log->current_io = NULL;
 533
 534                 ppl_submit_iounit(io);
 535         }
 536 }
 537
 538 void ppl_write_stripe_run(struct r5conf *conf)
 539 {
 540         struct ppl_conf *ppl_conf = conf->log_private;
 541         struct ppl_log *log;
 542         int i;
 543
 544         for (i = 0; i < ppl_conf->count; i++) {
 545                 log = &ppl_conf->child_logs[i];
 546
 547                 mutex_lock(&log->io_mutex);
 548                 ppl_submit_current_io(log);
 549                 mutex_unlock(&log->io_mutex);
 550         }
 551 }
 552
 553 static void ppl_io_unit_finished(struct ppl_io_unit *io)
 554 {
 555         struct ppl_log *log = io->log;
 556         struct ppl_conf *ppl_conf = log->ppl_conf;
 557         struct r5conf *conf = ppl_conf->mddev->private;
 558         unsigned long flags;
 559
 560         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 561
 562         local_irq_save(flags);
 563
 564         spin_lock(&log->io_list_lock);
 565         list_del(&io->log_sibling);
 566         spin_unlock(&log->io_list_lock);
 567
 568         mempool_free(io, &ppl_conf->io_pool);
 569
 570         spin_lock(&ppl_conf->no_mem_stripes_lock);
 571         if (!list_empty(&ppl_conf->no_mem_stripes)) {
 572                 struct stripe_head *sh;
 573
 574                 sh = list_first_entry(&ppl_conf->no_mem_stripes,
 575                                       struct stripe_head, log_list);
 576                 list_del_init(&sh->log_list);
 577                 set_bit(STRIPE_HANDLE, &sh->state);
 578                 raid5_release_stripe(sh);
 579         }
 580         spin_unlock(&ppl_conf->no_mem_stripes_lock);
 581
 582         local_irq_restore(flags);
 583
 584         wake_up(&conf->wait_for_quiescent);
 585 }
 586
 587 static void ppl_flush_endio(struct bio *bio)
 588 {
 589         struct ppl_io_unit *io = bio->bi_private;
 590         struct ppl_log *log = io->log;
 591         struct ppl_conf *ppl_conf = log->ppl_conf;
 592         struct r5conf *conf = ppl_conf->mddev->private;
 593         char b[BDEVNAME_SIZE];
 594
 595         pr_debug("%s: dev: %s\n", __func__, bio_devname(bio, b));
 596
 597         if (bio->bi_status) {
 598                 struct md_rdev *rdev;
 599
 600                 rcu_read_lock();
 601                 rdev = md_find_rdev_rcu(conf->mddev, bio_dev(bio));
 602                 if (rdev)
 603                         md_error(rdev->mddev, rdev);
 604                 rcu_read_unlock();
 605         }
 606
 607         bio_put(bio);
 608
 609         if (atomic_dec_and_test(&io->pending_flushes)) {
 610                 ppl_io_unit_finished(io);
 611                 md_wakeup_thread(conf->mddev->thread);
 612         }
 613 }
 614
 615 static void ppl_do_flush(struct ppl_io_unit *io)
 616 {
 617         struct ppl_log *log = io->log;
 618         struct ppl_conf *ppl_conf = log->ppl_conf;
 619         struct r5conf *conf = ppl_conf->mddev->private;
 620         int raid_disks = conf->raid_disks;
 621         int flushed_disks = 0;
 622         int i;
 623
 624         atomic_set(&io->pending_flushes, raid_disks);
 625
 626         for_each_set_bit(i, &log->disk_flush_bitmap, raid_disks) {
 627                 struct md_rdev *rdev;
 628                 struct block_device *bdev = NULL;
 629
 630                 rcu_read_lock();
 631                 rdev = rcu_dereference(conf->disks[i].rdev);
 632                 if (rdev && !test_bit(Faulty, &rdev->flags))
 633                         bdev = rdev->bdev;
 634                 rcu_read_unlock();
 635
 636                 if (bdev) {
 637                         struct bio *bio;
 638                         char b[BDEVNAME_SIZE];
 639
 640                         bio = bio_alloc_bioset(GFP_NOIO, 0, &ppl_conf->flush_bs);
 641                         bio_set_dev(bio, bdev);
 642                         bio->bi_private = io;
 643                         bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
 644                         bio->bi_end_io = ppl_flush_endio;
 645
 646                         pr_debug("%s: dev: %s\n", __func__,
 647                                  bio_devname(bio, b));
 648
 649                         submit_bio(bio);
 650                         flushed_disks++;
 651                 }
 652         }
 653
 654         log->disk_flush_bitmap = 0;
 655
 656         for (i = flushed_disks ; i < raid_disks; i++) {
 657                 if (atomic_dec_and_test(&io->pending_flushes))
 658                         ppl_io_unit_finished(io);
 659         }
 660 }
 661
 662 static inline bool ppl_no_io_unit_submitted(struct r5conf *conf,
 663                                             struct ppl_log *log)
 664 {
 665         struct ppl_io_unit *io;
 666
 667         io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
 668                                       log_sibling);
 669
 670         return !io || !io->submitted;
 671 }
 672
 673 void ppl_quiesce(struct r5conf *conf, int quiesce)
 674 {
 675         struct ppl_conf *ppl_conf = conf->log_private;
 676         int i;
 677
 678         if (quiesce) {
 679                 for (i = 0; i < ppl_conf->count; i++) {
 680                         struct ppl_log *log = &ppl_conf->child_logs[i];
 681
 682                         spin_lock_irq(&log->io_list_lock);
 683                         wait_event_lock_irq(conf->wait_for_quiescent,
 684                                             ppl_no_io_unit_submitted(conf, log),
 685                                             log->io_list_lock);
 686                         spin_unlock_irq(&log->io_list_lock);
 687                 }
 688         }
 689 }
 690
 691 int ppl_handle_flush_request(struct r5l_log *log, struct bio *bio)
 692 {
 693         if (bio->bi_iter.bi_size == 0) {
 694                 bio_endio(bio);
 695                 return 0;
 696         }
 697         bio->bi_opf &= ~REQ_PREFLUSH;
 698         return -EAGAIN;
 699 }
 700
 701 void ppl_stripe_write_finished(struct stripe_head *sh)
 702 {
 703         struct ppl_io_unit *io;
 704
 705         io = sh->ppl_io;
 706         sh->ppl_io = NULL;
 707
 708         if (io && atomic_dec_and_test(&io->pending_stripes)) {
 709                 if (io->log->disk_flush_bitmap)
 710                         ppl_do_flush(io);
 711                 else
 712                         ppl_io_unit_finished(io);
 713         }
 714 }
 715
 716 static void ppl_xor(int size, struct page *page1, struct page *page2)
 717 {
 718         struct async_submit_ctl submit;
 719         struct dma_async_tx_descriptor *tx;
 720         struct page *xor_srcs[] = { page1, page2 };
 721
 722         init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST,
 723                           NULL, NULL, NULL, NULL);
 724         tx = async_xor(page1, xor_srcs, 0, 2, size, &submit);
 725
 726         async_tx_quiesce(&tx);
 727 }
 728
 729 /*
 730  * PPL recovery strategy: xor partial parity and data from all modified data
 731  * disks within a stripe and write the result as the new stripe parity. If all
 732  * stripe data disks are modified (full stripe write), no partial parity is
 733  * available, so just xor the data disks.
 734  *
 735  * Recovery of a PPL entry shall occur only if all modified data disks are
 736  * available and read from all of them succeeds.
 737  *
 738  * A PPL entry applies to a stripe, partial parity size for an entry is at most
 739  * the size of the chunk. Examples of possible cases for a single entry:
 740  *
 741  * case 0: single data disk write:
 742  *   data0    data1    data2     ppl        parity
 743  * +--------+--------+--------+           +--------------------+
 744  * | ------ | ------ | ------ | +----+    | (no change)        |
 745  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 746  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 747  * | ------ | ------ | ------ | +----+    | (no change)        |
 748  * +--------+--------+--------+           +--------------------+
 749  * pp_size = data_size
 750  *
 751  * case 1: more than one data disk write:
 752  *   data0    data1    data2     ppl        parity
 753  * +--------+--------+--------+           +--------------------+
 754  * | ------ | ------ | ------ | +----+    | (no change)        |
 755  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 756  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 757  * | ------ | ------ | ------ | +----+    | (no change)        |
 758  * +--------+--------+--------+           +--------------------+
 759  * pp_size = data_size / modified_data_disks
 760  *
 761  * case 2: write to all data disks (also full stripe write):
 762  *   data0    data1    data2                parity
 763  * +--------+--------+--------+           +--------------------+
 764  * | ------ | ------ | ------ |           | (no change)        |
 765  * | -data- | -data- | -data- | --------> | xor all data       |
 766  * | ------ | ------ | ------ | --------> | (no change)        |
 767  * | ------ | ------ | ------ |           | (no change)        |
 768  * +--------+--------+--------+           +--------------------+
 769  * pp_size = 0
 770  *
 771  * The following cases are possible only in other implementations. The recovery
 772  * code can handle them, but they are not generated at runtime because they can
 773  * be reduced to cases 0, 1 and 2:
 774  *
 775  * case 3:
 776  *   data0    data1    data2     ppl        parity
 777  * +--------+--------+--------+ +----+    +--------------------+
 778  * | ------ | -data- | -data- | | pp |    | data1 ^ data2 ^ pp |
 779  * | ------ | -data- | -data- | | pp | -> | data1 ^ data2 ^ pp |
 780  * | -data- | -data- | -data- | | -- | -> | xor all data       |
 781  * | -data- | -data- | ------ | | pp |    | data0 ^ data1 ^ pp |
 782  * +--------+--------+--------+ +----+    +--------------------+
 783  * pp_size = chunk_size
 784  *
 785  * case 4:
 786  *   data0    data1    data2     ppl        parity
 787  * +--------+--------+--------+ +----+    +--------------------+
 788  * | ------ | -data- | ------ | | pp |    | data1 ^ pp         |
 789  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 790  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 791  * | -data- | ------ | ------ | | pp |    | data0 ^ pp         |
 792  * +--------+--------+--------+ +----+    +--------------------+
 793  * pp_size = chunk_size
 794  */
 795 static int ppl_recover_entry(struct ppl_log *log, struct ppl_header_entry *e,
 796                              sector_t ppl_sector)
 797 {
 798         struct ppl_conf *ppl_conf = log->ppl_conf;
 799         struct mddev *mddev = ppl_conf->mddev;
 800         struct r5conf *conf = mddev->private;
 801         int block_size = ppl_conf->block_size;
 802         struct page *page1;
 803         struct page *page2;
 804         sector_t r_sector_first;
 805         sector_t r_sector_last;
 806         int strip_sectors;
 807         int data_disks;
 808         int i;
 809         int ret = 0;
 810         char b[BDEVNAME_SIZE];
 811         unsigned int pp_size = le32_to_cpu(e->pp_size);
 812         unsigned int data_size = le32_to_cpu(e->data_size);
 813
 814         page1 = alloc_page(GFP_KERNEL);
 815         page2 = alloc_page(GFP_KERNEL);
 816
 817         if (!page1 || !page2) {
 818                 ret = -ENOMEM;
 819                 goto out;
 820         }
 821
 822         r_sector_first = le64_to_cpu(e->data_sector) * (block_size >> 9);
 823
 824         if ((pp_size >> 9) < conf->chunk_sectors) {
 825                 if (pp_size > 0) {
 826                         data_disks = data_size / pp_size;
 827                         strip_sectors = pp_size >> 9;
 828                 } else {
 829                         data_disks = conf->raid_disks - conf->max_degraded;
 830                         strip_sectors = (data_size >> 9) / data_disks;
 831                 }
 832                 r_sector_last = r_sector_first +
 833                                 (data_disks - 1) * conf->chunk_sectors +
 834                                 strip_sectors;
 835         } else {
 836                 data_disks = conf->raid_disks - conf->max_degraded;
 837                 strip_sectors = conf->chunk_sectors;
 838                 r_sector_last = r_sector_first + (data_size >> 9);
 839         }
 840
 841         pr_debug("%s: array sector first: %llu last: %llu\n", __func__,
 842                  (unsigned long long)r_sector_first,
 843                  (unsigned long long)r_sector_last);
 844
 845         /* if start and end is 4k aligned, use a 4k block */
 846         if (block_size == 512 &&
 847             (r_sector_first & (RAID5_STRIPE_SECTORS(conf) - 1)) == 0 &&
 848             (r_sector_last & (RAID5_STRIPE_SECTORS(conf) - 1)) == 0)
 849                 block_size = RAID5_STRIPE_SIZE(conf);
 850
 851         /* iterate through blocks in strip */
 852         for (i = 0; i < strip_sectors; i += (block_size >> 9)) {
 853                 bool update_parity = false;
 854                 sector_t parity_sector;
 855                 struct md_rdev *parity_rdev;
 856                 struct stripe_head sh;
 857                 int disk;
 858                 int indent = 0;
 859
 860                 pr_debug("%s:%*s iter %d start\n", __func__, indent, "", i);
 861                 indent += 2;
 862
 863                 memset(page_address(page1), 0, PAGE_SIZE);
 864
 865                 /* iterate through data member disks */
 866                 for (disk = 0; disk < data_disks; disk++) {
 867                         int dd_idx;
 868                         struct md_rdev *rdev;
 869                         sector_t sector;
 870                         sector_t r_sector = r_sector_first + i +
 871                                             (disk * conf->chunk_sectors);
 872
 873                         pr_debug("%s:%*s data member disk %d start\n",
 874                                  __func__, indent, "", disk);
 875                         indent += 2;
 876
 877                         if (r_sector >= r_sector_last) {
 878                                 pr_debug("%s:%*s array sector %llu doesn't need parity update\n",
 879                                          __func__, indent, "",
 880                                          (unsigned long long)r_sector);
 881                                 indent -= 2;
 882                                 continue;
 883                         }
 884
 885                         update_parity = true;
 886
 887                         /* map raid sector to member disk */
 888                         sector = raid5_compute_sector(conf, r_sector, 0,
 889                                                       &dd_idx, NULL);
 890                         pr_debug("%s:%*s processing array sector %llu => data member disk %d, sector %llu\n",
 891                                  __func__, indent, "",
 892                                  (unsigned long long)r_sector, dd_idx,
 893                                  (unsigned long long)sector);
 894
 895                         rdev = conf->disks[dd_idx].rdev;
 896                         if (!rdev || (!test_bit(In_sync, &rdev->flags) &&
 897                                       sector >= rdev->recovery_offset)) {
 898                                 pr_debug("%s:%*s data member disk %d missing\n",
 899                                          __func__, indent, "", dd_idx);
 900                                 update_parity = false;
 901                                 break;
 902                         }
 903
 904                         pr_debug("%s:%*s reading data member disk %s sector %llu\n",
 905                                  __func__, indent, "", bdevname(rdev->bdev, b),
 906                                  (unsigned long long)sector);
 907                         if (!sync_page_io(rdev, sector, block_size, page2,
 908                                         REQ_OP_READ, 0, false)) {
 909                                 md_error(mddev, rdev);
 910                                 pr_debug("%s:%*s read failed!\n", __func__,
 911                                          indent, "");
 912                                 ret = -EIO;
 913                                 goto out;
 914                         }
 915
 916                         ppl_xor(block_size, page1, page2);
 917
 918                         indent -= 2;
 919                 }
 920
 921                 if (!update_parity)
 922                         continue;
 923
 924                 if (pp_size > 0) {
 925                         pr_debug("%s:%*s reading pp disk sector %llu\n",
 926                                  __func__, indent, "",
 927                                  (unsigned long long)(ppl_sector + i));
 928                         if (!sync_page_io(log->rdev,
 929                                         ppl_sector - log->rdev->data_offset + i,
 930                                         block_size, page2, REQ_OP_READ, 0,
 931                                         false)) {
 932                                 pr_debug("%s:%*s read failed!\n", __func__,
 933                                          indent, "");
 934                                 md_error(mddev, log->rdev);
 935                                 ret = -EIO;
 936                                 goto out;
 937                         }
 938
 939                         ppl_xor(block_size, page1, page2);
 940                 }
 941
 942                 /* map raid sector to parity disk */
 943                 parity_sector = raid5_compute_sector(conf, r_sector_first + i,
 944                                 0, &disk, &sh);
 945                 BUG_ON(sh.pd_idx != le32_to_cpu(e->parity_disk));
 946                 parity_rdev = conf->disks[sh.pd_idx].rdev;
 947
 948                 BUG_ON(parity_rdev->bdev->bd_dev != log->rdev->bdev->bd_dev);
 949                 pr_debug("%s:%*s write parity at sector %llu, disk %s\n",
 950                          __func__, indent, "",
 951                          (unsigned long long)parity_sector,
 952                          bdevname(parity_rdev->bdev, b));
 953                 if (!sync_page_io(parity_rdev, parity_sector, block_size,
 954                                 page1, REQ_OP_WRITE, 0, false)) {
 955                         pr_debug("%s:%*s parity write error!\n", __func__,
 956                                  indent, "");
 957                         md_error(mddev, parity_rdev);
 958                         ret = -EIO;
 959                         goto out;
 960                 }
 961         }
 962 out:
 963         if (page1)
 964                 __free_page(page1);
 965         if (page2)
 966                 __free_page(page2);
 967         return ret;
 968 }
 969
 970 static int ppl_recover(struct ppl_log *log, struct ppl_header *pplhdr,
 971                        sector_t offset)
 972 {
 973         struct ppl_conf *ppl_conf = log->ppl_conf;
 974         struct md_rdev *rdev = log->rdev;
 975         struct mddev *mddev = rdev->mddev;
 976         sector_t ppl_sector = rdev->ppl.sector + offset +
 977                               (PPL_HEADER_SIZE >> 9);
 978         struct page *page;
 979         int i;
 980         int ret = 0;
 981
 982         page = alloc_page(GFP_KERNEL);
 983         if (!page)
 984                 return -ENOMEM;
 985
 986         /* iterate through all PPL entries saved */
 987         for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++) {
 988                 struct ppl_header_entry *e = &pplhdr->entries[i];
 989                 u32 pp_size = le32_to_cpu(e->pp_size);
 990                 sector_t sector = ppl_sector;
 991                 int ppl_entry_sectors = pp_size >> 9;
 992                 u32 crc, crc_stored;
 993
 994                 pr_debug("%s: disk: %d entry: %d ppl_sector: %llu pp_size: %u\n",
 995                          __func__, rdev->raid_disk, i,
 996                          (unsigned long long)ppl_sector, pp_size);
 997
 998                 crc = ~0;
 999                 crc_stored = le32_to_cpu(e->checksum);
1000
1001                 /* read parial parity for this entry and calculate its checksum */
1002                 while (pp_size) {
1003                         int s = pp_size > PAGE_SIZE ? PAGE_SIZE : pp_size;
1004
1005                         if (!sync_page_io(rdev, sector - rdev->data_offset,
1006                                         s, page, REQ_OP_READ, 0, false)) {
1007                                 md_error(mddev, rdev);
1008                                 ret = -EIO;
1009                                 goto out;
1010                         }
1011
1012                         crc = crc32c_le(crc, page_address(page), s);
1013
1014                         pp_size -= s;
1015                         sector += s >> 9;
1016                 }
1017
1018                 crc = ~crc;
1019
1020                 if (crc != crc_stored) {
1021                         /*
1022                          * Don't recover this entry if the checksum does not
1023                          * match, but keep going and try to recover other
1024                          * entries.
1025                          */
1026                         pr_debug("%s: ppl entry crc does not match: stored: 0x%x calculated: 0x%x\n",
1027                                  __func__, crc_stored, crc);
1028                         ppl_conf->mismatch_count++;
1029                 } else {
1030                         ret = ppl_recover_entry(log, e, ppl_sector);
1031                         if (ret)
1032                                 goto out;
1033                         ppl_conf->recovered_entries++;
1034                 }
1035
1036                 ppl_sector += ppl_entry_sectors;
1037         }
1038
1039         /* flush the disk cache after recovery if necessary */
1040         ret = blkdev_issue_flush(rdev->bdev, GFP_KERNEL);
1041 out:
1042         __free_page(page);
1043         return ret;
1044 }
1045
1046 static int ppl_write_empty_header(struct ppl_log *log)
1047 {
1048         struct page *page;
1049         struct ppl_header *pplhdr;
1050         struct md_rdev *rdev = log->rdev;
1051         int ret = 0;
1052
1053         pr_debug("%s: disk: %d ppl_sector: %llu\n", __func__,
1054                  rdev->raid_disk, (unsigned long long)rdev->ppl.sector);
1055
1056         page = alloc_page(GFP_NOIO | __GFP_ZERO);
1057         if (!page)
1058                 return -ENOMEM;
1059
1060         pplhdr = page_address(page);
1061         /* zero out PPL space to avoid collision with old PPLs */
1062         blkdev_issue_zeroout(rdev->bdev, rdev->ppl.sector,
1063                             log->rdev->ppl.size, GFP_NOIO, 0);
1064         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
1065         pplhdr->signature = cpu_to_le32(log->ppl_conf->signature);
1066         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PAGE_SIZE));
1067
1068         if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset,
1069                           PPL_HEADER_SIZE, page, REQ_OP_WRITE | REQ_SYNC |
1070                           REQ_FUA, 0, false)) {
1071                 md_error(rdev->mddev, rdev);
1072                 ret = -EIO;
1073         }
1074
1075         __free_page(page);
1076         return ret;
1077 }
1078
1079 static int ppl_load_distributed(struct ppl_log *log)
1080 {
1081         struct ppl_conf *ppl_conf = log->ppl_conf;
1082         struct md_rdev *rdev = log->rdev;
1083         struct mddev *mddev = rdev->mddev;
1084         struct page *page, *page2, *tmp;
1085         struct ppl_header *pplhdr = NULL, *prev_pplhdr = NULL;
1086         u32 crc, crc_stored;
1087         u32 signature;
1088         int ret = 0, i;
1089         sector_t pplhdr_offset = 0, prev_pplhdr_offset = 0;
1090
1091         pr_debug("%s: disk: %d\n", __func__, rdev->raid_disk);
1092         /* read PPL headers, find the recent one */
1093         page = alloc_page(GFP_KERNEL);
1094         if (!page)
1095                 return -ENOMEM;
1096
1097         page2 = alloc_page(GFP_KERNEL);
1098         if (!page2) {
1099                 __free_page(page);
1100                 return -ENOMEM;
1101         }
1102
1103         /* searching ppl area for latest ppl */
1104         while (pplhdr_offset < rdev->ppl.size - (PPL_HEADER_SIZE >> 9)) {
1105                 if (!sync_page_io(rdev,
1106                                   rdev->ppl.sector - rdev->data_offset +
1107                                   pplhdr_offset, PAGE_SIZE, page, REQ_OP_READ,
1108                                   0, false)) {
1109                         md_error(mddev, rdev);
1110                         ret = -EIO;
1111                         /* if not able to read - don't recover any PPL */
1112                         pplhdr = NULL;
1113                         break;
1114                 }
1115                 pplhdr = page_address(page);
1116
1117                 /* check header validity */
1118                 crc_stored = le32_to_cpu(pplhdr->checksum);
1119                 pplhdr->checksum = 0;
1120                 crc = ~crc32c_le(~0, pplhdr, PAGE_SIZE);
1121
1122                 if (crc_stored != crc) {
1123                         pr_debug("%s: ppl header crc does not match: stored: 0x%x calculated: 0x%x (offset: %llu)\n",
1124                                  __func__, crc_stored, crc,
1125                                  (unsigned long long)pplhdr_offset);
1126                         pplhdr = prev_pplhdr;
1127                         pplhdr_offset = prev_pplhdr_offset;
1128                         break;
1129                 }
1130
1131                 signature = le32_to_cpu(pplhdr->signature);
1132
1133                 if (mddev->external) {
1134                         /*
1135                          * For external metadata the header signature is set and
1136                          * validated in userspace.
1137                          */
1138                         ppl_conf->signature = signature;
1139                 } else if (ppl_conf->signature != signature) {
1140                         pr_debug("%s: ppl header signature does not match: stored: 0x%x configured: 0x%x (offset: %llu)\n",
1141                                  __func__, signature, ppl_conf->signature,
1142                                  (unsigned long long)pplhdr_offset);
1143                         pplhdr = prev_pplhdr;
1144                         pplhdr_offset = prev_pplhdr_offset;
1145                         break;
1146                 }
1147
1148                 if (prev_pplhdr && le64_to_cpu(prev_pplhdr->generation) >
1149                     le64_to_cpu(pplhdr->generation)) {
1150                         /* previous was newest */
1151                         pplhdr = prev_pplhdr;
1152                         pplhdr_offset = prev_pplhdr_offset;
1153                         break;
1154                 }
1155
1156                 prev_pplhdr_offset = pplhdr_offset;
1157                 prev_pplhdr = pplhdr;
1158
1159                 tmp = page;
1160                 page = page2;
1161                 page2 = tmp;
1162
1163                 /* calculate next potential ppl offset */
1164                 for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++)
1165                         pplhdr_offset +=
1166                             le32_to_cpu(pplhdr->entries[i].pp_size) >> 9;
1167                 pplhdr_offset += PPL_HEADER_SIZE >> 9;
1168         }
1169
1170         /* no valid ppl found */
1171         if (!pplhdr)
1172                 ppl_conf->mismatch_count++;
1173         else
1174                 pr_debug("%s: latest PPL found at offset: %llu, with generation: %llu\n",
1175                     __func__, (unsigned long long)pplhdr_offset,
1176                     le64_to_cpu(pplhdr->generation));
1177
1178         /* attempt to recover from log if we are starting a dirty array */
1179         if (pplhdr && !mddev->pers && mddev->recovery_cp != MaxSector)
1180                 ret = ppl_recover(log, pplhdr, pplhdr_offset);
1181
1182         /* write empty header if we are starting the array */
1183         if (!ret && !mddev->pers)
1184                 ret = ppl_write_empty_header(log);
1185
1186         __free_page(page);
1187         __free_page(page2);
1188
1189         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1190                  __func__, ret, ppl_conf->mismatch_count,
1191                  ppl_conf->recovered_entries);
1192         return ret;
1193 }
1194
1195 static int ppl_load(struct ppl_conf *ppl_conf)
1196 {
1197         int ret = 0;
1198         u32 signature = 0;
1199         bool signature_set = false;
1200         int i;
1201
1202         for (i = 0; i < ppl_conf->count; i++) {
1203                 struct ppl_log *log = &ppl_conf->child_logs[i];
1204
1205                 /* skip missing drive */
1206                 if (!log->rdev)
1207                         continue;
1208
1209                 ret = ppl_load_distributed(log);
1210                 if (ret)
1211                         break;
1212
1213                 /*
1214                  * For external metadata we can't check if the signature is
1215                  * correct on a single drive, but we can check if it is the same
1216                  * on all drives.
1217                  */
1218                 if (ppl_conf->mddev->external) {
1219                         if (!signature_set) {
1220                                 signature = ppl_conf->signature;
1221                                 signature_set = true;
1222                         } else if (signature != ppl_conf->signature) {
1223                                 pr_warn("md/raid:%s: PPL header signature does not match on all member drives\n",
1224                                         mdname(ppl_conf->mddev));
1225                                 ret = -EINVAL;
1226                                 break;
1227                         }
1228                 }
1229         }
1230
1231         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1232                  __func__, ret, ppl_conf->mismatch_count,
1233                  ppl_conf->recovered_entries);
1234         return ret;
1235 }
1236
1237 static void __ppl_exit_log(struct ppl_conf *ppl_conf)
1238 {
1239         clear_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1240         clear_bit(MD_HAS_MULTIPLE_PPLS, &ppl_conf->mddev->flags);
1241
1242         kfree(ppl_conf->child_logs);
1243
1244         bioset_exit(&ppl_conf->bs);
1245         bioset_exit(&ppl_conf->flush_bs);
1246         mempool_exit(&ppl_conf->io_pool);
1247         kmem_cache_destroy(ppl_conf->io_kc);
1248
1249         kfree(ppl_conf);
1250 }
1251
1252 void ppl_exit_log(struct r5conf *conf)
1253 {
1254         struct ppl_conf *ppl_conf = conf->log_private;
1255
1256         if (ppl_conf) {
1257                 __ppl_exit_log(ppl_conf);
1258                 conf->log_private = NULL;
1259         }
1260 }
1261
1262 static int ppl_validate_rdev(struct md_rdev *rdev)
1263 {
1264         char b[BDEVNAME_SIZE];
1265         int ppl_data_sectors;
1266         int ppl_size_new;
1267
1268         /*
1269          * The configured PPL size must be enough to store
1270          * the header and (at the very least) partial parity
1271          * for one stripe. Round it down to ensure the data
1272          * space is cleanly divisible by stripe size.
1273          */
1274         ppl_data_sectors = rdev->ppl.size - (PPL_HEADER_SIZE >> 9);
1275
1276         if (ppl_data_sectors > 0)
1277                 ppl_data_sectors = rounddown(ppl_data_sectors,
1278                                 RAID5_STRIPE_SECTORS((struct r5conf *)rdev->mddev->private));
1279
1280         if (ppl_data_sectors <= 0) {
1281                 pr_warn("md/raid:%s: PPL space too small on %s\n",
1282                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1283                 return -ENOSPC;
1284         }
1285
1286         ppl_size_new = ppl_data_sectors + (PPL_HEADER_SIZE >> 9);
1287
1288         if ((rdev->ppl.sector < rdev->data_offset &&
1289              rdev->ppl.sector + ppl_size_new > rdev->data_offset) ||
1290             (rdev->ppl.sector >= rdev->data_offset &&
1291              rdev->data_offset + rdev->sectors > rdev->ppl.sector)) {
1292                 pr_warn("md/raid:%s: PPL space overlaps with data on %s\n",
1293                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1294                 return -EINVAL;
1295         }
1296
1297         if (!rdev->mddev->external &&
1298             ((rdev->ppl.offset > 0 && rdev->ppl.offset < (rdev->sb_size >> 9)) ||
1299              (rdev->ppl.offset <= 0 && rdev->ppl.offset + ppl_size_new > 0))) {
1300                 pr_warn("md/raid:%s: PPL space overlaps with superblock on %s\n",
1301                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1302                 return -EINVAL;
1303         }
1304
1305         rdev->ppl.size = ppl_size_new;
1306
1307         return 0;
1308 }
1309
1310 static void ppl_init_child_log(struct ppl_log *log, struct md_rdev *rdev)
1311 {
1312         struct request_queue *q;
1313
1314         if ((rdev->ppl.size << 9) >= (PPL_SPACE_SIZE +
1315                                       PPL_HEADER_SIZE) * 2) {
1316                 log->use_multippl = true;
1317                 set_bit(MD_HAS_MULTIPLE_PPLS,
1318                         &log->ppl_conf->mddev->flags);
1319                 log->entry_space = PPL_SPACE_SIZE;
1320         } else {
1321                 log->use_multippl = false;
1322                 log->entry_space = (log->rdev->ppl.size << 9) -
1323                                    PPL_HEADER_SIZE;
1324         }
1325         log->next_io_sector = rdev->ppl.sector;
1326
1327         q = bdev_get_queue(rdev->bdev);
1328         if (test_bit(QUEUE_FLAG_WC, &q->queue_flags))
1329                 log->wb_cache_on = true;
1330 }
1331
1332 int ppl_init_log(struct r5conf *conf)
1333 {
1334         struct ppl_conf *ppl_conf;
1335         struct mddev *mddev = conf->mddev;
1336         int ret = 0;
1337         int max_disks;
1338         int i;
1339
1340         pr_debug("md/raid:%s: enabling distributed Partial Parity Log\n",
1341                  mdname(conf->mddev));
1342
1343         if (PAGE_SIZE != 4096)
1344                 return -EINVAL;
1345
1346         if (mddev->level != 5) {
1347                 pr_warn("md/raid:%s PPL is not compatible with raid level %d\n",
1348                         mdname(mddev), mddev->level);
1349                 return -EINVAL;
1350         }
1351
1352         if (mddev->bitmap_info.file || mddev->bitmap_info.offset) {
1353                 pr_warn("md/raid:%s PPL is not compatible with bitmap\n",
1354                         mdname(mddev));
1355                 return -EINVAL;
1356         }
1357
1358         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
1359                 pr_warn("md/raid:%s PPL is not compatible with journal\n",
1360                         mdname(mddev));
1361                 return -EINVAL;
1362         }
1363
1364         max_disks = sizeof_field(struct ppl_log, disk_flush_bitmap) *
1365                 BITS_PER_BYTE;
1366         if (conf->raid_disks > max_disks) {
1367                 pr_warn("md/raid:%s PPL doesn't support over %d disks in the array\n",
1368                         mdname(mddev), max_disks);
1369                 return -EINVAL;
1370         }
1371
1372         ppl_conf = kzalloc(sizeof(struct ppl_conf), GFP_KERNEL);
1373         if (!ppl_conf)
1374                 return -ENOMEM;
1375
1376         ppl_conf->mddev = mddev;
1377
1378         ppl_conf->io_kc = KMEM_CACHE(ppl_io_unit, 0);
1379         if (!ppl_conf->io_kc) {
1380                 ret = -ENOMEM;
1381                 goto err;
1382         }
1383
1384         ret = mempool_init(&ppl_conf->io_pool, conf->raid_disks, ppl_io_pool_alloc,
1385                            ppl_io_pool_free, ppl_conf->io_kc);
1386         if (ret)
1387                 goto err;
1388
1389         ret = bioset_init(&ppl_conf->bs, conf->raid_disks, 0, BIOSET_NEED_BVECS);
1390         if (ret)
1391                 goto err;
1392
1393         ret = bioset_init(&ppl_conf->flush_bs, conf->raid_disks, 0, 0);
1394         if (ret)
1395                 goto err;
1396
1397         ppl_conf->count = conf->raid_disks;
1398         ppl_conf->child_logs = kcalloc(ppl_conf->count, sizeof(struct ppl_log),
1399                                        GFP_KERNEL);
1400         if (!ppl_conf->child_logs) {
1401                 ret = -ENOMEM;
1402                 goto err;
1403         }
1404
1405         atomic64_set(&ppl_conf->seq, 0);
1406         INIT_LIST_HEAD(&ppl_conf->no_mem_stripes);
1407         spin_lock_init(&ppl_conf->no_mem_stripes_lock);
1408         ppl_conf->write_hint = RWH_WRITE_LIFE_NOT_SET;
1409
1410         if (!mddev->external) {
1411                 ppl_conf->signature = ~crc32c_le(~0, mddev->uuid, sizeof(mddev->uuid));
1412                 ppl_conf->block_size = 512;
1413         } else {
1414                 ppl_conf->block_size = queue_logical_block_size(mddev->queue);
1415         }
1416
1417         for (i = 0; i < ppl_conf->count; i++) {
1418                 struct ppl_log *log = &ppl_conf->child_logs[i];
1419                 struct md_rdev *rdev = conf->disks[i].rdev;
1420
1421                 mutex_init(&log->io_mutex);
1422                 spin_lock_init(&log->io_list_lock);
1423                 INIT_LIST_HEAD(&log->io_list);
1424
1425                 log->ppl_conf = ppl_conf;
1426                 log->rdev = rdev;
1427
1428                 if (rdev) {
1429                         ret = ppl_validate_rdev(rdev);
1430                         if (ret)
1431                                 goto err;
1432
1433                         ppl_init_child_log(log, rdev);
1434                 }
1435         }
1436
1437         /* load and possibly recover the logs from the member disks */
1438         ret = ppl_load(ppl_conf);
1439
1440         if (ret) {
1441                 goto err;
1442         } else if (!mddev->pers && mddev->recovery_cp == 0 &&
1443                    ppl_conf->recovered_entries > 0 &&
1444                    ppl_conf->mismatch_count == 0) {
1445                 /*
1446                  * If we are starting a dirty array and the recovery succeeds
1447                  * without any issues, set the array as clean.
1448                  */
1449                 mddev->recovery_cp = MaxSector;
1450                 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
1451         } else if (mddev->pers && ppl_conf->mismatch_count > 0) {
1452                 /* no mismatch allowed when enabling PPL for a running array */
1453                 ret = -EINVAL;
1454                 goto err;
1455         }
1456
1457         conf->log_private = ppl_conf;
1458         set_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1459
1460         return 0;
1461 err:
1462         __ppl_exit_log(ppl_conf);
1463         return ret;
1464 }
1465
1466 int ppl_modify_log(struct r5conf *conf, struct md_rdev *rdev, bool add)
1467 {
1468         struct ppl_conf *ppl_conf = conf->log_private;
1469         struct ppl_log *log;
1470         int ret = 0;
1471         char b[BDEVNAME_SIZE];
1472
1473         if (!rdev)
1474                 return -EINVAL;
1475
1476         pr_debug("%s: disk: %d operation: %s dev: %s\n",
1477                  __func__, rdev->raid_disk, add ? "add" : "remove",
1478                  bdevname(rdev->bdev, b));
1479
1480         if (rdev->raid_disk < 0)
1481                 return 0;
1482
1483         if (rdev->raid_disk >= ppl_conf->count)
1484                 return -ENODEV;
1485
1486         log = &ppl_conf->child_logs[rdev->raid_disk];
1487
1488         mutex_lock(&log->io_mutex);
1489         if (add) {
1490                 ret = ppl_validate_rdev(rdev);
1491                 if (!ret) {
1492                         log->rdev = rdev;
1493                         ret = ppl_write_empty_header(log);
1494                         ppl_init_child_log(log, rdev);
1495                 }
1496         } else {
1497                 log->rdev = NULL;
1498         }
1499         mutex_unlock(&log->io_mutex);
1500
1501         return ret;
1502 }
1503
1504 static ssize_t
1505 ppl_write_hint_show(struct mddev *mddev, char *buf)
1506 {
1507         size_t ret = 0;
1508         struct r5conf *conf;
1509         struct ppl_conf *ppl_conf = NULL;
1510
1511         spin_lock(&mddev->lock);
1512         conf = mddev->private;
1513         if (conf && raid5_has_ppl(conf))
1514                 ppl_conf = conf->log_private;
1515         ret = sprintf(buf, "%d\n", ppl_conf ? ppl_conf->write_hint : 0);
1516         spin_unlock(&mddev->lock);
1517
1518         return ret;
1519 }
1520
1521 static ssize_t
1522 ppl_write_hint_store(struct mddev *mddev, const char *page, size_t len)
1523 {
1524         struct r5conf *conf;
1525         struct ppl_conf *ppl_conf;
1526         int err = 0;
1527         unsigned short new;
1528
1529         if (len >= PAGE_SIZE)
1530                 return -EINVAL;
1531         if (kstrtou16(page, 10, &new))
1532                 return -EINVAL;
1533
1534         err = mddev_lock(mddev);
1535         if (err)
1536                 return err;
1537
1538         conf = mddev->private;
1539         if (!conf) {
1540                 err = -ENODEV;
1541         } else if (raid5_has_ppl(conf)) {
1542                 ppl_conf = conf->log_private;
1543                 if (!ppl_conf)
1544                         err = -EINVAL;
1545                 else
1546                         ppl_conf->write_hint = new;
1547         } else {
1548                 err = -EINVAL;
1549         }
1550
1551         mddev_unlock(mddev);
1552
1553         return err ?: len;
1554 }
1555
1556 struct md_sysfs_entry
1557 ppl_write_hint = __ATTR(ppl_write_hint, S_IRUGO | S_IWUSR,
1558                         ppl_write_hint_show,
1559                         ppl_write_hint_store);