2 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
3 * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
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.
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
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>
27 #include "raid5-log.h"
30 * metadata/data stored in disk with 4k size unit (a block) regardless
31 * underneath hardware sector size. only works with PAGE_SIZE == 4096
33 #define BLOCK_SECTORS (8)
34 #define BLOCK_SECTOR_SHIFT (3)
37 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
39 * In write through mode, the reclaim runs every log->max_free_space.
40 * This can prevent the recovery scans for too long
42 #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
43 #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
45 /* wake up reclaim thread periodically */
46 #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
47 /* start flush with these full stripes */
48 #define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
49 /* reclaim stripes in groups */
50 #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
53 * We only need 2 bios per I/O unit to make progress, but ensure we
54 * have a few more available to not get too tight.
56 #define R5L_POOL_SIZE 4
58 static char *r5c_journal_mode_str[] = {"write-through",
61 * raid5 cache state machine
63 * With the RAID cache, each stripe works in two phases:
67 * These two phases are controlled by bit STRIPE_R5C_CACHING:
68 * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
69 * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
71 * When there is no journal, or the journal is in write-through mode,
72 * the stripe is always in writing-out phase.
74 * For write-back journal, the stripe is sent to caching phase on write
75 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
76 * the write-out phase by clearing STRIPE_R5C_CACHING.
78 * Stripes in caching phase do not write the raid disks. Instead, all
79 * writes are committed from the log device. Therefore, a stripe in
80 * caching phase handles writes as:
81 * - write to log device
84 * Stripes in writing-out phase handle writes as:
86 * - write pending data and parity to journal
87 * - write data and parity to raid disks
88 * - return IO for pending writes
96 sector_t device_size; /* log device size, round to
98 sector_t max_free_space; /* reclaim run if free space is at
101 sector_t last_checkpoint; /* log tail. where recovery scan
103 u64 last_cp_seq; /* log tail sequence */
105 sector_t log_start; /* log head. where new data appends */
106 u64 seq; /* log head sequence */
108 sector_t next_checkpoint;
110 struct mutex io_mutex;
111 struct r5l_io_unit *current_io; /* current io_unit accepting new data */
113 spinlock_t io_list_lock;
114 struct list_head running_ios; /* io_units which are still running,
115 * and have not yet been completely
116 * written to the log */
117 struct list_head io_end_ios; /* io_units which have been completely
118 * written to the log but not yet written
120 struct list_head flushing_ios; /* io_units which are waiting for log
122 struct list_head finished_ios; /* io_units which settle down in log disk */
123 struct bio flush_bio;
125 struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */
127 struct kmem_cache *io_kc;
130 mempool_t *meta_pool;
132 struct md_thread *reclaim_thread;
133 unsigned long reclaim_target; /* number of space that need to be
134 * reclaimed. if it's 0, reclaim spaces
135 * used by io_units which are in
136 * IO_UNIT_STRIPE_END state (eg, reclaim
137 * dones't wait for specific io_unit
138 * switching to IO_UNIT_STRIPE_END
140 wait_queue_head_t iounit_wait;
142 struct list_head no_space_stripes; /* pending stripes, log has no space */
143 spinlock_t no_space_stripes_lock;
145 bool need_cache_flush;
148 enum r5c_journal_mode r5c_journal_mode;
150 /* all stripes in r5cache, in the order of seq at sh->log_start */
151 struct list_head stripe_in_journal_list;
153 spinlock_t stripe_in_journal_lock;
154 atomic_t stripe_in_journal_count;
156 /* to submit async io_units, to fulfill ordering of flush */
157 struct work_struct deferred_io_work;
158 /* to disable write back during in degraded mode */
159 struct work_struct disable_writeback_work;
161 /* to for chunk_aligned_read in writeback mode, details below */
162 spinlock_t tree_lock;
163 struct radix_tree_root big_stripe_tree;
167 * Enable chunk_aligned_read() with write back cache.
169 * Each chunk may contain more than one stripe (for example, a 256kB
170 * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
171 * chunk_aligned_read, these stripes are grouped into one "big_stripe".
172 * For each big_stripe, we count how many stripes of this big_stripe
173 * are in the write back cache. These data are tracked in a radix tree
174 * (big_stripe_tree). We use radix_tree item pointer as the counter.
175 * r5c_tree_index() is used to calculate keys for the radix tree.
177 * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
178 * big_stripe of each chunk in the tree. If this big_stripe is in the
179 * tree, chunk_aligned_read() aborts. This look up is protected by
182 * It is necessary to remember whether a stripe is counted in
183 * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
184 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
185 * two flags are set, the stripe is counted in big_stripe_tree. This
186 * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
187 * r5c_try_caching_write(); and moving clear_bit of
188 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
189 * r5c_finish_stripe_write_out().
193 * radix tree requests lowest 2 bits of data pointer to be 2b'00.
194 * So it is necessary to left shift the counter by 2 bits before using it
195 * as data pointer of the tree.
197 #define R5C_RADIX_COUNT_SHIFT 2
200 * calculate key for big_stripe_tree
202 * sect: align_bi->bi_iter.bi_sector or sh->sector
204 static inline sector_t r5c_tree_index(struct r5conf *conf,
209 offset = sector_div(sect, conf->chunk_sectors);
214 * an IO range starts from a meta data block and end at the next meta data
215 * block. The io unit's the meta data block tracks data/parity followed it. io
216 * unit is written to log disk with normal write, as we always flush log disk
217 * first and then start move data to raid disks, there is no requirement to
218 * write io unit with FLUSH/FUA
223 struct page *meta_page; /* store meta block */
224 int meta_offset; /* current offset in meta_page */
226 struct bio *current_bio;/* current_bio accepting new data */
228 atomic_t pending_stripe;/* how many stripes not flushed to raid */
229 u64 seq; /* seq number of the metablock */
230 sector_t log_start; /* where the io_unit starts */
231 sector_t log_end; /* where the io_unit ends */
232 struct list_head log_sibling; /* log->running_ios */
233 struct list_head stripe_list; /* stripes added to the io_unit */
237 struct bio *split_bio;
239 unsigned int has_flush:1; /* include flush request */
240 unsigned int has_fua:1; /* include fua request */
241 unsigned int has_null_flush:1; /* include null flush request */
242 unsigned int has_flush_payload:1; /* include flush payload */
244 * io isn't sent yet, flush/fua request can only be submitted till it's
245 * the first IO in running_ios list
247 unsigned int io_deferred:1;
249 struct bio_list flush_barriers; /* size == 0 flush bios */
252 /* r5l_io_unit state */
253 enum r5l_io_unit_state {
254 IO_UNIT_RUNNING = 0, /* accepting new IO */
255 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
256 * don't accepting new bio */
257 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
258 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
261 bool r5c_is_writeback(struct r5l_log *log)
263 return (log != NULL &&
264 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
267 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
270 if (start >= log->device_size)
271 start = start - log->device_size;
275 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
281 return end + log->device_size - start;
284 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
288 used_size = r5l_ring_distance(log, log->last_checkpoint,
291 return log->device_size > used_size + size;
294 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
295 enum r5l_io_unit_state state)
297 if (WARN_ON(io->state >= state))
303 r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
305 struct bio *wbi, *wbi2;
309 while (wbi && wbi->bi_iter.bi_sector <
310 dev->sector + STRIPE_SECTORS) {
311 wbi2 = r5_next_bio(wbi, dev->sector);
312 md_write_end(conf->mddev);
318 void r5c_handle_cached_data_endio(struct r5conf *conf,
319 struct stripe_head *sh, int disks)
323 for (i = sh->disks; i--; ) {
324 if (sh->dev[i].written) {
325 set_bit(R5_UPTODATE, &sh->dev[i].flags);
326 r5c_return_dev_pending_writes(conf, &sh->dev[i]);
327 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
329 !test_bit(STRIPE_DEGRADED, &sh->state),
335 void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
337 /* Check whether we should flush some stripes to free up stripe cache */
338 void r5c_check_stripe_cache_usage(struct r5conf *conf)
342 if (!r5c_is_writeback(conf->log))
345 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
346 atomic_read(&conf->r5c_cached_full_stripes);
349 * The following condition is true for either of the following:
350 * - stripe cache pressure high:
351 * total_cached > 3/4 min_nr_stripes ||
352 * empty_inactive_list_nr > 0
353 * - stripe cache pressure moderate:
354 * total_cached > 1/2 min_nr_stripes
356 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
357 atomic_read(&conf->empty_inactive_list_nr) > 0)
358 r5l_wake_reclaim(conf->log, 0);
362 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
363 * stripes in the cache
365 void r5c_check_cached_full_stripe(struct r5conf *conf)
367 if (!r5c_is_writeback(conf->log))
371 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
372 * or a full stripe (chunk size / 4k stripes).
374 if (atomic_read(&conf->r5c_cached_full_stripes) >=
375 min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
376 conf->chunk_sectors >> STRIPE_SHIFT))
377 r5l_wake_reclaim(conf->log, 0);
381 * Total log space (in sectors) needed to flush all data in cache
383 * To avoid deadlock due to log space, it is necessary to reserve log
384 * space to flush critical stripes (stripes that occupying log space near
385 * last_checkpoint). This function helps check how much log space is
386 * required to flush all cached stripes.
388 * To reduce log space requirements, two mechanisms are used to give cache
389 * flush higher priorities:
390 * 1. In handle_stripe_dirtying() and schedule_reconstruction(),
391 * stripes ALREADY in journal can be flushed w/o pending writes;
392 * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
393 * can be delayed (r5l_add_no_space_stripe).
395 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
396 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
397 * pages of journal space. For stripes that has not passed 1, flushing it
398 * requires (conf->raid_disks + 1) pages of journal space. There are at
399 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
400 * required to flush all cached stripes (in pages) is:
402 * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
403 * (group_cnt + 1) * (raid_disks + 1)
405 * (stripe_in_journal_count) * (max_degraded + 1) +
406 * (group_cnt + 1) * (raid_disks - max_degraded)
408 static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
410 struct r5l_log *log = conf->log;
412 if (!r5c_is_writeback(log))
415 return BLOCK_SECTORS *
416 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
417 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
421 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
423 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
424 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
425 * device is less than 2x of reclaim_required_space.
427 static inline void r5c_update_log_state(struct r5l_log *log)
429 struct r5conf *conf = log->rdev->mddev->private;
431 sector_t reclaim_space;
432 bool wake_reclaim = false;
434 if (!r5c_is_writeback(log))
437 free_space = r5l_ring_distance(log, log->log_start,
438 log->last_checkpoint);
439 reclaim_space = r5c_log_required_to_flush_cache(conf);
440 if (free_space < 2 * reclaim_space)
441 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
443 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
445 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
447 if (free_space < 3 * reclaim_space)
448 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
450 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
453 r5l_wake_reclaim(log, 0);
457 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
458 * This function should only be called in write-back mode.
460 void r5c_make_stripe_write_out(struct stripe_head *sh)
462 struct r5conf *conf = sh->raid_conf;
463 struct r5l_log *log = conf->log;
465 BUG_ON(!r5c_is_writeback(log));
467 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
468 clear_bit(STRIPE_R5C_CACHING, &sh->state);
470 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
471 atomic_inc(&conf->preread_active_stripes);
474 static void r5c_handle_data_cached(struct stripe_head *sh)
478 for (i = sh->disks; i--; )
479 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
480 set_bit(R5_InJournal, &sh->dev[i].flags);
481 clear_bit(R5_LOCKED, &sh->dev[i].flags);
483 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
487 * this journal write must contain full parity,
488 * it may also contain some data pages
490 static void r5c_handle_parity_cached(struct stripe_head *sh)
494 for (i = sh->disks; i--; )
495 if (test_bit(R5_InJournal, &sh->dev[i].flags))
496 set_bit(R5_Wantwrite, &sh->dev[i].flags);
500 * Setting proper flags after writing (or flushing) data and/or parity to the
501 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
503 static void r5c_finish_cache_stripe(struct stripe_head *sh)
505 struct r5l_log *log = sh->raid_conf->log;
507 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
508 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
510 * Set R5_InJournal for parity dev[pd_idx]. This means
511 * all data AND parity in the journal. For RAID 6, it is
512 * NOT necessary to set the flag for dev[qd_idx], as the
513 * two parities are written out together.
515 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
516 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
517 r5c_handle_data_cached(sh);
519 r5c_handle_parity_cached(sh);
520 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
524 static void r5l_io_run_stripes(struct r5l_io_unit *io)
526 struct stripe_head *sh, *next;
528 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
529 list_del_init(&sh->log_list);
531 r5c_finish_cache_stripe(sh);
533 set_bit(STRIPE_HANDLE, &sh->state);
534 raid5_release_stripe(sh);
538 static void r5l_log_run_stripes(struct r5l_log *log)
540 struct r5l_io_unit *io, *next;
542 assert_spin_locked(&log->io_list_lock);
544 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
545 /* don't change list order */
546 if (io->state < IO_UNIT_IO_END)
549 list_move_tail(&io->log_sibling, &log->finished_ios);
550 r5l_io_run_stripes(io);
554 static void r5l_move_to_end_ios(struct r5l_log *log)
556 struct r5l_io_unit *io, *next;
558 assert_spin_locked(&log->io_list_lock);
560 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
561 /* don't change list order */
562 if (io->state < IO_UNIT_IO_END)
564 list_move_tail(&io->log_sibling, &log->io_end_ios);
568 static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
569 static void r5l_log_endio(struct bio *bio)
571 struct r5l_io_unit *io = bio->bi_private;
572 struct r5l_io_unit *io_deferred;
573 struct r5l_log *log = io->log;
576 bool has_flush_payload;
579 md_error(log->rdev->mddev, log->rdev);
582 mempool_free(io->meta_page, log->meta_pool);
584 spin_lock_irqsave(&log->io_list_lock, flags);
585 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
588 * if the io doesn't not have null_flush or flush payload,
589 * it is not safe to access it after releasing io_list_lock.
590 * Therefore, it is necessary to check the condition with
593 has_null_flush = io->has_null_flush;
594 has_flush_payload = io->has_flush_payload;
596 if (log->need_cache_flush && !list_empty(&io->stripe_list))
597 r5l_move_to_end_ios(log);
599 r5l_log_run_stripes(log);
600 if (!list_empty(&log->running_ios)) {
602 * FLUSH/FUA io_unit is deferred because of ordering, now we
605 io_deferred = list_first_entry(&log->running_ios,
606 struct r5l_io_unit, log_sibling);
607 if (io_deferred->io_deferred)
608 schedule_work(&log->deferred_io_work);
611 spin_unlock_irqrestore(&log->io_list_lock, flags);
613 if (log->need_cache_flush)
614 md_wakeup_thread(log->rdev->mddev->thread);
616 /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
617 if (has_null_flush) {
620 WARN_ON(bio_list_empty(&io->flush_barriers));
621 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
623 if (atomic_dec_and_test(&io->pending_stripe)) {
624 __r5l_stripe_write_finished(io);
629 /* decrease pending_stripe for flush payload */
630 if (has_flush_payload)
631 if (atomic_dec_and_test(&io->pending_stripe))
632 __r5l_stripe_write_finished(io);
635 static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
639 spin_lock_irqsave(&log->io_list_lock, flags);
640 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
641 spin_unlock_irqrestore(&log->io_list_lock, flags);
644 * In case of journal device failures, submit_bio will get error
645 * and calls endio, then active stripes will continue write
646 * process. Therefore, it is not necessary to check Faulty bit
647 * of journal device here.
649 * We can't check split_bio after current_bio is submitted. If
650 * io->split_bio is null, after current_bio is submitted, current_bio
651 * might already be completed and the io_unit is freed. We submit
652 * split_bio first to avoid the issue.
656 io->split_bio->bi_opf |= REQ_PREFLUSH;
658 io->split_bio->bi_opf |= REQ_FUA;
659 submit_bio(io->split_bio);
663 io->current_bio->bi_opf |= REQ_PREFLUSH;
665 io->current_bio->bi_opf |= REQ_FUA;
666 submit_bio(io->current_bio);
669 /* deferred io_unit will be dispatched here */
670 static void r5l_submit_io_async(struct work_struct *work)
672 struct r5l_log *log = container_of(work, struct r5l_log,
674 struct r5l_io_unit *io = NULL;
677 spin_lock_irqsave(&log->io_list_lock, flags);
678 if (!list_empty(&log->running_ios)) {
679 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
681 if (!io->io_deferred)
686 spin_unlock_irqrestore(&log->io_list_lock, flags);
688 r5l_do_submit_io(log, io);
691 static void r5c_disable_writeback_async(struct work_struct *work)
693 struct r5l_log *log = container_of(work, struct r5l_log,
694 disable_writeback_work);
695 struct mddev *mddev = log->rdev->mddev;
697 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
699 pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
702 /* wait superblock change before suspend */
703 wait_event(mddev->sb_wait,
704 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
706 mddev_suspend(mddev);
707 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
711 static void r5l_submit_current_io(struct r5l_log *log)
713 struct r5l_io_unit *io = log->current_io;
715 struct r5l_meta_block *block;
718 bool do_submit = true;
723 block = page_address(io->meta_page);
724 block->meta_size = cpu_to_le32(io->meta_offset);
725 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
726 block->checksum = cpu_to_le32(crc);
727 bio = io->current_bio;
729 log->current_io = NULL;
730 spin_lock_irqsave(&log->io_list_lock, flags);
731 if (io->has_flush || io->has_fua) {
732 if (io != list_first_entry(&log->running_ios,
733 struct r5l_io_unit, log_sibling)) {
738 spin_unlock_irqrestore(&log->io_list_lock, flags);
740 r5l_do_submit_io(log, io);
743 static struct bio *r5l_bio_alloc(struct r5l_log *log)
745 struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
747 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
748 bio->bi_bdev = log->rdev->bdev;
749 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
754 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
756 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
758 r5c_update_log_state(log);
760 * If we filled up the log device start from the beginning again,
761 * which will require a new bio.
763 * Note: for this to work properly the log size needs to me a multiple
766 if (log->log_start == 0)
767 io->need_split_bio = true;
769 io->log_end = log->log_start;
772 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
774 struct r5l_io_unit *io;
775 struct r5l_meta_block *block;
777 io = mempool_alloc(log->io_pool, GFP_ATOMIC);
780 memset(io, 0, sizeof(*io));
783 INIT_LIST_HEAD(&io->log_sibling);
784 INIT_LIST_HEAD(&io->stripe_list);
785 bio_list_init(&io->flush_barriers);
786 io->state = IO_UNIT_RUNNING;
788 io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
789 block = page_address(io->meta_page);
791 block->magic = cpu_to_le32(R5LOG_MAGIC);
792 block->version = R5LOG_VERSION;
793 block->seq = cpu_to_le64(log->seq);
794 block->position = cpu_to_le64(log->log_start);
796 io->log_start = log->log_start;
797 io->meta_offset = sizeof(struct r5l_meta_block);
798 io->seq = log->seq++;
800 io->current_bio = r5l_bio_alloc(log);
801 io->current_bio->bi_end_io = r5l_log_endio;
802 io->current_bio->bi_private = io;
803 bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
805 r5_reserve_log_entry(log, io);
807 spin_lock_irq(&log->io_list_lock);
808 list_add_tail(&io->log_sibling, &log->running_ios);
809 spin_unlock_irq(&log->io_list_lock);
814 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
816 if (log->current_io &&
817 log->current_io->meta_offset + payload_size > PAGE_SIZE)
818 r5l_submit_current_io(log);
820 if (!log->current_io) {
821 log->current_io = r5l_new_meta(log);
822 if (!log->current_io)
829 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
831 u32 checksum1, u32 checksum2,
832 bool checksum2_valid)
834 struct r5l_io_unit *io = log->current_io;
835 struct r5l_payload_data_parity *payload;
837 payload = page_address(io->meta_page) + io->meta_offset;
838 payload->header.type = cpu_to_le16(type);
839 payload->header.flags = cpu_to_le16(0);
840 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
842 payload->location = cpu_to_le64(location);
843 payload->checksum[0] = cpu_to_le32(checksum1);
845 payload->checksum[1] = cpu_to_le32(checksum2);
847 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
848 sizeof(__le32) * (1 + !!checksum2_valid);
851 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
853 struct r5l_io_unit *io = log->current_io;
855 if (io->need_split_bio) {
856 BUG_ON(io->split_bio);
857 io->split_bio = io->current_bio;
858 io->current_bio = r5l_bio_alloc(log);
859 bio_chain(io->current_bio, io->split_bio);
860 io->need_split_bio = false;
863 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
866 r5_reserve_log_entry(log, io);
869 static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
871 struct mddev *mddev = log->rdev->mddev;
872 struct r5conf *conf = mddev->private;
873 struct r5l_io_unit *io;
874 struct r5l_payload_flush *payload;
878 * payload_flush requires extra writes to the journal.
879 * To avoid handling the extra IO in quiesce, just skip
885 mutex_lock(&log->io_mutex);
886 meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
888 if (r5l_get_meta(log, meta_size)) {
889 mutex_unlock(&log->io_mutex);
893 /* current implementation is one stripe per flush payload */
894 io = log->current_io;
895 payload = page_address(io->meta_page) + io->meta_offset;
896 payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
897 payload->header.flags = cpu_to_le16(0);
898 payload->size = cpu_to_le32(sizeof(__le64));
899 payload->flush_stripes[0] = cpu_to_le64(sect);
900 io->meta_offset += meta_size;
901 /* multiple flush payloads count as one pending_stripe */
902 if (!io->has_flush_payload) {
903 io->has_flush_payload = 1;
904 atomic_inc(&io->pending_stripe);
906 mutex_unlock(&log->io_mutex);
909 static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
910 int data_pages, int parity_pages)
915 struct r5l_io_unit *io;
918 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
920 sizeof(struct r5l_payload_data_parity) +
921 sizeof(__le32) * parity_pages;
923 ret = r5l_get_meta(log, meta_size);
927 io = log->current_io;
929 if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
932 for (i = 0; i < sh->disks; i++) {
933 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
934 test_bit(R5_InJournal, &sh->dev[i].flags))
936 if (i == sh->pd_idx || i == sh->qd_idx)
938 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
939 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
942 * we need to flush journal to make sure recovery can
943 * reach the data with fua flag
947 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
948 raid5_compute_blocknr(sh, i, 0),
949 sh->dev[i].log_checksum, 0, false);
950 r5l_append_payload_page(log, sh->dev[i].page);
953 if (parity_pages == 2) {
954 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
955 sh->sector, sh->dev[sh->pd_idx].log_checksum,
956 sh->dev[sh->qd_idx].log_checksum, true);
957 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
958 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
959 } else if (parity_pages == 1) {
960 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
961 sh->sector, sh->dev[sh->pd_idx].log_checksum,
963 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
964 } else /* Just writing data, not parity, in caching phase */
965 BUG_ON(parity_pages != 0);
967 list_add_tail(&sh->log_list, &io->stripe_list);
968 atomic_inc(&io->pending_stripe);
971 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
974 if (sh->log_start == MaxSector) {
975 BUG_ON(!list_empty(&sh->r5c));
976 sh->log_start = io->log_start;
977 spin_lock_irq(&log->stripe_in_journal_lock);
978 list_add_tail(&sh->r5c,
979 &log->stripe_in_journal_list);
980 spin_unlock_irq(&log->stripe_in_journal_lock);
981 atomic_inc(&log->stripe_in_journal_count);
986 /* add stripe to no_space_stripes, and then wake up reclaim */
987 static inline void r5l_add_no_space_stripe(struct r5l_log *log,
988 struct stripe_head *sh)
990 spin_lock(&log->no_space_stripes_lock);
991 list_add_tail(&sh->log_list, &log->no_space_stripes);
992 spin_unlock(&log->no_space_stripes_lock);
996 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
997 * data from log to raid disks), so we shouldn't wait for reclaim here
999 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
1001 struct r5conf *conf = sh->raid_conf;
1002 int write_disks = 0;
1003 int data_pages, parity_pages;
1007 bool wake_reclaim = false;
1011 /* Don't support stripe batch */
1012 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1013 test_bit(STRIPE_SYNCING, &sh->state)) {
1014 /* the stripe is written to log, we start writing it to raid */
1015 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1019 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1021 for (i = 0; i < sh->disks; i++) {
1024 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1025 test_bit(R5_InJournal, &sh->dev[i].flags))
1029 /* checksum is already calculated in last run */
1030 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1032 addr = kmap_atomic(sh->dev[i].page);
1033 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1035 kunmap_atomic(addr);
1037 parity_pages = 1 + !!(sh->qd_idx >= 0);
1038 data_pages = write_disks - parity_pages;
1040 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1042 * The stripe must enter state machine again to finish the write, so
1045 clear_bit(STRIPE_DELAYED, &sh->state);
1046 atomic_inc(&sh->count);
1048 mutex_lock(&log->io_mutex);
1050 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1052 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1053 if (!r5l_has_free_space(log, reserve)) {
1054 r5l_add_no_space_stripe(log, sh);
1055 wake_reclaim = true;
1057 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1059 spin_lock_irq(&log->io_list_lock);
1060 list_add_tail(&sh->log_list,
1061 &log->no_mem_stripes);
1062 spin_unlock_irq(&log->io_list_lock);
1065 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1067 * log space critical, do not process stripes that are
1068 * not in cache yet (sh->log_start == MaxSector).
1070 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1071 sh->log_start == MaxSector) {
1072 r5l_add_no_space_stripe(log, sh);
1073 wake_reclaim = true;
1075 } else if (!r5l_has_free_space(log, reserve)) {
1076 if (sh->log_start == log->last_checkpoint)
1079 r5l_add_no_space_stripe(log, sh);
1081 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1083 spin_lock_irq(&log->io_list_lock);
1084 list_add_tail(&sh->log_list,
1085 &log->no_mem_stripes);
1086 spin_unlock_irq(&log->io_list_lock);
1091 mutex_unlock(&log->io_mutex);
1093 r5l_wake_reclaim(log, reserve);
1097 void r5l_write_stripe_run(struct r5l_log *log)
1101 mutex_lock(&log->io_mutex);
1102 r5l_submit_current_io(log);
1103 mutex_unlock(&log->io_mutex);
1106 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1111 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1113 * in write through (journal only)
1114 * we flush log disk cache first, then write stripe data to
1115 * raid disks. So if bio is finished, the log disk cache is
1116 * flushed already. The recovery guarantees we can recovery
1117 * the bio from log disk, so we don't need to flush again
1119 if (bio->bi_iter.bi_size == 0) {
1123 bio->bi_opf &= ~REQ_PREFLUSH;
1125 /* write back (with cache) */
1126 if (bio->bi_iter.bi_size == 0) {
1127 mutex_lock(&log->io_mutex);
1128 r5l_get_meta(log, 0);
1129 bio_list_add(&log->current_io->flush_barriers, bio);
1130 log->current_io->has_flush = 1;
1131 log->current_io->has_null_flush = 1;
1132 atomic_inc(&log->current_io->pending_stripe);
1133 r5l_submit_current_io(log);
1134 mutex_unlock(&log->io_mutex);
1141 /* This will run after log space is reclaimed */
1142 static void r5l_run_no_space_stripes(struct r5l_log *log)
1144 struct stripe_head *sh;
1146 spin_lock(&log->no_space_stripes_lock);
1147 while (!list_empty(&log->no_space_stripes)) {
1148 sh = list_first_entry(&log->no_space_stripes,
1149 struct stripe_head, log_list);
1150 list_del_init(&sh->log_list);
1151 set_bit(STRIPE_HANDLE, &sh->state);
1152 raid5_release_stripe(sh);
1154 spin_unlock(&log->no_space_stripes_lock);
1158 * calculate new last_checkpoint
1159 * for write through mode, returns log->next_checkpoint
1160 * for write back, returns log_start of first sh in stripe_in_journal_list
1162 static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1164 struct stripe_head *sh;
1165 struct r5l_log *log = conf->log;
1167 unsigned long flags;
1169 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1170 return log->next_checkpoint;
1172 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1173 if (list_empty(&conf->log->stripe_in_journal_list)) {
1174 /* all stripes flushed */
1175 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1176 return log->next_checkpoint;
1178 sh = list_first_entry(&conf->log->stripe_in_journal_list,
1179 struct stripe_head, r5c);
1180 new_cp = sh->log_start;
1181 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1185 static sector_t r5l_reclaimable_space(struct r5l_log *log)
1187 struct r5conf *conf = log->rdev->mddev->private;
1189 return r5l_ring_distance(log, log->last_checkpoint,
1190 r5c_calculate_new_cp(conf));
1193 static void r5l_run_no_mem_stripe(struct r5l_log *log)
1195 struct stripe_head *sh;
1197 assert_spin_locked(&log->io_list_lock);
1199 if (!list_empty(&log->no_mem_stripes)) {
1200 sh = list_first_entry(&log->no_mem_stripes,
1201 struct stripe_head, log_list);
1202 list_del_init(&sh->log_list);
1203 set_bit(STRIPE_HANDLE, &sh->state);
1204 raid5_release_stripe(sh);
1208 static bool r5l_complete_finished_ios(struct r5l_log *log)
1210 struct r5l_io_unit *io, *next;
1213 assert_spin_locked(&log->io_list_lock);
1215 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1216 /* don't change list order */
1217 if (io->state < IO_UNIT_STRIPE_END)
1220 log->next_checkpoint = io->log_start;
1222 list_del(&io->log_sibling);
1223 mempool_free(io, log->io_pool);
1224 r5l_run_no_mem_stripe(log);
1232 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1234 struct r5l_log *log = io->log;
1235 struct r5conf *conf = log->rdev->mddev->private;
1236 unsigned long flags;
1238 spin_lock_irqsave(&log->io_list_lock, flags);
1239 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1241 if (!r5l_complete_finished_ios(log)) {
1242 spin_unlock_irqrestore(&log->io_list_lock, flags);
1246 if (r5l_reclaimable_space(log) > log->max_free_space ||
1247 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1248 r5l_wake_reclaim(log, 0);
1250 spin_unlock_irqrestore(&log->io_list_lock, flags);
1251 wake_up(&log->iounit_wait);
1254 void r5l_stripe_write_finished(struct stripe_head *sh)
1256 struct r5l_io_unit *io;
1261 if (io && atomic_dec_and_test(&io->pending_stripe))
1262 __r5l_stripe_write_finished(io);
1265 static void r5l_log_flush_endio(struct bio *bio)
1267 struct r5l_log *log = container_of(bio, struct r5l_log,
1269 unsigned long flags;
1270 struct r5l_io_unit *io;
1273 md_error(log->rdev->mddev, log->rdev);
1275 spin_lock_irqsave(&log->io_list_lock, flags);
1276 list_for_each_entry(io, &log->flushing_ios, log_sibling)
1277 r5l_io_run_stripes(io);
1278 list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1279 spin_unlock_irqrestore(&log->io_list_lock, flags);
1283 * Starting dispatch IO to raid.
1284 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1285 * broken meta in the middle of a log causes recovery can't find meta at the
1286 * head of log. If operations require meta at the head persistent in log, we
1287 * must make sure meta before it persistent in log too. A case is:
1289 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1290 * data/parity must be persistent in log before we do the write to raid disks.
1292 * The solution is we restrictly maintain io_unit list order. In this case, we
1293 * only write stripes of an io_unit to raid disks till the io_unit is the first
1294 * one whose data/parity is in log.
1296 void r5l_flush_stripe_to_raid(struct r5l_log *log)
1300 if (!log || !log->need_cache_flush)
1303 spin_lock_irq(&log->io_list_lock);
1304 /* flush bio is running */
1305 if (!list_empty(&log->flushing_ios)) {
1306 spin_unlock_irq(&log->io_list_lock);
1309 list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1310 do_flush = !list_empty(&log->flushing_ios);
1311 spin_unlock_irq(&log->io_list_lock);
1315 bio_reset(&log->flush_bio);
1316 log->flush_bio.bi_bdev = log->rdev->bdev;
1317 log->flush_bio.bi_end_io = r5l_log_flush_endio;
1318 log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1319 submit_bio(&log->flush_bio);
1322 static void r5l_write_super(struct r5l_log *log, sector_t cp);
1323 static void r5l_write_super_and_discard_space(struct r5l_log *log,
1326 struct block_device *bdev = log->rdev->bdev;
1327 struct mddev *mddev;
1329 r5l_write_super(log, end);
1331 if (!blk_queue_discard(bdev_get_queue(bdev)))
1334 mddev = log->rdev->mddev;
1336 * Discard could zero data, so before discard we must make sure
1337 * superblock is updated to new log tail. Updating superblock (either
1338 * directly call md_update_sb() or depend on md thread) must hold
1339 * reconfig mutex. On the other hand, raid5_quiesce is called with
1340 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1341 * for all IO finish, hence waitting for reclaim thread, while reclaim
1342 * thread is calling this function and waitting for reconfig mutex. So
1343 * there is a deadlock. We workaround this issue with a trylock.
1344 * FIXME: we could miss discard if we can't take reconfig mutex
1346 set_mask_bits(&mddev->sb_flags, 0,
1347 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1348 if (!mddev_trylock(mddev))
1350 md_update_sb(mddev, 1);
1351 mddev_unlock(mddev);
1353 /* discard IO error really doesn't matter, ignore it */
1354 if (log->last_checkpoint < end) {
1355 blkdev_issue_discard(bdev,
1356 log->last_checkpoint + log->rdev->data_offset,
1357 end - log->last_checkpoint, GFP_NOIO, 0);
1359 blkdev_issue_discard(bdev,
1360 log->last_checkpoint + log->rdev->data_offset,
1361 log->device_size - log->last_checkpoint,
1363 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1369 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1370 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1372 * must hold conf->device_lock
1374 static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1376 BUG_ON(list_empty(&sh->lru));
1377 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1378 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1381 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1382 * raid5_release_stripe() while holding conf->device_lock
1384 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1385 assert_spin_locked(&conf->device_lock);
1387 list_del_init(&sh->lru);
1388 atomic_inc(&sh->count);
1390 set_bit(STRIPE_HANDLE, &sh->state);
1391 atomic_inc(&conf->active_stripes);
1392 r5c_make_stripe_write_out(sh);
1394 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1395 atomic_inc(&conf->r5c_flushing_partial_stripes);
1397 atomic_inc(&conf->r5c_flushing_full_stripes);
1398 raid5_release_stripe(sh);
1402 * if num == 0, flush all full stripes
1403 * if num > 0, flush all full stripes. If less than num full stripes are
1404 * flushed, flush some partial stripes until totally num stripes are
1405 * flushed or there is no more cached stripes.
1407 void r5c_flush_cache(struct r5conf *conf, int num)
1410 struct stripe_head *sh, *next;
1412 assert_spin_locked(&conf->device_lock);
1417 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1418 r5c_flush_stripe(conf, sh);
1424 list_for_each_entry_safe(sh, next,
1425 &conf->r5c_partial_stripe_list, lru) {
1426 r5c_flush_stripe(conf, sh);
1432 static void r5c_do_reclaim(struct r5conf *conf)
1434 struct r5l_log *log = conf->log;
1435 struct stripe_head *sh;
1437 unsigned long flags;
1439 int stripes_to_flush;
1440 int flushing_partial, flushing_full;
1442 if (!r5c_is_writeback(log))
1445 flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1446 flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1447 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1448 atomic_read(&conf->r5c_cached_full_stripes) -
1449 flushing_full - flushing_partial;
1451 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1452 atomic_read(&conf->empty_inactive_list_nr) > 0)
1454 * if stripe cache pressure high, flush all full stripes and
1455 * some partial stripes
1457 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1458 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1459 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1460 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1462 * if stripe cache pressure moderate, or if there is many full
1463 * stripes,flush all full stripes
1465 stripes_to_flush = 0;
1467 /* no need to flush */
1468 stripes_to_flush = -1;
1470 if (stripes_to_flush >= 0) {
1471 spin_lock_irqsave(&conf->device_lock, flags);
1472 r5c_flush_cache(conf, stripes_to_flush);
1473 spin_unlock_irqrestore(&conf->device_lock, flags);
1476 /* if log space is tight, flush stripes on stripe_in_journal_list */
1477 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1478 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1479 spin_lock(&conf->device_lock);
1480 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1482 * stripes on stripe_in_journal_list could be in any
1483 * state of the stripe_cache state machine. In this
1484 * case, we only want to flush stripe on
1485 * r5c_cached_full/partial_stripes. The following
1486 * condition makes sure the stripe is on one of the
1489 if (!list_empty(&sh->lru) &&
1490 !test_bit(STRIPE_HANDLE, &sh->state) &&
1491 atomic_read(&sh->count) == 0) {
1492 r5c_flush_stripe(conf, sh);
1493 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1497 spin_unlock(&conf->device_lock);
1498 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1501 if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1502 r5l_run_no_space_stripes(log);
1504 md_wakeup_thread(conf->mddev->thread);
1507 static void r5l_do_reclaim(struct r5l_log *log)
1509 struct r5conf *conf = log->rdev->mddev->private;
1510 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1511 sector_t reclaimable;
1512 sector_t next_checkpoint;
1515 spin_lock_irq(&log->io_list_lock);
1516 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1517 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1519 * move proper io_unit to reclaim list. We should not change the order.
1520 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1521 * shouldn't reuse space of an unreclaimable io_unit
1524 reclaimable = r5l_reclaimable_space(log);
1525 if (reclaimable >= reclaim_target ||
1526 (list_empty(&log->running_ios) &&
1527 list_empty(&log->io_end_ios) &&
1528 list_empty(&log->flushing_ios) &&
1529 list_empty(&log->finished_ios)))
1532 md_wakeup_thread(log->rdev->mddev->thread);
1533 wait_event_lock_irq(log->iounit_wait,
1534 r5l_reclaimable_space(log) > reclaimable,
1538 next_checkpoint = r5c_calculate_new_cp(conf);
1539 spin_unlock_irq(&log->io_list_lock);
1541 if (reclaimable == 0 || !write_super)
1545 * write_super will flush cache of each raid disk. We must write super
1546 * here, because the log area might be reused soon and we don't want to
1549 r5l_write_super_and_discard_space(log, next_checkpoint);
1551 mutex_lock(&log->io_mutex);
1552 log->last_checkpoint = next_checkpoint;
1553 r5c_update_log_state(log);
1554 mutex_unlock(&log->io_mutex);
1556 r5l_run_no_space_stripes(log);
1559 static void r5l_reclaim_thread(struct md_thread *thread)
1561 struct mddev *mddev = thread->mddev;
1562 struct r5conf *conf = mddev->private;
1563 struct r5l_log *log = conf->log;
1567 r5c_do_reclaim(conf);
1568 r5l_do_reclaim(log);
1571 void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1573 unsigned long target;
1574 unsigned long new = (unsigned long)space; /* overflow in theory */
1579 target = log->reclaim_target;
1582 } while (cmpxchg(&log->reclaim_target, target, new) != target);
1583 md_wakeup_thread(log->reclaim_thread);
1586 void r5l_quiesce(struct r5l_log *log, int state)
1588 struct mddev *mddev;
1589 if (!log || state == 2)
1592 kthread_unpark(log->reclaim_thread->tsk);
1593 else if (state == 1) {
1594 /* make sure r5l_write_super_and_discard_space exits */
1595 mddev = log->rdev->mddev;
1596 wake_up(&mddev->sb_wait);
1597 kthread_park(log->reclaim_thread->tsk);
1598 r5l_wake_reclaim(log, MaxSector);
1599 r5l_do_reclaim(log);
1603 bool r5l_log_disk_error(struct r5conf *conf)
1605 struct r5l_log *log;
1607 /* don't allow write if journal disk is missing */
1609 log = rcu_dereference(conf->log);
1612 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1614 ret = test_bit(Faulty, &log->rdev->flags);
1619 #define R5L_RECOVERY_PAGE_POOL_SIZE 256
1621 struct r5l_recovery_ctx {
1622 struct page *meta_page; /* current meta */
1623 sector_t meta_total_blocks; /* total size of current meta and data */
1624 sector_t pos; /* recovery position */
1625 u64 seq; /* recovery position seq */
1626 int data_parity_stripes; /* number of data_parity stripes */
1627 int data_only_stripes; /* number of data_only stripes */
1628 struct list_head cached_list;
1631 * read ahead page pool (ra_pool)
1632 * in recovery, log is read sequentially. It is not efficient to
1633 * read every page with sync_page_io(). The read ahead page pool
1634 * reads multiple pages with one IO, so further log read can
1635 * just copy data from the pool.
1637 struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1638 sector_t pool_offset; /* offset of first page in the pool */
1639 int total_pages; /* total allocated pages */
1640 int valid_pages; /* pages with valid data */
1641 struct bio *ra_bio; /* bio to do the read ahead */
1644 static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1645 struct r5l_recovery_ctx *ctx)
1649 ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs);
1653 ctx->valid_pages = 0;
1654 ctx->total_pages = 0;
1655 while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1656 page = alloc_page(GFP_KERNEL);
1660 ctx->ra_pool[ctx->total_pages] = page;
1661 ctx->total_pages += 1;
1664 if (ctx->total_pages == 0) {
1665 bio_put(ctx->ra_bio);
1669 ctx->pool_offset = 0;
1673 static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1674 struct r5l_recovery_ctx *ctx)
1678 for (i = 0; i < ctx->total_pages; ++i)
1679 put_page(ctx->ra_pool[i]);
1680 bio_put(ctx->ra_bio);
1684 * fetch ctx->valid_pages pages from offset
1685 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1686 * However, if the offset is close to the end of the journal device,
1687 * ctx->valid_pages could be smaller than ctx->total_pages
1689 static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1690 struct r5l_recovery_ctx *ctx,
1693 bio_reset(ctx->ra_bio);
1694 ctx->ra_bio->bi_bdev = log->rdev->bdev;
1695 bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1696 ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1698 ctx->valid_pages = 0;
1699 ctx->pool_offset = offset;
1701 while (ctx->valid_pages < ctx->total_pages) {
1702 bio_add_page(ctx->ra_bio,
1703 ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1704 ctx->valid_pages += 1;
1706 offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1708 if (offset == 0) /* reached end of the device */
1712 return submit_bio_wait(ctx->ra_bio);
1716 * try read a page from the read ahead page pool, if the page is not in the
1717 * pool, call r5l_recovery_fetch_ra_pool
1719 static int r5l_recovery_read_page(struct r5l_log *log,
1720 struct r5l_recovery_ctx *ctx,
1726 if (offset < ctx->pool_offset ||
1727 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1728 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1733 BUG_ON(offset < ctx->pool_offset ||
1734 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1736 memcpy(page_address(page),
1737 page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1738 BLOCK_SECTOR_SHIFT]),
1743 static int r5l_recovery_read_meta_block(struct r5l_log *log,
1744 struct r5l_recovery_ctx *ctx)
1746 struct page *page = ctx->meta_page;
1747 struct r5l_meta_block *mb;
1748 u32 crc, stored_crc;
1751 ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1755 mb = page_address(page);
1756 stored_crc = le32_to_cpu(mb->checksum);
1759 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1760 le64_to_cpu(mb->seq) != ctx->seq ||
1761 mb->version != R5LOG_VERSION ||
1762 le64_to_cpu(mb->position) != ctx->pos)
1765 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1766 if (stored_crc != crc)
1769 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1772 ctx->meta_total_blocks = BLOCK_SECTORS;
1778 r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1780 sector_t pos, u64 seq)
1782 struct r5l_meta_block *mb;
1784 mb = page_address(page);
1786 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1787 mb->version = R5LOG_VERSION;
1788 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1789 mb->seq = cpu_to_le64(seq);
1790 mb->position = cpu_to_le64(pos);
1793 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1797 struct r5l_meta_block *mb;
1799 page = alloc_page(GFP_KERNEL);
1802 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1803 mb = page_address(page);
1804 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1806 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1807 REQ_SYNC | REQ_FUA, false)) {
1816 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1817 * to mark valid (potentially not flushed) data in the journal.
1819 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1820 * so there should not be any mismatch here.
1822 static void r5l_recovery_load_data(struct r5l_log *log,
1823 struct stripe_head *sh,
1824 struct r5l_recovery_ctx *ctx,
1825 struct r5l_payload_data_parity *payload,
1826 sector_t log_offset)
1828 struct mddev *mddev = log->rdev->mddev;
1829 struct r5conf *conf = mddev->private;
1832 raid5_compute_sector(conf,
1833 le64_to_cpu(payload->location), 0,
1835 r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1836 sh->dev[dd_idx].log_checksum =
1837 le32_to_cpu(payload->checksum[0]);
1838 ctx->meta_total_blocks += BLOCK_SECTORS;
1840 set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1841 set_bit(STRIPE_R5C_CACHING, &sh->state);
1844 static void r5l_recovery_load_parity(struct r5l_log *log,
1845 struct stripe_head *sh,
1846 struct r5l_recovery_ctx *ctx,
1847 struct r5l_payload_data_parity *payload,
1848 sector_t log_offset)
1850 struct mddev *mddev = log->rdev->mddev;
1851 struct r5conf *conf = mddev->private;
1853 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1854 r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1855 sh->dev[sh->pd_idx].log_checksum =
1856 le32_to_cpu(payload->checksum[0]);
1857 set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1859 if (sh->qd_idx >= 0) {
1860 r5l_recovery_read_page(
1861 log, ctx, sh->dev[sh->qd_idx].page,
1862 r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1863 sh->dev[sh->qd_idx].log_checksum =
1864 le32_to_cpu(payload->checksum[1]);
1865 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1867 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1870 static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1875 sh->log_start = MaxSector;
1876 for (i = sh->disks; i--; )
1877 sh->dev[i].flags = 0;
1881 r5l_recovery_replay_one_stripe(struct r5conf *conf,
1882 struct stripe_head *sh,
1883 struct r5l_recovery_ctx *ctx)
1885 struct md_rdev *rdev, *rrdev;
1889 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1890 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1892 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1898 * stripes that only have parity must have been flushed
1899 * before the crash that we are now recovering from, so
1900 * there is nothing more to recovery.
1902 if (data_count == 0)
1905 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1906 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1909 /* in case device is broken */
1911 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1913 atomic_inc(&rdev->nr_pending);
1915 sync_page_io(rdev, sh->sector, PAGE_SIZE,
1916 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1918 rdev_dec_pending(rdev, rdev->mddev);
1921 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1923 atomic_inc(&rrdev->nr_pending);
1925 sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1926 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1928 rdev_dec_pending(rrdev, rrdev->mddev);
1933 ctx->data_parity_stripes++;
1935 r5l_recovery_reset_stripe(sh);
1938 static struct stripe_head *
1939 r5c_recovery_alloc_stripe(struct r5conf *conf,
1940 sector_t stripe_sect)
1942 struct stripe_head *sh;
1944 sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1946 return NULL; /* no more stripe available */
1948 r5l_recovery_reset_stripe(sh);
1953 static struct stripe_head *
1954 r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1956 struct stripe_head *sh;
1958 list_for_each_entry(sh, list, lru)
1959 if (sh->sector == sect)
1965 r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1966 struct r5l_recovery_ctx *ctx)
1968 struct stripe_head *sh, *next;
1970 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1971 r5l_recovery_reset_stripe(sh);
1972 list_del_init(&sh->lru);
1973 raid5_release_stripe(sh);
1978 r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1979 struct r5l_recovery_ctx *ctx)
1981 struct stripe_head *sh, *next;
1983 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1984 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1985 r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1986 list_del_init(&sh->lru);
1987 raid5_release_stripe(sh);
1991 /* if matches return 0; otherwise return -EINVAL */
1993 r5l_recovery_verify_data_checksum(struct r5l_log *log,
1994 struct r5l_recovery_ctx *ctx,
1996 sector_t log_offset, __le32 log_checksum)
2001 r5l_recovery_read_page(log, ctx, page, log_offset);
2002 addr = kmap_atomic(page);
2003 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
2004 kunmap_atomic(addr);
2005 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
2009 * before loading data to stripe cache, we need verify checksum for all data,
2010 * if there is mismatch for any data page, we drop all data in the mata block
2013 r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
2014 struct r5l_recovery_ctx *ctx)
2016 struct mddev *mddev = log->rdev->mddev;
2017 struct r5conf *conf = mddev->private;
2018 struct r5l_meta_block *mb = page_address(ctx->meta_page);
2019 sector_t mb_offset = sizeof(struct r5l_meta_block);
2020 sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2022 struct r5l_payload_data_parity *payload;
2023 struct r5l_payload_flush *payload_flush;
2025 page = alloc_page(GFP_KERNEL);
2029 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2030 payload = (void *)mb + mb_offset;
2031 payload_flush = (void *)mb + mb_offset;
2033 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2034 if (r5l_recovery_verify_data_checksum(
2035 log, ctx, page, log_offset,
2036 payload->checksum[0]) < 0)
2038 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2039 if (r5l_recovery_verify_data_checksum(
2040 log, ctx, page, log_offset,
2041 payload->checksum[0]) < 0)
2043 if (conf->max_degraded == 2 && /* q for RAID 6 */
2044 r5l_recovery_verify_data_checksum(
2046 r5l_ring_add(log, log_offset,
2048 payload->checksum[1]) < 0)
2050 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2051 /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2052 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2055 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2056 mb_offset += sizeof(struct r5l_payload_flush) +
2057 le32_to_cpu(payload_flush->size);
2059 /* DATA or PARITY payload */
2060 log_offset = r5l_ring_add(log, log_offset,
2061 le32_to_cpu(payload->size));
2062 mb_offset += sizeof(struct r5l_payload_data_parity) +
2064 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2078 * Analyze all data/parity pages in one meta block
2081 * -EINVAL for unknown playload type
2082 * -EAGAIN for checksum mismatch of data page
2083 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2086 r5c_recovery_analyze_meta_block(struct r5l_log *log,
2087 struct r5l_recovery_ctx *ctx,
2088 struct list_head *cached_stripe_list)
2090 struct mddev *mddev = log->rdev->mddev;
2091 struct r5conf *conf = mddev->private;
2092 struct r5l_meta_block *mb;
2093 struct r5l_payload_data_parity *payload;
2094 struct r5l_payload_flush *payload_flush;
2096 sector_t log_offset;
2097 sector_t stripe_sect;
2098 struct stripe_head *sh;
2102 * for mismatch in data blocks, we will drop all data in this mb, but
2103 * we will still read next mb for other data with FLUSH flag, as
2104 * io_unit could finish out of order.
2106 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2110 return ret; /* -ENOMEM duo to alloc_page() failed */
2112 mb = page_address(ctx->meta_page);
2113 mb_offset = sizeof(struct r5l_meta_block);
2114 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2116 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2119 payload = (void *)mb + mb_offset;
2120 payload_flush = (void *)mb + mb_offset;
2122 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2125 count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2126 for (i = 0; i < count; ++i) {
2127 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2128 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2131 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2132 r5l_recovery_reset_stripe(sh);
2133 list_del_init(&sh->lru);
2134 raid5_release_stripe(sh);
2138 mb_offset += sizeof(struct r5l_payload_flush) +
2139 le32_to_cpu(payload_flush->size);
2143 /* DATA or PARITY payload */
2144 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2145 raid5_compute_sector(
2146 conf, le64_to_cpu(payload->location), 0, &dd,
2148 : le64_to_cpu(payload->location);
2150 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2154 sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
2156 * cannot get stripe from raid5_get_active_stripe
2157 * try replay some stripes
2160 r5c_recovery_replay_stripes(
2161 cached_stripe_list, ctx);
2162 sh = r5c_recovery_alloc_stripe(
2166 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2168 conf->min_nr_stripes * 2);
2169 raid5_set_cache_size(mddev,
2170 conf->min_nr_stripes * 2);
2171 sh = r5c_recovery_alloc_stripe(conf,
2175 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2179 list_add_tail(&sh->lru, cached_stripe_list);
2182 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2183 if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2184 test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2185 r5l_recovery_replay_one_stripe(conf, sh, ctx);
2186 list_move_tail(&sh->lru, cached_stripe_list);
2188 r5l_recovery_load_data(log, sh, ctx, payload,
2190 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2191 r5l_recovery_load_parity(log, sh, ctx, payload,
2196 log_offset = r5l_ring_add(log, log_offset,
2197 le32_to_cpu(payload->size));
2199 mb_offset += sizeof(struct r5l_payload_data_parity) +
2201 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2208 * Load the stripe into cache. The stripe will be written out later by
2209 * the stripe cache state machine.
2211 static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2212 struct stripe_head *sh)
2217 for (i = sh->disks; i--; ) {
2219 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2220 set_bit(R5_InJournal, &dev->flags);
2221 set_bit(R5_UPTODATE, &dev->flags);
2227 * Scan through the log for all to-be-flushed data
2229 * For stripes with data and parity, namely Data-Parity stripe
2230 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2232 * For stripes with only data, namely Data-Only stripe
2233 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2235 * For a stripe, if we see data after parity, we should discard all previous
2236 * data and parity for this stripe, as these data are already flushed to
2239 * At the end of the scan, we return the new journal_tail, which points to
2240 * first data-only stripe on the journal device, or next invalid meta block.
2242 static int r5c_recovery_flush_log(struct r5l_log *log,
2243 struct r5l_recovery_ctx *ctx)
2245 struct stripe_head *sh;
2248 /* scan through the log */
2250 if (r5l_recovery_read_meta_block(log, ctx))
2253 ret = r5c_recovery_analyze_meta_block(log, ctx,
2256 * -EAGAIN means mismatch in data block, in this case, we still
2257 * try scan the next metablock
2259 if (ret && ret != -EAGAIN)
2260 break; /* ret == -EINVAL or -ENOMEM */
2262 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2265 if (ret == -ENOMEM) {
2266 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2270 /* replay data-parity stripes */
2271 r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2273 /* load data-only stripes to stripe cache */
2274 list_for_each_entry(sh, &ctx->cached_list, lru) {
2275 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2276 r5c_recovery_load_one_stripe(log, sh);
2277 ctx->data_only_stripes++;
2284 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2285 * log will start here. but we can't let superblock point to last valid
2286 * meta block. The log might looks like:
2287 * | meta 1| meta 2| meta 3|
2288 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2289 * superblock points to meta 1, we write a new valid meta 2n. if crash
2290 * happens again, new recovery will start from meta 1. Since meta 2n is
2291 * valid now, recovery will think meta 3 is valid, which is wrong.
2292 * The solution is we create a new meta in meta2 with its seq == meta
2293 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2294 * will not think meta 3 is a valid meta, because its seq doesn't match
2298 * Before recovery, the log looks like the following
2300 * ---------------------------------------------
2301 * | valid log | invalid log |
2302 * ---------------------------------------------
2304 * |- log->last_checkpoint
2305 * |- log->last_cp_seq
2307 * Now we scan through the log until we see invalid entry
2309 * ---------------------------------------------
2310 * | valid log | invalid log |
2311 * ---------------------------------------------
2313 * |- log->last_checkpoint |- ctx->pos
2314 * |- log->last_cp_seq |- ctx->seq
2316 * From this point, we need to increase seq number by 10 to avoid
2317 * confusing next recovery.
2319 * ---------------------------------------------
2320 * | valid log | invalid log |
2321 * ---------------------------------------------
2323 * |- log->last_checkpoint |- ctx->pos+1
2324 * |- log->last_cp_seq |- ctx->seq+10001
2326 * However, it is not safe to start the state machine yet, because data only
2327 * parities are not yet secured in RAID. To save these data only parities, we
2328 * rewrite them from seq+11.
2330 * -----------------------------------------------------------------
2331 * | valid log | data only stripes | invalid log |
2332 * -----------------------------------------------------------------
2334 * |- log->last_checkpoint |- ctx->pos+n
2335 * |- log->last_cp_seq |- ctx->seq+10000+n
2337 * If failure happens again during this process, the recovery can safe start
2338 * again from log->last_checkpoint.
2340 * Once data only stripes are rewritten to journal, we move log_tail
2342 * -----------------------------------------------------------------
2343 * | old log | data only stripes | invalid log |
2344 * -----------------------------------------------------------------
2346 * |- log->last_checkpoint |- ctx->pos+n
2347 * |- log->last_cp_seq |- ctx->seq+10000+n
2349 * Then we can safely start the state machine. If failure happens from this
2350 * point on, the recovery will start from new log->last_checkpoint.
2353 r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2354 struct r5l_recovery_ctx *ctx)
2356 struct stripe_head *sh;
2357 struct mddev *mddev = log->rdev->mddev;
2359 sector_t next_checkpoint = MaxSector;
2361 page = alloc_page(GFP_KERNEL);
2363 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2368 WARN_ON(list_empty(&ctx->cached_list));
2370 list_for_each_entry(sh, &ctx->cached_list, lru) {
2371 struct r5l_meta_block *mb;
2376 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2377 r5l_recovery_create_empty_meta_block(log, page,
2378 ctx->pos, ctx->seq);
2379 mb = page_address(page);
2380 offset = le32_to_cpu(mb->meta_size);
2381 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2383 for (i = sh->disks; i--; ) {
2384 struct r5dev *dev = &sh->dev[i];
2385 struct r5l_payload_data_parity *payload;
2388 if (test_bit(R5_InJournal, &dev->flags)) {
2389 payload = (void *)mb + offset;
2390 payload->header.type = cpu_to_le16(
2391 R5LOG_PAYLOAD_DATA);
2392 payload->size = cpu_to_le32(BLOCK_SECTORS);
2393 payload->location = cpu_to_le64(
2394 raid5_compute_blocknr(sh, i, 0));
2395 addr = kmap_atomic(dev->page);
2396 payload->checksum[0] = cpu_to_le32(
2397 crc32c_le(log->uuid_checksum, addr,
2399 kunmap_atomic(addr);
2400 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2401 dev->page, REQ_OP_WRITE, 0, false);
2402 write_pos = r5l_ring_add(log, write_pos,
2404 offset += sizeof(__le32) +
2405 sizeof(struct r5l_payload_data_parity);
2409 mb->meta_size = cpu_to_le32(offset);
2410 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2412 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2413 REQ_OP_WRITE, REQ_SYNC | REQ_FUA, false);
2414 sh->log_start = ctx->pos;
2415 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2416 atomic_inc(&log->stripe_in_journal_count);
2417 ctx->pos = write_pos;
2419 next_checkpoint = sh->log_start;
2421 log->next_checkpoint = next_checkpoint;
2426 static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2427 struct r5l_recovery_ctx *ctx)
2429 struct mddev *mddev = log->rdev->mddev;
2430 struct r5conf *conf = mddev->private;
2431 struct stripe_head *sh, *next;
2433 if (ctx->data_only_stripes == 0)
2436 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2438 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2439 r5c_make_stripe_write_out(sh);
2440 set_bit(STRIPE_HANDLE, &sh->state);
2441 list_del_init(&sh->lru);
2442 raid5_release_stripe(sh);
2445 md_wakeup_thread(conf->mddev->thread);
2446 /* reuse conf->wait_for_quiescent in recovery */
2447 wait_event(conf->wait_for_quiescent,
2448 atomic_read(&conf->active_stripes) == 0);
2450 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2453 static int r5l_recovery_log(struct r5l_log *log)
2455 struct mddev *mddev = log->rdev->mddev;
2456 struct r5l_recovery_ctx *ctx;
2460 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2464 ctx->pos = log->last_checkpoint;
2465 ctx->seq = log->last_cp_seq;
2466 INIT_LIST_HEAD(&ctx->cached_list);
2467 ctx->meta_page = alloc_page(GFP_KERNEL);
2469 if (!ctx->meta_page) {
2474 if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2479 ret = r5c_recovery_flush_log(log, ctx);
2487 if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2488 pr_debug("md/raid:%s: starting from clean shutdown\n",
2491 pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2492 mdname(mddev), ctx->data_only_stripes,
2493 ctx->data_parity_stripes);
2495 if (ctx->data_only_stripes == 0) {
2496 log->next_checkpoint = ctx->pos;
2497 r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2498 ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2499 } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2500 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2506 log->log_start = ctx->pos;
2507 log->seq = ctx->seq;
2508 log->last_checkpoint = pos;
2509 r5l_write_super(log, pos);
2511 r5c_recovery_flush_data_only_stripes(log, ctx);
2514 r5l_recovery_free_ra_pool(log, ctx);
2516 __free_page(ctx->meta_page);
2522 static void r5l_write_super(struct r5l_log *log, sector_t cp)
2524 struct mddev *mddev = log->rdev->mddev;
2526 log->rdev->journal_tail = cp;
2527 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2530 static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2532 struct r5conf *conf = mddev->private;
2538 switch (conf->log->r5c_journal_mode) {
2539 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2541 page, PAGE_SIZE, "[%s] %s\n",
2542 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2543 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2545 case R5C_JOURNAL_MODE_WRITE_BACK:
2547 page, PAGE_SIZE, "%s [%s]\n",
2548 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2549 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2558 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2560 * @mode as defined in 'enum r5c_journal_mode'.
2563 int r5c_journal_mode_set(struct mddev *mddev, int mode)
2565 struct r5conf *conf;
2568 if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2569 mode > R5C_JOURNAL_MODE_WRITE_BACK)
2572 err = mddev_lock(mddev);
2575 conf = mddev->private;
2576 if (!conf || !conf->log) {
2577 mddev_unlock(mddev);
2581 if (raid5_calc_degraded(conf) > 0 &&
2582 mode == R5C_JOURNAL_MODE_WRITE_BACK) {
2583 mddev_unlock(mddev);
2587 mddev_suspend(mddev);
2588 conf->log->r5c_journal_mode = mode;
2589 mddev_resume(mddev);
2590 mddev_unlock(mddev);
2592 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2593 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2596 EXPORT_SYMBOL(r5c_journal_mode_set);
2598 static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2599 const char *page, size_t length)
2601 int mode = ARRAY_SIZE(r5c_journal_mode_str);
2602 size_t len = length;
2607 if (page[len - 1] == '\n')
2611 if (strlen(r5c_journal_mode_str[mode]) == len &&
2612 !strncmp(page, r5c_journal_mode_str[mode], len))
2615 return r5c_journal_mode_set(mddev, mode) ?: length;
2618 struct md_sysfs_entry
2619 r5c_journal_mode = __ATTR(journal_mode, 0644,
2620 r5c_journal_mode_show, r5c_journal_mode_store);
2623 * Try handle write operation in caching phase. This function should only
2624 * be called in write-back mode.
2626 * If all outstanding writes can be handled in caching phase, returns 0
2627 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2628 * and returns -EAGAIN
2630 int r5c_try_caching_write(struct r5conf *conf,
2631 struct stripe_head *sh,
2632 struct stripe_head_state *s,
2635 struct r5l_log *log = conf->log;
2640 sector_t tree_index;
2644 BUG_ON(!r5c_is_writeback(log));
2646 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2648 * There are two different scenarios here:
2649 * 1. The stripe has some data cached, and it is sent to
2650 * write-out phase for reclaim
2651 * 2. The stripe is clean, and this is the first write
2653 * For 1, return -EAGAIN, so we continue with
2654 * handle_stripe_dirtying().
2656 * For 2, set STRIPE_R5C_CACHING and continue with caching
2660 /* case 1: anything injournal or anything in written */
2661 if (s->injournal > 0 || s->written > 0)
2664 set_bit(STRIPE_R5C_CACHING, &sh->state);
2668 * When run in degraded mode, array is set to write-through mode.
2669 * This check helps drain pending write safely in the transition to
2670 * write-through mode.
2672 * When a stripe is syncing, the write is also handled in write
2675 if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2676 r5c_make_stripe_write_out(sh);
2680 for (i = disks; i--; ) {
2682 /* if non-overwrite, use writing-out phase */
2683 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2684 !test_bit(R5_InJournal, &dev->flags)) {
2685 r5c_make_stripe_write_out(sh);
2690 /* if the stripe is not counted in big_stripe_tree, add it now */
2691 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2692 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2693 tree_index = r5c_tree_index(conf, sh->sector);
2694 spin_lock(&log->tree_lock);
2695 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2698 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2699 pslot, &log->tree_lock) >>
2700 R5C_RADIX_COUNT_SHIFT;
2701 radix_tree_replace_slot(
2702 &log->big_stripe_tree, pslot,
2703 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2706 * this radix_tree_insert can fail safely, so no
2707 * need to call radix_tree_preload()
2709 ret = radix_tree_insert(
2710 &log->big_stripe_tree, tree_index,
2711 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2713 spin_unlock(&log->tree_lock);
2714 r5c_make_stripe_write_out(sh);
2718 spin_unlock(&log->tree_lock);
2721 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2722 * counted in the radix tree
2724 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2725 atomic_inc(&conf->r5c_cached_partial_stripes);
2728 for (i = disks; i--; ) {
2731 set_bit(R5_Wantwrite, &dev->flags);
2732 set_bit(R5_Wantdrain, &dev->flags);
2733 set_bit(R5_LOCKED, &dev->flags);
2739 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2741 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2742 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2743 * r5c_handle_data_cached()
2745 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2752 * free extra pages (orig_page) we allocated for prexor
2754 void r5c_release_extra_page(struct stripe_head *sh)
2756 struct r5conf *conf = sh->raid_conf;
2758 bool using_disk_info_extra_page;
2760 using_disk_info_extra_page =
2761 sh->dev[0].orig_page == conf->disks[0].extra_page;
2763 for (i = sh->disks; i--; )
2764 if (sh->dev[i].page != sh->dev[i].orig_page) {
2765 struct page *p = sh->dev[i].orig_page;
2767 sh->dev[i].orig_page = sh->dev[i].page;
2768 clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2770 if (!using_disk_info_extra_page)
2774 if (using_disk_info_extra_page) {
2775 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2776 md_wakeup_thread(conf->mddev->thread);
2780 void r5c_use_extra_page(struct stripe_head *sh)
2782 struct r5conf *conf = sh->raid_conf;
2786 for (i = sh->disks; i--; ) {
2788 if (dev->orig_page != dev->page)
2789 put_page(dev->orig_page);
2790 dev->orig_page = conf->disks[i].extra_page;
2795 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2796 * stripe is committed to RAID disks.
2798 void r5c_finish_stripe_write_out(struct r5conf *conf,
2799 struct stripe_head *sh,
2800 struct stripe_head_state *s)
2802 struct r5l_log *log = conf->log;
2805 sector_t tree_index;
2809 if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2812 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2813 clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2815 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2818 for (i = sh->disks; i--; ) {
2819 clear_bit(R5_InJournal, &sh->dev[i].flags);
2820 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2825 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2826 * We updated R5_InJournal, so we also update s->injournal.
2830 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2831 if (atomic_dec_and_test(&conf->pending_full_writes))
2832 md_wakeup_thread(conf->mddev->thread);
2835 wake_up(&conf->wait_for_overlap);
2837 spin_lock_irq(&log->stripe_in_journal_lock);
2838 list_del_init(&sh->r5c);
2839 spin_unlock_irq(&log->stripe_in_journal_lock);
2840 sh->log_start = MaxSector;
2842 atomic_dec(&log->stripe_in_journal_count);
2843 r5c_update_log_state(log);
2845 /* stop counting this stripe in big_stripe_tree */
2846 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2847 test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2848 tree_index = r5c_tree_index(conf, sh->sector);
2849 spin_lock(&log->tree_lock);
2850 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2852 BUG_ON(pslot == NULL);
2853 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2854 pslot, &log->tree_lock) >>
2855 R5C_RADIX_COUNT_SHIFT;
2857 radix_tree_delete(&log->big_stripe_tree, tree_index);
2859 radix_tree_replace_slot(
2860 &log->big_stripe_tree, pslot,
2861 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2862 spin_unlock(&log->tree_lock);
2865 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2866 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2867 atomic_dec(&conf->r5c_flushing_partial_stripes);
2868 atomic_dec(&conf->r5c_cached_partial_stripes);
2871 if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2872 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2873 atomic_dec(&conf->r5c_flushing_full_stripes);
2874 atomic_dec(&conf->r5c_cached_full_stripes);
2877 r5l_append_flush_payload(log, sh->sector);
2878 /* stripe is flused to raid disks, we can do resync now */
2879 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2880 set_bit(STRIPE_HANDLE, &sh->state);
2883 int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2885 struct r5conf *conf = sh->raid_conf;
2893 for (i = 0; i < sh->disks; i++) {
2896 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2898 addr = kmap_atomic(sh->dev[i].page);
2899 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2901 kunmap_atomic(addr);
2904 WARN_ON(pages == 0);
2907 * The stripe must enter state machine again to call endio, so
2910 clear_bit(STRIPE_DELAYED, &sh->state);
2911 atomic_inc(&sh->count);
2913 mutex_lock(&log->io_mutex);
2915 reserve = (1 + pages) << (PAGE_SHIFT - 9);
2917 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2918 sh->log_start == MaxSector)
2919 r5l_add_no_space_stripe(log, sh);
2920 else if (!r5l_has_free_space(log, reserve)) {
2921 if (sh->log_start == log->last_checkpoint)
2924 r5l_add_no_space_stripe(log, sh);
2926 ret = r5l_log_stripe(log, sh, pages, 0);
2928 spin_lock_irq(&log->io_list_lock);
2929 list_add_tail(&sh->log_list, &log->no_mem_stripes);
2930 spin_unlock_irq(&log->io_list_lock);
2934 mutex_unlock(&log->io_mutex);
2938 /* check whether this big stripe is in write back cache. */
2939 bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2941 struct r5l_log *log = conf->log;
2942 sector_t tree_index;
2948 WARN_ON_ONCE(!rcu_read_lock_held());
2949 tree_index = r5c_tree_index(conf, sect);
2950 slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2951 return slot != NULL;
2954 static int r5l_load_log(struct r5l_log *log)
2956 struct md_rdev *rdev = log->rdev;
2958 struct r5l_meta_block *mb;
2959 sector_t cp = log->rdev->journal_tail;
2960 u32 stored_crc, expected_crc;
2961 bool create_super = false;
2964 /* Make sure it's valid */
2965 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2967 page = alloc_page(GFP_KERNEL);
2971 if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2975 mb = page_address(page);
2977 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2978 mb->version != R5LOG_VERSION) {
2979 create_super = true;
2982 stored_crc = le32_to_cpu(mb->checksum);
2984 expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2985 if (stored_crc != expected_crc) {
2986 create_super = true;
2989 if (le64_to_cpu(mb->position) != cp) {
2990 create_super = true;
2995 log->last_cp_seq = prandom_u32();
2997 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2999 * Make sure super points to correct address. Log might have
3000 * data very soon. If super hasn't correct log tail address,
3001 * recovery can't find the log
3003 r5l_write_super(log, cp);
3005 log->last_cp_seq = le64_to_cpu(mb->seq);
3007 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3008 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3009 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3010 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3011 log->last_checkpoint = cp;
3016 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3017 log->seq = log->last_cp_seq + 1;
3018 log->next_checkpoint = cp;
3020 ret = r5l_recovery_log(log);
3022 r5c_update_log_state(log);
3029 void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3031 struct r5conf *conf = mddev->private;
3032 struct r5l_log *log = conf->log;
3037 if ((raid5_calc_degraded(conf) > 0 ||
3038 test_bit(Journal, &rdev->flags)) &&
3039 conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3040 schedule_work(&log->disable_writeback_work);
3043 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3045 struct request_queue *q = bdev_get_queue(rdev->bdev);
3046 struct r5l_log *log;
3047 char b[BDEVNAME_SIZE];
3049 pr_debug("md/raid:%s: using device %s as journal\n",
3050 mdname(conf->mddev), bdevname(rdev->bdev, b));
3052 if (PAGE_SIZE != 4096)
3056 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3057 * raid_disks r5l_payload_data_parity.
3059 * Write journal and cache does not work for very big array
3060 * (raid_disks > 203)
3062 if (sizeof(struct r5l_meta_block) +
3063 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3064 conf->raid_disks) > PAGE_SIZE) {
3065 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3066 mdname(conf->mddev), conf->raid_disks);
3070 log = kzalloc(sizeof(*log), GFP_KERNEL);
3075 log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
3077 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3078 sizeof(rdev->mddev->uuid));
3080 mutex_init(&log->io_mutex);
3082 spin_lock_init(&log->io_list_lock);
3083 INIT_LIST_HEAD(&log->running_ios);
3084 INIT_LIST_HEAD(&log->io_end_ios);
3085 INIT_LIST_HEAD(&log->flushing_ios);
3086 INIT_LIST_HEAD(&log->finished_ios);
3087 bio_init(&log->flush_bio, NULL, 0);
3089 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3093 log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
3097 log->bs = bioset_create(R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3101 log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
3102 if (!log->meta_pool)
3105 spin_lock_init(&log->tree_lock);
3106 INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3108 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3109 log->rdev->mddev, "reclaim");
3110 if (!log->reclaim_thread)
3111 goto reclaim_thread;
3112 log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3114 init_waitqueue_head(&log->iounit_wait);
3116 INIT_LIST_HEAD(&log->no_mem_stripes);
3118 INIT_LIST_HEAD(&log->no_space_stripes);
3119 spin_lock_init(&log->no_space_stripes_lock);
3121 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3122 INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3124 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3125 INIT_LIST_HEAD(&log->stripe_in_journal_list);
3126 spin_lock_init(&log->stripe_in_journal_lock);
3127 atomic_set(&log->stripe_in_journal_count, 0);
3129 rcu_assign_pointer(conf->log, log);
3131 if (r5l_load_log(log))
3134 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3138 rcu_assign_pointer(conf->log, NULL);
3139 md_unregister_thread(&log->reclaim_thread);
3141 mempool_destroy(log->meta_pool);
3143 bioset_free(log->bs);
3145 mempool_destroy(log->io_pool);
3147 kmem_cache_destroy(log->io_kc);
3153 void r5l_exit_log(struct r5conf *conf)
3155 struct r5l_log *log = conf->log;
3160 flush_work(&log->disable_writeback_work);
3161 md_unregister_thread(&log->reclaim_thread);
3162 mempool_destroy(log->meta_pool);
3163 bioset_free(log->bs);
3164 mempool_destroy(log->io_pool);
3165 kmem_cache_destroy(log->io_kc);