drivers/md/raid5-ppl.c

   1 /*
   2  * Partial Parity Log for closing the RAID5 write hole
   3  * Copyright (c) 2017, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14
  15 #include <linux/kernel.h>
  16 #include <linux/blkdev.h>
  17 #include <linux/slab.h>
  18 #include <linux/crc32c.h>
  19 #include <linux/async_tx.h>
  20 #include <linux/raid/md_p.h>
  21 #include "md.h"
  22 #include "raid5.h"
  23 #include "raid5-log.h"
  24
  25 /*
  26  * PPL consists of a 4KB header (struct ppl_header) and at least 128KB for
  27  * partial parity data. The header contains an array of entries
  28  * (struct ppl_header_entry) which describe the logged write requests.
  29  * Partial parity for the entries comes after the header, written in the same
  30  * sequence as the entries:
  31  *
  32  * Header
  33  *   entry0
  34  *   ...
  35  *   entryN
  36  * PP data
  37  *   PP for entry0
  38  *   ...
  39  *   PP for entryN
  40  *
  41  * An entry describes one or more consecutive stripe_heads, up to a full
  42  * stripe. The modifed raid data chunks form an m-by-n matrix, where m is the
  43  * number of stripe_heads in the entry and n is the number of modified data
  44  * disks. Every stripe_head in the entry must write to the same data disks.
  45  * An example of a valid case described by a single entry (writes to the first
  46  * stripe of a 4 disk array, 16k chunk size):
  47  *
  48  * sh->sector   dd0   dd1   dd2    ppl
  49  *            +-----+-----+-----+
  50  * 0          | --- | --- | --- | +----+
  51  * 8          | -W- | -W- | --- | | pp |   data_sector = 8
  52  * 16         | -W- | -W- | --- | | pp |   data_size = 3 * 2 * 4k
  53  * 24         | -W- | -W- | --- | | pp |   pp_size = 3 * 4k
  54  *            +-----+-----+-----+ +----+
  55  *
  56  * data_sector is the first raid sector of the modified data, data_size is the
  57  * total size of modified data and pp_size is the size of partial parity for
  58  * this entry. Entries for full stripe writes contain no partial parity
  59  * (pp_size = 0), they only mark the stripes for which parity should be
  60  * recalculated after an unclean shutdown. Every entry holds a checksum of its
  61  * partial parity, the header also has a checksum of the header itself.
  62  *
  63  * A write request is always logged to the PPL instance stored on the parity
  64  * disk of the corresponding stripe. For each member disk there is one ppl_log
  65  * used to handle logging for this disk, independently from others. They are
  66  * grouped in child_logs array in struct ppl_conf, which is assigned to
  67  * r5conf->log_private.
  68  *
  69  * ppl_io_unit represents a full PPL write, header_page contains the ppl_header.
  70  * PPL entries for logged stripes are added in ppl_log_stripe(). A stripe_head
  71  * can be appended to the last entry if it meets the conditions for a valid
  72  * entry described above, otherwise a new entry is added. Checksums of entries
  73  * are calculated incrementally as stripes containing partial parity are being
  74  * added. ppl_submit_iounit() calculates the checksum of the header and submits
  75  * a bio containing the header page and partial parity pages (sh->ppl_page) for
  76  * all stripes of the io_unit. When the PPL write completes, the stripes
  77  * associated with the io_unit are released and raid5d starts writing their data
  78  * and parity. When all stripes are written, the io_unit is freed and the next
  79  * can be submitted.
  80  *
  81  * An io_unit is used to gather stripes until it is submitted or becomes full
  82  * (if the maximum number of entries or size of PPL is reached). Another io_unit
  83  * can't be submitted until the previous has completed (PPL and stripe
  84  * data+parity is written). The log->io_list tracks all io_units of a log
  85  * (for a single member disk). New io_units are added to the end of the list
  86  * and the first io_unit is submitted, if it is not submitted already.
  87  * The current io_unit accepting new stripes is always at the end of the list.
  88  *
  89  * If write-back cache is enabled for any of the disks in the array, its data
  90  * must be flushed before next io_unit is submitted.
  91  */
  92
  93 #define PPL_SPACE_SIZE (128 * 1024)
  94
  95 struct ppl_conf {
  96         struct mddev *mddev;
  97
  98         /* array of child logs, one for each raid disk */
  99         struct ppl_log *child_logs;
 100         int count;
 101
 102         int block_size;         /* the logical block size used for data_sector
 103                                  * in ppl_header_entry */
 104         u32 signature;          /* raid array identifier */
 105         atomic64_t seq;         /* current log write sequence number */
 106
 107         struct kmem_cache *io_kc;
 108         mempool_t io_pool;
 109         struct bio_set bs;
 110         struct bio_set flush_bs;
 111
 112         /* used only for recovery */
 113         int recovered_entries;
 114         int mismatch_count;
 115
 116         /* stripes to retry if failed to allocate io_unit */
 117         struct list_head no_mem_stripes;
 118         spinlock_t no_mem_stripes_lock;
 119
 120         unsigned short write_hint;
 121 };
 122
 123 struct ppl_log {
 124         struct ppl_conf *ppl_conf;      /* shared between all log instances */
 125
 126         struct md_rdev *rdev;           /* array member disk associated with
 127                                          * this log instance */
 128         struct mutex io_mutex;
 129         struct ppl_io_unit *current_io; /* current io_unit accepting new data
 130                                          * always at the end of io_list */
 131         spinlock_t io_list_lock;
 132         struct list_head io_list;       /* all io_units of this log */
 133
 134         sector_t next_io_sector;
 135         unsigned int entry_space;
 136         bool use_multippl;
 137         bool wb_cache_on;
 138         unsigned long disk_flush_bitmap;
 139 };
 140
 141 #define PPL_IO_INLINE_BVECS 32
 142
 143 struct ppl_io_unit {
 144         struct ppl_log *log;
 145
 146         struct page *header_page;       /* for ppl_header */
 147
 148         unsigned int entries_count;     /* number of entries in ppl_header */
 149         unsigned int pp_size;           /* total size current of partial parity */
 150
 151         u64 seq;                        /* sequence number of this log write */
 152         struct list_head log_sibling;   /* log->io_list */
 153
 154         struct list_head stripe_list;   /* stripes added to the io_unit */
 155         atomic_t pending_stripes;       /* how many stripes not written to raid */
 156         atomic_t pending_flushes;       /* how many disk flushes are in progress */
 157
 158         bool submitted;                 /* true if write to log started */
 159
 160         /* inline bio and its biovec for submitting the iounit */
 161         struct bio bio;
 162         struct bio_vec biovec[PPL_IO_INLINE_BVECS];
 163 };
 164
 165 struct dma_async_tx_descriptor *
 166 ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu,
 167                        struct dma_async_tx_descriptor *tx)
 168 {
 169         int disks = sh->disks;
 170         struct page **srcs = percpu->scribble;
 171         int count = 0, pd_idx = sh->pd_idx, i;
 172         struct async_submit_ctl submit;
 173
 174         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
 175
 176         /*
 177          * Partial parity is the XOR of stripe data chunks that are not changed
 178          * during the write request. Depending on available data
 179          * (read-modify-write vs. reconstruct-write case) we calculate it
 180          * differently.
 181          */
 182         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
 183                 /*
 184                  * rmw: xor old data and parity from updated disks
 185                  * This is calculated earlier by ops_run_prexor5() so just copy
 186                  * the parity dev page.
 187                  */
 188                 srcs[count++] = sh->dev[pd_idx].page;
 189         } else if (sh->reconstruct_state == reconstruct_state_drain_run) {
 190                 /* rcw: xor data from all not updated disks */
 191                 for (i = disks; i--;) {
 192                         struct r5dev *dev = &sh->dev[i];
 193                         if (test_bit(R5_UPTODATE, &dev->flags))
 194                                 srcs[count++] = dev->page;
 195                 }
 196         } else {
 197                 return tx;
 198         }
 199
 200         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, tx,
 201                           NULL, sh, (void *) (srcs + sh->disks + 2));
 202
 203         if (count == 1)
 204                 tx = async_memcpy(sh->ppl_page, srcs[0], 0, 0, PAGE_SIZE,
 205                                   &submit);
 206         else
 207                 tx = async_xor(sh->ppl_page, srcs, 0, count, PAGE_SIZE,
 208                                &submit);
 209
 210         return tx;
 211 }
 212
 213 static void *ppl_io_pool_alloc(gfp_t gfp_mask, void *pool_data)
 214 {
 215         struct kmem_cache *kc = pool_data;
 216         struct ppl_io_unit *io;
 217
 218         io = kmem_cache_alloc(kc, gfp_mask);
 219         if (!io)
 220                 return NULL;
 221
 222         io->header_page = alloc_page(gfp_mask);
 223         if (!io->header_page) {
 224                 kmem_cache_free(kc, io);
 225                 return NULL;
 226         }
 227
 228         return io;
 229 }
 230
 231 static void ppl_io_pool_free(void *element, void *pool_data)
 232 {
 233         struct kmem_cache *kc = pool_data;
 234         struct ppl_io_unit *io = element;
 235
 236         __free_page(io->header_page);
 237         kmem_cache_free(kc, io);
 238 }
 239
 240 static struct ppl_io_unit *ppl_new_iounit(struct ppl_log *log,
 241                                           struct stripe_head *sh)
 242 {
 243         struct ppl_conf *ppl_conf = log->ppl_conf;
 244         struct ppl_io_unit *io;
 245         struct ppl_header *pplhdr;
 246         struct page *header_page;
 247
 248         io = mempool_alloc(&ppl_conf->io_pool, GFP_NOWAIT);
 249         if (!io)
 250                 return NULL;
 251
 252         header_page = io->header_page;
 253         memset(io, 0, sizeof(*io));
 254         io->header_page = header_page;
 255
 256         io->log = log;
 257         INIT_LIST_HEAD(&io->log_sibling);
 258         INIT_LIST_HEAD(&io->stripe_list);
 259         atomic_set(&io->pending_stripes, 0);
 260         atomic_set(&io->pending_flushes, 0);
 261         bio_init(&io->bio, io->biovec, PPL_IO_INLINE_BVECS);
 262
 263         pplhdr = page_address(io->header_page);
 264         clear_page(pplhdr);
 265         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
 266         pplhdr->signature = cpu_to_le32(ppl_conf->signature);
 267
 268         io->seq = atomic64_add_return(1, &ppl_conf->seq);
 269         pplhdr->generation = cpu_to_le64(io->seq);
 270
 271         return io;
 272 }
 273
 274 static int ppl_log_stripe(struct ppl_log *log, struct stripe_head *sh)
 275 {
 276         struct ppl_io_unit *io = log->current_io;
 277         struct ppl_header_entry *e = NULL;
 278         struct ppl_header *pplhdr;
 279         int i;
 280         sector_t data_sector = 0;
 281         int data_disks = 0;
 282         struct r5conf *conf = sh->raid_conf;
 283
 284         pr_debug("%s: stripe: %llu\n", __func__, (unsigned long long)sh->sector);
 285
 286         /* check if current io_unit is full */
 287         if (io && (io->pp_size == log->entry_space ||
 288                    io->entries_count == PPL_HDR_MAX_ENTRIES)) {
 289                 pr_debug("%s: add io_unit blocked by seq: %llu\n",
 290                          __func__, io->seq);
 291                 io = NULL;
 292         }
 293
 294         /* add a new unit if there is none or the current is full */
 295         if (!io) {
 296                 io = ppl_new_iounit(log, sh);
 297                 if (!io)
 298                         return -ENOMEM;
 299                 spin_lock_irq(&log->io_list_lock);
 300                 list_add_tail(&io->log_sibling, &log->io_list);
 301                 spin_unlock_irq(&log->io_list_lock);
 302
 303                 log->current_io = io;
 304         }
 305
 306         for (i = 0; i < sh->disks; i++) {
 307                 struct r5dev *dev = &sh->dev[i];
 308
 309                 if (i != sh->pd_idx && test_bit(R5_Wantwrite, &dev->flags)) {
 310                         if (!data_disks || dev->sector < data_sector)
 311                                 data_sector = dev->sector;
 312                         data_disks++;
 313                 }
 314         }
 315         BUG_ON(!data_disks);
 316
 317         pr_debug("%s: seq: %llu data_sector: %llu data_disks: %d\n", __func__,
 318                  io->seq, (unsigned long long)data_sector, data_disks);
 319
 320         pplhdr = page_address(io->header_page);
 321
 322         if (io->entries_count > 0) {
 323                 struct ppl_header_entry *last =
 324                                 &pplhdr->entries[io->entries_count - 1];
 325                 struct stripe_head *sh_last = list_last_entry(
 326                                 &io->stripe_list, struct stripe_head, log_list);
 327                 u64 data_sector_last = le64_to_cpu(last->data_sector);
 328                 u32 data_size_last = le32_to_cpu(last->data_size);
 329
 330                 /*
 331                  * Check if we can append the stripe to the last entry. It must
 332                  * be just after the last logged stripe and write to the same
 333                  * disks. Use bit shift and logarithm to avoid 64-bit division.
 334                  */
 335                 if ((sh->sector == sh_last->sector + STRIPE_SECTORS) &&
 336                     (data_sector >> ilog2(conf->chunk_sectors) ==
 337                      data_sector_last >> ilog2(conf->chunk_sectors)) &&
 338                     ((data_sector - data_sector_last) * data_disks ==
 339                      data_size_last >> 9))
 340                         e = last;
 341         }
 342
 343         if (!e) {
 344                 e = &pplhdr->entries[io->entries_count++];
 345                 e->data_sector = cpu_to_le64(data_sector);
 346                 e->parity_disk = cpu_to_le32(sh->pd_idx);
 347                 e->checksum = cpu_to_le32(~0);
 348         }
 349
 350         le32_add_cpu(&e->data_size, data_disks << PAGE_SHIFT);
 351
 352         /* don't write any PP if full stripe write */
 353         if (!test_bit(STRIPE_FULL_WRITE, &sh->state)) {
 354                 le32_add_cpu(&e->pp_size, PAGE_SIZE);
 355                 io->pp_size += PAGE_SIZE;
 356                 e->checksum = cpu_to_le32(crc32c_le(le32_to_cpu(e->checksum),
 357                                                     page_address(sh->ppl_page),
 358                                                     PAGE_SIZE));
 359         }
 360
 361         list_add_tail(&sh->log_list, &io->stripe_list);
 362         atomic_inc(&io->pending_stripes);
 363         sh->ppl_io = io;
 364
 365         return 0;
 366 }
 367
 368 int ppl_write_stripe(struct r5conf *conf, struct stripe_head *sh)
 369 {
 370         struct ppl_conf *ppl_conf = conf->log_private;
 371         struct ppl_io_unit *io = sh->ppl_io;
 372         struct ppl_log *log;
 373
 374         if (io || test_bit(STRIPE_SYNCING, &sh->state) || !sh->ppl_page ||
 375             !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 376             !test_bit(R5_Insync, &sh->dev[sh->pd_idx].flags)) {
 377                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 378                 return -EAGAIN;
 379         }
 380
 381         log = &ppl_conf->child_logs[sh->pd_idx];
 382
 383         mutex_lock(&log->io_mutex);
 384
 385         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 386                 mutex_unlock(&log->io_mutex);
 387                 return -EAGAIN;
 388         }
 389
 390         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 391         clear_bit(STRIPE_DELAYED, &sh->state);
 392         atomic_inc(&sh->count);
 393
 394         if (ppl_log_stripe(log, sh)) {
 395                 spin_lock_irq(&ppl_conf->no_mem_stripes_lock);
 396                 list_add_tail(&sh->log_list, &ppl_conf->no_mem_stripes);
 397                 spin_unlock_irq(&ppl_conf->no_mem_stripes_lock);
 398         }
 399
 400         mutex_unlock(&log->io_mutex);
 401
 402         return 0;
 403 }
 404
 405 static void ppl_log_endio(struct bio *bio)
 406 {
 407         struct ppl_io_unit *io = bio->bi_private;
 408         struct ppl_log *log = io->log;
 409         struct ppl_conf *ppl_conf = log->ppl_conf;
 410         struct stripe_head *sh, *next;
 411
 412         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 413
 414         if (bio->bi_status)
 415                 md_error(ppl_conf->mddev, log->rdev);
 416
 417         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 418                 list_del_init(&sh->log_list);
 419
 420                 set_bit(STRIPE_HANDLE, &sh->state);
 421                 raid5_release_stripe(sh);
 422         }
 423 }
 424
 425 static void ppl_submit_iounit_bio(struct ppl_io_unit *io, struct bio *bio)
 426 {
 427         char b[BDEVNAME_SIZE];
 428
 429         pr_debug("%s: seq: %llu size: %u sector: %llu dev: %s\n",
 430                  __func__, io->seq, bio->bi_iter.bi_size,
 431                  (unsigned long long)bio->bi_iter.bi_sector,
 432                  bio_devname(bio, b));
 433
 434         submit_bio(bio);
 435 }
 436
 437 static void ppl_submit_iounit(struct ppl_io_unit *io)
 438 {
 439         struct ppl_log *log = io->log;
 440         struct ppl_conf *ppl_conf = log->ppl_conf;
 441         struct ppl_header *pplhdr = page_address(io->header_page);
 442         struct bio *bio = &io->bio;
 443         struct stripe_head *sh;
 444         int i;
 445
 446         bio->bi_private = io;
 447
 448         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 449                 ppl_log_endio(bio);
 450                 return;
 451         }
 452
 453         for (i = 0; i < io->entries_count; i++) {
 454                 struct ppl_header_entry *e = &pplhdr->entries[i];
 455
 456                 pr_debug("%s: seq: %llu entry: %d data_sector: %llu pp_size: %u data_size: %u\n",
 457                          __func__, io->seq, i, le64_to_cpu(e->data_sector),
 458                          le32_to_cpu(e->pp_size), le32_to_cpu(e->data_size));
 459
 460                 e->data_sector = cpu_to_le64(le64_to_cpu(e->data_sector) >>
 461                                              ilog2(ppl_conf->block_size >> 9));
 462                 e->checksum = cpu_to_le32(~le32_to_cpu(e->checksum));
 463         }
 464
 465         pplhdr->entries_count = cpu_to_le32(io->entries_count);
 466         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PPL_HEADER_SIZE));
 467
 468         /* Rewind the buffer if current PPL is larger then remaining space */
 469         if (log->use_multippl &&
 470             log->rdev->ppl.sector + log->rdev->ppl.size - log->next_io_sector <
 471             (PPL_HEADER_SIZE + io->pp_size) >> 9)
 472                 log->next_io_sector = log->rdev->ppl.sector;
 473
 474
 475         bio->bi_end_io = ppl_log_endio;
 476         bio->bi_opf = REQ_OP_WRITE | REQ_FUA;
 477         bio_set_dev(bio, log->rdev->bdev);
 478         bio->bi_iter.bi_sector = log->next_io_sector;
 479         bio_add_page(bio, io->header_page, PAGE_SIZE, 0);
 480         bio->bi_write_hint = ppl_conf->write_hint;
 481
 482         pr_debug("%s: log->current_io_sector: %llu\n", __func__,
 483             (unsigned long long)log->next_io_sector);
 484
 485         if (log->use_multippl)
 486                 log->next_io_sector += (PPL_HEADER_SIZE + io->pp_size) >> 9;
 487
 488         WARN_ON(log->disk_flush_bitmap != 0);
 489
 490         list_for_each_entry(sh, &io->stripe_list, log_list) {
 491                 for (i = 0; i < sh->disks; i++) {
 492                         struct r5dev *dev = &sh->dev[i];
 493
 494                         if ((ppl_conf->child_logs[i].wb_cache_on) &&
 495                             (test_bit(R5_Wantwrite, &dev->flags))) {
 496                                 set_bit(i, &log->disk_flush_bitmap);
 497                         }
 498                 }
 499
 500                 /* entries for full stripe writes have no partial parity */
 501                 if (test_bit(STRIPE_FULL_WRITE, &sh->state))
 502                         continue;
 503
 504                 if (!bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0)) {
 505                         struct bio *prev = bio;
 506
 507                         bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES,
 508                                                &ppl_conf->bs);
 509                         bio->bi_opf = prev->bi_opf;
 510                         bio->bi_write_hint = prev->bi_write_hint;
 511                         bio_copy_dev(bio, prev);
 512                         bio->bi_iter.bi_sector = bio_end_sector(prev);
 513                         bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0);
 514
 515                         bio_chain(bio, prev);
 516                         ppl_submit_iounit_bio(io, prev);
 517                 }
 518         }
 519
 520         ppl_submit_iounit_bio(io, bio);
 521 }
 522
 523 static void ppl_submit_current_io(struct ppl_log *log)
 524 {
 525         struct ppl_io_unit *io;
 526
 527         spin_lock_irq(&log->io_list_lock);
 528
 529         io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
 530                                       log_sibling);
 531         if (io && io->submitted)
 532                 io = NULL;
 533
 534         spin_unlock_irq(&log->io_list_lock);
 535
 536         if (io) {
 537                 io->submitted = true;
 538
 539                 if (io == log->current_io)
 540                         log->current_io = NULL;
 541
 542                 ppl_submit_iounit(io);
 543         }
 544 }
 545
 546 void ppl_write_stripe_run(struct r5conf *conf)
 547 {
 548         struct ppl_conf *ppl_conf = conf->log_private;
 549         struct ppl_log *log;
 550         int i;
 551
 552         for (i = 0; i < ppl_conf->count; i++) {
 553                 log = &ppl_conf->child_logs[i];
 554
 555                 mutex_lock(&log->io_mutex);
 556                 ppl_submit_current_io(log);
 557                 mutex_unlock(&log->io_mutex);
 558         }
 559 }
 560
 561 static void ppl_io_unit_finished(struct ppl_io_unit *io)
 562 {
 563         struct ppl_log *log = io->log;
 564         struct ppl_conf *ppl_conf = log->ppl_conf;
 565         struct r5conf *conf = ppl_conf->mddev->private;
 566         unsigned long flags;
 567
 568         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 569
 570         local_irq_save(flags);
 571
 572         spin_lock(&log->io_list_lock);
 573         list_del(&io->log_sibling);
 574         spin_unlock(&log->io_list_lock);
 575
 576         mempool_free(io, &ppl_conf->io_pool);
 577
 578         spin_lock(&ppl_conf->no_mem_stripes_lock);
 579         if (!list_empty(&ppl_conf->no_mem_stripes)) {
 580                 struct stripe_head *sh;
 581
 582                 sh = list_first_entry(&ppl_conf->no_mem_stripes,
 583                                       struct stripe_head, log_list);
 584                 list_del_init(&sh->log_list);
 585                 set_bit(STRIPE_HANDLE, &sh->state);
 586                 raid5_release_stripe(sh);
 587         }
 588         spin_unlock(&ppl_conf->no_mem_stripes_lock);
 589
 590         local_irq_restore(flags);
 591
 592         wake_up(&conf->wait_for_quiescent);
 593 }
 594
 595 static void ppl_flush_endio(struct bio *bio)
 596 {
 597         struct ppl_io_unit *io = bio->bi_private;
 598         struct ppl_log *log = io->log;
 599         struct ppl_conf *ppl_conf = log->ppl_conf;
 600         struct r5conf *conf = ppl_conf->mddev->private;
 601         char b[BDEVNAME_SIZE];
 602
 603         pr_debug("%s: dev: %s\n", __func__, bio_devname(bio, b));
 604
 605         if (bio->bi_status) {
 606                 struct md_rdev *rdev;
 607
 608                 rcu_read_lock();
 609                 rdev = md_find_rdev_rcu(conf->mddev, bio_dev(bio));
 610                 if (rdev)
 611                         md_error(rdev->mddev, rdev);
 612                 rcu_read_unlock();
 613         }
 614
 615         bio_put(bio);
 616
 617         if (atomic_dec_and_test(&io->pending_flushes)) {
 618                 ppl_io_unit_finished(io);
 619                 md_wakeup_thread(conf->mddev->thread);
 620         }
 621 }
 622
 623 static void ppl_do_flush(struct ppl_io_unit *io)
 624 {
 625         struct ppl_log *log = io->log;
 626         struct ppl_conf *ppl_conf = log->ppl_conf;
 627         struct r5conf *conf = ppl_conf->mddev->private;
 628         int raid_disks = conf->raid_disks;
 629         int flushed_disks = 0;
 630         int i;
 631
 632         atomic_set(&io->pending_flushes, raid_disks);
 633
 634         for_each_set_bit(i, &log->disk_flush_bitmap, raid_disks) {
 635                 struct md_rdev *rdev;
 636                 struct block_device *bdev = NULL;
 637
 638                 rcu_read_lock();
 639                 rdev = rcu_dereference(conf->disks[i].rdev);
 640                 if (rdev && !test_bit(Faulty, &rdev->flags))
 641                         bdev = rdev->bdev;
 642                 rcu_read_unlock();
 643
 644                 if (bdev) {
 645                         struct bio *bio;
 646                         char b[BDEVNAME_SIZE];
 647
 648                         bio = bio_alloc_bioset(GFP_NOIO, 0, &ppl_conf->flush_bs);
 649                         bio_set_dev(bio, bdev);
 650                         bio->bi_private = io;
 651                         bio->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
 652                         bio->bi_end_io = ppl_flush_endio;
 653
 654                         pr_debug("%s: dev: %s\n", __func__,
 655                                  bio_devname(bio, b));
 656
 657                         submit_bio(bio);
 658                         flushed_disks++;
 659                 }
 660         }
 661
 662         log->disk_flush_bitmap = 0;
 663
 664         for (i = flushed_disks ; i < raid_disks; i++) {
 665                 if (atomic_dec_and_test(&io->pending_flushes))
 666                         ppl_io_unit_finished(io);
 667         }
 668 }
 669
 670 static inline bool ppl_no_io_unit_submitted(struct r5conf *conf,
 671                                             struct ppl_log *log)
 672 {
 673         struct ppl_io_unit *io;
 674
 675         io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
 676                                       log_sibling);
 677
 678         return !io || !io->submitted;
 679 }
 680
 681 void ppl_quiesce(struct r5conf *conf, int quiesce)
 682 {
 683         struct ppl_conf *ppl_conf = conf->log_private;
 684         int i;
 685
 686         if (quiesce) {
 687                 for (i = 0; i < ppl_conf->count; i++) {
 688                         struct ppl_log *log = &ppl_conf->child_logs[i];
 689
 690                         spin_lock_irq(&log->io_list_lock);
 691                         wait_event_lock_irq(conf->wait_for_quiescent,
 692                                             ppl_no_io_unit_submitted(conf, log),
 693                                             log->io_list_lock);
 694                         spin_unlock_irq(&log->io_list_lock);
 695                 }
 696         }
 697 }
 698
 699 int ppl_handle_flush_request(struct r5l_log *log, struct bio *bio)
 700 {
 701         if (bio->bi_iter.bi_size == 0) {
 702                 bio_endio(bio);
 703                 return 0;
 704         }
 705         bio->bi_opf &= ~REQ_PREFLUSH;
 706         return -EAGAIN;
 707 }
 708
 709 void ppl_stripe_write_finished(struct stripe_head *sh)
 710 {
 711         struct ppl_io_unit *io;
 712
 713         io = sh->ppl_io;
 714         sh->ppl_io = NULL;
 715
 716         if (io && atomic_dec_and_test(&io->pending_stripes)) {
 717                 if (io->log->disk_flush_bitmap)
 718                         ppl_do_flush(io);
 719                 else
 720                         ppl_io_unit_finished(io);
 721         }
 722 }
 723
 724 static void ppl_xor(int size, struct page *page1, struct page *page2)
 725 {
 726         struct async_submit_ctl submit;
 727         struct dma_async_tx_descriptor *tx;
 728         struct page *xor_srcs[] = { page1, page2 };
 729
 730         init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST,
 731                           NULL, NULL, NULL, NULL);
 732         tx = async_xor(page1, xor_srcs, 0, 2, size, &submit);
 733
 734         async_tx_quiesce(&tx);
 735 }
 736
 737 /*
 738  * PPL recovery strategy: xor partial parity and data from all modified data
 739  * disks within a stripe and write the result as the new stripe parity. If all
 740  * stripe data disks are modified (full stripe write), no partial parity is
 741  * available, so just xor the data disks.
 742  *
 743  * Recovery of a PPL entry shall occur only if all modified data disks are
 744  * available and read from all of them succeeds.
 745  *
 746  * A PPL entry applies to a stripe, partial parity size for an entry is at most
 747  * the size of the chunk. Examples of possible cases for a single entry:
 748  *
 749  * case 0: single data disk write:
 750  *   data0    data1    data2     ppl        parity
 751  * +--------+--------+--------+           +--------------------+
 752  * | ------ | ------ | ------ | +----+    | (no change)        |
 753  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 754  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 755  * | ------ | ------ | ------ | +----+    | (no change)        |
 756  * +--------+--------+--------+           +--------------------+
 757  * pp_size = data_size
 758  *
 759  * case 1: more than one data disk write:
 760  *   data0    data1    data2     ppl        parity
 761  * +--------+--------+--------+           +--------------------+
 762  * | ------ | ------ | ------ | +----+    | (no change)        |
 763  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 764  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 765  * | ------ | ------ | ------ | +----+    | (no change)        |
 766  * +--------+--------+--------+           +--------------------+
 767  * pp_size = data_size / modified_data_disks
 768  *
 769  * case 2: write to all data disks (also full stripe write):
 770  *   data0    data1    data2                parity
 771  * +--------+--------+--------+           +--------------------+
 772  * | ------ | ------ | ------ |           | (no change)        |
 773  * | -data- | -data- | -data- | --------> | xor all data       |
 774  * | ------ | ------ | ------ | --------> | (no change)        |
 775  * | ------ | ------ | ------ |           | (no change)        |
 776  * +--------+--------+--------+           +--------------------+
 777  * pp_size = 0
 778  *
 779  * The following cases are possible only in other implementations. The recovery
 780  * code can handle them, but they are not generated at runtime because they can
 781  * be reduced to cases 0, 1 and 2:
 782  *
 783  * case 3:
 784  *   data0    data1    data2     ppl        parity
 785  * +--------+--------+--------+ +----+    +--------------------+
 786  * | ------ | -data- | -data- | | pp |    | data1 ^ data2 ^ pp |
 787  * | ------ | -data- | -data- | | pp | -> | data1 ^ data2 ^ pp |
 788  * | -data- | -data- | -data- | | -- | -> | xor all data       |
 789  * | -data- | -data- | ------ | | pp |    | data0 ^ data1 ^ pp |
 790  * +--------+--------+--------+ +----+    +--------------------+
 791  * pp_size = chunk_size
 792  *
 793  * case 4:
 794  *   data0    data1    data2     ppl        parity
 795  * +--------+--------+--------+ +----+    +--------------------+
 796  * | ------ | -data- | ------ | | pp |    | data1 ^ pp         |
 797  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 798  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 799  * | -data- | ------ | ------ | | pp |    | data0 ^ pp         |
 800  * +--------+--------+--------+ +----+    +--------------------+
 801  * pp_size = chunk_size
 802  */
 803 static int ppl_recover_entry(struct ppl_log *log, struct ppl_header_entry *e,
 804                              sector_t ppl_sector)
 805 {
 806         struct ppl_conf *ppl_conf = log->ppl_conf;
 807         struct mddev *mddev = ppl_conf->mddev;
 808         struct r5conf *conf = mddev->private;
 809         int block_size = ppl_conf->block_size;
 810         struct page *page1;
 811         struct page *page2;
 812         sector_t r_sector_first;
 813         sector_t r_sector_last;
 814         int strip_sectors;
 815         int data_disks;
 816         int i;
 817         int ret = 0;
 818         char b[BDEVNAME_SIZE];
 819         unsigned int pp_size = le32_to_cpu(e->pp_size);
 820         unsigned int data_size = le32_to_cpu(e->data_size);
 821
 822         page1 = alloc_page(GFP_KERNEL);
 823         page2 = alloc_page(GFP_KERNEL);
 824
 825         if (!page1 || !page2) {
 826                 ret = -ENOMEM;
 827                 goto out;
 828         }
 829
 830         r_sector_first = le64_to_cpu(e->data_sector) * (block_size >> 9);
 831
 832         if ((pp_size >> 9) < conf->chunk_sectors) {
 833                 if (pp_size > 0) {
 834                         data_disks = data_size / pp_size;
 835                         strip_sectors = pp_size >> 9;
 836                 } else {
 837                         data_disks = conf->raid_disks - conf->max_degraded;
 838                         strip_sectors = (data_size >> 9) / data_disks;
 839                 }
 840                 r_sector_last = r_sector_first +
 841                                 (data_disks - 1) * conf->chunk_sectors +
 842                                 strip_sectors;
 843         } else {
 844                 data_disks = conf->raid_disks - conf->max_degraded;
 845                 strip_sectors = conf->chunk_sectors;
 846                 r_sector_last = r_sector_first + (data_size >> 9);
 847         }
 848
 849         pr_debug("%s: array sector first: %llu last: %llu\n", __func__,
 850                  (unsigned long long)r_sector_first,
 851                  (unsigned long long)r_sector_last);
 852
 853         /* if start and end is 4k aligned, use a 4k block */
 854         if (block_size == 512 &&
 855             (r_sector_first & (STRIPE_SECTORS - 1)) == 0 &&
 856             (r_sector_last & (STRIPE_SECTORS - 1)) == 0)
 857                 block_size = STRIPE_SIZE;
 858
 859         /* iterate through blocks in strip */
 860         for (i = 0; i < strip_sectors; i += (block_size >> 9)) {
 861                 bool update_parity = false;
 862                 sector_t parity_sector;
 863                 struct md_rdev *parity_rdev;
 864                 struct stripe_head sh;
 865                 int disk;
 866                 int indent = 0;
 867
 868                 pr_debug("%s:%*s iter %d start\n", __func__, indent, "", i);
 869                 indent += 2;
 870
 871                 memset(page_address(page1), 0, PAGE_SIZE);
 872
 873                 /* iterate through data member disks */
 874                 for (disk = 0; disk < data_disks; disk++) {
 875                         int dd_idx;
 876                         struct md_rdev *rdev;
 877                         sector_t sector;
 878                         sector_t r_sector = r_sector_first + i +
 879                                             (disk * conf->chunk_sectors);
 880
 881                         pr_debug("%s:%*s data member disk %d start\n",
 882                                  __func__, indent, "", disk);
 883                         indent += 2;
 884
 885                         if (r_sector >= r_sector_last) {
 886                                 pr_debug("%s:%*s array sector %llu doesn't need parity update\n",
 887                                          __func__, indent, "",
 888                                          (unsigned long long)r_sector);
 889                                 indent -= 2;
 890                                 continue;
 891                         }
 892
 893                         update_parity = true;
 894
 895                         /* map raid sector to member disk */
 896                         sector = raid5_compute_sector(conf, r_sector, 0,
 897                                                       &dd_idx, NULL);
 898                         pr_debug("%s:%*s processing array sector %llu => data member disk %d, sector %llu\n",
 899                                  __func__, indent, "",
 900                                  (unsigned long long)r_sector, dd_idx,
 901                                  (unsigned long long)sector);
 902
 903                         rdev = conf->disks[dd_idx].rdev;
 904                         if (!rdev || (!test_bit(In_sync, &rdev->flags) &&
 905                                       sector >= rdev->recovery_offset)) {
 906                                 pr_debug("%s:%*s data member disk %d missing\n",
 907                                          __func__, indent, "", dd_idx);
 908                                 update_parity = false;
 909                                 break;
 910                         }
 911
 912                         pr_debug("%s:%*s reading data member disk %s sector %llu\n",
 913                                  __func__, indent, "", bdevname(rdev->bdev, b),
 914                                  (unsigned long long)sector);
 915                         if (!sync_page_io(rdev, sector, block_size, page2,
 916                                         REQ_OP_READ, 0, false)) {
 917                                 md_error(mddev, rdev);
 918                                 pr_debug("%s:%*s read failed!\n", __func__,
 919                                          indent, "");
 920                                 ret = -EIO;
 921                                 goto out;
 922                         }
 923
 924                         ppl_xor(block_size, page1, page2);
 925
 926                         indent -= 2;
 927                 }
 928
 929                 if (!update_parity)
 930                         continue;
 931
 932                 if (pp_size > 0) {
 933                         pr_debug("%s:%*s reading pp disk sector %llu\n",
 934                                  __func__, indent, "",
 935                                  (unsigned long long)(ppl_sector + i));
 936                         if (!sync_page_io(log->rdev,
 937                                         ppl_sector - log->rdev->data_offset + i,
 938                                         block_size, page2, REQ_OP_READ, 0,
 939                                         false)) {
 940                                 pr_debug("%s:%*s read failed!\n", __func__,
 941                                          indent, "");
 942                                 md_error(mddev, log->rdev);
 943                                 ret = -EIO;
 944                                 goto out;
 945                         }
 946
 947                         ppl_xor(block_size, page1, page2);
 948                 }
 949
 950                 /* map raid sector to parity disk */
 951                 parity_sector = raid5_compute_sector(conf, r_sector_first + i,
 952                                 0, &disk, &sh);
 953                 BUG_ON(sh.pd_idx != le32_to_cpu(e->parity_disk));
 954                 parity_rdev = conf->disks[sh.pd_idx].rdev;
 955
 956                 BUG_ON(parity_rdev->bdev->bd_dev != log->rdev->bdev->bd_dev);
 957                 pr_debug("%s:%*s write parity at sector %llu, disk %s\n",
 958                          __func__, indent, "",
 959                          (unsigned long long)parity_sector,
 960                          bdevname(parity_rdev->bdev, b));
 961                 if (!sync_page_io(parity_rdev, parity_sector, block_size,
 962                                 page1, REQ_OP_WRITE, 0, false)) {
 963                         pr_debug("%s:%*s parity write error!\n", __func__,
 964                                  indent, "");
 965                         md_error(mddev, parity_rdev);
 966                         ret = -EIO;
 967                         goto out;
 968                 }
 969         }
 970 out:
 971         if (page1)
 972                 __free_page(page1);
 973         if (page2)
 974                 __free_page(page2);
 975         return ret;
 976 }
 977
 978 static int ppl_recover(struct ppl_log *log, struct ppl_header *pplhdr,
 979                        sector_t offset)
 980 {
 981         struct ppl_conf *ppl_conf = log->ppl_conf;
 982         struct md_rdev *rdev = log->rdev;
 983         struct mddev *mddev = rdev->mddev;
 984         sector_t ppl_sector = rdev->ppl.sector + offset +
 985                               (PPL_HEADER_SIZE >> 9);
 986         struct page *page;
 987         int i;
 988         int ret = 0;
 989
 990         page = alloc_page(GFP_KERNEL);
 991         if (!page)
 992                 return -ENOMEM;
 993
 994         /* iterate through all PPL entries saved */
 995         for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++) {
 996                 struct ppl_header_entry *e = &pplhdr->entries[i];
 997                 u32 pp_size = le32_to_cpu(e->pp_size);
 998                 sector_t sector = ppl_sector;
 999                 int ppl_entry_sectors = pp_size >> 9;
1000                 u32 crc, crc_stored;
1001
1002                 pr_debug("%s: disk: %d entry: %d ppl_sector: %llu pp_size: %u\n",
1003                          __func__, rdev->raid_disk, i,
1004                          (unsigned long long)ppl_sector, pp_size);
1005
1006                 crc = ~0;
1007                 crc_stored = le32_to_cpu(e->checksum);
1008
1009                 /* read parial parity for this entry and calculate its checksum */
1010                 while (pp_size) {
1011                         int s = pp_size > PAGE_SIZE ? PAGE_SIZE : pp_size;
1012
1013                         if (!sync_page_io(rdev, sector - rdev->data_offset,
1014                                         s, page, REQ_OP_READ, 0, false)) {
1015                                 md_error(mddev, rdev);
1016                                 ret = -EIO;
1017                                 goto out;
1018                         }
1019
1020                         crc = crc32c_le(crc, page_address(page), s);
1021
1022                         pp_size -= s;
1023                         sector += s >> 9;
1024                 }
1025
1026                 crc = ~crc;
1027
1028                 if (crc != crc_stored) {
1029                         /*
1030                          * Don't recover this entry if the checksum does not
1031                          * match, but keep going and try to recover other
1032                          * entries.
1033                          */
1034                         pr_debug("%s: ppl entry crc does not match: stored: 0x%x calculated: 0x%x\n",
1035                                  __func__, crc_stored, crc);
1036                         ppl_conf->mismatch_count++;
1037                 } else {
1038                         ret = ppl_recover_entry(log, e, ppl_sector);
1039                         if (ret)
1040                                 goto out;
1041                         ppl_conf->recovered_entries++;
1042                 }
1043
1044                 ppl_sector += ppl_entry_sectors;
1045         }
1046
1047         /* flush the disk cache after recovery if necessary */
1048         ret = blkdev_issue_flush(rdev->bdev, GFP_KERNEL, NULL);
1049 out:
1050         __free_page(page);
1051         return ret;
1052 }
1053
1054 static int ppl_write_empty_header(struct ppl_log *log)
1055 {
1056         struct page *page;
1057         struct ppl_header *pplhdr;
1058         struct md_rdev *rdev = log->rdev;
1059         int ret = 0;
1060
1061         pr_debug("%s: disk: %d ppl_sector: %llu\n", __func__,
1062                  rdev->raid_disk, (unsigned long long)rdev->ppl.sector);
1063
1064         page = alloc_page(GFP_NOIO | __GFP_ZERO);
1065         if (!page)
1066                 return -ENOMEM;
1067
1068         pplhdr = page_address(page);
1069         /* zero out PPL space to avoid collision with old PPLs */
1070         blkdev_issue_zeroout(rdev->bdev, rdev->ppl.sector,
1071                             log->rdev->ppl.size, GFP_NOIO, 0);
1072         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
1073         pplhdr->signature = cpu_to_le32(log->ppl_conf->signature);
1074         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PAGE_SIZE));
1075
1076         if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset,
1077                           PPL_HEADER_SIZE, page, REQ_OP_WRITE | REQ_SYNC |
1078                           REQ_FUA, 0, false)) {
1079                 md_error(rdev->mddev, rdev);
1080                 ret = -EIO;
1081         }
1082
1083         __free_page(page);
1084         return ret;
1085 }
1086
1087 static int ppl_load_distributed(struct ppl_log *log)
1088 {
1089         struct ppl_conf *ppl_conf = log->ppl_conf;
1090         struct md_rdev *rdev = log->rdev;
1091         struct mddev *mddev = rdev->mddev;
1092         struct page *page, *page2, *tmp;
1093         struct ppl_header *pplhdr = NULL, *prev_pplhdr = NULL;
1094         u32 crc, crc_stored;
1095         u32 signature;
1096         int ret = 0, i;
1097         sector_t pplhdr_offset = 0, prev_pplhdr_offset = 0;
1098
1099         pr_debug("%s: disk: %d\n", __func__, rdev->raid_disk);
1100         /* read PPL headers, find the recent one */
1101         page = alloc_page(GFP_KERNEL);
1102         if (!page)
1103                 return -ENOMEM;
1104
1105         page2 = alloc_page(GFP_KERNEL);
1106         if (!page2) {
1107                 __free_page(page);
1108                 return -ENOMEM;
1109         }
1110
1111         /* searching ppl area for latest ppl */
1112         while (pplhdr_offset < rdev->ppl.size - (PPL_HEADER_SIZE >> 9)) {
1113                 if (!sync_page_io(rdev,
1114                                   rdev->ppl.sector - rdev->data_offset +
1115                                   pplhdr_offset, PAGE_SIZE, page, REQ_OP_READ,
1116                                   0, false)) {
1117                         md_error(mddev, rdev);
1118                         ret = -EIO;
1119                         /* if not able to read - don't recover any PPL */
1120                         pplhdr = NULL;
1121                         break;
1122                 }
1123                 pplhdr = page_address(page);
1124
1125                 /* check header validity */
1126                 crc_stored = le32_to_cpu(pplhdr->checksum);
1127                 pplhdr->checksum = 0;
1128                 crc = ~crc32c_le(~0, pplhdr, PAGE_SIZE);
1129
1130                 if (crc_stored != crc) {
1131                         pr_debug("%s: ppl header crc does not match: stored: 0x%x calculated: 0x%x (offset: %llu)\n",
1132                                  __func__, crc_stored, crc,
1133                                  (unsigned long long)pplhdr_offset);
1134                         pplhdr = prev_pplhdr;
1135                         pplhdr_offset = prev_pplhdr_offset;
1136                         break;
1137                 }
1138
1139                 signature = le32_to_cpu(pplhdr->signature);
1140
1141                 if (mddev->external) {
1142                         /*
1143                          * For external metadata the header signature is set and
1144                          * validated in userspace.
1145                          */
1146                         ppl_conf->signature = signature;
1147                 } else if (ppl_conf->signature != signature) {
1148                         pr_debug("%s: ppl header signature does not match: stored: 0x%x configured: 0x%x (offset: %llu)\n",
1149                                  __func__, signature, ppl_conf->signature,
1150                                  (unsigned long long)pplhdr_offset);
1151                         pplhdr = prev_pplhdr;
1152                         pplhdr_offset = prev_pplhdr_offset;
1153                         break;
1154                 }
1155
1156                 if (prev_pplhdr && le64_to_cpu(prev_pplhdr->generation) >
1157                     le64_to_cpu(pplhdr->generation)) {
1158                         /* previous was newest */
1159                         pplhdr = prev_pplhdr;
1160                         pplhdr_offset = prev_pplhdr_offset;
1161                         break;
1162                 }
1163
1164                 prev_pplhdr_offset = pplhdr_offset;
1165                 prev_pplhdr = pplhdr;
1166
1167                 tmp = page;
1168                 page = page2;
1169                 page2 = tmp;
1170
1171                 /* calculate next potential ppl offset */
1172                 for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++)
1173                         pplhdr_offset +=
1174                             le32_to_cpu(pplhdr->entries[i].pp_size) >> 9;
1175                 pplhdr_offset += PPL_HEADER_SIZE >> 9;
1176         }
1177
1178         /* no valid ppl found */
1179         if (!pplhdr)
1180                 ppl_conf->mismatch_count++;
1181         else
1182                 pr_debug("%s: latest PPL found at offset: %llu, with generation: %llu\n",
1183                     __func__, (unsigned long long)pplhdr_offset,
1184                     le64_to_cpu(pplhdr->generation));
1185
1186         /* attempt to recover from log if we are starting a dirty array */
1187         if (pplhdr && !mddev->pers && mddev->recovery_cp != MaxSector)
1188                 ret = ppl_recover(log, pplhdr, pplhdr_offset);
1189
1190         /* write empty header if we are starting the array */
1191         if (!ret && !mddev->pers)
1192                 ret = ppl_write_empty_header(log);
1193
1194         __free_page(page);
1195         __free_page(page2);
1196
1197         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1198                  __func__, ret, ppl_conf->mismatch_count,
1199                  ppl_conf->recovered_entries);
1200         return ret;
1201 }
1202
1203 static int ppl_load(struct ppl_conf *ppl_conf)
1204 {
1205         int ret = 0;
1206         u32 signature = 0;
1207         bool signature_set = false;
1208         int i;
1209
1210         for (i = 0; i < ppl_conf->count; i++) {
1211                 struct ppl_log *log = &ppl_conf->child_logs[i];
1212
1213                 /* skip missing drive */
1214                 if (!log->rdev)
1215                         continue;
1216
1217                 ret = ppl_load_distributed(log);
1218                 if (ret)
1219                         break;
1220
1221                 /*
1222                  * For external metadata we can't check if the signature is
1223                  * correct on a single drive, but we can check if it is the same
1224                  * on all drives.
1225                  */
1226                 if (ppl_conf->mddev->external) {
1227                         if (!signature_set) {
1228                                 signature = ppl_conf->signature;
1229                                 signature_set = true;
1230                         } else if (signature != ppl_conf->signature) {
1231                                 pr_warn("md/raid:%s: PPL header signature does not match on all member drives\n",
1232                                         mdname(ppl_conf->mddev));
1233                                 ret = -EINVAL;
1234                                 break;
1235                         }
1236                 }
1237         }
1238
1239         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1240                  __func__, ret, ppl_conf->mismatch_count,
1241                  ppl_conf->recovered_entries);
1242         return ret;
1243 }
1244
1245 static void __ppl_exit_log(struct ppl_conf *ppl_conf)
1246 {
1247         clear_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1248         clear_bit(MD_HAS_MULTIPLE_PPLS, &ppl_conf->mddev->flags);
1249
1250         kfree(ppl_conf->child_logs);
1251
1252         bioset_exit(&ppl_conf->bs);
1253         bioset_exit(&ppl_conf->flush_bs);
1254         mempool_exit(&ppl_conf->io_pool);
1255         kmem_cache_destroy(ppl_conf->io_kc);
1256
1257         kfree(ppl_conf);
1258 }
1259
1260 void ppl_exit_log(struct r5conf *conf)
1261 {
1262         struct ppl_conf *ppl_conf = conf->log_private;
1263
1264         if (ppl_conf) {
1265                 __ppl_exit_log(ppl_conf);
1266                 conf->log_private = NULL;
1267         }
1268 }
1269
1270 static int ppl_validate_rdev(struct md_rdev *rdev)
1271 {
1272         char b[BDEVNAME_SIZE];
1273         int ppl_data_sectors;
1274         int ppl_size_new;
1275
1276         /*
1277          * The configured PPL size must be enough to store
1278          * the header and (at the very least) partial parity
1279          * for one stripe. Round it down to ensure the data
1280          * space is cleanly divisible by stripe size.
1281          */
1282         ppl_data_sectors = rdev->ppl.size - (PPL_HEADER_SIZE >> 9);
1283
1284         if (ppl_data_sectors > 0)
1285                 ppl_data_sectors = rounddown(ppl_data_sectors, STRIPE_SECTORS);
1286
1287         if (ppl_data_sectors <= 0) {
1288                 pr_warn("md/raid:%s: PPL space too small on %s\n",
1289                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1290                 return -ENOSPC;
1291         }
1292
1293         ppl_size_new = ppl_data_sectors + (PPL_HEADER_SIZE >> 9);
1294
1295         if ((rdev->ppl.sector < rdev->data_offset &&
1296              rdev->ppl.sector + ppl_size_new > rdev->data_offset) ||
1297             (rdev->ppl.sector >= rdev->data_offset &&
1298              rdev->data_offset + rdev->sectors > rdev->ppl.sector)) {
1299                 pr_warn("md/raid:%s: PPL space overlaps with data on %s\n",
1300                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1301                 return -EINVAL;
1302         }
1303
1304         if (!rdev->mddev->external &&
1305             ((rdev->ppl.offset > 0 && rdev->ppl.offset < (rdev->sb_size >> 9)) ||
1306              (rdev->ppl.offset <= 0 && rdev->ppl.offset + ppl_size_new > 0))) {
1307                 pr_warn("md/raid:%s: PPL space overlaps with superblock on %s\n",
1308                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1309                 return -EINVAL;
1310         }
1311
1312         rdev->ppl.size = ppl_size_new;
1313
1314         return 0;
1315 }
1316
1317 static void ppl_init_child_log(struct ppl_log *log, struct md_rdev *rdev)
1318 {
1319         struct request_queue *q;
1320
1321         if ((rdev->ppl.size << 9) >= (PPL_SPACE_SIZE +
1322                                       PPL_HEADER_SIZE) * 2) {
1323                 log->use_multippl = true;
1324                 set_bit(MD_HAS_MULTIPLE_PPLS,
1325                         &log->ppl_conf->mddev->flags);
1326                 log->entry_space = PPL_SPACE_SIZE;
1327         } else {
1328                 log->use_multippl = false;
1329                 log->entry_space = (log->rdev->ppl.size << 9) -
1330                                    PPL_HEADER_SIZE;
1331         }
1332         log->next_io_sector = rdev->ppl.sector;
1333
1334         q = bdev_get_queue(rdev->bdev);
1335         if (test_bit(QUEUE_FLAG_WC, &q->queue_flags))
1336                 log->wb_cache_on = true;
1337 }
1338
1339 int ppl_init_log(struct r5conf *conf)
1340 {
1341         struct ppl_conf *ppl_conf;
1342         struct mddev *mddev = conf->mddev;
1343         int ret = 0;
1344         int max_disks;
1345         int i;
1346
1347         pr_debug("md/raid:%s: enabling distributed Partial Parity Log\n",
1348                  mdname(conf->mddev));
1349
1350         if (PAGE_SIZE != 4096)
1351                 return -EINVAL;
1352
1353         if (mddev->level != 5) {
1354                 pr_warn("md/raid:%s PPL is not compatible with raid level %d\n",
1355                         mdname(mddev), mddev->level);
1356                 return -EINVAL;
1357         }
1358
1359         if (mddev->bitmap_info.file || mddev->bitmap_info.offset) {
1360                 pr_warn("md/raid:%s PPL is not compatible with bitmap\n",
1361                         mdname(mddev));
1362                 return -EINVAL;
1363         }
1364
1365         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
1366                 pr_warn("md/raid:%s PPL is not compatible with journal\n",
1367                         mdname(mddev));
1368                 return -EINVAL;
1369         }
1370
1371         max_disks = FIELD_SIZEOF(struct ppl_log, disk_flush_bitmap) *
1372                 BITS_PER_BYTE;
1373         if (conf->raid_disks > max_disks) {
1374                 pr_warn("md/raid:%s PPL doesn't support over %d disks in the array\n",
1375                         mdname(mddev), max_disks);
1376                 return -EINVAL;
1377         }
1378
1379         ppl_conf = kzalloc(sizeof(struct ppl_conf), GFP_KERNEL);
1380         if (!ppl_conf)
1381                 return -ENOMEM;
1382
1383         ppl_conf->mddev = mddev;
1384
1385         ppl_conf->io_kc = KMEM_CACHE(ppl_io_unit, 0);
1386         if (!ppl_conf->io_kc) {
1387                 ret = -ENOMEM;
1388                 goto err;
1389         }
1390
1391         ret = mempool_init(&ppl_conf->io_pool, conf->raid_disks, ppl_io_pool_alloc,
1392                            ppl_io_pool_free, ppl_conf->io_kc);
1393         if (ret)
1394                 goto err;
1395
1396         ret = bioset_init(&ppl_conf->bs, conf->raid_disks, 0, BIOSET_NEED_BVECS);
1397         if (ret)
1398                 goto err;
1399
1400         ret = bioset_init(&ppl_conf->flush_bs, conf->raid_disks, 0, 0);
1401         if (ret)
1402                 goto err;
1403
1404         ppl_conf->count = conf->raid_disks;
1405         ppl_conf->child_logs = kcalloc(ppl_conf->count, sizeof(struct ppl_log),
1406                                        GFP_KERNEL);
1407         if (!ppl_conf->child_logs) {
1408                 ret = -ENOMEM;
1409                 goto err;
1410         }
1411
1412         atomic64_set(&ppl_conf->seq, 0);
1413         INIT_LIST_HEAD(&ppl_conf->no_mem_stripes);
1414         spin_lock_init(&ppl_conf->no_mem_stripes_lock);
1415         ppl_conf->write_hint = RWF_WRITE_LIFE_NOT_SET;
1416
1417         if (!mddev->external) {
1418                 ppl_conf->signature = ~crc32c_le(~0, mddev->uuid, sizeof(mddev->uuid));
1419                 ppl_conf->block_size = 512;
1420         } else {
1421                 ppl_conf->block_size = queue_logical_block_size(mddev->queue);
1422         }
1423
1424         for (i = 0; i < ppl_conf->count; i++) {
1425                 struct ppl_log *log = &ppl_conf->child_logs[i];
1426                 struct md_rdev *rdev = conf->disks[i].rdev;
1427
1428                 mutex_init(&log->io_mutex);
1429                 spin_lock_init(&log->io_list_lock);
1430                 INIT_LIST_HEAD(&log->io_list);
1431
1432                 log->ppl_conf = ppl_conf;
1433                 log->rdev = rdev;
1434
1435                 if (rdev) {
1436                         ret = ppl_validate_rdev(rdev);
1437                         if (ret)
1438                                 goto err;
1439
1440                         ppl_init_child_log(log, rdev);
1441                 }
1442         }
1443
1444         /* load and possibly recover the logs from the member disks */
1445         ret = ppl_load(ppl_conf);
1446
1447         if (ret) {
1448                 goto err;
1449         } else if (!mddev->pers && mddev->recovery_cp == 0 &&
1450                    ppl_conf->recovered_entries > 0 &&
1451                    ppl_conf->mismatch_count == 0) {
1452                 /*
1453                  * If we are starting a dirty array and the recovery succeeds
1454                  * without any issues, set the array as clean.
1455                  */
1456                 mddev->recovery_cp = MaxSector;
1457                 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
1458         } else if (mddev->pers && ppl_conf->mismatch_count > 0) {
1459                 /* no mismatch allowed when enabling PPL for a running array */
1460                 ret = -EINVAL;
1461                 goto err;
1462         }
1463
1464         conf->log_private = ppl_conf;
1465         set_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1466
1467         return 0;
1468 err:
1469         __ppl_exit_log(ppl_conf);
1470         return ret;
1471 }
1472
1473 int ppl_modify_log(struct r5conf *conf, struct md_rdev *rdev, bool add)
1474 {
1475         struct ppl_conf *ppl_conf = conf->log_private;
1476         struct ppl_log *log;
1477         int ret = 0;
1478         char b[BDEVNAME_SIZE];
1479
1480         if (!rdev)
1481                 return -EINVAL;
1482
1483         pr_debug("%s: disk: %d operation: %s dev: %s\n",
1484                  __func__, rdev->raid_disk, add ? "add" : "remove",
1485                  bdevname(rdev->bdev, b));
1486
1487         if (rdev->raid_disk < 0)
1488                 return 0;
1489
1490         if (rdev->raid_disk >= ppl_conf->count)
1491                 return -ENODEV;
1492
1493         log = &ppl_conf->child_logs[rdev->raid_disk];
1494
1495         mutex_lock(&log->io_mutex);
1496         if (add) {
1497                 ret = ppl_validate_rdev(rdev);
1498                 if (!ret) {
1499                         log->rdev = rdev;
1500                         ret = ppl_write_empty_header(log);
1501                         ppl_init_child_log(log, rdev);
1502                 }
1503         } else {
1504                 log->rdev = NULL;
1505         }
1506         mutex_unlock(&log->io_mutex);
1507
1508         return ret;
1509 }
1510
1511 static ssize_t
1512 ppl_write_hint_show(struct mddev *mddev, char *buf)
1513 {
1514         size_t ret = 0;
1515         struct r5conf *conf;
1516         struct ppl_conf *ppl_conf = NULL;
1517
1518         spin_lock(&mddev->lock);
1519         conf = mddev->private;
1520         if (conf && raid5_has_ppl(conf))
1521                 ppl_conf = conf->log_private;
1522         ret = sprintf(buf, "%d\n", ppl_conf ? ppl_conf->write_hint : 0);
1523         spin_unlock(&mddev->lock);
1524
1525         return ret;
1526 }
1527
1528 static ssize_t
1529 ppl_write_hint_store(struct mddev *mddev, const char *page, size_t len)
1530 {
1531         struct r5conf *conf;
1532         struct ppl_conf *ppl_conf;
1533         int err = 0;
1534         unsigned short new;
1535
1536         if (len >= PAGE_SIZE)
1537                 return -EINVAL;
1538         if (kstrtou16(page, 10, &new))
1539                 return -EINVAL;
1540
1541         err = mddev_lock(mddev);
1542         if (err)
1543                 return err;
1544
1545         conf = mddev->private;
1546         if (!conf) {
1547                 err = -ENODEV;
1548         } else if (raid5_has_ppl(conf)) {
1549                 ppl_conf = conf->log_private;
1550                 if (!ppl_conf)
1551                         err = -EINVAL;
1552                 else
1553                         ppl_conf->write_hint = new;
1554         } else {
1555                 err = -EINVAL;
1556         }
1557
1558         mddev_unlock(mddev);
1559
1560         return err ?: len;
1561 }
1562
1563 struct md_sysfs_entry
1564 ppl_write_hint = __ATTR(ppl_write_hint, S_IRUGO | S_IWUSR,
1565                         ppl_write_hint_show,
1566                         ppl_write_hint_store);