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/flex_array.h>
  20 #include <linux/async_tx.h>
  21 #include <linux/raid/md_p.h>
  22 #include "md.h"
  23 #include "raid5.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
  90 #define PPL_SPACE_SIZE (128 * 1024)
  91
  92 struct ppl_conf {
  93         struct mddev *mddev;
  94
  95         /* array of child logs, one for each raid disk */
  96         struct ppl_log *child_logs;
  97         int count;
  98
  99         int block_size;         /* the logical block size used for data_sector
 100                                  * in ppl_header_entry */
 101         u32 signature;          /* raid array identifier */
 102         atomic64_t seq;         /* current log write sequence number */
 103
 104         struct kmem_cache *io_kc;
 105         mempool_t *io_pool;
 106         struct bio_set *bs;
 107
 108         /* used only for recovery */
 109         int recovered_entries;
 110         int mismatch_count;
 111
 112         /* stripes to retry if failed to allocate io_unit */
 113         struct list_head no_mem_stripes;
 114         spinlock_t no_mem_stripes_lock;
 115 };
 116
 117 struct ppl_log {
 118         struct ppl_conf *ppl_conf;      /* shared between all log instances */
 119
 120         struct md_rdev *rdev;           /* array member disk associated with
 121                                          * this log instance */
 122         struct mutex io_mutex;
 123         struct ppl_io_unit *current_io; /* current io_unit accepting new data
 124                                          * always at the end of io_list */
 125         spinlock_t io_list_lock;
 126         struct list_head io_list;       /* all io_units of this log */
 127
 128         sector_t next_io_sector;
 129         unsigned int entry_space;
 130         bool use_multippl;
 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
 149         bool submitted;                 /* true if write to log started */
 150
 151         /* inline bio and its biovec for submitting the iounit */
 152         struct bio bio;
 153         struct bio_vec biovec[PPL_IO_INLINE_BVECS];
 154 };
 155
 156 struct dma_async_tx_descriptor *
 157 ops_run_partial_parity(struct stripe_head *sh, struct raid5_percpu *percpu,
 158                        struct dma_async_tx_descriptor *tx)
 159 {
 160         int disks = sh->disks;
 161         struct page **srcs = flex_array_get(percpu->scribble, 0);
 162         int count = 0, pd_idx = sh->pd_idx, i;
 163         struct async_submit_ctl submit;
 164
 165         pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
 166
 167         /*
 168          * Partial parity is the XOR of stripe data chunks that are not changed
 169          * during the write request. Depending on available data
 170          * (read-modify-write vs. reconstruct-write case) we calculate it
 171          * differently.
 172          */
 173         if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
 174                 /*
 175                  * rmw: xor old data and parity from updated disks
 176                  * This is calculated earlier by ops_run_prexor5() so just copy
 177                  * the parity dev page.
 178                  */
 179                 srcs[count++] = sh->dev[pd_idx].page;
 180         } else if (sh->reconstruct_state == reconstruct_state_drain_run) {
 181                 /* rcw: xor data from all not updated disks */
 182                 for (i = disks; i--;) {
 183                         struct r5dev *dev = &sh->dev[i];
 184                         if (test_bit(R5_UPTODATE, &dev->flags))
 185                                 srcs[count++] = dev->page;
 186                 }
 187         } else {
 188                 return tx;
 189         }
 190
 191         init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, tx,
 192                           NULL, sh, flex_array_get(percpu->scribble, 0)
 193                           + sizeof(struct page *) * (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         bio_init(&io->bio, io->biovec, PPL_IO_INLINE_BVECS);
 253
 254         pplhdr = page_address(io->header_page);
 255         clear_page(pplhdr);
 256         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
 257         pplhdr->signature = cpu_to_le32(ppl_conf->signature);
 258
 259         io->seq = atomic64_add_return(1, &ppl_conf->seq);
 260         pplhdr->generation = cpu_to_le64(io->seq);
 261
 262         return io;
 263 }
 264
 265 static int ppl_log_stripe(struct ppl_log *log, struct stripe_head *sh)
 266 {
 267         struct ppl_io_unit *io = log->current_io;
 268         struct ppl_header_entry *e = NULL;
 269         struct ppl_header *pplhdr;
 270         int i;
 271         sector_t data_sector = 0;
 272         int data_disks = 0;
 273         struct r5conf *conf = sh->raid_conf;
 274
 275         pr_debug("%s: stripe: %llu\n", __func__, (unsigned long long)sh->sector);
 276
 277         /* check if current io_unit is full */
 278         if (io && (io->pp_size == log->entry_space ||
 279                    io->entries_count == PPL_HDR_MAX_ENTRIES)) {
 280                 pr_debug("%s: add io_unit blocked by seq: %llu\n",
 281                          __func__, io->seq);
 282                 io = NULL;
 283         }
 284
 285         /* add a new unit if there is none or the current is full */
 286         if (!io) {
 287                 io = ppl_new_iounit(log, sh);
 288                 if (!io)
 289                         return -ENOMEM;
 290                 spin_lock_irq(&log->io_list_lock);
 291                 list_add_tail(&io->log_sibling, &log->io_list);
 292                 spin_unlock_irq(&log->io_list_lock);
 293
 294                 log->current_io = io;
 295         }
 296
 297         for (i = 0; i < sh->disks; i++) {
 298                 struct r5dev *dev = &sh->dev[i];
 299
 300                 if (i != sh->pd_idx && test_bit(R5_Wantwrite, &dev->flags)) {
 301                         if (!data_disks || dev->sector < data_sector)
 302                                 data_sector = dev->sector;
 303                         data_disks++;
 304                 }
 305         }
 306         BUG_ON(!data_disks);
 307
 308         pr_debug("%s: seq: %llu data_sector: %llu data_disks: %d\n", __func__,
 309                  io->seq, (unsigned long long)data_sector, data_disks);
 310
 311         pplhdr = page_address(io->header_page);
 312
 313         if (io->entries_count > 0) {
 314                 struct ppl_header_entry *last =
 315                                 &pplhdr->entries[io->entries_count - 1];
 316                 struct stripe_head *sh_last = list_last_entry(
 317                                 &io->stripe_list, struct stripe_head, log_list);
 318                 u64 data_sector_last = le64_to_cpu(last->data_sector);
 319                 u32 data_size_last = le32_to_cpu(last->data_size);
 320
 321                 /*
 322                  * Check if we can append the stripe to the last entry. It must
 323                  * be just after the last logged stripe and write to the same
 324                  * disks. Use bit shift and logarithm to avoid 64-bit division.
 325                  */
 326                 if ((sh->sector == sh_last->sector + STRIPE_SECTORS) &&
 327                     (data_sector >> ilog2(conf->chunk_sectors) ==
 328                      data_sector_last >> ilog2(conf->chunk_sectors)) &&
 329                     ((data_sector - data_sector_last) * data_disks ==
 330                      data_size_last >> 9))
 331                         e = last;
 332         }
 333
 334         if (!e) {
 335                 e = &pplhdr->entries[io->entries_count++];
 336                 e->data_sector = cpu_to_le64(data_sector);
 337                 e->parity_disk = cpu_to_le32(sh->pd_idx);
 338                 e->checksum = cpu_to_le32(~0);
 339         }
 340
 341         le32_add_cpu(&e->data_size, data_disks << PAGE_SHIFT);
 342
 343         /* don't write any PP if full stripe write */
 344         if (!test_bit(STRIPE_FULL_WRITE, &sh->state)) {
 345                 le32_add_cpu(&e->pp_size, PAGE_SIZE);
 346                 io->pp_size += PAGE_SIZE;
 347                 e->checksum = cpu_to_le32(crc32c_le(le32_to_cpu(e->checksum),
 348                                                     page_address(sh->ppl_page),
 349                                                     PAGE_SIZE));
 350         }
 351
 352         list_add_tail(&sh->log_list, &io->stripe_list);
 353         atomic_inc(&io->pending_stripes);
 354         sh->ppl_io = io;
 355
 356         return 0;
 357 }
 358
 359 int ppl_write_stripe(struct r5conf *conf, struct stripe_head *sh)
 360 {
 361         struct ppl_conf *ppl_conf = conf->log_private;
 362         struct ppl_io_unit *io = sh->ppl_io;
 363         struct ppl_log *log;
 364
 365         if (io || test_bit(STRIPE_SYNCING, &sh->state) || !sh->ppl_page ||
 366             !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 367             !test_bit(R5_Insync, &sh->dev[sh->pd_idx].flags)) {
 368                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 369                 return -EAGAIN;
 370         }
 371
 372         log = &ppl_conf->child_logs[sh->pd_idx];
 373
 374         mutex_lock(&log->io_mutex);
 375
 376         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 377                 mutex_unlock(&log->io_mutex);
 378                 return -EAGAIN;
 379         }
 380
 381         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 382         clear_bit(STRIPE_DELAYED, &sh->state);
 383         atomic_inc(&sh->count);
 384
 385         if (ppl_log_stripe(log, sh)) {
 386                 spin_lock_irq(&ppl_conf->no_mem_stripes_lock);
 387                 list_add_tail(&sh->log_list, &ppl_conf->no_mem_stripes);
 388                 spin_unlock_irq(&ppl_conf->no_mem_stripes_lock);
 389         }
 390
 391         mutex_unlock(&log->io_mutex);
 392
 393         return 0;
 394 }
 395
 396 static void ppl_log_endio(struct bio *bio)
 397 {
 398         struct ppl_io_unit *io = bio->bi_private;
 399         struct ppl_log *log = io->log;
 400         struct ppl_conf *ppl_conf = log->ppl_conf;
 401         struct stripe_head *sh, *next;
 402
 403         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 404
 405         if (bio->bi_status)
 406                 md_error(ppl_conf->mddev, log->rdev);
 407
 408         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 409                 list_del_init(&sh->log_list);
 410
 411                 set_bit(STRIPE_HANDLE, &sh->state);
 412                 raid5_release_stripe(sh);
 413         }
 414 }
 415
 416 static void ppl_submit_iounit_bio(struct ppl_io_unit *io, struct bio *bio)
 417 {
 418         char b[BDEVNAME_SIZE];
 419
 420         pr_debug("%s: seq: %llu size: %u sector: %llu dev: %s\n",
 421                  __func__, io->seq, bio->bi_iter.bi_size,
 422                  (unsigned long long)bio->bi_iter.bi_sector,
 423                  bio_devname(bio, b));
 424
 425         submit_bio(bio);
 426 }
 427
 428 static void ppl_submit_iounit(struct ppl_io_unit *io)
 429 {
 430         struct ppl_log *log = io->log;
 431         struct ppl_conf *ppl_conf = log->ppl_conf;
 432         struct ppl_header *pplhdr = page_address(io->header_page);
 433         struct bio *bio = &io->bio;
 434         struct stripe_head *sh;
 435         int i;
 436
 437         bio->bi_private = io;
 438
 439         if (!log->rdev || test_bit(Faulty, &log->rdev->flags)) {
 440                 ppl_log_endio(bio);
 441                 return;
 442         }
 443
 444         for (i = 0; i < io->entries_count; i++) {
 445                 struct ppl_header_entry *e = &pplhdr->entries[i];
 446
 447                 pr_debug("%s: seq: %llu entry: %d data_sector: %llu pp_size: %u data_size: %u\n",
 448                          __func__, io->seq, i, le64_to_cpu(e->data_sector),
 449                          le32_to_cpu(e->pp_size), le32_to_cpu(e->data_size));
 450
 451                 e->data_sector = cpu_to_le64(le64_to_cpu(e->data_sector) >>
 452                                              ilog2(ppl_conf->block_size >> 9));
 453                 e->checksum = cpu_to_le32(~le32_to_cpu(e->checksum));
 454         }
 455
 456         pplhdr->entries_count = cpu_to_le32(io->entries_count);
 457         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PPL_HEADER_SIZE));
 458
 459         /* Rewind the buffer if current PPL is larger then remaining space */
 460         if (log->use_multippl &&
 461             log->rdev->ppl.sector + log->rdev->ppl.size - log->next_io_sector <
 462             (PPL_HEADER_SIZE + io->pp_size) >> 9)
 463                 log->next_io_sector = log->rdev->ppl.sector;
 464
 465
 466         bio->bi_end_io = ppl_log_endio;
 467         bio->bi_opf = REQ_OP_WRITE | REQ_FUA;
 468         bio_set_dev(bio, log->rdev->bdev);
 469         bio->bi_iter.bi_sector = log->next_io_sector;
 470         bio_add_page(bio, io->header_page, PAGE_SIZE, 0);
 471
 472         pr_debug("%s: log->current_io_sector: %llu\n", __func__,
 473             (unsigned long long)log->next_io_sector);
 474
 475         if (log->use_multippl)
 476                 log->next_io_sector += (PPL_HEADER_SIZE + io->pp_size) >> 9;
 477
 478         list_for_each_entry(sh, &io->stripe_list, log_list) {
 479                 /* entries for full stripe writes have no partial parity */
 480                 if (test_bit(STRIPE_FULL_WRITE, &sh->state))
 481                         continue;
 482
 483                 if (!bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0)) {
 484                         struct bio *prev = bio;
 485
 486                         bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES,
 487                                                ppl_conf->bs);
 488                         bio->bi_opf = prev->bi_opf;
 489                         bio_copy_dev(bio, prev);
 490                         bio->bi_iter.bi_sector = bio_end_sector(prev);
 491                         bio_add_page(bio, sh->ppl_page, PAGE_SIZE, 0);
 492
 493                         bio_chain(bio, prev);
 494                         ppl_submit_iounit_bio(io, prev);
 495                 }
 496         }
 497
 498         ppl_submit_iounit_bio(io, bio);
 499 }
 500
 501 static void ppl_submit_current_io(struct ppl_log *log)
 502 {
 503         struct ppl_io_unit *io;
 504
 505         spin_lock_irq(&log->io_list_lock);
 506
 507         io = list_first_entry_or_null(&log->io_list, struct ppl_io_unit,
 508                                       log_sibling);
 509         if (io && io->submitted)
 510                 io = NULL;
 511
 512         spin_unlock_irq(&log->io_list_lock);
 513
 514         if (io) {
 515                 io->submitted = true;
 516
 517                 if (io == log->current_io)
 518                         log->current_io = NULL;
 519
 520                 ppl_submit_iounit(io);
 521         }
 522 }
 523
 524 void ppl_write_stripe_run(struct r5conf *conf)
 525 {
 526         struct ppl_conf *ppl_conf = conf->log_private;
 527         struct ppl_log *log;
 528         int i;
 529
 530         for (i = 0; i < ppl_conf->count; i++) {
 531                 log = &ppl_conf->child_logs[i];
 532
 533                 mutex_lock(&log->io_mutex);
 534                 ppl_submit_current_io(log);
 535                 mutex_unlock(&log->io_mutex);
 536         }
 537 }
 538
 539 static void ppl_io_unit_finished(struct ppl_io_unit *io)
 540 {
 541         struct ppl_log *log = io->log;
 542         struct ppl_conf *ppl_conf = log->ppl_conf;
 543         unsigned long flags;
 544
 545         pr_debug("%s: seq: %llu\n", __func__, io->seq);
 546
 547         local_irq_save(flags);
 548
 549         spin_lock(&log->io_list_lock);
 550         list_del(&io->log_sibling);
 551         spin_unlock(&log->io_list_lock);
 552
 553         mempool_free(io, ppl_conf->io_pool);
 554
 555         spin_lock(&ppl_conf->no_mem_stripes_lock);
 556         if (!list_empty(&ppl_conf->no_mem_stripes)) {
 557                 struct stripe_head *sh;
 558
 559                 sh = list_first_entry(&ppl_conf->no_mem_stripes,
 560                                       struct stripe_head, log_list);
 561                 list_del_init(&sh->log_list);
 562                 set_bit(STRIPE_HANDLE, &sh->state);
 563                 raid5_release_stripe(sh);
 564         }
 565         spin_unlock(&ppl_conf->no_mem_stripes_lock);
 566
 567         local_irq_restore(flags);
 568 }
 569
 570 void ppl_stripe_write_finished(struct stripe_head *sh)
 571 {
 572         struct ppl_io_unit *io;
 573
 574         io = sh->ppl_io;
 575         sh->ppl_io = NULL;
 576
 577         if (io && atomic_dec_and_test(&io->pending_stripes))
 578                 ppl_io_unit_finished(io);
 579 }
 580
 581 static void ppl_xor(int size, struct page *page1, struct page *page2)
 582 {
 583         struct async_submit_ctl submit;
 584         struct dma_async_tx_descriptor *tx;
 585         struct page *xor_srcs[] = { page1, page2 };
 586
 587         init_async_submit(&submit, ASYNC_TX_ACK|ASYNC_TX_XOR_DROP_DST,
 588                           NULL, NULL, NULL, NULL);
 589         tx = async_xor(page1, xor_srcs, 0, 2, size, &submit);
 590
 591         async_tx_quiesce(&tx);
 592 }
 593
 594 /*
 595  * PPL recovery strategy: xor partial parity and data from all modified data
 596  * disks within a stripe and write the result as the new stripe parity. If all
 597  * stripe data disks are modified (full stripe write), no partial parity is
 598  * available, so just xor the data disks.
 599  *
 600  * Recovery of a PPL entry shall occur only if all modified data disks are
 601  * available and read from all of them succeeds.
 602  *
 603  * A PPL entry applies to a stripe, partial parity size for an entry is at most
 604  * the size of the chunk. Examples of possible cases for a single entry:
 605  *
 606  * case 0: single data disk write:
 607  *   data0    data1    data2     ppl        parity
 608  * +--------+--------+--------+           +--------------------+
 609  * | ------ | ------ | ------ | +----+    | (no change)        |
 610  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 611  * | ------ | -data- | ------ | | pp | -> | data1 ^ pp         |
 612  * | ------ | ------ | ------ | +----+    | (no change)        |
 613  * +--------+--------+--------+           +--------------------+
 614  * pp_size = data_size
 615  *
 616  * case 1: more than one data disk write:
 617  *   data0    data1    data2     ppl        parity
 618  * +--------+--------+--------+           +--------------------+
 619  * | ------ | ------ | ------ | +----+    | (no change)        |
 620  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 621  * | -data- | -data- | ------ | | pp | -> | data0 ^ data1 ^ pp |
 622  * | ------ | ------ | ------ | +----+    | (no change)        |
 623  * +--------+--------+--------+           +--------------------+
 624  * pp_size = data_size / modified_data_disks
 625  *
 626  * case 2: write to all data disks (also full stripe write):
 627  *   data0    data1    data2                parity
 628  * +--------+--------+--------+           +--------------------+
 629  * | ------ | ------ | ------ |           | (no change)        |
 630  * | -data- | -data- | -data- | --------> | xor all data       |
 631  * | ------ | ------ | ------ | --------> | (no change)        |
 632  * | ------ | ------ | ------ |           | (no change)        |
 633  * +--------+--------+--------+           +--------------------+
 634  * pp_size = 0
 635  *
 636  * The following cases are possible only in other implementations. The recovery
 637  * code can handle them, but they are not generated at runtime because they can
 638  * be reduced to cases 0, 1 and 2:
 639  *
 640  * case 3:
 641  *   data0    data1    data2     ppl        parity
 642  * +--------+--------+--------+ +----+    +--------------------+
 643  * | ------ | -data- | -data- | | pp |    | data1 ^ data2 ^ pp |
 644  * | ------ | -data- | -data- | | pp | -> | data1 ^ data2 ^ pp |
 645  * | -data- | -data- | -data- | | -- | -> | xor all data       |
 646  * | -data- | -data- | ------ | | pp |    | data0 ^ data1 ^ pp |
 647  * +--------+--------+--------+ +----+    +--------------------+
 648  * pp_size = chunk_size
 649  *
 650  * case 4:
 651  *   data0    data1    data2     ppl        parity
 652  * +--------+--------+--------+ +----+    +--------------------+
 653  * | ------ | -data- | ------ | | pp |    | data1 ^ pp         |
 654  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 655  * | ------ | ------ | ------ | | -- | -> | (no change)        |
 656  * | -data- | ------ | ------ | | pp |    | data0 ^ pp         |
 657  * +--------+--------+--------+ +----+    +--------------------+
 658  * pp_size = chunk_size
 659  */
 660 static int ppl_recover_entry(struct ppl_log *log, struct ppl_header_entry *e,
 661                              sector_t ppl_sector)
 662 {
 663         struct ppl_conf *ppl_conf = log->ppl_conf;
 664         struct mddev *mddev = ppl_conf->mddev;
 665         struct r5conf *conf = mddev->private;
 666         int block_size = ppl_conf->block_size;
 667         struct page *page1;
 668         struct page *page2;
 669         sector_t r_sector_first;
 670         sector_t r_sector_last;
 671         int strip_sectors;
 672         int data_disks;
 673         int i;
 674         int ret = 0;
 675         char b[BDEVNAME_SIZE];
 676         unsigned int pp_size = le32_to_cpu(e->pp_size);
 677         unsigned int data_size = le32_to_cpu(e->data_size);
 678
 679         page1 = alloc_page(GFP_KERNEL);
 680         page2 = alloc_page(GFP_KERNEL);
 681
 682         if (!page1 || !page2) {
 683                 ret = -ENOMEM;
 684                 goto out;
 685         }
 686
 687         r_sector_first = le64_to_cpu(e->data_sector) * (block_size >> 9);
 688
 689         if ((pp_size >> 9) < conf->chunk_sectors) {
 690                 if (pp_size > 0) {
 691                         data_disks = data_size / pp_size;
 692                         strip_sectors = pp_size >> 9;
 693                 } else {
 694                         data_disks = conf->raid_disks - conf->max_degraded;
 695                         strip_sectors = (data_size >> 9) / data_disks;
 696                 }
 697                 r_sector_last = r_sector_first +
 698                                 (data_disks - 1) * conf->chunk_sectors +
 699                                 strip_sectors;
 700         } else {
 701                 data_disks = conf->raid_disks - conf->max_degraded;
 702                 strip_sectors = conf->chunk_sectors;
 703                 r_sector_last = r_sector_first + (data_size >> 9);
 704         }
 705
 706         pr_debug("%s: array sector first: %llu last: %llu\n", __func__,
 707                  (unsigned long long)r_sector_first,
 708                  (unsigned long long)r_sector_last);
 709
 710         /* if start and end is 4k aligned, use a 4k block */
 711         if (block_size == 512 &&
 712             (r_sector_first & (STRIPE_SECTORS - 1)) == 0 &&
 713             (r_sector_last & (STRIPE_SECTORS - 1)) == 0)
 714                 block_size = STRIPE_SIZE;
 715
 716         /* iterate through blocks in strip */
 717         for (i = 0; i < strip_sectors; i += (block_size >> 9)) {
 718                 bool update_parity = false;
 719                 sector_t parity_sector;
 720                 struct md_rdev *parity_rdev;
 721                 struct stripe_head sh;
 722                 int disk;
 723                 int indent = 0;
 724
 725                 pr_debug("%s:%*s iter %d start\n", __func__, indent, "", i);
 726                 indent += 2;
 727
 728                 memset(page_address(page1), 0, PAGE_SIZE);
 729
 730                 /* iterate through data member disks */
 731                 for (disk = 0; disk < data_disks; disk++) {
 732                         int dd_idx;
 733                         struct md_rdev *rdev;
 734                         sector_t sector;
 735                         sector_t r_sector = r_sector_first + i +
 736                                             (disk * conf->chunk_sectors);
 737
 738                         pr_debug("%s:%*s data member disk %d start\n",
 739                                  __func__, indent, "", disk);
 740                         indent += 2;
 741
 742                         if (r_sector >= r_sector_last) {
 743                                 pr_debug("%s:%*s array sector %llu doesn't need parity update\n",
 744                                          __func__, indent, "",
 745                                          (unsigned long long)r_sector);
 746                                 indent -= 2;
 747                                 continue;
 748                         }
 749
 750                         update_parity = true;
 751
 752                         /* map raid sector to member disk */
 753                         sector = raid5_compute_sector(conf, r_sector, 0,
 754                                                       &dd_idx, NULL);
 755                         pr_debug("%s:%*s processing array sector %llu => data member disk %d, sector %llu\n",
 756                                  __func__, indent, "",
 757                                  (unsigned long long)r_sector, dd_idx,
 758                                  (unsigned long long)sector);
 759
 760                         rdev = conf->disks[dd_idx].rdev;
 761                         if (!rdev || (!test_bit(In_sync, &rdev->flags) &&
 762                                       sector >= rdev->recovery_offset)) {
 763                                 pr_debug("%s:%*s data member disk %d missing\n",
 764                                          __func__, indent, "", dd_idx);
 765                                 update_parity = false;
 766                                 break;
 767                         }
 768
 769                         pr_debug("%s:%*s reading data member disk %s sector %llu\n",
 770                                  __func__, indent, "", bdevname(rdev->bdev, b),
 771                                  (unsigned long long)sector);
 772                         if (!sync_page_io(rdev, sector, block_size, page2,
 773                                         REQ_OP_READ, 0, false)) {
 774                                 md_error(mddev, rdev);
 775                                 pr_debug("%s:%*s read failed!\n", __func__,
 776                                          indent, "");
 777                                 ret = -EIO;
 778                                 goto out;
 779                         }
 780
 781                         ppl_xor(block_size, page1, page2);
 782
 783                         indent -= 2;
 784                 }
 785
 786                 if (!update_parity)
 787                         continue;
 788
 789                 if (pp_size > 0) {
 790                         pr_debug("%s:%*s reading pp disk sector %llu\n",
 791                                  __func__, indent, "",
 792                                  (unsigned long long)(ppl_sector + i));
 793                         if (!sync_page_io(log->rdev,
 794                                         ppl_sector - log->rdev->data_offset + i,
 795                                         block_size, page2, REQ_OP_READ, 0,
 796                                         false)) {
 797                                 pr_debug("%s:%*s read failed!\n", __func__,
 798                                          indent, "");
 799                                 md_error(mddev, log->rdev);
 800                                 ret = -EIO;
 801                                 goto out;
 802                         }
 803
 804                         ppl_xor(block_size, page1, page2);
 805                 }
 806
 807                 /* map raid sector to parity disk */
 808                 parity_sector = raid5_compute_sector(conf, r_sector_first + i,
 809                                 0, &disk, &sh);
 810                 BUG_ON(sh.pd_idx != le32_to_cpu(e->parity_disk));
 811                 parity_rdev = conf->disks[sh.pd_idx].rdev;
 812
 813                 BUG_ON(parity_rdev->bdev->bd_dev != log->rdev->bdev->bd_dev);
 814                 pr_debug("%s:%*s write parity at sector %llu, disk %s\n",
 815                          __func__, indent, "",
 816                          (unsigned long long)parity_sector,
 817                          bdevname(parity_rdev->bdev, b));
 818                 if (!sync_page_io(parity_rdev, parity_sector, block_size,
 819                                 page1, REQ_OP_WRITE, 0, false)) {
 820                         pr_debug("%s:%*s parity write error!\n", __func__,
 821                                  indent, "");
 822                         md_error(mddev, parity_rdev);
 823                         ret = -EIO;
 824                         goto out;
 825                 }
 826         }
 827 out:
 828         if (page1)
 829                 __free_page(page1);
 830         if (page2)
 831                 __free_page(page2);
 832         return ret;
 833 }
 834
 835 static int ppl_recover(struct ppl_log *log, struct ppl_header *pplhdr,
 836                        sector_t offset)
 837 {
 838         struct ppl_conf *ppl_conf = log->ppl_conf;
 839         struct md_rdev *rdev = log->rdev;
 840         struct mddev *mddev = rdev->mddev;
 841         sector_t ppl_sector = rdev->ppl.sector + offset +
 842                               (PPL_HEADER_SIZE >> 9);
 843         struct page *page;
 844         int i;
 845         int ret = 0;
 846
 847         page = alloc_page(GFP_KERNEL);
 848         if (!page)
 849                 return -ENOMEM;
 850
 851         /* iterate through all PPL entries saved */
 852         for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++) {
 853                 struct ppl_header_entry *e = &pplhdr->entries[i];
 854                 u32 pp_size = le32_to_cpu(e->pp_size);
 855                 sector_t sector = ppl_sector;
 856                 int ppl_entry_sectors = pp_size >> 9;
 857                 u32 crc, crc_stored;
 858
 859                 pr_debug("%s: disk: %d entry: %d ppl_sector: %llu pp_size: %u\n",
 860                          __func__, rdev->raid_disk, i,
 861                          (unsigned long long)ppl_sector, pp_size);
 862
 863                 crc = ~0;
 864                 crc_stored = le32_to_cpu(e->checksum);
 865
 866                 /* read parial parity for this entry and calculate its checksum */
 867                 while (pp_size) {
 868                         int s = pp_size > PAGE_SIZE ? PAGE_SIZE : pp_size;
 869
 870                         if (!sync_page_io(rdev, sector - rdev->data_offset,
 871                                         s, page, REQ_OP_READ, 0, false)) {
 872                                 md_error(mddev, rdev);
 873                                 ret = -EIO;
 874                                 goto out;
 875                         }
 876
 877                         crc = crc32c_le(crc, page_address(page), s);
 878
 879                         pp_size -= s;
 880                         sector += s >> 9;
 881                 }
 882
 883                 crc = ~crc;
 884
 885                 if (crc != crc_stored) {
 886                         /*
 887                          * Don't recover this entry if the checksum does not
 888                          * match, but keep going and try to recover other
 889                          * entries.
 890                          */
 891                         pr_debug("%s: ppl entry crc does not match: stored: 0x%x calculated: 0x%x\n",
 892                                  __func__, crc_stored, crc);
 893                         ppl_conf->mismatch_count++;
 894                 } else {
 895                         ret = ppl_recover_entry(log, e, ppl_sector);
 896                         if (ret)
 897                                 goto out;
 898                         ppl_conf->recovered_entries++;
 899                 }
 900
 901                 ppl_sector += ppl_entry_sectors;
 902         }
 903
 904         /* flush the disk cache after recovery if necessary */
 905         ret = blkdev_issue_flush(rdev->bdev, GFP_KERNEL, NULL);
 906 out:
 907         __free_page(page);
 908         return ret;
 909 }
 910
 911 static int ppl_write_empty_header(struct ppl_log *log)
 912 {
 913         struct page *page;
 914         struct ppl_header *pplhdr;
 915         struct md_rdev *rdev = log->rdev;
 916         int ret = 0;
 917
 918         pr_debug("%s: disk: %d ppl_sector: %llu\n", __func__,
 919                  rdev->raid_disk, (unsigned long long)rdev->ppl.sector);
 920
 921         page = alloc_page(GFP_NOIO | __GFP_ZERO);
 922         if (!page)
 923                 return -ENOMEM;
 924
 925         pplhdr = page_address(page);
 926         /* zero out PPL space to avoid collision with old PPLs */
 927         blkdev_issue_zeroout(rdev->bdev, rdev->ppl.sector,
 928                             log->rdev->ppl.size, GFP_NOIO, 0);
 929         memset(pplhdr->reserved, 0xff, PPL_HDR_RESERVED);
 930         pplhdr->signature = cpu_to_le32(log->ppl_conf->signature);
 931         pplhdr->checksum = cpu_to_le32(~crc32c_le(~0, pplhdr, PAGE_SIZE));
 932
 933         if (!sync_page_io(rdev, rdev->ppl.sector - rdev->data_offset,
 934                           PPL_HEADER_SIZE, page, REQ_OP_WRITE | REQ_SYNC |
 935                           REQ_FUA, 0, false)) {
 936                 md_error(rdev->mddev, rdev);
 937                 ret = -EIO;
 938         }
 939
 940         __free_page(page);
 941         return ret;
 942 }
 943
 944 static int ppl_load_distributed(struct ppl_log *log)
 945 {
 946         struct ppl_conf *ppl_conf = log->ppl_conf;
 947         struct md_rdev *rdev = log->rdev;
 948         struct mddev *mddev = rdev->mddev;
 949         struct page *page, *page2, *tmp;
 950         struct ppl_header *pplhdr = NULL, *prev_pplhdr = NULL;
 951         u32 crc, crc_stored;
 952         u32 signature;
 953         int ret = 0, i;
 954         sector_t pplhdr_offset = 0, prev_pplhdr_offset = 0;
 955
 956         pr_debug("%s: disk: %d\n", __func__, rdev->raid_disk);
 957         /* read PPL headers, find the recent one */
 958         page = alloc_page(GFP_KERNEL);
 959         if (!page)
 960                 return -ENOMEM;
 961
 962         page2 = alloc_page(GFP_KERNEL);
 963         if (!page2) {
 964                 __free_page(page);
 965                 return -ENOMEM;
 966         }
 967
 968         /* searching ppl area for latest ppl */
 969         while (pplhdr_offset < rdev->ppl.size - (PPL_HEADER_SIZE >> 9)) {
 970                 if (!sync_page_io(rdev,
 971                                   rdev->ppl.sector - rdev->data_offset +
 972                                   pplhdr_offset, PAGE_SIZE, page, REQ_OP_READ,
 973                                   0, false)) {
 974                         md_error(mddev, rdev);
 975                         ret = -EIO;
 976                         /* if not able to read - don't recover any PPL */
 977                         pplhdr = NULL;
 978                         break;
 979                 }
 980                 pplhdr = page_address(page);
 981
 982                 /* check header validity */
 983                 crc_stored = le32_to_cpu(pplhdr->checksum);
 984                 pplhdr->checksum = 0;
 985                 crc = ~crc32c_le(~0, pplhdr, PAGE_SIZE);
 986
 987                 if (crc_stored != crc) {
 988                         pr_debug("%s: ppl header crc does not match: stored: 0x%x calculated: 0x%x (offset: %llu)\n",
 989                                  __func__, crc_stored, crc,
 990                                  (unsigned long long)pplhdr_offset);
 991                         pplhdr = prev_pplhdr;
 992                         pplhdr_offset = prev_pplhdr_offset;
 993                         break;
 994                 }
 995
 996                 signature = le32_to_cpu(pplhdr->signature);
 997
 998                 if (mddev->external) {
 999                         /*
1000                          * For external metadata the header signature is set and
1001                          * validated in userspace.
1002                          */
1003                         ppl_conf->signature = signature;
1004                 } else if (ppl_conf->signature != signature) {
1005                         pr_debug("%s: ppl header signature does not match: stored: 0x%x configured: 0x%x (offset: %llu)\n",
1006                                  __func__, signature, ppl_conf->signature,
1007                                  (unsigned long long)pplhdr_offset);
1008                         pplhdr = prev_pplhdr;
1009                         pplhdr_offset = prev_pplhdr_offset;
1010                         break;
1011                 }
1012
1013                 if (prev_pplhdr && le64_to_cpu(prev_pplhdr->generation) >
1014                     le64_to_cpu(pplhdr->generation)) {
1015                         /* previous was newest */
1016                         pplhdr = prev_pplhdr;
1017                         pplhdr_offset = prev_pplhdr_offset;
1018                         break;
1019                 }
1020
1021                 prev_pplhdr_offset = pplhdr_offset;
1022                 prev_pplhdr = pplhdr;
1023
1024                 tmp = page;
1025                 page = page2;
1026                 page2 = tmp;
1027
1028                 /* calculate next potential ppl offset */
1029                 for (i = 0; i < le32_to_cpu(pplhdr->entries_count); i++)
1030                         pplhdr_offset +=
1031                             le32_to_cpu(pplhdr->entries[i].pp_size) >> 9;
1032                 pplhdr_offset += PPL_HEADER_SIZE >> 9;
1033         }
1034
1035         /* no valid ppl found */
1036         if (!pplhdr)
1037                 ppl_conf->mismatch_count++;
1038         else
1039                 pr_debug("%s: latest PPL found at offset: %llu, with generation: %llu\n",
1040                     __func__, (unsigned long long)pplhdr_offset,
1041                     le64_to_cpu(pplhdr->generation));
1042
1043         /* attempt to recover from log if we are starting a dirty array */
1044         if (pplhdr && !mddev->pers && mddev->recovery_cp != MaxSector)
1045                 ret = ppl_recover(log, pplhdr, pplhdr_offset);
1046
1047         /* write empty header if we are starting the array */
1048         if (!ret && !mddev->pers)
1049                 ret = ppl_write_empty_header(log);
1050
1051         __free_page(page);
1052         __free_page(page2);
1053
1054         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1055                  __func__, ret, ppl_conf->mismatch_count,
1056                  ppl_conf->recovered_entries);
1057         return ret;
1058 }
1059
1060 static int ppl_load(struct ppl_conf *ppl_conf)
1061 {
1062         int ret = 0;
1063         u32 signature = 0;
1064         bool signature_set = false;
1065         int i;
1066
1067         for (i = 0; i < ppl_conf->count; i++) {
1068                 struct ppl_log *log = &ppl_conf->child_logs[i];
1069
1070                 /* skip missing drive */
1071                 if (!log->rdev)
1072                         continue;
1073
1074                 ret = ppl_load_distributed(log);
1075                 if (ret)
1076                         break;
1077
1078                 /*
1079                  * For external metadata we can't check if the signature is
1080                  * correct on a single drive, but we can check if it is the same
1081                  * on all drives.
1082                  */
1083                 if (ppl_conf->mddev->external) {
1084                         if (!signature_set) {
1085                                 signature = ppl_conf->signature;
1086                                 signature_set = true;
1087                         } else if (signature != ppl_conf->signature) {
1088                                 pr_warn("md/raid:%s: PPL header signature does not match on all member drives\n",
1089                                         mdname(ppl_conf->mddev));
1090                                 ret = -EINVAL;
1091                                 break;
1092                         }
1093                 }
1094         }
1095
1096         pr_debug("%s: return: %d mismatch_count: %d recovered_entries: %d\n",
1097                  __func__, ret, ppl_conf->mismatch_count,
1098                  ppl_conf->recovered_entries);
1099         return ret;
1100 }
1101
1102 static void __ppl_exit_log(struct ppl_conf *ppl_conf)
1103 {
1104         clear_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1105         clear_bit(MD_HAS_MULTIPLE_PPLS, &ppl_conf->mddev->flags);
1106
1107         kfree(ppl_conf->child_logs);
1108
1109         if (ppl_conf->bs)
1110                 bioset_free(ppl_conf->bs);
1111         mempool_destroy(ppl_conf->io_pool);
1112         kmem_cache_destroy(ppl_conf->io_kc);
1113
1114         kfree(ppl_conf);
1115 }
1116
1117 void ppl_exit_log(struct r5conf *conf)
1118 {
1119         struct ppl_conf *ppl_conf = conf->log_private;
1120
1121         if (ppl_conf) {
1122                 __ppl_exit_log(ppl_conf);
1123                 conf->log_private = NULL;
1124         }
1125 }
1126
1127 static int ppl_validate_rdev(struct md_rdev *rdev)
1128 {
1129         char b[BDEVNAME_SIZE];
1130         int ppl_data_sectors;
1131         int ppl_size_new;
1132
1133         /*
1134          * The configured PPL size must be enough to store
1135          * the header and (at the very least) partial parity
1136          * for one stripe. Round it down to ensure the data
1137          * space is cleanly divisible by stripe size.
1138          */
1139         ppl_data_sectors = rdev->ppl.size - (PPL_HEADER_SIZE >> 9);
1140
1141         if (ppl_data_sectors > 0)
1142                 ppl_data_sectors = rounddown(ppl_data_sectors, STRIPE_SECTORS);
1143
1144         if (ppl_data_sectors <= 0) {
1145                 pr_warn("md/raid:%s: PPL space too small on %s\n",
1146                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1147                 return -ENOSPC;
1148         }
1149
1150         ppl_size_new = ppl_data_sectors + (PPL_HEADER_SIZE >> 9);
1151
1152         if ((rdev->ppl.sector < rdev->data_offset &&
1153              rdev->ppl.sector + ppl_size_new > rdev->data_offset) ||
1154             (rdev->ppl.sector >= rdev->data_offset &&
1155              rdev->data_offset + rdev->sectors > rdev->ppl.sector)) {
1156                 pr_warn("md/raid:%s: PPL space overlaps with data on %s\n",
1157                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1158                 return -EINVAL;
1159         }
1160
1161         if (!rdev->mddev->external &&
1162             ((rdev->ppl.offset > 0 && rdev->ppl.offset < (rdev->sb_size >> 9)) ||
1163              (rdev->ppl.offset <= 0 && rdev->ppl.offset + ppl_size_new > 0))) {
1164                 pr_warn("md/raid:%s: PPL space overlaps with superblock on %s\n",
1165                         mdname(rdev->mddev), bdevname(rdev->bdev, b));
1166                 return -EINVAL;
1167         }
1168
1169         rdev->ppl.size = ppl_size_new;
1170
1171         return 0;
1172 }
1173
1174 static void ppl_init_child_log(struct ppl_log *log, struct md_rdev *rdev)
1175 {
1176         if ((rdev->ppl.size << 9) >= (PPL_SPACE_SIZE +
1177                                       PPL_HEADER_SIZE) * 2) {
1178                 log->use_multippl = true;
1179                 set_bit(MD_HAS_MULTIPLE_PPLS,
1180                         &log->ppl_conf->mddev->flags);
1181                 log->entry_space = PPL_SPACE_SIZE;
1182         } else {
1183                 log->use_multippl = false;
1184                 log->entry_space = (log->rdev->ppl.size << 9) -
1185                                    PPL_HEADER_SIZE;
1186         }
1187         log->next_io_sector = rdev->ppl.sector;
1188 }
1189
1190 int ppl_init_log(struct r5conf *conf)
1191 {
1192         struct ppl_conf *ppl_conf;
1193         struct mddev *mddev = conf->mddev;
1194         int ret = 0;
1195         int i;
1196         bool need_cache_flush = false;
1197
1198         pr_debug("md/raid:%s: enabling distributed Partial Parity Log\n",
1199                  mdname(conf->mddev));
1200
1201         if (PAGE_SIZE != 4096)
1202                 return -EINVAL;
1203
1204         if (mddev->level != 5) {
1205                 pr_warn("md/raid:%s PPL is not compatible with raid level %d\n",
1206                         mdname(mddev), mddev->level);
1207                 return -EINVAL;
1208         }
1209
1210         if (mddev->bitmap_info.file || mddev->bitmap_info.offset) {
1211                 pr_warn("md/raid:%s PPL is not compatible with bitmap\n",
1212                         mdname(mddev));
1213                 return -EINVAL;
1214         }
1215
1216         if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
1217                 pr_warn("md/raid:%s PPL is not compatible with journal\n",
1218                         mdname(mddev));
1219                 return -EINVAL;
1220         }
1221
1222         ppl_conf = kzalloc(sizeof(struct ppl_conf), GFP_KERNEL);
1223         if (!ppl_conf)
1224                 return -ENOMEM;
1225
1226         ppl_conf->mddev = mddev;
1227
1228         ppl_conf->io_kc = KMEM_CACHE(ppl_io_unit, 0);
1229         if (!ppl_conf->io_kc) {
1230                 ret = -ENOMEM;
1231                 goto err;
1232         }
1233
1234         ppl_conf->io_pool = mempool_create(conf->raid_disks, ppl_io_pool_alloc,
1235                                            ppl_io_pool_free, ppl_conf->io_kc);
1236         if (!ppl_conf->io_pool) {
1237                 ret = -ENOMEM;
1238                 goto err;
1239         }
1240
1241         ppl_conf->bs = bioset_create(conf->raid_disks, 0, BIOSET_NEED_BVECS);
1242         if (!ppl_conf->bs) {
1243                 ret = -ENOMEM;
1244                 goto err;
1245         }
1246
1247         ppl_conf->count = conf->raid_disks;
1248         ppl_conf->child_logs = kcalloc(ppl_conf->count, sizeof(struct ppl_log),
1249                                        GFP_KERNEL);
1250         if (!ppl_conf->child_logs) {
1251                 ret = -ENOMEM;
1252                 goto err;
1253         }
1254
1255         atomic64_set(&ppl_conf->seq, 0);
1256         INIT_LIST_HEAD(&ppl_conf->no_mem_stripes);
1257         spin_lock_init(&ppl_conf->no_mem_stripes_lock);
1258
1259         if (!mddev->external) {
1260                 ppl_conf->signature = ~crc32c_le(~0, mddev->uuid, sizeof(mddev->uuid));
1261                 ppl_conf->block_size = 512;
1262         } else {
1263                 ppl_conf->block_size = queue_logical_block_size(mddev->queue);
1264         }
1265
1266         for (i = 0; i < ppl_conf->count; i++) {
1267                 struct ppl_log *log = &ppl_conf->child_logs[i];
1268                 struct md_rdev *rdev = conf->disks[i].rdev;
1269
1270                 mutex_init(&log->io_mutex);
1271                 spin_lock_init(&log->io_list_lock);
1272                 INIT_LIST_HEAD(&log->io_list);
1273
1274                 log->ppl_conf = ppl_conf;
1275                 log->rdev = rdev;
1276
1277                 if (rdev) {
1278                         struct request_queue *q;
1279
1280                         ret = ppl_validate_rdev(rdev);
1281                         if (ret)
1282                                 goto err;
1283
1284                         q = bdev_get_queue(rdev->bdev);
1285                         if (test_bit(QUEUE_FLAG_WC, &q->queue_flags))
1286                                 need_cache_flush = true;
1287                         ppl_init_child_log(log, rdev);
1288                 }
1289         }
1290
1291         if (need_cache_flush)
1292                 pr_warn("md/raid:%s: Volatile write-back cache should be disabled on all member drives when using PPL!\n",
1293                         mdname(mddev));
1294
1295         /* load and possibly recover the logs from the member disks */
1296         ret = ppl_load(ppl_conf);
1297
1298         if (ret) {
1299                 goto err;
1300         } else if (!mddev->pers &&
1301                    mddev->recovery_cp == 0 && !mddev->degraded &&
1302                    ppl_conf->recovered_entries > 0 &&
1303                    ppl_conf->mismatch_count == 0) {
1304                 /*
1305                  * If we are starting a dirty array and the recovery succeeds
1306                  * without any issues, set the array as clean.
1307                  */
1308                 mddev->recovery_cp = MaxSector;
1309                 set_bit(MD_SB_CHANGE_CLEAN, &mddev->sb_flags);
1310         } else if (mddev->pers && ppl_conf->mismatch_count > 0) {
1311                 /* no mismatch allowed when enabling PPL for a running array */
1312                 ret = -EINVAL;
1313                 goto err;
1314         }
1315
1316         conf->log_private = ppl_conf;
1317         set_bit(MD_HAS_PPL, &ppl_conf->mddev->flags);
1318
1319         return 0;
1320 err:
1321         __ppl_exit_log(ppl_conf);
1322         return ret;
1323 }
1324
1325 int ppl_modify_log(struct r5conf *conf, struct md_rdev *rdev, bool add)
1326 {
1327         struct ppl_conf *ppl_conf = conf->log_private;
1328         struct ppl_log *log;
1329         int ret = 0;
1330         char b[BDEVNAME_SIZE];
1331
1332         if (!rdev)
1333                 return -EINVAL;
1334
1335         pr_debug("%s: disk: %d operation: %s dev: %s\n",
1336                  __func__, rdev->raid_disk, add ? "add" : "remove",
1337                  bdevname(rdev->bdev, b));
1338
1339         if (rdev->raid_disk < 0)
1340                 return 0;
1341
1342         if (rdev->raid_disk >= ppl_conf->count)
1343                 return -ENODEV;
1344
1345         log = &ppl_conf->child_logs[rdev->raid_disk];
1346
1347         mutex_lock(&log->io_mutex);
1348         if (add) {
1349                 ret = ppl_validate_rdev(rdev);
1350                 if (!ret) {
1351                         log->rdev = rdev;
1352                         ret = ppl_write_empty_header(log);
1353                         ppl_init_child_log(log, rdev);
1354                 }
1355         } else {
1356                 log->rdev = NULL;
1357         }
1358         mutex_unlock(&log->io_mutex);
1359
1360         return ret;
1361 }