drivers/md/raid5-cache.c

   1 /*
   2  * Copyright (C) 2015 Shaohua Li <[email protected]>
   3  * Copyright (C) 2016 Song Liu <[email protected]>
   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/wait.h>
  17 #include <linux/blkdev.h>
  18 #include <linux/slab.h>
  19 #include <linux/raid/md_p.h>
  20 #include <linux/crc32c.h>
  21 #include <linux/random.h>
  22 #include <linux/kthread.h>
  23 #include <linux/types.h>
  24 #include "md.h"
  25 #include "raid5.h"
  26 #include "bitmap.h"
  27
  28 /*
  29  * metadata/data stored in disk with 4k size unit (a block) regardless
  30  * underneath hardware sector size. only works with PAGE_SIZE == 4096
  31  */
  32 #define BLOCK_SECTORS (8)
  33
  34 /*
  35  * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
  36  *
  37  * In write through mode, the reclaim runs every log->max_free_space.
  38  * This can prevent the recovery scans for too long
  39  */
  40 #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
  41 #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
  42
  43 /* wake up reclaim thread periodically */
  44 #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
  45 /* start flush with these full stripes */
  46 #define R5C_FULL_STRIPE_FLUSH_BATCH 256
  47 /* reclaim stripes in groups */
  48 #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
  49
  50 /*
  51  * We only need 2 bios per I/O unit to make progress, but ensure we
  52  * have a few more available to not get too tight.
  53  */
  54 #define R5L_POOL_SIZE   4
  55
  56 /*
  57  * r5c journal modes of the array: write-back or write-through.
  58  * write-through mode has identical behavior as existing log only
  59  * implementation.
  60  */
  61 enum r5c_journal_mode {
  62         R5C_JOURNAL_MODE_WRITE_THROUGH = 0,
  63         R5C_JOURNAL_MODE_WRITE_BACK = 1,
  64 };
  65
  66 static char *r5c_journal_mode_str[] = {"write-through",
  67                                        "write-back"};
  68 /*
  69  * raid5 cache state machine
  70  *
  71  * With the RAID cache, each stripe works in two phases:
  72  *      - caching phase
  73  *      - writing-out phase
  74  *
  75  * These two phases are controlled by bit STRIPE_R5C_CACHING:
  76  *   if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
  77  *   if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
  78  *
  79  * When there is no journal, or the journal is in write-through mode,
  80  * the stripe is always in writing-out phase.
  81  *
  82  * For write-back journal, the stripe is sent to caching phase on write
  83  * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
  84  * the write-out phase by clearing STRIPE_R5C_CACHING.
  85  *
  86  * Stripes in caching phase do not write the raid disks. Instead, all
  87  * writes are committed from the log device. Therefore, a stripe in
  88  * caching phase handles writes as:
  89  *      - write to log device
  90  *      - return IO
  91  *
  92  * Stripes in writing-out phase handle writes as:
  93  *      - calculate parity
  94  *      - write pending data and parity to journal
  95  *      - write data and parity to raid disks
  96  *      - return IO for pending writes
  97  */
  98
  99 struct r5l_log {
 100         struct md_rdev *rdev;
 101
 102         u32 uuid_checksum;
 103
 104         sector_t device_size;           /* log device size, round to
 105                                          * BLOCK_SECTORS */
 106         sector_t max_free_space;        /* reclaim run if free space is at
 107                                          * this size */
 108
 109         sector_t last_checkpoint;       /* log tail. where recovery scan
 110                                          * starts from */
 111         u64 last_cp_seq;                /* log tail sequence */
 112
 113         sector_t log_start;             /* log head. where new data appends */
 114         u64 seq;                        /* log head sequence */
 115
 116         sector_t next_checkpoint;
 117
 118         struct mutex io_mutex;
 119         struct r5l_io_unit *current_io; /* current io_unit accepting new data */
 120
 121         spinlock_t io_list_lock;
 122         struct list_head running_ios;   /* io_units which are still running,
 123                                          * and have not yet been completely
 124                                          * written to the log */
 125         struct list_head io_end_ios;    /* io_units which have been completely
 126                                          * written to the log but not yet written
 127                                          * to the RAID */
 128         struct list_head flushing_ios;  /* io_units which are waiting for log
 129                                          * cache flush */
 130         struct list_head finished_ios;  /* io_units which settle down in log disk */
 131         struct bio flush_bio;
 132
 133         struct list_head no_mem_stripes;   /* pending stripes, -ENOMEM */
 134
 135         struct kmem_cache *io_kc;
 136         mempool_t *io_pool;
 137         struct bio_set *bs;
 138         mempool_t *meta_pool;
 139
 140         struct md_thread *reclaim_thread;
 141         unsigned long reclaim_target;   /* number of space that need to be
 142                                          * reclaimed.  if it's 0, reclaim spaces
 143                                          * used by io_units which are in
 144                                          * IO_UNIT_STRIPE_END state (eg, reclaim
 145                                          * dones't wait for specific io_unit
 146                                          * switching to IO_UNIT_STRIPE_END
 147                                          * state) */
 148         wait_queue_head_t iounit_wait;
 149
 150         struct list_head no_space_stripes; /* pending stripes, log has no space */
 151         spinlock_t no_space_stripes_lock;
 152
 153         bool need_cache_flush;
 154
 155         /* for r5c_cache */
 156         enum r5c_journal_mode r5c_journal_mode;
 157
 158         /* all stripes in r5cache, in the order of seq at sh->log_start */
 159         struct list_head stripe_in_journal_list;
 160
 161         spinlock_t stripe_in_journal_lock;
 162         atomic_t stripe_in_journal_count;
 163
 164         /* to submit async io_units, to fulfill ordering of flush */
 165         struct work_struct deferred_io_work;
 166         /* to disable write back during in degraded mode */
 167         struct work_struct disable_writeback_work;
 168
 169         /* to for chunk_aligned_read in writeback mode, details below */
 170         spinlock_t tree_lock;
 171         struct radix_tree_root big_stripe_tree;
 172 };
 173
 174 /*
 175  * Enable chunk_aligned_read() with write back cache.
 176  *
 177  * Each chunk may contain more than one stripe (for example, a 256kB
 178  * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
 179  * chunk_aligned_read, these stripes are grouped into one "big_stripe".
 180  * For each big_stripe, we count how many stripes of this big_stripe
 181  * are in the write back cache. These data are tracked in a radix tree
 182  * (big_stripe_tree). We use radix_tree item pointer as the counter.
 183  * r5c_tree_index() is used to calculate keys for the radix tree.
 184  *
 185  * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
 186  * big_stripe of each chunk in the tree. If this big_stripe is in the
 187  * tree, chunk_aligned_read() aborts. This look up is protected by
 188  * rcu_read_lock().
 189  *
 190  * It is necessary to remember whether a stripe is counted in
 191  * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
 192  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
 193  * two flags are set, the stripe is counted in big_stripe_tree. This
 194  * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
 195  * r5c_try_caching_write(); and moving clear_bit of
 196  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
 197  * r5c_finish_stripe_write_out().
 198  */
 199
 200 /*
 201  * radix tree requests lowest 2 bits of data pointer to be 2b'00.
 202  * So it is necessary to left shift the counter by 2 bits before using it
 203  * as data pointer of the tree.
 204  */
 205 #define R5C_RADIX_COUNT_SHIFT 2
 206
 207 /*
 208  * calculate key for big_stripe_tree
 209  *
 210  * sect: align_bi->bi_iter.bi_sector or sh->sector
 211  */
 212 static inline sector_t r5c_tree_index(struct r5conf *conf,
 213                                       sector_t sect)
 214 {
 215         sector_t offset;
 216
 217         offset = sector_div(sect, conf->chunk_sectors);
 218         return sect;
 219 }
 220
 221 /*
 222  * an IO range starts from a meta data block and end at the next meta data
 223  * block. The io unit's the meta data block tracks data/parity followed it. io
 224  * unit is written to log disk with normal write, as we always flush log disk
 225  * first and then start move data to raid disks, there is no requirement to
 226  * write io unit with FLUSH/FUA
 227  */
 228 struct r5l_io_unit {
 229         struct r5l_log *log;
 230
 231         struct page *meta_page; /* store meta block */
 232         int meta_offset;        /* current offset in meta_page */
 233
 234         struct bio *current_bio;/* current_bio accepting new data */
 235
 236         atomic_t pending_stripe;/* how many stripes not flushed to raid */
 237         u64 seq;                /* seq number of the metablock */
 238         sector_t log_start;     /* where the io_unit starts */
 239         sector_t log_end;       /* where the io_unit ends */
 240         struct list_head log_sibling; /* log->running_ios */
 241         struct list_head stripe_list; /* stripes added to the io_unit */
 242
 243         int state;
 244         bool need_split_bio;
 245         struct bio *split_bio;
 246
 247         unsigned int has_flush:1;      /* include flush request */
 248         unsigned int has_fua:1;        /* include fua request */
 249         unsigned int has_null_flush:1; /* include empty flush request */
 250         /*
 251          * io isn't sent yet, flush/fua request can only be submitted till it's
 252          * the first IO in running_ios list
 253          */
 254         unsigned int io_deferred:1;
 255
 256         struct bio_list flush_barriers;   /* size == 0 flush bios */
 257 };
 258
 259 /* r5l_io_unit state */
 260 enum r5l_io_unit_state {
 261         IO_UNIT_RUNNING = 0,    /* accepting new IO */
 262         IO_UNIT_IO_START = 1,   /* io_unit bio start writing to log,
 263                                  * don't accepting new bio */
 264         IO_UNIT_IO_END = 2,     /* io_unit bio finish writing to log */
 265         IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
 266 };
 267
 268 bool r5c_is_writeback(struct r5l_log *log)
 269 {
 270         return (log != NULL &&
 271                 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
 272 }
 273
 274 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
 275 {
 276         start += inc;
 277         if (start >= log->device_size)
 278                 start = start - log->device_size;
 279         return start;
 280 }
 281
 282 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
 283                                   sector_t end)
 284 {
 285         if (end >= start)
 286                 return end - start;
 287         else
 288                 return end + log->device_size - start;
 289 }
 290
 291 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
 292 {
 293         sector_t used_size;
 294
 295         used_size = r5l_ring_distance(log, log->last_checkpoint,
 296                                         log->log_start);
 297
 298         return log->device_size > used_size + size;
 299 }
 300
 301 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
 302                                     enum r5l_io_unit_state state)
 303 {
 304         if (WARN_ON(io->state >= state))
 305                 return;
 306         io->state = state;
 307 }
 308
 309 static void
 310 r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev,
 311                               struct bio_list *return_bi)
 312 {
 313         struct bio *wbi, *wbi2;
 314
 315         wbi = dev->written;
 316         dev->written = NULL;
 317         while (wbi && wbi->bi_iter.bi_sector <
 318                dev->sector + STRIPE_SECTORS) {
 319                 wbi2 = r5_next_bio(wbi, dev->sector);
 320                 if (!raid5_dec_bi_active_stripes(wbi)) {
 321                         md_write_end(conf->mddev);
 322                         bio_list_add(return_bi, wbi);
 323                 }
 324                 wbi = wbi2;
 325         }
 326 }
 327
 328 void r5c_handle_cached_data_endio(struct r5conf *conf,
 329           struct stripe_head *sh, int disks, struct bio_list *return_bi)
 330 {
 331         int i;
 332
 333         for (i = sh->disks; i--; ) {
 334                 if (sh->dev[i].written) {
 335                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
 336                         r5c_return_dev_pending_writes(conf, &sh->dev[i],
 337                                                       return_bi);
 338                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
 339                                         STRIPE_SECTORS,
 340                                         !test_bit(STRIPE_DEGRADED, &sh->state),
 341                                         0);
 342                 }
 343         }
 344 }
 345
 346 /* Check whether we should flush some stripes to free up stripe cache */
 347 void r5c_check_stripe_cache_usage(struct r5conf *conf)
 348 {
 349         int total_cached;
 350
 351         if (!r5c_is_writeback(conf->log))
 352                 return;
 353
 354         total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
 355                 atomic_read(&conf->r5c_cached_full_stripes);
 356
 357         /*
 358          * The following condition is true for either of the following:
 359          *   - stripe cache pressure high:
 360          *          total_cached > 3/4 min_nr_stripes ||
 361          *          empty_inactive_list_nr > 0
 362          *   - stripe cache pressure moderate:
 363          *          total_cached > 1/2 min_nr_stripes
 364          */
 365         if (total_cached > conf->min_nr_stripes * 1 / 2 ||
 366             atomic_read(&conf->empty_inactive_list_nr) > 0)
 367                 r5l_wake_reclaim(conf->log, 0);
 368 }
 369
 370 /*
 371  * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
 372  * stripes in the cache
 373  */
 374 void r5c_check_cached_full_stripe(struct r5conf *conf)
 375 {
 376         if (!r5c_is_writeback(conf->log))
 377                 return;
 378
 379         /*
 380          * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
 381          * or a full stripe (chunk size / 4k stripes).
 382          */
 383         if (atomic_read(&conf->r5c_cached_full_stripes) >=
 384             min(R5C_FULL_STRIPE_FLUSH_BATCH,
 385                 conf->chunk_sectors >> STRIPE_SHIFT))
 386                 r5l_wake_reclaim(conf->log, 0);
 387 }
 388
 389 /*
 390  * Total log space (in sectors) needed to flush all data in cache
 391  *
 392  * To avoid deadlock due to log space, it is necessary to reserve log
 393  * space to flush critical stripes (stripes that occupying log space near
 394  * last_checkpoint). This function helps check how much log space is
 395  * required to flush all cached stripes.
 396  *
 397  * To reduce log space requirements, two mechanisms are used to give cache
 398  * flush higher priorities:
 399  *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
 400  *       stripes ALREADY in journal can be flushed w/o pending writes;
 401  *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
 402  *       can be delayed (r5l_add_no_space_stripe).
 403  *
 404  * In cache flush, the stripe goes through 1 and then 2. For a stripe that
 405  * already passed 1, flushing it requires at most (conf->max_degraded + 1)
 406  * pages of journal space. For stripes that has not passed 1, flushing it
 407  * requires (conf->raid_disks + 1) pages of journal space. There are at
 408  * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
 409  * required to flush all cached stripes (in pages) is:
 410  *
 411  *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
 412  *     (group_cnt + 1) * (raid_disks + 1)
 413  * or
 414  *     (stripe_in_journal_count) * (max_degraded + 1) +
 415  *     (group_cnt + 1) * (raid_disks - max_degraded)
 416  */
 417 static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
 418 {
 419         struct r5l_log *log = conf->log;
 420
 421         if (!r5c_is_writeback(log))
 422                 return 0;
 423
 424         return BLOCK_SECTORS *
 425                 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
 426                  (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
 427 }
 428
 429 /*
 430  * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
 431  *
 432  * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
 433  * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
 434  * device is less than 2x of reclaim_required_space.
 435  */
 436 static inline void r5c_update_log_state(struct r5l_log *log)
 437 {
 438         struct r5conf *conf = log->rdev->mddev->private;
 439         sector_t free_space;
 440         sector_t reclaim_space;
 441         bool wake_reclaim = false;
 442
 443         if (!r5c_is_writeback(log))
 444                 return;
 445
 446         free_space = r5l_ring_distance(log, log->log_start,
 447                                        log->last_checkpoint);
 448         reclaim_space = r5c_log_required_to_flush_cache(conf);
 449         if (free_space < 2 * reclaim_space)
 450                 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 451         else {
 452                 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
 453                         wake_reclaim = true;
 454                 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 455         }
 456         if (free_space < 3 * reclaim_space)
 457                 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
 458         else
 459                 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
 460
 461         if (wake_reclaim)
 462                 r5l_wake_reclaim(log, 0);
 463 }
 464
 465 /*
 466  * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
 467  * This function should only be called in write-back mode.
 468  */
 469 void r5c_make_stripe_write_out(struct stripe_head *sh)
 470 {
 471         struct r5conf *conf = sh->raid_conf;
 472         struct r5l_log *log = conf->log;
 473
 474         BUG_ON(!r5c_is_writeback(log));
 475
 476         WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
 477         clear_bit(STRIPE_R5C_CACHING, &sh->state);
 478
 479         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 480                 atomic_inc(&conf->preread_active_stripes);
 481 }
 482
 483 static void r5c_handle_data_cached(struct stripe_head *sh)
 484 {
 485         int i;
 486
 487         for (i = sh->disks; i--; )
 488                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 489                         set_bit(R5_InJournal, &sh->dev[i].flags);
 490                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
 491                 }
 492         clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 493 }
 494
 495 /*
 496  * this journal write must contain full parity,
 497  * it may also contain some data pages
 498  */
 499 static void r5c_handle_parity_cached(struct stripe_head *sh)
 500 {
 501         int i;
 502
 503         for (i = sh->disks; i--; )
 504                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
 505                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
 506 }
 507
 508 /*
 509  * Setting proper flags after writing (or flushing) data and/or parity to the
 510  * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
 511  */
 512 static void r5c_finish_cache_stripe(struct stripe_head *sh)
 513 {
 514         struct r5l_log *log = sh->raid_conf->log;
 515
 516         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
 517                 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
 518                 /*
 519                  * Set R5_InJournal for parity dev[pd_idx]. This means
 520                  * all data AND parity in the journal. For RAID 6, it is
 521                  * NOT necessary to set the flag for dev[qd_idx], as the
 522                  * two parities are written out together.
 523                  */
 524                 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 525         } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
 526                 r5c_handle_data_cached(sh);
 527         } else {
 528                 r5c_handle_parity_cached(sh);
 529                 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 530         }
 531 }
 532
 533 static void r5l_io_run_stripes(struct r5l_io_unit *io)
 534 {
 535         struct stripe_head *sh, *next;
 536
 537         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 538                 list_del_init(&sh->log_list);
 539
 540                 r5c_finish_cache_stripe(sh);
 541
 542                 set_bit(STRIPE_HANDLE, &sh->state);
 543                 raid5_release_stripe(sh);
 544         }
 545 }
 546
 547 static void r5l_log_run_stripes(struct r5l_log *log)
 548 {
 549         struct r5l_io_unit *io, *next;
 550
 551         assert_spin_locked(&log->io_list_lock);
 552
 553         list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 554                 /* don't change list order */
 555                 if (io->state < IO_UNIT_IO_END)
 556                         break;
 557
 558                 list_move_tail(&io->log_sibling, &log->finished_ios);
 559                 r5l_io_run_stripes(io);
 560         }
 561 }
 562
 563 static void r5l_move_to_end_ios(struct r5l_log *log)
 564 {
 565         struct r5l_io_unit *io, *next;
 566
 567         assert_spin_locked(&log->io_list_lock);
 568
 569         list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 570                 /* don't change list order */
 571                 if (io->state < IO_UNIT_IO_END)
 572                         break;
 573                 list_move_tail(&io->log_sibling, &log->io_end_ios);
 574         }
 575 }
 576
 577 static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
 578 static void r5l_log_endio(struct bio *bio)
 579 {
 580         struct r5l_io_unit *io = bio->bi_private;
 581         struct r5l_io_unit *io_deferred;
 582         struct r5l_log *log = io->log;
 583         unsigned long flags;
 584
 585         if (bio->bi_error)
 586                 md_error(log->rdev->mddev, log->rdev);
 587
 588         bio_put(bio);
 589         mempool_free(io->meta_page, log->meta_pool);
 590
 591         spin_lock_irqsave(&log->io_list_lock, flags);
 592         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 593         if (log->need_cache_flush)
 594                 r5l_move_to_end_ios(log);
 595         else
 596                 r5l_log_run_stripes(log);
 597         if (!list_empty(&log->running_ios)) {
 598                 /*
 599                  * FLUSH/FUA io_unit is deferred because of ordering, now we
 600                  * can dispatch it
 601                  */
 602                 io_deferred = list_first_entry(&log->running_ios,
 603                                                struct r5l_io_unit, log_sibling);
 604                 if (io_deferred->io_deferred)
 605                         schedule_work(&log->deferred_io_work);
 606         }
 607
 608         spin_unlock_irqrestore(&log->io_list_lock, flags);
 609
 610         if (log->need_cache_flush)
 611                 md_wakeup_thread(log->rdev->mddev->thread);
 612
 613         if (io->has_null_flush) {
 614                 struct bio *bi;
 615
 616                 WARN_ON(bio_list_empty(&io->flush_barriers));
 617                 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
 618                         bio_endio(bi);
 619                         atomic_dec(&io->pending_stripe);
 620                 }
 621                 if (atomic_read(&io->pending_stripe) == 0)
 622                         __r5l_stripe_write_finished(io);
 623         }
 624 }
 625
 626 static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
 627 {
 628         unsigned long flags;
 629
 630         spin_lock_irqsave(&log->io_list_lock, flags);
 631         __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 632         spin_unlock_irqrestore(&log->io_list_lock, flags);
 633
 634         if (io->has_flush)
 635                 io->current_bio->bi_opf |= REQ_PREFLUSH;
 636         if (io->has_fua)
 637                 io->current_bio->bi_opf |= REQ_FUA;
 638         submit_bio(io->current_bio);
 639
 640         if (!io->split_bio)
 641                 return;
 642
 643         if (io->has_flush)
 644                 io->split_bio->bi_opf |= REQ_PREFLUSH;
 645         if (io->has_fua)
 646                 io->split_bio->bi_opf |= REQ_FUA;
 647         submit_bio(io->split_bio);
 648 }
 649
 650 /* deferred io_unit will be dispatched here */
 651 static void r5l_submit_io_async(struct work_struct *work)
 652 {
 653         struct r5l_log *log = container_of(work, struct r5l_log,
 654                                            deferred_io_work);
 655         struct r5l_io_unit *io = NULL;
 656         unsigned long flags;
 657
 658         spin_lock_irqsave(&log->io_list_lock, flags);
 659         if (!list_empty(&log->running_ios)) {
 660                 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
 661                                       log_sibling);
 662                 if (!io->io_deferred)
 663                         io = NULL;
 664                 else
 665                         io->io_deferred = 0;
 666         }
 667         spin_unlock_irqrestore(&log->io_list_lock, flags);
 668         if (io)
 669                 r5l_do_submit_io(log, io);
 670 }
 671
 672 static void r5c_disable_writeback_async(struct work_struct *work)
 673 {
 674         struct r5l_log *log = container_of(work, struct r5l_log,
 675                                            disable_writeback_work);
 676         struct mddev *mddev = log->rdev->mddev;
 677
 678         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 679                 return;
 680         pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
 681                 mdname(mddev));
 682         mddev_suspend(mddev);
 683         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
 684         mddev_resume(mddev);
 685 }
 686
 687 static void r5l_submit_current_io(struct r5l_log *log)
 688 {
 689         struct r5l_io_unit *io = log->current_io;
 690         struct bio *bio;
 691         struct r5l_meta_block *block;
 692         unsigned long flags;
 693         u32 crc;
 694         bool do_submit = true;
 695
 696         if (!io)
 697                 return;
 698
 699         block = page_address(io->meta_page);
 700         block->meta_size = cpu_to_le32(io->meta_offset);
 701         crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 702         block->checksum = cpu_to_le32(crc);
 703         bio = io->current_bio;
 704
 705         log->current_io = NULL;
 706         spin_lock_irqsave(&log->io_list_lock, flags);
 707         if (io->has_flush || io->has_fua) {
 708                 if (io != list_first_entry(&log->running_ios,
 709                                            struct r5l_io_unit, log_sibling)) {
 710                         io->io_deferred = 1;
 711                         do_submit = false;
 712                 }
 713         }
 714         spin_unlock_irqrestore(&log->io_list_lock, flags);
 715         if (do_submit)
 716                 r5l_do_submit_io(log, io);
 717 }
 718
 719 static struct bio *r5l_bio_alloc(struct r5l_log *log)
 720 {
 721         struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
 722
 723         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 724         bio->bi_bdev = log->rdev->bdev;
 725         bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
 726
 727         return bio;
 728 }
 729
 730 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
 731 {
 732         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 733
 734         r5c_update_log_state(log);
 735         /*
 736          * If we filled up the log device start from the beginning again,
 737          * which will require a new bio.
 738          *
 739          * Note: for this to work properly the log size needs to me a multiple
 740          * of BLOCK_SECTORS.
 741          */
 742         if (log->log_start == 0)
 743                 io->need_split_bio = true;
 744
 745         io->log_end = log->log_start;
 746 }
 747
 748 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 749 {
 750         struct r5l_io_unit *io;
 751         struct r5l_meta_block *block;
 752
 753         io = mempool_alloc(log->io_pool, GFP_ATOMIC);
 754         if (!io)
 755                 return NULL;
 756         memset(io, 0, sizeof(*io));
 757
 758         io->log = log;
 759         INIT_LIST_HEAD(&io->log_sibling);
 760         INIT_LIST_HEAD(&io->stripe_list);
 761         bio_list_init(&io->flush_barriers);
 762         io->state = IO_UNIT_RUNNING;
 763
 764         io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
 765         block = page_address(io->meta_page);
 766         clear_page(block);
 767         block->magic = cpu_to_le32(R5LOG_MAGIC);
 768         block->version = R5LOG_VERSION;
 769         block->seq = cpu_to_le64(log->seq);
 770         block->position = cpu_to_le64(log->log_start);
 771
 772         io->log_start = log->log_start;
 773         io->meta_offset = sizeof(struct r5l_meta_block);
 774         io->seq = log->seq++;
 775
 776         io->current_bio = r5l_bio_alloc(log);
 777         io->current_bio->bi_end_io = r5l_log_endio;
 778         io->current_bio->bi_private = io;
 779         bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
 780
 781         r5_reserve_log_entry(log, io);
 782
 783         spin_lock_irq(&log->io_list_lock);
 784         list_add_tail(&io->log_sibling, &log->running_ios);
 785         spin_unlock_irq(&log->io_list_lock);
 786
 787         return io;
 788 }
 789
 790 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 791 {
 792         if (log->current_io &&
 793             log->current_io->meta_offset + payload_size > PAGE_SIZE)
 794                 r5l_submit_current_io(log);
 795
 796         if (!log->current_io) {
 797                 log->current_io = r5l_new_meta(log);
 798                 if (!log->current_io)
 799                         return -ENOMEM;
 800         }
 801
 802         return 0;
 803 }
 804
 805 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 806                                     sector_t location,
 807                                     u32 checksum1, u32 checksum2,
 808                                     bool checksum2_valid)
 809 {
 810         struct r5l_io_unit *io = log->current_io;
 811         struct r5l_payload_data_parity *payload;
 812
 813         payload = page_address(io->meta_page) + io->meta_offset;
 814         payload->header.type = cpu_to_le16(type);
 815         payload->header.flags = cpu_to_le16(0);
 816         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 817                                     (PAGE_SHIFT - 9));
 818         payload->location = cpu_to_le64(location);
 819         payload->checksum[0] = cpu_to_le32(checksum1);
 820         if (checksum2_valid)
 821                 payload->checksum[1] = cpu_to_le32(checksum2);
 822
 823         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 824                 sizeof(__le32) * (1 + !!checksum2_valid);
 825 }
 826
 827 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 828 {
 829         struct r5l_io_unit *io = log->current_io;
 830
 831         if (io->need_split_bio) {
 832                 BUG_ON(io->split_bio);
 833                 io->split_bio = io->current_bio;
 834                 io->current_bio = r5l_bio_alloc(log);
 835                 bio_chain(io->current_bio, io->split_bio);
 836                 io->need_split_bio = false;
 837         }
 838
 839         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
 840                 BUG();
 841
 842         r5_reserve_log_entry(log, io);
 843 }
 844
 845 static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 846                            int data_pages, int parity_pages)
 847 {
 848         int i;
 849         int meta_size;
 850         int ret;
 851         struct r5l_io_unit *io;
 852
 853         meta_size =
 854                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 855                  * data_pages) +
 856                 sizeof(struct r5l_payload_data_parity) +
 857                 sizeof(__le32) * parity_pages;
 858
 859         ret = r5l_get_meta(log, meta_size);
 860         if (ret)
 861                 return ret;
 862
 863         io = log->current_io;
 864
 865         if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
 866                 io->has_flush = 1;
 867
 868         for (i = 0; i < sh->disks; i++) {
 869                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
 870                     test_bit(R5_InJournal, &sh->dev[i].flags))
 871                         continue;
 872                 if (i == sh->pd_idx || i == sh->qd_idx)
 873                         continue;
 874                 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
 875                     log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
 876                         io->has_fua = 1;
 877                         /*
 878                          * we need to flush journal to make sure recovery can
 879                          * reach the data with fua flag
 880                          */
 881                         io->has_flush = 1;
 882                 }
 883                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 884                                         raid5_compute_blocknr(sh, i, 0),
 885                                         sh->dev[i].log_checksum, 0, false);
 886                 r5l_append_payload_page(log, sh->dev[i].page);
 887         }
 888
 889         if (parity_pages == 2) {
 890                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 891                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 892                                         sh->dev[sh->qd_idx].log_checksum, true);
 893                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 894                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 895         } else if (parity_pages == 1) {
 896                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 897                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 898                                         0, false);
 899                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 900         } else  /* Just writing data, not parity, in caching phase */
 901                 BUG_ON(parity_pages != 0);
 902
 903         list_add_tail(&sh->log_list, &io->stripe_list);
 904         atomic_inc(&io->pending_stripe);
 905         sh->log_io = io;
 906
 907         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 908                 return 0;
 909
 910         if (sh->log_start == MaxSector) {
 911                 BUG_ON(!list_empty(&sh->r5c));
 912                 sh->log_start = io->log_start;
 913                 spin_lock_irq(&log->stripe_in_journal_lock);
 914                 list_add_tail(&sh->r5c,
 915                               &log->stripe_in_journal_list);
 916                 spin_unlock_irq(&log->stripe_in_journal_lock);
 917                 atomic_inc(&log->stripe_in_journal_count);
 918         }
 919         return 0;
 920 }
 921
 922 /* add stripe to no_space_stripes, and then wake up reclaim */
 923 static inline void r5l_add_no_space_stripe(struct r5l_log *log,
 924                                            struct stripe_head *sh)
 925 {
 926         spin_lock(&log->no_space_stripes_lock);
 927         list_add_tail(&sh->log_list, &log->no_space_stripes);
 928         spin_unlock(&log->no_space_stripes_lock);
 929 }
 930
 931 /*
 932  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 933  * data from log to raid disks), so we shouldn't wait for reclaim here
 934  */
 935 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 936 {
 937         struct r5conf *conf = sh->raid_conf;
 938         int write_disks = 0;
 939         int data_pages, parity_pages;
 940         int reserve;
 941         int i;
 942         int ret = 0;
 943         bool wake_reclaim = false;
 944
 945         if (!log)
 946                 return -EAGAIN;
 947         /* Don't support stripe batch */
 948         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 949             test_bit(STRIPE_SYNCING, &sh->state)) {
 950                 /* the stripe is written to log, we start writing it to raid */
 951                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 952                 return -EAGAIN;
 953         }
 954
 955         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
 956
 957         for (i = 0; i < sh->disks; i++) {
 958                 void *addr;
 959
 960                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
 961                     test_bit(R5_InJournal, &sh->dev[i].flags))
 962                         continue;
 963
 964                 write_disks++;
 965                 /* checksum is already calculated in last run */
 966                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
 967                         continue;
 968                 addr = kmap_atomic(sh->dev[i].page);
 969                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
 970                                                     addr, PAGE_SIZE);
 971                 kunmap_atomic(addr);
 972         }
 973         parity_pages = 1 + !!(sh->qd_idx >= 0);
 974         data_pages = write_disks - parity_pages;
 975
 976         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 977         /*
 978          * The stripe must enter state machine again to finish the write, so
 979          * don't delay.
 980          */
 981         clear_bit(STRIPE_DELAYED, &sh->state);
 982         atomic_inc(&sh->count);
 983
 984         mutex_lock(&log->io_mutex);
 985         /* meta + data */
 986         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
 987
 988         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
 989                 if (!r5l_has_free_space(log, reserve)) {
 990                         r5l_add_no_space_stripe(log, sh);
 991                         wake_reclaim = true;
 992                 } else {
 993                         ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
 994                         if (ret) {
 995                                 spin_lock_irq(&log->io_list_lock);
 996                                 list_add_tail(&sh->log_list,
 997                                               &log->no_mem_stripes);
 998                                 spin_unlock_irq(&log->io_list_lock);
 999                         }
1000                 }
1001         } else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1002                 /*
1003                  * log space critical, do not process stripes that are
1004                  * not in cache yet (sh->log_start == MaxSector).
1005                  */
1006                 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1007                     sh->log_start == MaxSector) {
1008                         r5l_add_no_space_stripe(log, sh);
1009                         wake_reclaim = true;
1010                         reserve = 0;
1011                 } else if (!r5l_has_free_space(log, reserve)) {
1012                         if (sh->log_start == log->last_checkpoint)
1013                                 BUG();
1014                         else
1015                                 r5l_add_no_space_stripe(log, sh);
1016                 } else {
1017                         ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1018                         if (ret) {
1019                                 spin_lock_irq(&log->io_list_lock);
1020                                 list_add_tail(&sh->log_list,
1021                                               &log->no_mem_stripes);
1022                                 spin_unlock_irq(&log->io_list_lock);
1023                         }
1024                 }
1025         }
1026
1027         mutex_unlock(&log->io_mutex);
1028         if (wake_reclaim)
1029                 r5l_wake_reclaim(log, reserve);
1030         return 0;
1031 }
1032
1033 void r5l_write_stripe_run(struct r5l_log *log)
1034 {
1035         if (!log)
1036                 return;
1037         mutex_lock(&log->io_mutex);
1038         r5l_submit_current_io(log);
1039         mutex_unlock(&log->io_mutex);
1040 }
1041
1042 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1043 {
1044         if (!log)
1045                 return -ENODEV;
1046
1047         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1048                 /*
1049                  * in write through (journal only)
1050                  * we flush log disk cache first, then write stripe data to
1051                  * raid disks. So if bio is finished, the log disk cache is
1052                  * flushed already. The recovery guarantees we can recovery
1053                  * the bio from log disk, so we don't need to flush again
1054                  */
1055                 if (bio->bi_iter.bi_size == 0) {
1056                         bio_endio(bio);
1057                         return 0;
1058                 }
1059                 bio->bi_opf &= ~REQ_PREFLUSH;
1060         } else {
1061                 /* write back (with cache) */
1062                 if (bio->bi_iter.bi_size == 0) {
1063                         mutex_lock(&log->io_mutex);
1064                         r5l_get_meta(log, 0);
1065                         bio_list_add(&log->current_io->flush_barriers, bio);
1066                         log->current_io->has_flush = 1;
1067                         log->current_io->has_null_flush = 1;
1068                         atomic_inc(&log->current_io->pending_stripe);
1069                         r5l_submit_current_io(log);
1070                         mutex_unlock(&log->io_mutex);
1071                         return 0;
1072                 }
1073         }
1074         return -EAGAIN;
1075 }
1076
1077 /* This will run after log space is reclaimed */
1078 static void r5l_run_no_space_stripes(struct r5l_log *log)
1079 {
1080         struct stripe_head *sh;
1081
1082         spin_lock(&log->no_space_stripes_lock);
1083         while (!list_empty(&log->no_space_stripes)) {
1084                 sh = list_first_entry(&log->no_space_stripes,
1085                                       struct stripe_head, log_list);
1086                 list_del_init(&sh->log_list);
1087                 set_bit(STRIPE_HANDLE, &sh->state);
1088                 raid5_release_stripe(sh);
1089         }
1090         spin_unlock(&log->no_space_stripes_lock);
1091 }
1092
1093 /*
1094  * calculate new last_checkpoint
1095  * for write through mode, returns log->next_checkpoint
1096  * for write back, returns log_start of first sh in stripe_in_journal_list
1097  */
1098 static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1099 {
1100         struct stripe_head *sh;
1101         struct r5l_log *log = conf->log;
1102         sector_t new_cp;
1103         unsigned long flags;
1104
1105         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1106                 return log->next_checkpoint;
1107
1108         spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1109         if (list_empty(&conf->log->stripe_in_journal_list)) {
1110                 /* all stripes flushed */
1111                 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1112                 return log->next_checkpoint;
1113         }
1114         sh = list_first_entry(&conf->log->stripe_in_journal_list,
1115                               struct stripe_head, r5c);
1116         new_cp = sh->log_start;
1117         spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1118         return new_cp;
1119 }
1120
1121 static sector_t r5l_reclaimable_space(struct r5l_log *log)
1122 {
1123         struct r5conf *conf = log->rdev->mddev->private;
1124
1125         return r5l_ring_distance(log, log->last_checkpoint,
1126                                  r5c_calculate_new_cp(conf));
1127 }
1128
1129 static void r5l_run_no_mem_stripe(struct r5l_log *log)
1130 {
1131         struct stripe_head *sh;
1132
1133         assert_spin_locked(&log->io_list_lock);
1134
1135         if (!list_empty(&log->no_mem_stripes)) {
1136                 sh = list_first_entry(&log->no_mem_stripes,
1137                                       struct stripe_head, log_list);
1138                 list_del_init(&sh->log_list);
1139                 set_bit(STRIPE_HANDLE, &sh->state);
1140                 raid5_release_stripe(sh);
1141         }
1142 }
1143
1144 static bool r5l_complete_finished_ios(struct r5l_log *log)
1145 {
1146         struct r5l_io_unit *io, *next;
1147         bool found = false;
1148
1149         assert_spin_locked(&log->io_list_lock);
1150
1151         list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1152                 /* don't change list order */
1153                 if (io->state < IO_UNIT_STRIPE_END)
1154                         break;
1155
1156                 log->next_checkpoint = io->log_start;
1157
1158                 list_del(&io->log_sibling);
1159                 mempool_free(io, log->io_pool);
1160                 r5l_run_no_mem_stripe(log);
1161
1162                 found = true;
1163         }
1164
1165         return found;
1166 }
1167
1168 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1169 {
1170         struct r5l_log *log = io->log;
1171         struct r5conf *conf = log->rdev->mddev->private;
1172         unsigned long flags;
1173
1174         spin_lock_irqsave(&log->io_list_lock, flags);
1175         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1176
1177         if (!r5l_complete_finished_ios(log)) {
1178                 spin_unlock_irqrestore(&log->io_list_lock, flags);
1179                 return;
1180         }
1181
1182         if (r5l_reclaimable_space(log) > log->max_free_space ||
1183             test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1184                 r5l_wake_reclaim(log, 0);
1185
1186         spin_unlock_irqrestore(&log->io_list_lock, flags);
1187         wake_up(&log->iounit_wait);
1188 }
1189
1190 void r5l_stripe_write_finished(struct stripe_head *sh)
1191 {
1192         struct r5l_io_unit *io;
1193
1194         io = sh->log_io;
1195         sh->log_io = NULL;
1196
1197         if (io && atomic_dec_and_test(&io->pending_stripe))
1198                 __r5l_stripe_write_finished(io);
1199 }
1200
1201 static void r5l_log_flush_endio(struct bio *bio)
1202 {
1203         struct r5l_log *log = container_of(bio, struct r5l_log,
1204                 flush_bio);
1205         unsigned long flags;
1206         struct r5l_io_unit *io;
1207
1208         if (bio->bi_error)
1209                 md_error(log->rdev->mddev, log->rdev);
1210
1211         spin_lock_irqsave(&log->io_list_lock, flags);
1212         list_for_each_entry(io, &log->flushing_ios, log_sibling)
1213                 r5l_io_run_stripes(io);
1214         list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1215         spin_unlock_irqrestore(&log->io_list_lock, flags);
1216 }
1217
1218 /*
1219  * Starting dispatch IO to raid.
1220  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1221  * broken meta in the middle of a log causes recovery can't find meta at the
1222  * head of log. If operations require meta at the head persistent in log, we
1223  * must make sure meta before it persistent in log too. A case is:
1224  *
1225  * stripe data/parity is in log, we start write stripe to raid disks. stripe
1226  * data/parity must be persistent in log before we do the write to raid disks.
1227  *
1228  * The solution is we restrictly maintain io_unit list order. In this case, we
1229  * only write stripes of an io_unit to raid disks till the io_unit is the first
1230  * one whose data/parity is in log.
1231  */
1232 void r5l_flush_stripe_to_raid(struct r5l_log *log)
1233 {
1234         bool do_flush;
1235
1236         if (!log || !log->need_cache_flush)
1237                 return;
1238
1239         spin_lock_irq(&log->io_list_lock);
1240         /* flush bio is running */
1241         if (!list_empty(&log->flushing_ios)) {
1242                 spin_unlock_irq(&log->io_list_lock);
1243                 return;
1244         }
1245         list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1246         do_flush = !list_empty(&log->flushing_ios);
1247         spin_unlock_irq(&log->io_list_lock);
1248
1249         if (!do_flush)
1250                 return;
1251         bio_reset(&log->flush_bio);
1252         log->flush_bio.bi_bdev = log->rdev->bdev;
1253         log->flush_bio.bi_end_io = r5l_log_flush_endio;
1254         log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1255         submit_bio(&log->flush_bio);
1256 }
1257
1258 static void r5l_write_super(struct r5l_log *log, sector_t cp);
1259 static void r5l_write_super_and_discard_space(struct r5l_log *log,
1260         sector_t end)
1261 {
1262         struct block_device *bdev = log->rdev->bdev;
1263         struct mddev *mddev;
1264
1265         r5l_write_super(log, end);
1266
1267         if (!blk_queue_discard(bdev_get_queue(bdev)))
1268                 return;
1269
1270         mddev = log->rdev->mddev;
1271         /*
1272          * Discard could zero data, so before discard we must make sure
1273          * superblock is updated to new log tail. Updating superblock (either
1274          * directly call md_update_sb() or depend on md thread) must hold
1275          * reconfig mutex. On the other hand, raid5_quiesce is called with
1276          * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1277          * for all IO finish, hence waitting for reclaim thread, while reclaim
1278          * thread is calling this function and waitting for reconfig mutex. So
1279          * there is a deadlock. We workaround this issue with a trylock.
1280          * FIXME: we could miss discard if we can't take reconfig mutex
1281          */
1282         set_mask_bits(&mddev->sb_flags, 0,
1283                 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1284         if (!mddev_trylock(mddev))
1285                 return;
1286         md_update_sb(mddev, 1);
1287         mddev_unlock(mddev);
1288
1289         /* discard IO error really doesn't matter, ignore it */
1290         if (log->last_checkpoint < end) {
1291                 blkdev_issue_discard(bdev,
1292                                 log->last_checkpoint + log->rdev->data_offset,
1293                                 end - log->last_checkpoint, GFP_NOIO, 0);
1294         } else {
1295                 blkdev_issue_discard(bdev,
1296                                 log->last_checkpoint + log->rdev->data_offset,
1297                                 log->device_size - log->last_checkpoint,
1298                                 GFP_NOIO, 0);
1299                 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1300                                 GFP_NOIO, 0);
1301         }
1302 }
1303
1304 /*
1305  * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1306  * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1307  *
1308  * must hold conf->device_lock
1309  */
1310 static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1311 {
1312         BUG_ON(list_empty(&sh->lru));
1313         BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1314         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1315
1316         /*
1317          * The stripe is not ON_RELEASE_LIST, so it is safe to call
1318          * raid5_release_stripe() while holding conf->device_lock
1319          */
1320         BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1321         assert_spin_locked(&conf->device_lock);
1322
1323         list_del_init(&sh->lru);
1324         atomic_inc(&sh->count);
1325
1326         set_bit(STRIPE_HANDLE, &sh->state);
1327         atomic_inc(&conf->active_stripes);
1328         r5c_make_stripe_write_out(sh);
1329
1330         if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1331                 atomic_inc(&conf->r5c_flushing_partial_stripes);
1332         else
1333                 atomic_inc(&conf->r5c_flushing_full_stripes);
1334         raid5_release_stripe(sh);
1335 }
1336
1337 /*
1338  * if num == 0, flush all full stripes
1339  * if num > 0, flush all full stripes. If less than num full stripes are
1340  *             flushed, flush some partial stripes until totally num stripes are
1341  *             flushed or there is no more cached stripes.
1342  */
1343 void r5c_flush_cache(struct r5conf *conf, int num)
1344 {
1345         int count;
1346         struct stripe_head *sh, *next;
1347
1348         assert_spin_locked(&conf->device_lock);
1349         if (!conf->log)
1350                 return;
1351
1352         count = 0;
1353         list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1354                 r5c_flush_stripe(conf, sh);
1355                 count++;
1356         }
1357
1358         if (count >= num)
1359                 return;
1360         list_for_each_entry_safe(sh, next,
1361                                  &conf->r5c_partial_stripe_list, lru) {
1362                 r5c_flush_stripe(conf, sh);
1363                 if (++count >= num)
1364                         break;
1365         }
1366 }
1367
1368 static void r5c_do_reclaim(struct r5conf *conf)
1369 {
1370         struct r5l_log *log = conf->log;
1371         struct stripe_head *sh;
1372         int count = 0;
1373         unsigned long flags;
1374         int total_cached;
1375         int stripes_to_flush;
1376         int flushing_partial, flushing_full;
1377
1378         if (!r5c_is_writeback(log))
1379                 return;
1380
1381         flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1382         flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1383         total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1384                 atomic_read(&conf->r5c_cached_full_stripes) -
1385                 flushing_full - flushing_partial;
1386
1387         if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1388             atomic_read(&conf->empty_inactive_list_nr) > 0)
1389                 /*
1390                  * if stripe cache pressure high, flush all full stripes and
1391                  * some partial stripes
1392                  */
1393                 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1394         else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1395                  atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1396                  R5C_FULL_STRIPE_FLUSH_BATCH)
1397                 /*
1398                  * if stripe cache pressure moderate, or if there is many full
1399                  * stripes,flush all full stripes
1400                  */
1401                 stripes_to_flush = 0;
1402         else
1403                 /* no need to flush */
1404                 stripes_to_flush = -1;
1405
1406         if (stripes_to_flush >= 0) {
1407                 spin_lock_irqsave(&conf->device_lock, flags);
1408                 r5c_flush_cache(conf, stripes_to_flush);
1409                 spin_unlock_irqrestore(&conf->device_lock, flags);
1410         }
1411
1412         /* if log space is tight, flush stripes on stripe_in_journal_list */
1413         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1414                 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1415                 spin_lock(&conf->device_lock);
1416                 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1417                         /*
1418                          * stripes on stripe_in_journal_list could be in any
1419                          * state of the stripe_cache state machine. In this
1420                          * case, we only want to flush stripe on
1421                          * r5c_cached_full/partial_stripes. The following
1422                          * condition makes sure the stripe is on one of the
1423                          * two lists.
1424                          */
1425                         if (!list_empty(&sh->lru) &&
1426                             !test_bit(STRIPE_HANDLE, &sh->state) &&
1427                             atomic_read(&sh->count) == 0) {
1428                                 r5c_flush_stripe(conf, sh);
1429                                 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1430                                         break;
1431                         }
1432                 }
1433                 spin_unlock(&conf->device_lock);
1434                 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1435         }
1436
1437         if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1438                 r5l_run_no_space_stripes(log);
1439
1440         md_wakeup_thread(conf->mddev->thread);
1441 }
1442
1443 static void r5l_do_reclaim(struct r5l_log *log)
1444 {
1445         struct r5conf *conf = log->rdev->mddev->private;
1446         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1447         sector_t reclaimable;
1448         sector_t next_checkpoint;
1449         bool write_super;
1450
1451         spin_lock_irq(&log->io_list_lock);
1452         write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1453                 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1454         /*
1455          * move proper io_unit to reclaim list. We should not change the order.
1456          * reclaimable/unreclaimable io_unit can be mixed in the list, we
1457          * shouldn't reuse space of an unreclaimable io_unit
1458          */
1459         while (1) {
1460                 reclaimable = r5l_reclaimable_space(log);
1461                 if (reclaimable >= reclaim_target ||
1462                     (list_empty(&log->running_ios) &&
1463                      list_empty(&log->io_end_ios) &&
1464                      list_empty(&log->flushing_ios) &&
1465                      list_empty(&log->finished_ios)))
1466                         break;
1467
1468                 md_wakeup_thread(log->rdev->mddev->thread);
1469                 wait_event_lock_irq(log->iounit_wait,
1470                                     r5l_reclaimable_space(log) > reclaimable,
1471                                     log->io_list_lock);
1472         }
1473
1474         next_checkpoint = r5c_calculate_new_cp(conf);
1475         spin_unlock_irq(&log->io_list_lock);
1476
1477         if (reclaimable == 0 || !write_super)
1478                 return;
1479
1480         /*
1481          * write_super will flush cache of each raid disk. We must write super
1482          * here, because the log area might be reused soon and we don't want to
1483          * confuse recovery
1484          */
1485         r5l_write_super_and_discard_space(log, next_checkpoint);
1486
1487         mutex_lock(&log->io_mutex);
1488         log->last_checkpoint = next_checkpoint;
1489         r5c_update_log_state(log);
1490         mutex_unlock(&log->io_mutex);
1491
1492         r5l_run_no_space_stripes(log);
1493 }
1494
1495 static void r5l_reclaim_thread(struct md_thread *thread)
1496 {
1497         struct mddev *mddev = thread->mddev;
1498         struct r5conf *conf = mddev->private;
1499         struct r5l_log *log = conf->log;
1500
1501         if (!log)
1502                 return;
1503         r5c_do_reclaim(conf);
1504         r5l_do_reclaim(log);
1505 }
1506
1507 void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1508 {
1509         unsigned long target;
1510         unsigned long new = (unsigned long)space; /* overflow in theory */
1511
1512         if (!log)
1513                 return;
1514         do {
1515                 target = log->reclaim_target;
1516                 if (new < target)
1517                         return;
1518         } while (cmpxchg(&log->reclaim_target, target, new) != target);
1519         md_wakeup_thread(log->reclaim_thread);
1520 }
1521
1522 void r5l_quiesce(struct r5l_log *log, int state)
1523 {
1524         struct mddev *mddev;
1525         if (!log || state == 2)
1526                 return;
1527         if (state == 0)
1528                 kthread_unpark(log->reclaim_thread->tsk);
1529         else if (state == 1) {
1530                 /* make sure r5l_write_super_and_discard_space exits */
1531                 mddev = log->rdev->mddev;
1532                 wake_up(&mddev->sb_wait);
1533                 kthread_park(log->reclaim_thread->tsk);
1534                 r5l_wake_reclaim(log, MaxSector);
1535                 r5l_do_reclaim(log);
1536         }
1537 }
1538
1539 bool r5l_log_disk_error(struct r5conf *conf)
1540 {
1541         struct r5l_log *log;
1542         bool ret;
1543         /* don't allow write if journal disk is missing */
1544         rcu_read_lock();
1545         log = rcu_dereference(conf->log);
1546
1547         if (!log)
1548                 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1549         else
1550                 ret = test_bit(Faulty, &log->rdev->flags);
1551         rcu_read_unlock();
1552         return ret;
1553 }
1554
1555 struct r5l_recovery_ctx {
1556         struct page *meta_page;         /* current meta */
1557         sector_t meta_total_blocks;     /* total size of current meta and data */
1558         sector_t pos;                   /* recovery position */
1559         u64 seq;                        /* recovery position seq */
1560         int data_parity_stripes;        /* number of data_parity stripes */
1561         int data_only_stripes;          /* number of data_only stripes */
1562         struct list_head cached_list;
1563 };
1564
1565 static int r5l_recovery_read_meta_block(struct r5l_log *log,
1566                                         struct r5l_recovery_ctx *ctx)
1567 {
1568         struct page *page = ctx->meta_page;
1569         struct r5l_meta_block *mb;
1570         u32 crc, stored_crc;
1571
1572         if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, REQ_OP_READ, 0,
1573                           false))
1574                 return -EIO;
1575
1576         mb = page_address(page);
1577         stored_crc = le32_to_cpu(mb->checksum);
1578         mb->checksum = 0;
1579
1580         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1581             le64_to_cpu(mb->seq) != ctx->seq ||
1582             mb->version != R5LOG_VERSION ||
1583             le64_to_cpu(mb->position) != ctx->pos)
1584                 return -EINVAL;
1585
1586         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1587         if (stored_crc != crc)
1588                 return -EINVAL;
1589
1590         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1591                 return -EINVAL;
1592
1593         ctx->meta_total_blocks = BLOCK_SECTORS;
1594
1595         return 0;
1596 }
1597
1598 static void
1599 r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1600                                      struct page *page,
1601                                      sector_t pos, u64 seq)
1602 {
1603         struct r5l_meta_block *mb;
1604
1605         mb = page_address(page);
1606         clear_page(mb);
1607         mb->magic = cpu_to_le32(R5LOG_MAGIC);
1608         mb->version = R5LOG_VERSION;
1609         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1610         mb->seq = cpu_to_le64(seq);
1611         mb->position = cpu_to_le64(pos);
1612 }
1613
1614 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1615                                           u64 seq)
1616 {
1617         struct page *page;
1618         struct r5l_meta_block *mb;
1619
1620         page = alloc_page(GFP_KERNEL);
1621         if (!page)
1622                 return -ENOMEM;
1623         r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1624         mb = page_address(page);
1625         mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1626                                              mb, PAGE_SIZE));
1627         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1628                           REQ_FUA, false)) {
1629                 __free_page(page);
1630                 return -EIO;
1631         }
1632         __free_page(page);
1633         return 0;
1634 }
1635
1636 /*
1637  * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1638  * to mark valid (potentially not flushed) data in the journal.
1639  *
1640  * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1641  * so there should not be any mismatch here.
1642  */
1643 static void r5l_recovery_load_data(struct r5l_log *log,
1644                                    struct stripe_head *sh,
1645                                    struct r5l_recovery_ctx *ctx,
1646                                    struct r5l_payload_data_parity *payload,
1647                                    sector_t log_offset)
1648 {
1649         struct mddev *mddev = log->rdev->mddev;
1650         struct r5conf *conf = mddev->private;
1651         int dd_idx;
1652
1653         raid5_compute_sector(conf,
1654                              le64_to_cpu(payload->location), 0,
1655                              &dd_idx, sh);
1656         sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1657                      sh->dev[dd_idx].page, REQ_OP_READ, 0, false);
1658         sh->dev[dd_idx].log_checksum =
1659                 le32_to_cpu(payload->checksum[0]);
1660         ctx->meta_total_blocks += BLOCK_SECTORS;
1661
1662         set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1663         set_bit(STRIPE_R5C_CACHING, &sh->state);
1664 }
1665
1666 static void r5l_recovery_load_parity(struct r5l_log *log,
1667                                      struct stripe_head *sh,
1668                                      struct r5l_recovery_ctx *ctx,
1669                                      struct r5l_payload_data_parity *payload,
1670                                      sector_t log_offset)
1671 {
1672         struct mddev *mddev = log->rdev->mddev;
1673         struct r5conf *conf = mddev->private;
1674
1675         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1676         sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1677                      sh->dev[sh->pd_idx].page, REQ_OP_READ, 0, false);
1678         sh->dev[sh->pd_idx].log_checksum =
1679                 le32_to_cpu(payload->checksum[0]);
1680         set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1681
1682         if (sh->qd_idx >= 0) {
1683                 sync_page_io(log->rdev,
1684                              r5l_ring_add(log, log_offset, BLOCK_SECTORS),
1685                              PAGE_SIZE, sh->dev[sh->qd_idx].page,
1686                              REQ_OP_READ, 0, false);
1687                 sh->dev[sh->qd_idx].log_checksum =
1688                         le32_to_cpu(payload->checksum[1]);
1689                 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1690         }
1691         clear_bit(STRIPE_R5C_CACHING, &sh->state);
1692 }
1693
1694 static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1695 {
1696         int i;
1697
1698         sh->state = 0;
1699         sh->log_start = MaxSector;
1700         for (i = sh->disks; i--; )
1701                 sh->dev[i].flags = 0;
1702 }
1703
1704 static void
1705 r5l_recovery_replay_one_stripe(struct r5conf *conf,
1706                                struct stripe_head *sh,
1707                                struct r5l_recovery_ctx *ctx)
1708 {
1709         struct md_rdev *rdev, *rrdev;
1710         int disk_index;
1711         int data_count = 0;
1712
1713         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1714                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1715                         continue;
1716                 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1717                         continue;
1718                 data_count++;
1719         }
1720
1721         /*
1722          * stripes that only have parity must have been flushed
1723          * before the crash that we are now recovering from, so
1724          * there is nothing more to recovery.
1725          */
1726         if (data_count == 0)
1727                 goto out;
1728
1729         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1730                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1731                         continue;
1732
1733                 /* in case device is broken */
1734                 rcu_read_lock();
1735                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1736                 if (rdev) {
1737                         atomic_inc(&rdev->nr_pending);
1738                         rcu_read_unlock();
1739                         sync_page_io(rdev, sh->sector, PAGE_SIZE,
1740                                      sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1741                                      false);
1742                         rdev_dec_pending(rdev, rdev->mddev);
1743                         rcu_read_lock();
1744                 }
1745                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1746                 if (rrdev) {
1747                         atomic_inc(&rrdev->nr_pending);
1748                         rcu_read_unlock();
1749                         sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1750                                      sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1751                                      false);
1752                         rdev_dec_pending(rrdev, rrdev->mddev);
1753                         rcu_read_lock();
1754                 }
1755                 rcu_read_unlock();
1756         }
1757         ctx->data_parity_stripes++;
1758 out:
1759         r5l_recovery_reset_stripe(sh);
1760 }
1761
1762 static struct stripe_head *
1763 r5c_recovery_alloc_stripe(struct r5conf *conf,
1764                           sector_t stripe_sect)
1765 {
1766         struct stripe_head *sh;
1767
1768         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1769         if (!sh)
1770                 return NULL;  /* no more stripe available */
1771
1772         r5l_recovery_reset_stripe(sh);
1773
1774         return sh;
1775 }
1776
1777 static struct stripe_head *
1778 r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1779 {
1780         struct stripe_head *sh;
1781
1782         list_for_each_entry(sh, list, lru)
1783                 if (sh->sector == sect)
1784                         return sh;
1785         return NULL;
1786 }
1787
1788 static void
1789 r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1790                           struct r5l_recovery_ctx *ctx)
1791 {
1792         struct stripe_head *sh, *next;
1793
1794         list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1795                 r5l_recovery_reset_stripe(sh);
1796                 list_del_init(&sh->lru);
1797                 raid5_release_stripe(sh);
1798         }
1799 }
1800
1801 static void
1802 r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1803                             struct r5l_recovery_ctx *ctx)
1804 {
1805         struct stripe_head *sh, *next;
1806
1807         list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1808                 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1809                         r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1810                         list_del_init(&sh->lru);
1811                         raid5_release_stripe(sh);
1812                 }
1813 }
1814
1815 /* if matches return 0; otherwise return -EINVAL */
1816 static int
1817 r5l_recovery_verify_data_checksum(struct r5l_log *log, struct page *page,
1818                                   sector_t log_offset, __le32 log_checksum)
1819 {
1820         void *addr;
1821         u32 checksum;
1822
1823         sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1824                      page, REQ_OP_READ, 0, false);
1825         addr = kmap_atomic(page);
1826         checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1827         kunmap_atomic(addr);
1828         return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1829 }
1830
1831 /*
1832  * before loading data to stripe cache, we need verify checksum for all data,
1833  * if there is mismatch for any data page, we drop all data in the mata block
1834  */
1835 static int
1836 r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1837                                          struct r5l_recovery_ctx *ctx)
1838 {
1839         struct mddev *mddev = log->rdev->mddev;
1840         struct r5conf *conf = mddev->private;
1841         struct r5l_meta_block *mb = page_address(ctx->meta_page);
1842         sector_t mb_offset = sizeof(struct r5l_meta_block);
1843         sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1844         struct page *page;
1845         struct r5l_payload_data_parity *payload;
1846
1847         page = alloc_page(GFP_KERNEL);
1848         if (!page)
1849                 return -ENOMEM;
1850
1851         while (mb_offset < le32_to_cpu(mb->meta_size)) {
1852                 payload = (void *)mb + mb_offset;
1853
1854                 if (payload->header.type == R5LOG_PAYLOAD_DATA) {
1855                         if (r5l_recovery_verify_data_checksum(
1856                                     log, page, log_offset,
1857                                     payload->checksum[0]) < 0)
1858                                 goto mismatch;
1859                 } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) {
1860                         if (r5l_recovery_verify_data_checksum(
1861                                     log, page, log_offset,
1862                                     payload->checksum[0]) < 0)
1863                                 goto mismatch;
1864                         if (conf->max_degraded == 2 && /* q for RAID 6 */
1865                             r5l_recovery_verify_data_checksum(
1866                                     log, page,
1867                                     r5l_ring_add(log, log_offset,
1868                                                  BLOCK_SECTORS),
1869                                     payload->checksum[1]) < 0)
1870                                 goto mismatch;
1871                 } else /* not R5LOG_PAYLOAD_DATA or R5LOG_PAYLOAD_PARITY */
1872                         goto mismatch;
1873
1874                 log_offset = r5l_ring_add(log, log_offset,
1875                                           le32_to_cpu(payload->size));
1876
1877                 mb_offset += sizeof(struct r5l_payload_data_parity) +
1878                         sizeof(__le32) *
1879                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
1880         }
1881
1882         put_page(page);
1883         return 0;
1884
1885 mismatch:
1886         put_page(page);
1887         return -EINVAL;
1888 }
1889
1890 /*
1891  * Analyze all data/parity pages in one meta block
1892  * Returns:
1893  * 0 for success
1894  * -EINVAL for unknown playload type
1895  * -EAGAIN for checksum mismatch of data page
1896  * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
1897  */
1898 static int
1899 r5c_recovery_analyze_meta_block(struct r5l_log *log,
1900                                 struct r5l_recovery_ctx *ctx,
1901                                 struct list_head *cached_stripe_list)
1902 {
1903         struct mddev *mddev = log->rdev->mddev;
1904         struct r5conf *conf = mddev->private;
1905         struct r5l_meta_block *mb;
1906         struct r5l_payload_data_parity *payload;
1907         int mb_offset;
1908         sector_t log_offset;
1909         sector_t stripe_sect;
1910         struct stripe_head *sh;
1911         int ret;
1912
1913         /*
1914          * for mismatch in data blocks, we will drop all data in this mb, but
1915          * we will still read next mb for other data with FLUSH flag, as
1916          * io_unit could finish out of order.
1917          */
1918         ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
1919         if (ret == -EINVAL)
1920                 return -EAGAIN;
1921         else if (ret)
1922                 return ret;   /* -ENOMEM duo to alloc_page() failed */
1923
1924         mb = page_address(ctx->meta_page);
1925         mb_offset = sizeof(struct r5l_meta_block);
1926         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1927
1928         while (mb_offset < le32_to_cpu(mb->meta_size)) {
1929                 int dd;
1930
1931                 payload = (void *)mb + mb_offset;
1932                 stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ?
1933                         raid5_compute_sector(
1934                                 conf, le64_to_cpu(payload->location), 0, &dd,
1935                                 NULL)
1936                         : le64_to_cpu(payload->location);
1937
1938                 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
1939                                                 stripe_sect);
1940
1941                 if (!sh) {
1942                         sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
1943                         /*
1944                          * cannot get stripe from raid5_get_active_stripe
1945                          * try replay some stripes
1946                          */
1947                         if (!sh) {
1948                                 r5c_recovery_replay_stripes(
1949                                         cached_stripe_list, ctx);
1950                                 sh = r5c_recovery_alloc_stripe(
1951                                         conf, stripe_sect);
1952                         }
1953                         if (!sh) {
1954                                 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
1955                                         mdname(mddev),
1956                                         conf->min_nr_stripes * 2);
1957                                 raid5_set_cache_size(mddev,
1958                                                      conf->min_nr_stripes * 2);
1959                                 sh = r5c_recovery_alloc_stripe(conf,
1960                                                                stripe_sect);
1961                         }
1962                         if (!sh) {
1963                                 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
1964                                        mdname(mddev));
1965                                 return -ENOMEM;
1966                         }
1967                         list_add_tail(&sh->lru, cached_stripe_list);
1968                 }
1969
1970                 if (payload->header.type == R5LOG_PAYLOAD_DATA) {
1971                         if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
1972                             test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
1973                                 r5l_recovery_replay_one_stripe(conf, sh, ctx);
1974                                 list_move_tail(&sh->lru, cached_stripe_list);
1975                         }
1976                         r5l_recovery_load_data(log, sh, ctx, payload,
1977                                                log_offset);
1978                 } else if (payload->header.type == R5LOG_PAYLOAD_PARITY)
1979                         r5l_recovery_load_parity(log, sh, ctx, payload,
1980                                                  log_offset);
1981                 else
1982                         return -EINVAL;
1983
1984                 log_offset = r5l_ring_add(log, log_offset,
1985                                           le32_to_cpu(payload->size));
1986
1987                 mb_offset += sizeof(struct r5l_payload_data_parity) +
1988                         sizeof(__le32) *
1989                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
1990         }
1991
1992         return 0;
1993 }
1994
1995 /*
1996  * Load the stripe into cache. The stripe will be written out later by
1997  * the stripe cache state machine.
1998  */
1999 static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2000                                          struct stripe_head *sh)
2001 {
2002         struct r5dev *dev;
2003         int i;
2004
2005         for (i = sh->disks; i--; ) {
2006                 dev = sh->dev + i;
2007                 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2008                         set_bit(R5_InJournal, &dev->flags);
2009                         set_bit(R5_UPTODATE, &dev->flags);
2010                 }
2011         }
2012 }
2013
2014 /*
2015  * Scan through the log for all to-be-flushed data
2016  *
2017  * For stripes with data and parity, namely Data-Parity stripe
2018  * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2019  *
2020  * For stripes with only data, namely Data-Only stripe
2021  * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2022  *
2023  * For a stripe, if we see data after parity, we should discard all previous
2024  * data and parity for this stripe, as these data are already flushed to
2025  * the array.
2026  *
2027  * At the end of the scan, we return the new journal_tail, which points to
2028  * first data-only stripe on the journal device, or next invalid meta block.
2029  */
2030 static int r5c_recovery_flush_log(struct r5l_log *log,
2031                                   struct r5l_recovery_ctx *ctx)
2032 {
2033         struct stripe_head *sh;
2034         int ret = 0;
2035
2036         /* scan through the log */
2037         while (1) {
2038                 if (r5l_recovery_read_meta_block(log, ctx))
2039                         break;
2040
2041                 ret = r5c_recovery_analyze_meta_block(log, ctx,
2042                                                       &ctx->cached_list);
2043                 /*
2044                  * -EAGAIN means mismatch in data block, in this case, we still
2045                  * try scan the next metablock
2046                  */
2047                 if (ret && ret != -EAGAIN)
2048                         break;   /* ret == -EINVAL or -ENOMEM */
2049                 ctx->seq++;
2050                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2051         }
2052
2053         if (ret == -ENOMEM) {
2054                 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2055                 return ret;
2056         }
2057
2058         /* replay data-parity stripes */
2059         r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2060
2061         /* load data-only stripes to stripe cache */
2062         list_for_each_entry(sh, &ctx->cached_list, lru) {
2063                 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2064                 r5c_recovery_load_one_stripe(log, sh);
2065                 ctx->data_only_stripes++;
2066         }
2067
2068         return 0;
2069 }
2070
2071 /*
2072  * we did a recovery. Now ctx.pos points to an invalid meta block. New
2073  * log will start here. but we can't let superblock point to last valid
2074  * meta block. The log might looks like:
2075  * | meta 1| meta 2| meta 3|
2076  * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2077  * superblock points to meta 1, we write a new valid meta 2n.  if crash
2078  * happens again, new recovery will start from meta 1. Since meta 2n is
2079  * valid now, recovery will think meta 3 is valid, which is wrong.
2080  * The solution is we create a new meta in meta2 with its seq == meta
2081  * 1's seq + 10000 and let superblock points to meta2. The same recovery
2082  * will not think meta 3 is a valid meta, because its seq doesn't match
2083  */
2084
2085 /*
2086  * Before recovery, the log looks like the following
2087  *
2088  *   ---------------------------------------------
2089  *   |           valid log        | invalid log  |
2090  *   ---------------------------------------------
2091  *   ^
2092  *   |- log->last_checkpoint
2093  *   |- log->last_cp_seq
2094  *
2095  * Now we scan through the log until we see invalid entry
2096  *
2097  *   ---------------------------------------------
2098  *   |           valid log        | invalid log  |
2099  *   ---------------------------------------------
2100  *   ^                            ^
2101  *   |- log->last_checkpoint      |- ctx->pos
2102  *   |- log->last_cp_seq          |- ctx->seq
2103  *
2104  * From this point, we need to increase seq number by 10 to avoid
2105  * confusing next recovery.
2106  *
2107  *   ---------------------------------------------
2108  *   |           valid log        | invalid log  |
2109  *   ---------------------------------------------
2110  *   ^                              ^
2111  *   |- log->last_checkpoint        |- ctx->pos+1
2112  *   |- log->last_cp_seq            |- ctx->seq+10001
2113  *
2114  * However, it is not safe to start the state machine yet, because data only
2115  * parities are not yet secured in RAID. To save these data only parities, we
2116  * rewrite them from seq+11.
2117  *
2118  *   -----------------------------------------------------------------
2119  *   |           valid log        | data only stripes | invalid log  |
2120  *   -----------------------------------------------------------------
2121  *   ^                                                ^
2122  *   |- log->last_checkpoint                          |- ctx->pos+n
2123  *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2124  *
2125  * If failure happens again during this process, the recovery can safe start
2126  * again from log->last_checkpoint.
2127  *
2128  * Once data only stripes are rewritten to journal, we move log_tail
2129  *
2130  *   -----------------------------------------------------------------
2131  *   |     old log        |    data only stripes    | invalid log  |
2132  *   -----------------------------------------------------------------
2133  *                        ^                         ^
2134  *                        |- log->last_checkpoint   |- ctx->pos+n
2135  *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2136  *
2137  * Then we can safely start the state machine. If failure happens from this
2138  * point on, the recovery will start from new log->last_checkpoint.
2139  */
2140 static int
2141 r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2142                                        struct r5l_recovery_ctx *ctx)
2143 {
2144         struct stripe_head *sh;
2145         struct mddev *mddev = log->rdev->mddev;
2146         struct page *page;
2147         sector_t next_checkpoint = MaxSector;
2148
2149         page = alloc_page(GFP_KERNEL);
2150         if (!page) {
2151                 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2152                        mdname(mddev));
2153                 return -ENOMEM;
2154         }
2155
2156         WARN_ON(list_empty(&ctx->cached_list));
2157
2158         list_for_each_entry(sh, &ctx->cached_list, lru) {
2159                 struct r5l_meta_block *mb;
2160                 int i;
2161                 int offset;
2162                 sector_t write_pos;
2163
2164                 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2165                 r5l_recovery_create_empty_meta_block(log, page,
2166                                                      ctx->pos, ctx->seq);
2167                 mb = page_address(page);
2168                 offset = le32_to_cpu(mb->meta_size);
2169                 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2170
2171                 for (i = sh->disks; i--; ) {
2172                         struct r5dev *dev = &sh->dev[i];
2173                         struct r5l_payload_data_parity *payload;
2174                         void *addr;
2175
2176                         if (test_bit(R5_InJournal, &dev->flags)) {
2177                                 payload = (void *)mb + offset;
2178                                 payload->header.type = cpu_to_le16(
2179                                         R5LOG_PAYLOAD_DATA);
2180                                 payload->size = BLOCK_SECTORS;
2181                                 payload->location = cpu_to_le64(
2182                                         raid5_compute_blocknr(sh, i, 0));
2183                                 addr = kmap_atomic(dev->page);
2184                                 payload->checksum[0] = cpu_to_le32(
2185                                         crc32c_le(log->uuid_checksum, addr,
2186                                                   PAGE_SIZE));
2187                                 kunmap_atomic(addr);
2188                                 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2189                                              dev->page, REQ_OP_WRITE, 0, false);
2190                                 write_pos = r5l_ring_add(log, write_pos,
2191                                                          BLOCK_SECTORS);
2192                                 offset += sizeof(__le32) +
2193                                         sizeof(struct r5l_payload_data_parity);
2194
2195                         }
2196                 }
2197                 mb->meta_size = cpu_to_le32(offset);
2198                 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2199                                                      mb, PAGE_SIZE));
2200                 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2201                              REQ_OP_WRITE, REQ_FUA, false);
2202                 sh->log_start = ctx->pos;
2203                 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2204                 atomic_inc(&log->stripe_in_journal_count);
2205                 ctx->pos = write_pos;
2206                 ctx->seq += 1;
2207                 next_checkpoint = sh->log_start;
2208         }
2209         log->next_checkpoint = next_checkpoint;
2210         __free_page(page);
2211         return 0;
2212 }
2213
2214 static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2215                                                  struct r5l_recovery_ctx *ctx)
2216 {
2217         struct mddev *mddev = log->rdev->mddev;
2218         struct r5conf *conf = mddev->private;
2219         struct stripe_head *sh, *next;
2220
2221         if (ctx->data_only_stripes == 0)
2222                 return;
2223
2224         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2225
2226         list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2227                 r5c_make_stripe_write_out(sh);
2228                 set_bit(STRIPE_HANDLE, &sh->state);
2229                 list_del_init(&sh->lru);
2230                 raid5_release_stripe(sh);
2231         }
2232
2233         md_wakeup_thread(conf->mddev->thread);
2234         /* reuse conf->wait_for_quiescent in recovery */
2235         wait_event(conf->wait_for_quiescent,
2236                    atomic_read(&conf->active_stripes) == 0);
2237
2238         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2239 }
2240
2241 static int r5l_recovery_log(struct r5l_log *log)
2242 {
2243         struct mddev *mddev = log->rdev->mddev;
2244         struct r5l_recovery_ctx ctx;
2245         int ret;
2246         sector_t pos;
2247
2248         ctx.pos = log->last_checkpoint;
2249         ctx.seq = log->last_cp_seq;
2250         ctx.meta_page = alloc_page(GFP_KERNEL);
2251         ctx.data_only_stripes = 0;
2252         ctx.data_parity_stripes = 0;
2253         INIT_LIST_HEAD(&ctx.cached_list);
2254
2255         if (!ctx.meta_page)
2256                 return -ENOMEM;
2257
2258         ret = r5c_recovery_flush_log(log, &ctx);
2259         __free_page(ctx.meta_page);
2260
2261         if (ret)
2262                 return ret;
2263
2264         pos = ctx.pos;
2265         ctx.seq += 10000;
2266
2267
2268         if ((ctx.data_only_stripes == 0) && (ctx.data_parity_stripes == 0))
2269                 pr_debug("md/raid:%s: starting from clean shutdown\n",
2270                          mdname(mddev));
2271         else
2272                 pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2273                          mdname(mddev), ctx.data_only_stripes,
2274                          ctx.data_parity_stripes);
2275
2276         if (ctx.data_only_stripes == 0) {
2277                 log->next_checkpoint = ctx.pos;
2278                 r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq++);
2279                 ctx.pos = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
2280         } else if (r5c_recovery_rewrite_data_only_stripes(log, &ctx)) {
2281                 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2282                        mdname(mddev));
2283                 return -EIO;
2284         }
2285
2286         log->log_start = ctx.pos;
2287         log->seq = ctx.seq;
2288         log->last_checkpoint = pos;
2289         r5l_write_super(log, pos);
2290
2291         r5c_recovery_flush_data_only_stripes(log, &ctx);
2292         return 0;
2293 }
2294
2295 static void r5l_write_super(struct r5l_log *log, sector_t cp)
2296 {
2297         struct mddev *mddev = log->rdev->mddev;
2298
2299         log->rdev->journal_tail = cp;
2300         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2301 }
2302
2303 static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2304 {
2305         struct r5conf *conf = mddev->private;
2306         int ret;
2307
2308         if (!conf->log)
2309                 return 0;
2310
2311         switch (conf->log->r5c_journal_mode) {
2312         case R5C_JOURNAL_MODE_WRITE_THROUGH:
2313                 ret = snprintf(
2314                         page, PAGE_SIZE, "[%s] %s\n",
2315                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2316                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2317                 break;
2318         case R5C_JOURNAL_MODE_WRITE_BACK:
2319                 ret = snprintf(
2320                         page, PAGE_SIZE, "%s [%s]\n",
2321                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2322                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2323                 break;
2324         default:
2325                 ret = 0;
2326         }
2327         return ret;
2328 }
2329
2330 static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2331                                       const char *page, size_t length)
2332 {
2333         struct r5conf *conf = mddev->private;
2334         struct r5l_log *log = conf->log;
2335         int val = -1, i;
2336         int len = length;
2337
2338         if (!log)
2339                 return -ENODEV;
2340
2341         if (len && page[len - 1] == '\n')
2342                 len -= 1;
2343         for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++)
2344                 if (strlen(r5c_journal_mode_str[i]) == len &&
2345                     strncmp(page, r5c_journal_mode_str[i], len) == 0) {
2346                         val = i;
2347                         break;
2348                 }
2349         if (val < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2350             val > R5C_JOURNAL_MODE_WRITE_BACK)
2351                 return -EINVAL;
2352
2353         if (raid5_calc_degraded(conf) > 0 &&
2354             val == R5C_JOURNAL_MODE_WRITE_BACK)
2355                 return -EINVAL;
2356
2357         mddev_suspend(mddev);
2358         conf->log->r5c_journal_mode = val;
2359         mddev_resume(mddev);
2360
2361         pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2362                  mdname(mddev), val, r5c_journal_mode_str[val]);
2363         return length;
2364 }
2365
2366 struct md_sysfs_entry
2367 r5c_journal_mode = __ATTR(journal_mode, 0644,
2368                           r5c_journal_mode_show, r5c_journal_mode_store);
2369
2370 /*
2371  * Try handle write operation in caching phase. This function should only
2372  * be called in write-back mode.
2373  *
2374  * If all outstanding writes can be handled in caching phase, returns 0
2375  * If writes requires write-out phase, call r5c_make_stripe_write_out()
2376  * and returns -EAGAIN
2377  */
2378 int r5c_try_caching_write(struct r5conf *conf,
2379                           struct stripe_head *sh,
2380                           struct stripe_head_state *s,
2381                           int disks)
2382 {
2383         struct r5l_log *log = conf->log;
2384         int i;
2385         struct r5dev *dev;
2386         int to_cache = 0;
2387         void **pslot;
2388         sector_t tree_index;
2389         int ret;
2390         uintptr_t refcount;
2391
2392         BUG_ON(!r5c_is_writeback(log));
2393
2394         if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2395                 /*
2396                  * There are two different scenarios here:
2397                  *  1. The stripe has some data cached, and it is sent to
2398                  *     write-out phase for reclaim
2399                  *  2. The stripe is clean, and this is the first write
2400                  *
2401                  * For 1, return -EAGAIN, so we continue with
2402                  * handle_stripe_dirtying().
2403                  *
2404                  * For 2, set STRIPE_R5C_CACHING and continue with caching
2405                  * write.
2406                  */
2407
2408                 /* case 1: anything injournal or anything in written */
2409                 if (s->injournal > 0 || s->written > 0)
2410                         return -EAGAIN;
2411                 /* case 2 */
2412                 set_bit(STRIPE_R5C_CACHING, &sh->state);
2413         }
2414
2415         /*
2416          * When run in degraded mode, array is set to write-through mode.
2417          * This check helps drain pending write safely in the transition to
2418          * write-through mode.
2419          */
2420         if (s->failed) {
2421                 r5c_make_stripe_write_out(sh);
2422                 return -EAGAIN;
2423         }
2424
2425         for (i = disks; i--; ) {
2426                 dev = &sh->dev[i];
2427                 /* if non-overwrite, use writing-out phase */
2428                 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2429                     !test_bit(R5_InJournal, &dev->flags)) {
2430                         r5c_make_stripe_write_out(sh);
2431                         return -EAGAIN;
2432                 }
2433         }
2434
2435         /* if the stripe is not counted in big_stripe_tree, add it now */
2436         if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2437             !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2438                 tree_index = r5c_tree_index(conf, sh->sector);
2439                 spin_lock(&log->tree_lock);
2440                 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2441                                                tree_index);
2442                 if (pslot) {
2443                         refcount = (uintptr_t)radix_tree_deref_slot_protected(
2444                                 pslot, &log->tree_lock) >>
2445                                 R5C_RADIX_COUNT_SHIFT;
2446                         radix_tree_replace_slot(
2447                                 &log->big_stripe_tree, pslot,
2448                                 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2449                 } else {
2450                         /*
2451                          * this radix_tree_insert can fail safely, so no
2452                          * need to call radix_tree_preload()
2453                          */
2454                         ret = radix_tree_insert(
2455                                 &log->big_stripe_tree, tree_index,
2456                                 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2457                         if (ret) {
2458                                 spin_unlock(&log->tree_lock);
2459                                 r5c_make_stripe_write_out(sh);
2460                                 return -EAGAIN;
2461                         }
2462                 }
2463                 spin_unlock(&log->tree_lock);
2464
2465                 /*
2466                  * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2467                  * counted in the radix tree
2468                  */
2469                 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2470                 atomic_inc(&conf->r5c_cached_partial_stripes);
2471         }
2472
2473         for (i = disks; i--; ) {
2474                 dev = &sh->dev[i];
2475                 if (dev->towrite) {
2476                         set_bit(R5_Wantwrite, &dev->flags);
2477                         set_bit(R5_Wantdrain, &dev->flags);
2478                         set_bit(R5_LOCKED, &dev->flags);
2479                         to_cache++;
2480                 }
2481         }
2482
2483         if (to_cache) {
2484                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2485                 /*
2486                  * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2487                  * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2488                  * r5c_handle_data_cached()
2489                  */
2490                 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2491         }
2492
2493         return 0;
2494 }
2495
2496 /*
2497  * free extra pages (orig_page) we allocated for prexor
2498  */
2499 void r5c_release_extra_page(struct stripe_head *sh)
2500 {
2501         struct r5conf *conf = sh->raid_conf;
2502         int i;
2503         bool using_disk_info_extra_page;
2504
2505         using_disk_info_extra_page =
2506                 sh->dev[0].orig_page == conf->disks[0].extra_page;
2507
2508         for (i = sh->disks; i--; )
2509                 if (sh->dev[i].page != sh->dev[i].orig_page) {
2510                         struct page *p = sh->dev[i].orig_page;
2511
2512                         sh->dev[i].orig_page = sh->dev[i].page;
2513                         clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2514
2515                         if (!using_disk_info_extra_page)
2516                                 put_page(p);
2517                 }
2518
2519         if (using_disk_info_extra_page) {
2520                 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2521                 md_wakeup_thread(conf->mddev->thread);
2522         }
2523 }
2524
2525 void r5c_use_extra_page(struct stripe_head *sh)
2526 {
2527         struct r5conf *conf = sh->raid_conf;
2528         int i;
2529         struct r5dev *dev;
2530
2531         for (i = sh->disks; i--; ) {
2532                 dev = &sh->dev[i];
2533                 if (dev->orig_page != dev->page)
2534                         put_page(dev->orig_page);
2535                 dev->orig_page = conf->disks[i].extra_page;
2536         }
2537 }
2538
2539 /*
2540  * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2541  * stripe is committed to RAID disks.
2542  */
2543 void r5c_finish_stripe_write_out(struct r5conf *conf,
2544                                  struct stripe_head *sh,
2545                                  struct stripe_head_state *s)
2546 {
2547         struct r5l_log *log = conf->log;
2548         int i;
2549         int do_wakeup = 0;
2550         sector_t tree_index;
2551         void **pslot;
2552         uintptr_t refcount;
2553
2554         if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2555                 return;
2556
2557         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2558         clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2559
2560         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2561                 return;
2562
2563         for (i = sh->disks; i--; ) {
2564                 clear_bit(R5_InJournal, &sh->dev[i].flags);
2565                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2566                         do_wakeup = 1;
2567         }
2568
2569         /*
2570          * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2571          * We updated R5_InJournal, so we also update s->injournal.
2572          */
2573         s->injournal = 0;
2574
2575         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2576                 if (atomic_dec_and_test(&conf->pending_full_writes))
2577                         md_wakeup_thread(conf->mddev->thread);
2578
2579         if (do_wakeup)
2580                 wake_up(&conf->wait_for_overlap);
2581
2582         spin_lock_irq(&log->stripe_in_journal_lock);
2583         list_del_init(&sh->r5c);
2584         spin_unlock_irq(&log->stripe_in_journal_lock);
2585         sh->log_start = MaxSector;
2586
2587         atomic_dec(&log->stripe_in_journal_count);
2588         r5c_update_log_state(log);
2589
2590         /* stop counting this stripe in big_stripe_tree */
2591         if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2592             test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2593                 tree_index = r5c_tree_index(conf, sh->sector);
2594                 spin_lock(&log->tree_lock);
2595                 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2596                                                tree_index);
2597                 BUG_ON(pslot == NULL);
2598                 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2599                         pslot, &log->tree_lock) >>
2600                         R5C_RADIX_COUNT_SHIFT;
2601                 if (refcount == 1)
2602                         radix_tree_delete(&log->big_stripe_tree, tree_index);
2603                 else
2604                         radix_tree_replace_slot(
2605                                 &log->big_stripe_tree, pslot,
2606                                 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2607                 spin_unlock(&log->tree_lock);
2608         }
2609
2610         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2611                 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2612                 atomic_dec(&conf->r5c_flushing_partial_stripes);
2613                 atomic_dec(&conf->r5c_cached_partial_stripes);
2614         }
2615
2616         if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2617                 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2618                 atomic_dec(&conf->r5c_flushing_full_stripes);
2619                 atomic_dec(&conf->r5c_cached_full_stripes);
2620         }
2621 }
2622
2623 int
2624 r5c_cache_data(struct r5l_log *log, struct stripe_head *sh,
2625                struct stripe_head_state *s)
2626 {
2627         struct r5conf *conf = sh->raid_conf;
2628         int pages = 0;
2629         int reserve;
2630         int i;
2631         int ret = 0;
2632
2633         BUG_ON(!log);
2634
2635         for (i = 0; i < sh->disks; i++) {
2636                 void *addr;
2637
2638                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2639                         continue;
2640                 addr = kmap_atomic(sh->dev[i].page);
2641                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2642                                                     addr, PAGE_SIZE);
2643                 kunmap_atomic(addr);
2644                 pages++;
2645         }
2646         WARN_ON(pages == 0);
2647
2648         /*
2649          * The stripe must enter state machine again to call endio, so
2650          * don't delay.
2651          */
2652         clear_bit(STRIPE_DELAYED, &sh->state);
2653         atomic_inc(&sh->count);
2654
2655         mutex_lock(&log->io_mutex);
2656         /* meta + data */
2657         reserve = (1 + pages) << (PAGE_SHIFT - 9);
2658
2659         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2660             sh->log_start == MaxSector)
2661                 r5l_add_no_space_stripe(log, sh);
2662         else if (!r5l_has_free_space(log, reserve)) {
2663                 if (sh->log_start == log->last_checkpoint)
2664                         BUG();
2665                 else
2666                         r5l_add_no_space_stripe(log, sh);
2667         } else {
2668                 ret = r5l_log_stripe(log, sh, pages, 0);
2669                 if (ret) {
2670                         spin_lock_irq(&log->io_list_lock);
2671                         list_add_tail(&sh->log_list, &log->no_mem_stripes);
2672                         spin_unlock_irq(&log->io_list_lock);
2673                 }
2674         }
2675
2676         mutex_unlock(&log->io_mutex);
2677         return 0;
2678 }
2679
2680 /* check whether this big stripe is in write back cache. */
2681 bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2682 {
2683         struct r5l_log *log = conf->log;
2684         sector_t tree_index;
2685         void *slot;
2686
2687         if (!log)
2688                 return false;
2689
2690         WARN_ON_ONCE(!rcu_read_lock_held());
2691         tree_index = r5c_tree_index(conf, sect);
2692         slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2693         return slot != NULL;
2694 }
2695
2696 static int r5l_load_log(struct r5l_log *log)
2697 {
2698         struct md_rdev *rdev = log->rdev;
2699         struct page *page;
2700         struct r5l_meta_block *mb;
2701         sector_t cp = log->rdev->journal_tail;
2702         u32 stored_crc, expected_crc;
2703         bool create_super = false;
2704         int ret = 0;
2705
2706         /* Make sure it's valid */
2707         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2708                 cp = 0;
2709         page = alloc_page(GFP_KERNEL);
2710         if (!page)
2711                 return -ENOMEM;
2712
2713         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2714                 ret = -EIO;
2715                 goto ioerr;
2716         }
2717         mb = page_address(page);
2718
2719         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2720             mb->version != R5LOG_VERSION) {
2721                 create_super = true;
2722                 goto create;
2723         }
2724         stored_crc = le32_to_cpu(mb->checksum);
2725         mb->checksum = 0;
2726         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2727         if (stored_crc != expected_crc) {
2728                 create_super = true;
2729                 goto create;
2730         }
2731         if (le64_to_cpu(mb->position) != cp) {
2732                 create_super = true;
2733                 goto create;
2734         }
2735 create:
2736         if (create_super) {
2737                 log->last_cp_seq = prandom_u32();
2738                 cp = 0;
2739                 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2740                 /*
2741                  * Make sure super points to correct address. Log might have
2742                  * data very soon. If super hasn't correct log tail address,
2743                  * recovery can't find the log
2744                  */
2745                 r5l_write_super(log, cp);
2746         } else
2747                 log->last_cp_seq = le64_to_cpu(mb->seq);
2748
2749         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
2750         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
2751         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
2752                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
2753         log->last_checkpoint = cp;
2754
2755         __free_page(page);
2756
2757         if (create_super) {
2758                 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
2759                 log->seq = log->last_cp_seq + 1;
2760                 log->next_checkpoint = cp;
2761         } else
2762                 ret = r5l_recovery_log(log);
2763
2764         r5c_update_log_state(log);
2765         return ret;
2766 ioerr:
2767         __free_page(page);
2768         return ret;
2769 }
2770
2771 void r5c_update_on_rdev_error(struct mddev *mddev)
2772 {
2773         struct r5conf *conf = mddev->private;
2774         struct r5l_log *log = conf->log;
2775
2776         if (!log)
2777                 return;
2778
2779         if (raid5_calc_degraded(conf) > 0 &&
2780             conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
2781                 schedule_work(&log->disable_writeback_work);
2782 }
2783
2784 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
2785 {
2786         struct request_queue *q = bdev_get_queue(rdev->bdev);
2787         struct r5l_log *log;
2788
2789         if (PAGE_SIZE != 4096)
2790                 return -EINVAL;
2791
2792         /*
2793          * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
2794          * raid_disks r5l_payload_data_parity.
2795          *
2796          * Write journal and cache does not work for very big array
2797          * (raid_disks > 203)
2798          */
2799         if (sizeof(struct r5l_meta_block) +
2800             ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
2801              conf->raid_disks) > PAGE_SIZE) {
2802                 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
2803                        mdname(conf->mddev), conf->raid_disks);
2804                 return -EINVAL;
2805         }
2806
2807         log = kzalloc(sizeof(*log), GFP_KERNEL);
2808         if (!log)
2809                 return -ENOMEM;
2810         log->rdev = rdev;
2811
2812         log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
2813
2814         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
2815                                        sizeof(rdev->mddev->uuid));
2816
2817         mutex_init(&log->io_mutex);
2818
2819         spin_lock_init(&log->io_list_lock);
2820         INIT_LIST_HEAD(&log->running_ios);
2821         INIT_LIST_HEAD(&log->io_end_ios);
2822         INIT_LIST_HEAD(&log->flushing_ios);
2823         INIT_LIST_HEAD(&log->finished_ios);
2824         bio_init(&log->flush_bio, NULL, 0);
2825
2826         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
2827         if (!log->io_kc)
2828                 goto io_kc;
2829
2830         log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
2831         if (!log->io_pool)
2832                 goto io_pool;
2833
2834         log->bs = bioset_create(R5L_POOL_SIZE, 0);
2835         if (!log->bs)
2836                 goto io_bs;
2837
2838         log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
2839         if (!log->meta_pool)
2840                 goto out_mempool;
2841
2842         spin_lock_init(&log->tree_lock);
2843         INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
2844
2845         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
2846                                                  log->rdev->mddev, "reclaim");
2847         if (!log->reclaim_thread)
2848                 goto reclaim_thread;
2849         log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
2850
2851         init_waitqueue_head(&log->iounit_wait);
2852
2853         INIT_LIST_HEAD(&log->no_mem_stripes);
2854
2855         INIT_LIST_HEAD(&log->no_space_stripes);
2856         spin_lock_init(&log->no_space_stripes_lock);
2857
2858         INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
2859         INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
2860
2861         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2862         INIT_LIST_HEAD(&log->stripe_in_journal_list);
2863         spin_lock_init(&log->stripe_in_journal_lock);
2864         atomic_set(&log->stripe_in_journal_count, 0);
2865
2866         rcu_assign_pointer(conf->log, log);
2867
2868         if (r5l_load_log(log))
2869                 goto error;
2870
2871         set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
2872         return 0;
2873
2874 error:
2875         rcu_assign_pointer(conf->log, NULL);
2876         md_unregister_thread(&log->reclaim_thread);
2877 reclaim_thread:
2878         mempool_destroy(log->meta_pool);
2879 out_mempool:
2880         bioset_free(log->bs);
2881 io_bs:
2882         mempool_destroy(log->io_pool);
2883 io_pool:
2884         kmem_cache_destroy(log->io_kc);
2885 io_kc:
2886         kfree(log);
2887         return -EINVAL;
2888 }
2889
2890 void r5l_exit_log(struct r5l_log *log)
2891 {
2892         flush_work(&log->disable_writeback_work);
2893         md_unregister_thread(&log->reclaim_thread);
2894         mempool_destroy(log->meta_pool);
2895         bioset_free(log->bs);
2896         mempool_destroy(log->io_pool);
2897         kmem_cache_destroy(log->io_kc);
2898         kfree(log);
2899 }