1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
4 * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
5 */
6#include <linux/kernel.h>
7#include <linux/wait.h>
8#include <linux/blkdev.h>
9#include <linux/slab.h>
10#include <linux/raid/md_p.h>
11#include <linux/crc32c.h>
12#include <linux/random.h>
13#include <linux/kthread.h>
14#include <linux/types.h>
15#include "md.h"
16#include "raid5.h"
17#include "md-bitmap.h"
18#include "raid5-log.h"
19
20/*
21 * metadata/data stored in disk with 4k size unit (a block) regardless
22 * underneath hardware sector size. only works with PAGE_SIZE == 4096
23 */
24#define BLOCK_SECTORS (8)
25#define BLOCK_SECTOR_SHIFT (3)
26
27/*
28 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
29 *
30 * In write through mode, the reclaim runs every log->max_free_space.
31 * This can prevent the recovery scans for too long
32 */
33#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
34#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
35
36/* wake up reclaim thread periodically */
37#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
38/* start flush with these full stripes */
39#define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
40/* reclaim stripes in groups */
41#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
42
43/*
44 * We only need 2 bios per I/O unit to make progress, but ensure we
45 * have a few more available to not get too tight.
46 */
47#define R5L_POOL_SIZE 4
48
49static char *r5c_journal_mode_str[] = {"write-through",
50 "write-back"};
51/*
52 * raid5 cache state machine
53 *
54 * With the RAID cache, each stripe works in two phases:
55 * - caching phase
56 * - writing-out phase
57 *
58 * These two phases are controlled by bit STRIPE_R5C_CACHING:
59 * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
60 * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
61 *
62 * When there is no journal, or the journal is in write-through mode,
63 * the stripe is always in writing-out phase.
64 *
65 * For write-back journal, the stripe is sent to caching phase on write
66 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
67 * the write-out phase by clearing STRIPE_R5C_CACHING.
68 *
69 * Stripes in caching phase do not write the raid disks. Instead, all
70 * writes are committed from the log device. Therefore, a stripe in
71 * caching phase handles writes as:
72 * - write to log device
73 * - return IO
74 *
75 * Stripes in writing-out phase handle writes as:
76 * - calculate parity
77 * - write pending data and parity to journal
78 * - write data and parity to raid disks
79 * - return IO for pending writes
80 */
81
82struct r5l_log {
83 struct md_rdev *rdev;
84
85 u32 uuid_checksum;
86
87 sector_t device_size; /* log device size, round to
88 * BLOCK_SECTORS */
89 sector_t max_free_space; /* reclaim run if free space is at
90 * this size */
91
92 sector_t last_checkpoint; /* log tail. where recovery scan
93 * starts from */
94 u64 last_cp_seq; /* log tail sequence */
95
96 sector_t log_start; /* log head. where new data appends */
97 u64 seq; /* log head sequence */
98
99 sector_t next_checkpoint;
100
101 struct mutex io_mutex;
102 struct r5l_io_unit *current_io; /* current io_unit accepting new data */
103
104 spinlock_t io_list_lock;
105 struct list_head running_ios; /* io_units which are still running,
106 * and have not yet been completely
107 * written to the log */
108 struct list_head io_end_ios; /* io_units which have been completely
109 * written to the log but not yet written
110 * to the RAID */
111 struct list_head flushing_ios; /* io_units which are waiting for log
112 * cache flush */
113 struct list_head finished_ios; /* io_units which settle down in log disk */
114 struct bio flush_bio;
115
116 struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */
117
118 struct kmem_cache *io_kc;
119 mempool_t io_pool;
120 struct bio_set bs;
121 mempool_t meta_pool;
122
123 struct md_thread __rcu *reclaim_thread;
124 unsigned long reclaim_target; /* number of space that need to be
125 * reclaimed. if it's 0, reclaim spaces
126 * used by io_units which are in
127 * IO_UNIT_STRIPE_END state (eg, reclaim
128 * doesn't wait for specific io_unit
129 * switching to IO_UNIT_STRIPE_END
130 * state) */
131 wait_queue_head_t iounit_wait;
132
133 struct list_head no_space_stripes; /* pending stripes, log has no space */
134 spinlock_t no_space_stripes_lock;
135
136 bool need_cache_flush;
137
138 /* for r5c_cache */
139 enum r5c_journal_mode r5c_journal_mode;
140
141 /* all stripes in r5cache, in the order of seq at sh->log_start */
142 struct list_head stripe_in_journal_list;
143
144 spinlock_t stripe_in_journal_lock;
145 atomic_t stripe_in_journal_count;
146
147 /* to submit async io_units, to fulfill ordering of flush */
148 struct work_struct deferred_io_work;
149 /* to disable write back during in degraded mode */
150 struct work_struct disable_writeback_work;
151
152 /* to for chunk_aligned_read in writeback mode, details below */
153 spinlock_t tree_lock;
154 struct radix_tree_root big_stripe_tree;
155};
156
157/*
158 * Enable chunk_aligned_read() with write back cache.
159 *
160 * Each chunk may contain more than one stripe (for example, a 256kB
161 * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
162 * chunk_aligned_read, these stripes are grouped into one "big_stripe".
163 * For each big_stripe, we count how many stripes of this big_stripe
164 * are in the write back cache. These data are tracked in a radix tree
165 * (big_stripe_tree). We use radix_tree item pointer as the counter.
166 * r5c_tree_index() is used to calculate keys for the radix tree.
167 *
168 * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
169 * big_stripe of each chunk in the tree. If this big_stripe is in the
170 * tree, chunk_aligned_read() aborts. This look up is protected by
171 * rcu_read_lock().
172 *
173 * It is necessary to remember whether a stripe is counted in
174 * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
175 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
176 * two flags are set, the stripe is counted in big_stripe_tree. This
177 * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
178 * r5c_try_caching_write(); and moving clear_bit of
179 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
180 * r5c_finish_stripe_write_out().
181 */
182
183/*
184 * radix tree requests lowest 2 bits of data pointer to be 2b'00.
185 * So it is necessary to left shift the counter by 2 bits before using it
186 * as data pointer of the tree.
187 */
188#define R5C_RADIX_COUNT_SHIFT 2
189
190/*
191 * calculate key for big_stripe_tree
192 *
193 * sect: align_bi->bi_iter.bi_sector or sh->sector
194 */
195static inline sector_t r5c_tree_index(struct r5conf *conf,
196 sector_t sect)
197{
198 sector_div(sect, conf->chunk_sectors);
199 return sect;
200}
201
202/*
203 * an IO range starts from a meta data block and end at the next meta data
204 * block. The io unit's the meta data block tracks data/parity followed it. io
205 * unit is written to log disk with normal write, as we always flush log disk
206 * first and then start move data to raid disks, there is no requirement to
207 * write io unit with FLUSH/FUA
208 */
209struct r5l_io_unit {
210 struct r5l_log *log;
211
212 struct page *meta_page; /* store meta block */
213 int meta_offset; /* current offset in meta_page */
214
215 struct bio *current_bio;/* current_bio accepting new data */
216
217 atomic_t pending_stripe;/* how many stripes not flushed to raid */
218 u64 seq; /* seq number of the metablock */
219 sector_t log_start; /* where the io_unit starts */
220 sector_t log_end; /* where the io_unit ends */
221 struct list_head log_sibling; /* log->running_ios */
222 struct list_head stripe_list; /* stripes added to the io_unit */
223
224 int state;
225 bool need_split_bio;
226 struct bio *split_bio;
227
228 unsigned int has_flush:1; /* include flush request */
229 unsigned int has_fua:1; /* include fua request */
230 unsigned int has_null_flush:1; /* include null flush request */
231 unsigned int has_flush_payload:1; /* include flush payload */
232 /*
233 * io isn't sent yet, flush/fua request can only be submitted till it's
234 * the first IO in running_ios list
235 */
236 unsigned int io_deferred:1;
237
238 struct bio_list flush_barriers; /* size == 0 flush bios */
239};
240
241/* r5l_io_unit state */
242enum r5l_io_unit_state {
243 IO_UNIT_RUNNING = 0, /* accepting new IO */
244 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
245 * don't accepting new bio */
246 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
247 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
248};
249
250bool r5c_is_writeback(struct r5l_log *log)
251{
252 return (log != NULL &&
253 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
254}
255
256static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
257{
258 start += inc;
259 if (start >= log->device_size)
260 start = start - log->device_size;
261 return start;
262}
263
264static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
265 sector_t end)
266{
267 if (end >= start)
268 return end - start;
269 else
270 return end + log->device_size - start;
271}
272
273static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
274{
275 sector_t used_size;
276
277 used_size = r5l_ring_distance(log, start: log->last_checkpoint,
278 end: log->log_start);
279
280 return log->device_size > used_size + size;
281}
282
283static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
284 enum r5l_io_unit_state state)
285{
286 if (WARN_ON(io->state >= state))
287 return;
288 io->state = state;
289}
290
291static void
292r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
293{
294 struct bio *wbi, *wbi2;
295
296 wbi = dev->written;
297 dev->written = NULL;
298 while (wbi && wbi->bi_iter.bi_sector <
299 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
300 wbi2 = r5_next_bio(conf, bio: wbi, sector: dev->sector);
301 md_write_end(mddev: conf->mddev);
302 bio_endio(wbi);
303 wbi = wbi2;
304 }
305}
306
307void r5c_handle_cached_data_endio(struct r5conf *conf,
308 struct stripe_head *sh, int disks)
309{
310 int i;
311
312 for (i = sh->disks; i--; ) {
313 if (sh->dev[i].written) {
314 set_bit(nr: R5_UPTODATE, addr: &sh->dev[i].flags);
315 r5c_return_dev_pending_writes(conf, dev: &sh->dev[i]);
316 md_bitmap_endwrite(bitmap: conf->mddev->bitmap, offset: sh->sector,
317 RAID5_STRIPE_SECTORS(conf),
318 success: !test_bit(STRIPE_DEGRADED, &sh->state),
319 behind: 0);
320 }
321 }
322}
323
324void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
325
326/* Check whether we should flush some stripes to free up stripe cache */
327void r5c_check_stripe_cache_usage(struct r5conf *conf)
328{
329 int total_cached;
330 struct r5l_log *log = READ_ONCE(conf->log);
331
332 if (!r5c_is_writeback(log))
333 return;
334
335 total_cached = atomic_read(v: &conf->r5c_cached_partial_stripes) +
336 atomic_read(v: &conf->r5c_cached_full_stripes);
337
338 /*
339 * The following condition is true for either of the following:
340 * - stripe cache pressure high:
341 * total_cached > 3/4 min_nr_stripes ||
342 * empty_inactive_list_nr > 0
343 * - stripe cache pressure moderate:
344 * total_cached > 1/2 min_nr_stripes
345 */
346 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
347 atomic_read(v: &conf->empty_inactive_list_nr) > 0)
348 r5l_wake_reclaim(log, space: 0);
349}
350
351/*
352 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
353 * stripes in the cache
354 */
355void r5c_check_cached_full_stripe(struct r5conf *conf)
356{
357 struct r5l_log *log = READ_ONCE(conf->log);
358
359 if (!r5c_is_writeback(log))
360 return;
361
362 /*
363 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
364 * or a full stripe (chunk size / 4k stripes).
365 */
366 if (atomic_read(v: &conf->r5c_cached_full_stripes) >=
367 min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
368 conf->chunk_sectors >> RAID5_STRIPE_SHIFT(conf)))
369 r5l_wake_reclaim(log, space: 0);
370}
371
372/*
373 * Total log space (in sectors) needed to flush all data in cache
374 *
375 * To avoid deadlock due to log space, it is necessary to reserve log
376 * space to flush critical stripes (stripes that occupying log space near
377 * last_checkpoint). This function helps check how much log space is
378 * required to flush all cached stripes.
379 *
380 * To reduce log space requirements, two mechanisms are used to give cache
381 * flush higher priorities:
382 * 1. In handle_stripe_dirtying() and schedule_reconstruction(),
383 * stripes ALREADY in journal can be flushed w/o pending writes;
384 * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
385 * can be delayed (r5l_add_no_space_stripe).
386 *
387 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
388 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
389 * pages of journal space. For stripes that has not passed 1, flushing it
390 * requires (conf->raid_disks + 1) pages of journal space. There are at
391 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
392 * required to flush all cached stripes (in pages) is:
393 *
394 * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
395 * (group_cnt + 1) * (raid_disks + 1)
396 * or
397 * (stripe_in_journal_count) * (max_degraded + 1) +
398 * (group_cnt + 1) * (raid_disks - max_degraded)
399 */
400static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
401{
402 struct r5l_log *log = READ_ONCE(conf->log);
403
404 if (!r5c_is_writeback(log))
405 return 0;
406
407 return BLOCK_SECTORS *
408 ((conf->max_degraded + 1) * atomic_read(v: &log->stripe_in_journal_count) +
409 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
410}
411
412/*
413 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
414 *
415 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
416 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
417 * device is less than 2x of reclaim_required_space.
418 */
419static inline void r5c_update_log_state(struct r5l_log *log)
420{
421 struct r5conf *conf = log->rdev->mddev->private;
422 sector_t free_space;
423 sector_t reclaim_space;
424 bool wake_reclaim = false;
425
426 if (!r5c_is_writeback(log))
427 return;
428
429 free_space = r5l_ring_distance(log, start: log->log_start,
430 end: log->last_checkpoint);
431 reclaim_space = r5c_log_required_to_flush_cache(conf);
432 if (free_space < 2 * reclaim_space)
433 set_bit(nr: R5C_LOG_CRITICAL, addr: &conf->cache_state);
434 else {
435 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
436 wake_reclaim = true;
437 clear_bit(nr: R5C_LOG_CRITICAL, addr: &conf->cache_state);
438 }
439 if (free_space < 3 * reclaim_space)
440 set_bit(nr: R5C_LOG_TIGHT, addr: &conf->cache_state);
441 else
442 clear_bit(nr: R5C_LOG_TIGHT, addr: &conf->cache_state);
443
444 if (wake_reclaim)
445 r5l_wake_reclaim(log, space: 0);
446}
447
448/*
449 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
450 * This function should only be called in write-back mode.
451 */
452void r5c_make_stripe_write_out(struct stripe_head *sh)
453{
454 struct r5conf *conf = sh->raid_conf;
455 struct r5l_log *log = READ_ONCE(conf->log);
456
457 BUG_ON(!r5c_is_writeback(log));
458
459 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
460 clear_bit(nr: STRIPE_R5C_CACHING, addr: &sh->state);
461
462 if (!test_and_set_bit(nr: STRIPE_PREREAD_ACTIVE, addr: &sh->state))
463 atomic_inc(v: &conf->preread_active_stripes);
464}
465
466static void r5c_handle_data_cached(struct stripe_head *sh)
467{
468 int i;
469
470 for (i = sh->disks; i--; )
471 if (test_and_clear_bit(nr: R5_Wantwrite, addr: &sh->dev[i].flags)) {
472 set_bit(nr: R5_InJournal, addr: &sh->dev[i].flags);
473 clear_bit(nr: R5_LOCKED, addr: &sh->dev[i].flags);
474 }
475 clear_bit(nr: STRIPE_LOG_TRAPPED, addr: &sh->state);
476}
477
478/*
479 * this journal write must contain full parity,
480 * it may also contain some data pages
481 */
482static void r5c_handle_parity_cached(struct stripe_head *sh)
483{
484 int i;
485
486 for (i = sh->disks; i--; )
487 if (test_bit(R5_InJournal, &sh->dev[i].flags))
488 set_bit(nr: R5_Wantwrite, addr: &sh->dev[i].flags);
489}
490
491/*
492 * Setting proper flags after writing (or flushing) data and/or parity to the
493 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
494 */
495static void r5c_finish_cache_stripe(struct stripe_head *sh)
496{
497 struct r5l_log *log = READ_ONCE(sh->raid_conf->log);
498
499 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
500 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
501 /*
502 * Set R5_InJournal for parity dev[pd_idx]. This means
503 * all data AND parity in the journal. For RAID 6, it is
504 * NOT necessary to set the flag for dev[qd_idx], as the
505 * two parities are written out together.
506 */
507 set_bit(nr: R5_InJournal, addr: &sh->dev[sh->pd_idx].flags);
508 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
509 r5c_handle_data_cached(sh);
510 } else {
511 r5c_handle_parity_cached(sh);
512 set_bit(nr: R5_InJournal, addr: &sh->dev[sh->pd_idx].flags);
513 }
514}
515
516static void r5l_io_run_stripes(struct r5l_io_unit *io)
517{
518 struct stripe_head *sh, *next;
519
520 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
521 list_del_init(entry: &sh->log_list);
522
523 r5c_finish_cache_stripe(sh);
524
525 set_bit(nr: STRIPE_HANDLE, addr: &sh->state);
526 raid5_release_stripe(sh);
527 }
528}
529
530static void r5l_log_run_stripes(struct r5l_log *log)
531{
532 struct r5l_io_unit *io, *next;
533
534 lockdep_assert_held(&log->io_list_lock);
535
536 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
537 /* don't change list order */
538 if (io->state < IO_UNIT_IO_END)
539 break;
540
541 list_move_tail(list: &io->log_sibling, head: &log->finished_ios);
542 r5l_io_run_stripes(io);
543 }
544}
545
546static void r5l_move_to_end_ios(struct r5l_log *log)
547{
548 struct r5l_io_unit *io, *next;
549
550 lockdep_assert_held(&log->io_list_lock);
551
552 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
553 /* don't change list order */
554 if (io->state < IO_UNIT_IO_END)
555 break;
556 list_move_tail(list: &io->log_sibling, head: &log->io_end_ios);
557 }
558}
559
560static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
561static void r5l_log_endio(struct bio *bio)
562{
563 struct r5l_io_unit *io = bio->bi_private;
564 struct r5l_io_unit *io_deferred;
565 struct r5l_log *log = io->log;
566 unsigned long flags;
567 bool has_null_flush;
568 bool has_flush_payload;
569
570 if (bio->bi_status)
571 md_error(mddev: log->rdev->mddev, rdev: log->rdev);
572
573 bio_put(bio);
574 mempool_free(element: io->meta_page, pool: &log->meta_pool);
575
576 spin_lock_irqsave(&log->io_list_lock, flags);
577 __r5l_set_io_unit_state(io, state: IO_UNIT_IO_END);
578
579 /*
580 * if the io doesn't not have null_flush or flush payload,
581 * it is not safe to access it after releasing io_list_lock.
582 * Therefore, it is necessary to check the condition with
583 * the lock held.
584 */
585 has_null_flush = io->has_null_flush;
586 has_flush_payload = io->has_flush_payload;
587
588 if (log->need_cache_flush && !list_empty(head: &io->stripe_list))
589 r5l_move_to_end_ios(log);
590 else
591 r5l_log_run_stripes(log);
592 if (!list_empty(head: &log->running_ios)) {
593 /*
594 * FLUSH/FUA io_unit is deferred because of ordering, now we
595 * can dispatch it
596 */
597 io_deferred = list_first_entry(&log->running_ios,
598 struct r5l_io_unit, log_sibling);
599 if (io_deferred->io_deferred)
600 schedule_work(work: &log->deferred_io_work);
601 }
602
603 spin_unlock_irqrestore(lock: &log->io_list_lock, flags);
604
605 if (log->need_cache_flush)
606 md_wakeup_thread(thread: log->rdev->mddev->thread);
607
608 /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
609 if (has_null_flush) {
610 struct bio *bi;
611
612 WARN_ON(bio_list_empty(&io->flush_barriers));
613 while ((bi = bio_list_pop(bl: &io->flush_barriers)) != NULL) {
614 bio_endio(bi);
615 if (atomic_dec_and_test(v: &io->pending_stripe)) {
616 __r5l_stripe_write_finished(io);
617 return;
618 }
619 }
620 }
621 /* decrease pending_stripe for flush payload */
622 if (has_flush_payload)
623 if (atomic_dec_and_test(v: &io->pending_stripe))
624 __r5l_stripe_write_finished(io);
625}
626
627static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
628{
629 unsigned long flags;
630
631 spin_lock_irqsave(&log->io_list_lock, flags);
632 __r5l_set_io_unit_state(io, state: IO_UNIT_IO_START);
633 spin_unlock_irqrestore(lock: &log->io_list_lock, flags);
634
635 /*
636 * In case of journal device failures, submit_bio will get error
637 * and calls endio, then active stripes will continue write
638 * process. Therefore, it is not necessary to check Faulty bit
639 * of journal device here.
640 *
641 * We can't check split_bio after current_bio is submitted. If
642 * io->split_bio is null, after current_bio is submitted, current_bio
643 * might already be completed and the io_unit is freed. We submit
644 * split_bio first to avoid the issue.
645 */
646 if (io->split_bio) {
647 if (io->has_flush)
648 io->split_bio->bi_opf |= REQ_PREFLUSH;
649 if (io->has_fua)
650 io->split_bio->bi_opf |= REQ_FUA;
651 submit_bio(bio: io->split_bio);
652 }
653
654 if (io->has_flush)
655 io->current_bio->bi_opf |= REQ_PREFLUSH;
656 if (io->has_fua)
657 io->current_bio->bi_opf |= REQ_FUA;
658 submit_bio(bio: io->current_bio);
659}
660
661/* deferred io_unit will be dispatched here */
662static void r5l_submit_io_async(struct work_struct *work)
663{
664 struct r5l_log *log = container_of(work, struct r5l_log,
665 deferred_io_work);
666 struct r5l_io_unit *io = NULL;
667 unsigned long flags;
668
669 spin_lock_irqsave(&log->io_list_lock, flags);
670 if (!list_empty(head: &log->running_ios)) {
671 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
672 log_sibling);
673 if (!io->io_deferred)
674 io = NULL;
675 else
676 io->io_deferred = 0;
677 }
678 spin_unlock_irqrestore(lock: &log->io_list_lock, flags);
679 if (io)
680 r5l_do_submit_io(log, io);
681}
682
683static void r5c_disable_writeback_async(struct work_struct *work)
684{
685 struct r5l_log *log = container_of(work, struct r5l_log,
686 disable_writeback_work);
687 struct mddev *mddev = log->rdev->mddev;
688 struct r5conf *conf = mddev->private;
689
690 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
691 return;
692 pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
693 mdname(mddev));
694
695 /* wait superblock change before suspend */
696 wait_event(mddev->sb_wait,
697 !READ_ONCE(conf->log) ||
698 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
699
700 log = READ_ONCE(conf->log);
701 if (log) {
702 mddev_suspend(mddev, interruptible: false);
703 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
704 mddev_resume(mddev);
705 }
706}
707
708static void r5l_submit_current_io(struct r5l_log *log)
709{
710 struct r5l_io_unit *io = log->current_io;
711 struct r5l_meta_block *block;
712 unsigned long flags;
713 u32 crc;
714 bool do_submit = true;
715
716 if (!io)
717 return;
718
719 block = page_address(io->meta_page);
720 block->meta_size = cpu_to_le32(io->meta_offset);
721 crc = crc32c_le(crc: log->uuid_checksum, address: block, PAGE_SIZE);
722 block->checksum = cpu_to_le32(crc);
723
724 log->current_io = NULL;
725 spin_lock_irqsave(&log->io_list_lock, flags);
726 if (io->has_flush || io->has_fua) {
727 if (io != list_first_entry(&log->running_ios,
728 struct r5l_io_unit, log_sibling)) {
729 io->io_deferred = 1;
730 do_submit = false;
731 }
732 }
733 spin_unlock_irqrestore(lock: &log->io_list_lock, flags);
734 if (do_submit)
735 r5l_do_submit_io(log, io);
736}
737
738static struct bio *r5l_bio_alloc(struct r5l_log *log)
739{
740 struct bio *bio = bio_alloc_bioset(bdev: log->rdev->bdev, BIO_MAX_VECS,
741 opf: REQ_OP_WRITE, GFP_NOIO, bs: &log->bs);
742
743 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
744
745 return bio;
746}
747
748static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
749{
750 log->log_start = r5l_ring_add(log, start: log->log_start, BLOCK_SECTORS);
751
752 r5c_update_log_state(log);
753 /*
754 * If we filled up the log device start from the beginning again,
755 * which will require a new bio.
756 *
757 * Note: for this to work properly the log size needs to me a multiple
758 * of BLOCK_SECTORS.
759 */
760 if (log->log_start == 0)
761 io->need_split_bio = true;
762
763 io->log_end = log->log_start;
764}
765
766static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
767{
768 struct r5l_io_unit *io;
769 struct r5l_meta_block *block;
770
771 io = mempool_alloc(pool: &log->io_pool, GFP_ATOMIC);
772 if (!io)
773 return NULL;
774 memset(io, 0, sizeof(*io));
775
776 io->log = log;
777 INIT_LIST_HEAD(list: &io->log_sibling);
778 INIT_LIST_HEAD(list: &io->stripe_list);
779 bio_list_init(bl: &io->flush_barriers);
780 io->state = IO_UNIT_RUNNING;
781
782 io->meta_page = mempool_alloc(pool: &log->meta_pool, GFP_NOIO);
783 block = page_address(io->meta_page);
784 clear_page(page: block);
785 block->magic = cpu_to_le32(R5LOG_MAGIC);
786 block->version = R5LOG_VERSION;
787 block->seq = cpu_to_le64(log->seq);
788 block->position = cpu_to_le64(log->log_start);
789
790 io->log_start = log->log_start;
791 io->meta_offset = sizeof(struct r5l_meta_block);
792 io->seq = log->seq++;
793
794 io->current_bio = r5l_bio_alloc(log);
795 io->current_bio->bi_end_io = r5l_log_endio;
796 io->current_bio->bi_private = io;
797 __bio_add_page(bio: io->current_bio, page: io->meta_page, PAGE_SIZE, off: 0);
798
799 r5_reserve_log_entry(log, io);
800
801 spin_lock_irq(lock: &log->io_list_lock);
802 list_add_tail(new: &io->log_sibling, head: &log->running_ios);
803 spin_unlock_irq(lock: &log->io_list_lock);
804
805 return io;
806}
807
808static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
809{
810 if (log->current_io &&
811 log->current_io->meta_offset + payload_size > PAGE_SIZE)
812 r5l_submit_current_io(log);
813
814 if (!log->current_io) {
815 log->current_io = r5l_new_meta(log);
816 if (!log->current_io)
817 return -ENOMEM;
818 }
819
820 return 0;
821}
822
823static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
824 sector_t location,
825 u32 checksum1, u32 checksum2,
826 bool checksum2_valid)
827{
828 struct r5l_io_unit *io = log->current_io;
829 struct r5l_payload_data_parity *payload;
830
831 payload = page_address(io->meta_page) + io->meta_offset;
832 payload->header.type = cpu_to_le16(type);
833 payload->header.flags = cpu_to_le16(0);
834 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
835 (PAGE_SHIFT - 9));
836 payload->location = cpu_to_le64(location);
837 payload->checksum[0] = cpu_to_le32(checksum1);
838 if (checksum2_valid)
839 payload->checksum[1] = cpu_to_le32(checksum2);
840
841 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
842 sizeof(__le32) * (1 + !!checksum2_valid);
843}
844
845static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
846{
847 struct r5l_io_unit *io = log->current_io;
848
849 if (io->need_split_bio) {
850 BUG_ON(io->split_bio);
851 io->split_bio = io->current_bio;
852 io->current_bio = r5l_bio_alloc(log);
853 bio_chain(io->current_bio, io->split_bio);
854 io->need_split_bio = false;
855 }
856
857 if (!bio_add_page(bio: io->current_bio, page, PAGE_SIZE, off: 0))
858 BUG();
859
860 r5_reserve_log_entry(log, io);
861}
862
863static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
864{
865 struct mddev *mddev = log->rdev->mddev;
866 struct r5conf *conf = mddev->private;
867 struct r5l_io_unit *io;
868 struct r5l_payload_flush *payload;
869 int meta_size;
870
871 /*
872 * payload_flush requires extra writes to the journal.
873 * To avoid handling the extra IO in quiesce, just skip
874 * flush_payload
875 */
876 if (conf->quiesce)
877 return;
878
879 mutex_lock(&log->io_mutex);
880 meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
881
882 if (r5l_get_meta(log, payload_size: meta_size)) {
883 mutex_unlock(lock: &log->io_mutex);
884 return;
885 }
886
887 /* current implementation is one stripe per flush payload */
888 io = log->current_io;
889 payload = page_address(io->meta_page) + io->meta_offset;
890 payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
891 payload->header.flags = cpu_to_le16(0);
892 payload->size = cpu_to_le32(sizeof(__le64));
893 payload->flush_stripes[0] = cpu_to_le64(sect);
894 io->meta_offset += meta_size;
895 /* multiple flush payloads count as one pending_stripe */
896 if (!io->has_flush_payload) {
897 io->has_flush_payload = 1;
898 atomic_inc(v: &io->pending_stripe);
899 }
900 mutex_unlock(lock: &log->io_mutex);
901}
902
903static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
904 int data_pages, int parity_pages)
905{
906 int i;
907 int meta_size;
908 int ret;
909 struct r5l_io_unit *io;
910
911 meta_size =
912 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
913 * data_pages) +
914 sizeof(struct r5l_payload_data_parity) +
915 sizeof(__le32) * parity_pages;
916
917 ret = r5l_get_meta(log, payload_size: meta_size);
918 if (ret)
919 return ret;
920
921 io = log->current_io;
922
923 if (test_and_clear_bit(nr: STRIPE_R5C_PREFLUSH, addr: &sh->state))
924 io->has_flush = 1;
925
926 for (i = 0; i < sh->disks; i++) {
927 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
928 test_bit(R5_InJournal, &sh->dev[i].flags))
929 continue;
930 if (i == sh->pd_idx || i == sh->qd_idx)
931 continue;
932 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
933 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
934 io->has_fua = 1;
935 /*
936 * we need to flush journal to make sure recovery can
937 * reach the data with fua flag
938 */
939 io->has_flush = 1;
940 }
941 r5l_append_payload_meta(log, type: R5LOG_PAYLOAD_DATA,
942 location: raid5_compute_blocknr(sh, i, previous: 0),
943 checksum1: sh->dev[i].log_checksum, checksum2: 0, checksum2_valid: false);
944 r5l_append_payload_page(log, page: sh->dev[i].page);
945 }
946
947 if (parity_pages == 2) {
948 r5l_append_payload_meta(log, type: R5LOG_PAYLOAD_PARITY,
949 location: sh->sector, checksum1: sh->dev[sh->pd_idx].log_checksum,
950 checksum2: sh->dev[sh->qd_idx].log_checksum, checksum2_valid: true);
951 r5l_append_payload_page(log, page: sh->dev[sh->pd_idx].page);
952 r5l_append_payload_page(log, page: sh->dev[sh->qd_idx].page);
953 } else if (parity_pages == 1) {
954 r5l_append_payload_meta(log, type: R5LOG_PAYLOAD_PARITY,
955 location: sh->sector, checksum1: sh->dev[sh->pd_idx].log_checksum,
956 checksum2: 0, checksum2_valid: false);
957 r5l_append_payload_page(log, page: sh->dev[sh->pd_idx].page);
958 } else /* Just writing data, not parity, in caching phase */
959 BUG_ON(parity_pages != 0);
960
961 list_add_tail(new: &sh->log_list, head: &io->stripe_list);
962 atomic_inc(v: &io->pending_stripe);
963 sh->log_io = io;
964
965 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
966 return 0;
967
968 if (sh->log_start == MaxSector) {
969 BUG_ON(!list_empty(&sh->r5c));
970 sh->log_start = io->log_start;
971 spin_lock_irq(lock: &log->stripe_in_journal_lock);
972 list_add_tail(new: &sh->r5c,
973 head: &log->stripe_in_journal_list);
974 spin_unlock_irq(lock: &log->stripe_in_journal_lock);
975 atomic_inc(v: &log->stripe_in_journal_count);
976 }
977 return 0;
978}
979
980/* add stripe to no_space_stripes, and then wake up reclaim */
981static inline void r5l_add_no_space_stripe(struct r5l_log *log,
982 struct stripe_head *sh)
983{
984 spin_lock(lock: &log->no_space_stripes_lock);
985 list_add_tail(new: &sh->log_list, head: &log->no_space_stripes);
986 spin_unlock(lock: &log->no_space_stripes_lock);
987}
988
989/*
990 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
991 * data from log to raid disks), so we shouldn't wait for reclaim here
992 */
993int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
994{
995 struct r5conf *conf = sh->raid_conf;
996 int write_disks = 0;
997 int data_pages, parity_pages;
998 int reserve;
999 int i;
1000 int ret = 0;
1001 bool wake_reclaim = false;
1002
1003 if (!log)
1004 return -EAGAIN;
1005 /* Don't support stripe batch */
1006 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1007 test_bit(STRIPE_SYNCING, &sh->state)) {
1008 /* the stripe is written to log, we start writing it to raid */
1009 clear_bit(nr: STRIPE_LOG_TRAPPED, addr: &sh->state);
1010 return -EAGAIN;
1011 }
1012
1013 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1014
1015 for (i = 0; i < sh->disks; i++) {
1016 void *addr;
1017
1018 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1019 test_bit(R5_InJournal, &sh->dev[i].flags))
1020 continue;
1021
1022 write_disks++;
1023 /* checksum is already calculated in last run */
1024 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1025 continue;
1026 addr = kmap_atomic(page: sh->dev[i].page);
1027 sh->dev[i].log_checksum = crc32c_le(crc: log->uuid_checksum,
1028 address: addr, PAGE_SIZE);
1029 kunmap_atomic(addr);
1030 }
1031 parity_pages = 1 + !!(sh->qd_idx >= 0);
1032 data_pages = write_disks - parity_pages;
1033
1034 set_bit(nr: STRIPE_LOG_TRAPPED, addr: &sh->state);
1035 /*
1036 * The stripe must enter state machine again to finish the write, so
1037 * don't delay.
1038 */
1039 clear_bit(nr: STRIPE_DELAYED, addr: &sh->state);
1040 atomic_inc(v: &sh->count);
1041
1042 mutex_lock(&log->io_mutex);
1043 /* meta + data */
1044 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1045
1046 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1047 if (!r5l_has_free_space(log, size: reserve)) {
1048 r5l_add_no_space_stripe(log, sh);
1049 wake_reclaim = true;
1050 } else {
1051 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1052 if (ret) {
1053 spin_lock_irq(lock: &log->io_list_lock);
1054 list_add_tail(new: &sh->log_list,
1055 head: &log->no_mem_stripes);
1056 spin_unlock_irq(lock: &log->io_list_lock);
1057 }
1058 }
1059 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1060 /*
1061 * log space critical, do not process stripes that are
1062 * not in cache yet (sh->log_start == MaxSector).
1063 */
1064 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1065 sh->log_start == MaxSector) {
1066 r5l_add_no_space_stripe(log, sh);
1067 wake_reclaim = true;
1068 reserve = 0;
1069 } else if (!r5l_has_free_space(log, size: reserve)) {
1070 if (sh->log_start == log->last_checkpoint)
1071 BUG();
1072 else
1073 r5l_add_no_space_stripe(log, sh);
1074 } else {
1075 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1076 if (ret) {
1077 spin_lock_irq(lock: &log->io_list_lock);
1078 list_add_tail(new: &sh->log_list,
1079 head: &log->no_mem_stripes);
1080 spin_unlock_irq(lock: &log->io_list_lock);
1081 }
1082 }
1083 }
1084
1085 mutex_unlock(lock: &log->io_mutex);
1086 if (wake_reclaim)
1087 r5l_wake_reclaim(log, space: reserve);
1088 return 0;
1089}
1090
1091void r5l_write_stripe_run(struct r5l_log *log)
1092{
1093 if (!log)
1094 return;
1095 mutex_lock(&log->io_mutex);
1096 r5l_submit_current_io(log);
1097 mutex_unlock(lock: &log->io_mutex);
1098}
1099
1100int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1101{
1102 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1103 /*
1104 * in write through (journal only)
1105 * we flush log disk cache first, then write stripe data to
1106 * raid disks. So if bio is finished, the log disk cache is
1107 * flushed already. The recovery guarantees we can recovery
1108 * the bio from log disk, so we don't need to flush again
1109 */
1110 if (bio->bi_iter.bi_size == 0) {
1111 bio_endio(bio);
1112 return 0;
1113 }
1114 bio->bi_opf &= ~REQ_PREFLUSH;
1115 } else {
1116 /* write back (with cache) */
1117 if (bio->bi_iter.bi_size == 0) {
1118 mutex_lock(&log->io_mutex);
1119 r5l_get_meta(log, payload_size: 0);
1120 bio_list_add(bl: &log->current_io->flush_barriers, bio);
1121 log->current_io->has_flush = 1;
1122 log->current_io->has_null_flush = 1;
1123 atomic_inc(v: &log->current_io->pending_stripe);
1124 r5l_submit_current_io(log);
1125 mutex_unlock(lock: &log->io_mutex);
1126 return 0;
1127 }
1128 }
1129 return -EAGAIN;
1130}
1131
1132/* This will run after log space is reclaimed */
1133static void r5l_run_no_space_stripes(struct r5l_log *log)
1134{
1135 struct stripe_head *sh;
1136
1137 spin_lock(lock: &log->no_space_stripes_lock);
1138 while (!list_empty(head: &log->no_space_stripes)) {
1139 sh = list_first_entry(&log->no_space_stripes,
1140 struct stripe_head, log_list);
1141 list_del_init(entry: &sh->log_list);
1142 set_bit(nr: STRIPE_HANDLE, addr: &sh->state);
1143 raid5_release_stripe(sh);
1144 }
1145 spin_unlock(lock: &log->no_space_stripes_lock);
1146}
1147
1148/*
1149 * calculate new last_checkpoint
1150 * for write through mode, returns log->next_checkpoint
1151 * for write back, returns log_start of first sh in stripe_in_journal_list
1152 */
1153static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1154{
1155 struct stripe_head *sh;
1156 struct r5l_log *log = READ_ONCE(conf->log);
1157 sector_t new_cp;
1158 unsigned long flags;
1159
1160 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1161 return log->next_checkpoint;
1162
1163 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1164 if (list_empty(head: &log->stripe_in_journal_list)) {
1165 /* all stripes flushed */
1166 spin_unlock_irqrestore(lock: &log->stripe_in_journal_lock, flags);
1167 return log->next_checkpoint;
1168 }
1169 sh = list_first_entry(&log->stripe_in_journal_list,
1170 struct stripe_head, r5c);
1171 new_cp = sh->log_start;
1172 spin_unlock_irqrestore(lock: &log->stripe_in_journal_lock, flags);
1173 return new_cp;
1174}
1175
1176static sector_t r5l_reclaimable_space(struct r5l_log *log)
1177{
1178 struct r5conf *conf = log->rdev->mddev->private;
1179
1180 return r5l_ring_distance(log, start: log->last_checkpoint,
1181 end: r5c_calculate_new_cp(conf));
1182}
1183
1184static void r5l_run_no_mem_stripe(struct r5l_log *log)
1185{
1186 struct stripe_head *sh;
1187
1188 lockdep_assert_held(&log->io_list_lock);
1189
1190 if (!list_empty(head: &log->no_mem_stripes)) {
1191 sh = list_first_entry(&log->no_mem_stripes,
1192 struct stripe_head, log_list);
1193 list_del_init(entry: &sh->log_list);
1194 set_bit(nr: STRIPE_HANDLE, addr: &sh->state);
1195 raid5_release_stripe(sh);
1196 }
1197}
1198
1199static bool r5l_complete_finished_ios(struct r5l_log *log)
1200{
1201 struct r5l_io_unit *io, *next;
1202 bool found = false;
1203
1204 lockdep_assert_held(&log->io_list_lock);
1205
1206 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1207 /* don't change list order */
1208 if (io->state < IO_UNIT_STRIPE_END)
1209 break;
1210
1211 log->next_checkpoint = io->log_start;
1212
1213 list_del(entry: &io->log_sibling);
1214 mempool_free(element: io, pool: &log->io_pool);
1215 r5l_run_no_mem_stripe(log);
1216
1217 found = true;
1218 }
1219
1220 return found;
1221}
1222
1223static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1224{
1225 struct r5l_log *log = io->log;
1226 struct r5conf *conf = log->rdev->mddev->private;
1227 unsigned long flags;
1228
1229 spin_lock_irqsave(&log->io_list_lock, flags);
1230 __r5l_set_io_unit_state(io, state: IO_UNIT_STRIPE_END);
1231
1232 if (!r5l_complete_finished_ios(log)) {
1233 spin_unlock_irqrestore(lock: &log->io_list_lock, flags);
1234 return;
1235 }
1236
1237 if (r5l_reclaimable_space(log) > log->max_free_space ||
1238 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1239 r5l_wake_reclaim(log, space: 0);
1240
1241 spin_unlock_irqrestore(lock: &log->io_list_lock, flags);
1242 wake_up(&log->iounit_wait);
1243}
1244
1245void r5l_stripe_write_finished(struct stripe_head *sh)
1246{
1247 struct r5l_io_unit *io;
1248
1249 io = sh->log_io;
1250 sh->log_io = NULL;
1251
1252 if (io && atomic_dec_and_test(v: &io->pending_stripe))
1253 __r5l_stripe_write_finished(io);
1254}
1255
1256static void r5l_log_flush_endio(struct bio *bio)
1257{
1258 struct r5l_log *log = container_of(bio, struct r5l_log,
1259 flush_bio);
1260 unsigned long flags;
1261 struct r5l_io_unit *io;
1262
1263 if (bio->bi_status)
1264 md_error(mddev: log->rdev->mddev, rdev: log->rdev);
1265 bio_uninit(bio);
1266
1267 spin_lock_irqsave(&log->io_list_lock, flags);
1268 list_for_each_entry(io, &log->flushing_ios, log_sibling)
1269 r5l_io_run_stripes(io);
1270 list_splice_tail_init(list: &log->flushing_ios, head: &log->finished_ios);
1271 spin_unlock_irqrestore(lock: &log->io_list_lock, flags);
1272}
1273
1274/*
1275 * Starting dispatch IO to raid.
1276 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1277 * broken meta in the middle of a log causes recovery can't find meta at the
1278 * head of log. If operations require meta at the head persistent in log, we
1279 * must make sure meta before it persistent in log too. A case is:
1280 *
1281 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1282 * data/parity must be persistent in log before we do the write to raid disks.
1283 *
1284 * The solution is we restrictly maintain io_unit list order. In this case, we
1285 * only write stripes of an io_unit to raid disks till the io_unit is the first
1286 * one whose data/parity is in log.
1287 */
1288void r5l_flush_stripe_to_raid(struct r5l_log *log)
1289{
1290 bool do_flush;
1291
1292 if (!log || !log->need_cache_flush)
1293 return;
1294
1295 spin_lock_irq(lock: &log->io_list_lock);
1296 /* flush bio is running */
1297 if (!list_empty(head: &log->flushing_ios)) {
1298 spin_unlock_irq(lock: &log->io_list_lock);
1299 return;
1300 }
1301 list_splice_tail_init(list: &log->io_end_ios, head: &log->flushing_ios);
1302 do_flush = !list_empty(head: &log->flushing_ios);
1303 spin_unlock_irq(lock: &log->io_list_lock);
1304
1305 if (!do_flush)
1306 return;
1307 bio_init(bio: &log->flush_bio, bdev: log->rdev->bdev, NULL, max_vecs: 0,
1308 opf: REQ_OP_WRITE | REQ_PREFLUSH);
1309 log->flush_bio.bi_end_io = r5l_log_flush_endio;
1310 submit_bio(bio: &log->flush_bio);
1311}
1312
1313static void r5l_write_super(struct r5l_log *log, sector_t cp);
1314static void r5l_write_super_and_discard_space(struct r5l_log *log,
1315 sector_t end)
1316{
1317 struct block_device *bdev = log->rdev->bdev;
1318 struct mddev *mddev;
1319
1320 r5l_write_super(log, cp: end);
1321
1322 if (!bdev_max_discard_sectors(bdev))
1323 return;
1324
1325 mddev = log->rdev->mddev;
1326 /*
1327 * Discard could zero data, so before discard we must make sure
1328 * superblock is updated to new log tail. Updating superblock (either
1329 * directly call md_update_sb() or depend on md thread) must hold
1330 * reconfig mutex. On the other hand, raid5_quiesce is called with
1331 * reconfig_mutex hold. The first step of raid5_quiesce() is waiting
1332 * for all IO finish, hence waiting for reclaim thread, while reclaim
1333 * thread is calling this function and waiting for reconfig mutex. So
1334 * there is a deadlock. We workaround this issue with a trylock.
1335 * FIXME: we could miss discard if we can't take reconfig mutex
1336 */
1337 set_mask_bits(&mddev->sb_flags, 0,
1338 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1339 if (!mddev_trylock(mddev))
1340 return;
1341 md_update_sb(mddev, force: 1);
1342 mddev_unlock(mddev);
1343
1344 /* discard IO error really doesn't matter, ignore it */
1345 if (log->last_checkpoint < end) {
1346 blkdev_issue_discard(bdev,
1347 sector: log->last_checkpoint + log->rdev->data_offset,
1348 nr_sects: end - log->last_checkpoint, GFP_NOIO);
1349 } else {
1350 blkdev_issue_discard(bdev,
1351 sector: log->last_checkpoint + log->rdev->data_offset,
1352 nr_sects: log->device_size - log->last_checkpoint,
1353 GFP_NOIO);
1354 blkdev_issue_discard(bdev, sector: log->rdev->data_offset, nr_sects: end,
1355 GFP_NOIO);
1356 }
1357}
1358
1359/*
1360 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1361 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1362 *
1363 * must hold conf->device_lock
1364 */
1365static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1366{
1367 BUG_ON(list_empty(&sh->lru));
1368 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1369 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1370
1371 /*
1372 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1373 * raid5_release_stripe() while holding conf->device_lock
1374 */
1375 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1376 lockdep_assert_held(&conf->device_lock);
1377
1378 list_del_init(entry: &sh->lru);
1379 atomic_inc(v: &sh->count);
1380
1381 set_bit(nr: STRIPE_HANDLE, addr: &sh->state);
1382 atomic_inc(v: &conf->active_stripes);
1383 r5c_make_stripe_write_out(sh);
1384
1385 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1386 atomic_inc(v: &conf->r5c_flushing_partial_stripes);
1387 else
1388 atomic_inc(v: &conf->r5c_flushing_full_stripes);
1389 raid5_release_stripe(sh);
1390}
1391
1392/*
1393 * if num == 0, flush all full stripes
1394 * if num > 0, flush all full stripes. If less than num full stripes are
1395 * flushed, flush some partial stripes until totally num stripes are
1396 * flushed or there is no more cached stripes.
1397 */
1398void r5c_flush_cache(struct r5conf *conf, int num)
1399{
1400 int count;
1401 struct stripe_head *sh, *next;
1402
1403 lockdep_assert_held(&conf->device_lock);
1404 if (!READ_ONCE(conf->log))
1405 return;
1406
1407 count = 0;
1408 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1409 r5c_flush_stripe(conf, sh);
1410 count++;
1411 }
1412
1413 if (count >= num)
1414 return;
1415 list_for_each_entry_safe(sh, next,
1416 &conf->r5c_partial_stripe_list, lru) {
1417 r5c_flush_stripe(conf, sh);
1418 if (++count >= num)
1419 break;
1420 }
1421}
1422
1423static void r5c_do_reclaim(struct r5conf *conf)
1424{
1425 struct r5l_log *log = READ_ONCE(conf->log);
1426 struct stripe_head *sh;
1427 int count = 0;
1428 unsigned long flags;
1429 int total_cached;
1430 int stripes_to_flush;
1431 int flushing_partial, flushing_full;
1432
1433 if (!r5c_is_writeback(log))
1434 return;
1435
1436 flushing_partial = atomic_read(v: &conf->r5c_flushing_partial_stripes);
1437 flushing_full = atomic_read(v: &conf->r5c_flushing_full_stripes);
1438 total_cached = atomic_read(v: &conf->r5c_cached_partial_stripes) +
1439 atomic_read(v: &conf->r5c_cached_full_stripes) -
1440 flushing_full - flushing_partial;
1441
1442 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1443 atomic_read(v: &conf->empty_inactive_list_nr) > 0)
1444 /*
1445 * if stripe cache pressure high, flush all full stripes and
1446 * some partial stripes
1447 */
1448 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1449 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1450 atomic_read(v: &conf->r5c_cached_full_stripes) - flushing_full >
1451 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1452 /*
1453 * if stripe cache pressure moderate, or if there is many full
1454 * stripes,flush all full stripes
1455 */
1456 stripes_to_flush = 0;
1457 else
1458 /* no need to flush */
1459 stripes_to_flush = -1;
1460
1461 if (stripes_to_flush >= 0) {
1462 spin_lock_irqsave(&conf->device_lock, flags);
1463 r5c_flush_cache(conf, num: stripes_to_flush);
1464 spin_unlock_irqrestore(lock: &conf->device_lock, flags);
1465 }
1466
1467 /* if log space is tight, flush stripes on stripe_in_journal_list */
1468 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1469 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1470 spin_lock(lock: &conf->device_lock);
1471 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1472 /*
1473 * stripes on stripe_in_journal_list could be in any
1474 * state of the stripe_cache state machine. In this
1475 * case, we only want to flush stripe on
1476 * r5c_cached_full/partial_stripes. The following
1477 * condition makes sure the stripe is on one of the
1478 * two lists.
1479 */
1480 if (!list_empty(head: &sh->lru) &&
1481 !test_bit(STRIPE_HANDLE, &sh->state) &&
1482 atomic_read(v: &sh->count) == 0) {
1483 r5c_flush_stripe(conf, sh);
1484 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1485 break;
1486 }
1487 }
1488 spin_unlock(lock: &conf->device_lock);
1489 spin_unlock_irqrestore(lock: &log->stripe_in_journal_lock, flags);
1490 }
1491
1492 if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1493 r5l_run_no_space_stripes(log);
1494
1495 md_wakeup_thread(thread: conf->mddev->thread);
1496}
1497
1498static void r5l_do_reclaim(struct r5l_log *log)
1499{
1500 struct r5conf *conf = log->rdev->mddev->private;
1501 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1502 sector_t reclaimable;
1503 sector_t next_checkpoint;
1504 bool write_super;
1505
1506 spin_lock_irq(lock: &log->io_list_lock);
1507 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1508 reclaim_target != 0 || !list_empty(head: &log->no_space_stripes);
1509 /*
1510 * move proper io_unit to reclaim list. We should not change the order.
1511 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1512 * shouldn't reuse space of an unreclaimable io_unit
1513 */
1514 while (1) {
1515 reclaimable = r5l_reclaimable_space(log);
1516 if (reclaimable >= reclaim_target ||
1517 (list_empty(head: &log->running_ios) &&
1518 list_empty(head: &log->io_end_ios) &&
1519 list_empty(head: &log->flushing_ios) &&
1520 list_empty(head: &log->finished_ios)))
1521 break;
1522
1523 md_wakeup_thread(thread: log->rdev->mddev->thread);
1524 wait_event_lock_irq(log->iounit_wait,
1525 r5l_reclaimable_space(log) > reclaimable,
1526 log->io_list_lock);
1527 }
1528
1529 next_checkpoint = r5c_calculate_new_cp(conf);
1530 spin_unlock_irq(lock: &log->io_list_lock);
1531
1532 if (reclaimable == 0 || !write_super)
1533 return;
1534
1535 /*
1536 * write_super will flush cache of each raid disk. We must write super
1537 * here, because the log area might be reused soon and we don't want to
1538 * confuse recovery
1539 */
1540 r5l_write_super_and_discard_space(log, end: next_checkpoint);
1541
1542 mutex_lock(&log->io_mutex);
1543 log->last_checkpoint = next_checkpoint;
1544 r5c_update_log_state(log);
1545 mutex_unlock(lock: &log->io_mutex);
1546
1547 r5l_run_no_space_stripes(log);
1548}
1549
1550static void r5l_reclaim_thread(struct md_thread *thread)
1551{
1552 struct mddev *mddev = thread->mddev;
1553 struct r5conf *conf = mddev->private;
1554 struct r5l_log *log = READ_ONCE(conf->log);
1555
1556 if (!log)
1557 return;
1558 r5c_do_reclaim(conf);
1559 r5l_do_reclaim(log);
1560}
1561
1562void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1563{
1564 unsigned long target;
1565 unsigned long new = (unsigned long)space; /* overflow in theory */
1566
1567 if (!log)
1568 return;
1569
1570 target = READ_ONCE(log->reclaim_target);
1571 do {
1572 if (new < target)
1573 return;
1574 } while (!try_cmpxchg(&log->reclaim_target, &target, new));
1575 md_wakeup_thread(thread: log->reclaim_thread);
1576}
1577
1578void r5l_quiesce(struct r5l_log *log, int quiesce)
1579{
1580 struct mddev *mddev = log->rdev->mddev;
1581 struct md_thread *thread = rcu_dereference_protected(
1582 log->reclaim_thread, lockdep_is_held(&mddev->reconfig_mutex));
1583
1584 if (quiesce) {
1585 /* make sure r5l_write_super_and_discard_space exits */
1586 wake_up(&mddev->sb_wait);
1587 kthread_park(k: thread->tsk);
1588 r5l_wake_reclaim(log, MaxSector);
1589 r5l_do_reclaim(log);
1590 } else
1591 kthread_unpark(k: thread->tsk);
1592}
1593
1594bool r5l_log_disk_error(struct r5conf *conf)
1595{
1596 struct r5l_log *log = READ_ONCE(conf->log);
1597
1598 /* don't allow write if journal disk is missing */
1599 if (!log)
1600 return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1601 else
1602 return test_bit(Faulty, &log->rdev->flags);
1603}
1604
1605#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1606
1607struct r5l_recovery_ctx {
1608 struct page *meta_page; /* current meta */
1609 sector_t meta_total_blocks; /* total size of current meta and data */
1610 sector_t pos; /* recovery position */
1611 u64 seq; /* recovery position seq */
1612 int data_parity_stripes; /* number of data_parity stripes */
1613 int data_only_stripes; /* number of data_only stripes */
1614 struct list_head cached_list;
1615
1616 /*
1617 * read ahead page pool (ra_pool)
1618 * in recovery, log is read sequentially. It is not efficient to
1619 * read every page with sync_page_io(). The read ahead page pool
1620 * reads multiple pages with one IO, so further log read can
1621 * just copy data from the pool.
1622 */
1623 struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1624 struct bio_vec ra_bvec[R5L_RECOVERY_PAGE_POOL_SIZE];
1625 sector_t pool_offset; /* offset of first page in the pool */
1626 int total_pages; /* total allocated pages */
1627 int valid_pages; /* pages with valid data */
1628};
1629
1630static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1631 struct r5l_recovery_ctx *ctx)
1632{
1633 struct page *page;
1634
1635 ctx->valid_pages = 0;
1636 ctx->total_pages = 0;
1637 while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1638 page = alloc_page(GFP_KERNEL);
1639
1640 if (!page)
1641 break;
1642 ctx->ra_pool[ctx->total_pages] = page;
1643 ctx->total_pages += 1;
1644 }
1645
1646 if (ctx->total_pages == 0)
1647 return -ENOMEM;
1648
1649 ctx->pool_offset = 0;
1650 return 0;
1651}
1652
1653static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1654 struct r5l_recovery_ctx *ctx)
1655{
1656 int i;
1657
1658 for (i = 0; i < ctx->total_pages; ++i)
1659 put_page(page: ctx->ra_pool[i]);
1660}
1661
1662/*
1663 * fetch ctx->valid_pages pages from offset
1664 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1665 * However, if the offset is close to the end of the journal device,
1666 * ctx->valid_pages could be smaller than ctx->total_pages
1667 */
1668static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1669 struct r5l_recovery_ctx *ctx,
1670 sector_t offset)
1671{
1672 struct bio bio;
1673 int ret;
1674
1675 bio_init(bio: &bio, bdev: log->rdev->bdev, table: ctx->ra_bvec,
1676 R5L_RECOVERY_PAGE_POOL_SIZE, opf: REQ_OP_READ);
1677 bio.bi_iter.bi_sector = log->rdev->data_offset + offset;
1678
1679 ctx->valid_pages = 0;
1680 ctx->pool_offset = offset;
1681
1682 while (ctx->valid_pages < ctx->total_pages) {
1683 __bio_add_page(bio: &bio, page: ctx->ra_pool[ctx->valid_pages], PAGE_SIZE,
1684 off: 0);
1685 ctx->valid_pages += 1;
1686
1687 offset = r5l_ring_add(log, start: offset, BLOCK_SECTORS);
1688
1689 if (offset == 0) /* reached end of the device */
1690 break;
1691 }
1692
1693 ret = submit_bio_wait(bio: &bio);
1694 bio_uninit(&bio);
1695 return ret;
1696}
1697
1698/*
1699 * try read a page from the read ahead page pool, if the page is not in the
1700 * pool, call r5l_recovery_fetch_ra_pool
1701 */
1702static int r5l_recovery_read_page(struct r5l_log *log,
1703 struct r5l_recovery_ctx *ctx,
1704 struct page *page,
1705 sector_t offset)
1706{
1707 int ret;
1708
1709 if (offset < ctx->pool_offset ||
1710 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1711 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1712 if (ret)
1713 return ret;
1714 }
1715
1716 BUG_ON(offset < ctx->pool_offset ||
1717 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1718
1719 memcpy(page_address(page),
1720 page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1721 BLOCK_SECTOR_SHIFT]),
1722 PAGE_SIZE);
1723 return 0;
1724}
1725
1726static int r5l_recovery_read_meta_block(struct r5l_log *log,
1727 struct r5l_recovery_ctx *ctx)
1728{
1729 struct page *page = ctx->meta_page;
1730 struct r5l_meta_block *mb;
1731 u32 crc, stored_crc;
1732 int ret;
1733
1734 ret = r5l_recovery_read_page(log, ctx, page, offset: ctx->pos);
1735 if (ret != 0)
1736 return ret;
1737
1738 mb = page_address(page);
1739 stored_crc = le32_to_cpu(mb->checksum);
1740 mb->checksum = 0;
1741
1742 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1743 le64_to_cpu(mb->seq) != ctx->seq ||
1744 mb->version != R5LOG_VERSION ||
1745 le64_to_cpu(mb->position) != ctx->pos)
1746 return -EINVAL;
1747
1748 crc = crc32c_le(crc: log->uuid_checksum, address: mb, PAGE_SIZE);
1749 if (stored_crc != crc)
1750 return -EINVAL;
1751
1752 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1753 return -EINVAL;
1754
1755 ctx->meta_total_blocks = BLOCK_SECTORS;
1756
1757 return 0;
1758}
1759
1760static void
1761r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1762 struct page *page,
1763 sector_t pos, u64 seq)
1764{
1765 struct r5l_meta_block *mb;
1766
1767 mb = page_address(page);
1768 clear_page(page: mb);
1769 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1770 mb->version = R5LOG_VERSION;
1771 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1772 mb->seq = cpu_to_le64(seq);
1773 mb->position = cpu_to_le64(pos);
1774}
1775
1776static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1777 u64 seq)
1778{
1779 struct page *page;
1780 struct r5l_meta_block *mb;
1781
1782 page = alloc_page(GFP_KERNEL);
1783 if (!page)
1784 return -ENOMEM;
1785 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1786 mb = page_address(page);
1787 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1788 mb, PAGE_SIZE));
1789 if (!sync_page_io(rdev: log->rdev, sector: pos, PAGE_SIZE, page, opf: REQ_OP_WRITE |
1790 REQ_SYNC | REQ_FUA, metadata_op: false)) {
1791 __free_page(page);
1792 return -EIO;
1793 }
1794 __free_page(page);
1795 return 0;
1796}
1797
1798/*
1799 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1800 * to mark valid (potentially not flushed) data in the journal.
1801 *
1802 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1803 * so there should not be any mismatch here.
1804 */
1805static void r5l_recovery_load_data(struct r5l_log *log,
1806 struct stripe_head *sh,
1807 struct r5l_recovery_ctx *ctx,
1808 struct r5l_payload_data_parity *payload,
1809 sector_t log_offset)
1810{
1811 struct mddev *mddev = log->rdev->mddev;
1812 struct r5conf *conf = mddev->private;
1813 int dd_idx;
1814
1815 raid5_compute_sector(conf,
1816 le64_to_cpu(payload->location), previous: 0,
1817 dd_idx: &dd_idx, sh);
1818 r5l_recovery_read_page(log, ctx, page: sh->dev[dd_idx].page, offset: log_offset);
1819 sh->dev[dd_idx].log_checksum =
1820 le32_to_cpu(payload->checksum[0]);
1821 ctx->meta_total_blocks += BLOCK_SECTORS;
1822
1823 set_bit(nr: R5_Wantwrite, addr: &sh->dev[dd_idx].flags);
1824 set_bit(nr: STRIPE_R5C_CACHING, addr: &sh->state);
1825}
1826
1827static void r5l_recovery_load_parity(struct r5l_log *log,
1828 struct stripe_head *sh,
1829 struct r5l_recovery_ctx *ctx,
1830 struct r5l_payload_data_parity *payload,
1831 sector_t log_offset)
1832{
1833 struct mddev *mddev = log->rdev->mddev;
1834 struct r5conf *conf = mddev->private;
1835
1836 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1837 r5l_recovery_read_page(log, ctx, page: sh->dev[sh->pd_idx].page, offset: log_offset);
1838 sh->dev[sh->pd_idx].log_checksum =
1839 le32_to_cpu(payload->checksum[0]);
1840 set_bit(nr: R5_Wantwrite, addr: &sh->dev[sh->pd_idx].flags);
1841
1842 if (sh->qd_idx >= 0) {
1843 r5l_recovery_read_page(
1844 log, ctx, page: sh->dev[sh->qd_idx].page,
1845 offset: r5l_ring_add(log, start: log_offset, BLOCK_SECTORS));
1846 sh->dev[sh->qd_idx].log_checksum =
1847 le32_to_cpu(payload->checksum[1]);
1848 set_bit(nr: R5_Wantwrite, addr: &sh->dev[sh->qd_idx].flags);
1849 }
1850 clear_bit(nr: STRIPE_R5C_CACHING, addr: &sh->state);
1851}
1852
1853static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1854{
1855 int i;
1856
1857 sh->state = 0;
1858 sh->log_start = MaxSector;
1859 for (i = sh->disks; i--; )
1860 sh->dev[i].flags = 0;
1861}
1862
1863static void
1864r5l_recovery_replay_one_stripe(struct r5conf *conf,
1865 struct stripe_head *sh,
1866 struct r5l_recovery_ctx *ctx)
1867{
1868 struct md_rdev *rdev, *rrdev;
1869 int disk_index;
1870 int data_count = 0;
1871
1872 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1873 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1874 continue;
1875 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1876 continue;
1877 data_count++;
1878 }
1879
1880 /*
1881 * stripes that only have parity must have been flushed
1882 * before the crash that we are now recovering from, so
1883 * there is nothing more to recovery.
1884 */
1885 if (data_count == 0)
1886 goto out;
1887
1888 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1889 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1890 continue;
1891
1892 /* in case device is broken */
1893 rcu_read_lock();
1894 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1895 if (rdev) {
1896 atomic_inc(v: &rdev->nr_pending);
1897 rcu_read_unlock();
1898 sync_page_io(rdev, sector: sh->sector, PAGE_SIZE,
1899 page: sh->dev[disk_index].page, opf: REQ_OP_WRITE,
1900 metadata_op: false);
1901 rdev_dec_pending(rdev, mddev: rdev->mddev);
1902 rcu_read_lock();
1903 }
1904 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1905 if (rrdev) {
1906 atomic_inc(v: &rrdev->nr_pending);
1907 rcu_read_unlock();
1908 sync_page_io(rdev: rrdev, sector: sh->sector, PAGE_SIZE,
1909 page: sh->dev[disk_index].page, opf: REQ_OP_WRITE,
1910 metadata_op: false);
1911 rdev_dec_pending(rdev: rrdev, mddev: rrdev->mddev);
1912 rcu_read_lock();
1913 }
1914 rcu_read_unlock();
1915 }
1916 ctx->data_parity_stripes++;
1917out:
1918 r5l_recovery_reset_stripe(sh);
1919}
1920
1921static struct stripe_head *
1922r5c_recovery_alloc_stripe(
1923 struct r5conf *conf,
1924 sector_t stripe_sect,
1925 int noblock)
1926{
1927 struct stripe_head *sh;
1928
1929 sh = raid5_get_active_stripe(conf, NULL, sector: stripe_sect,
1930 flags: noblock ? R5_GAS_NOBLOCK : 0);
1931 if (!sh)
1932 return NULL; /* no more stripe available */
1933
1934 r5l_recovery_reset_stripe(sh);
1935
1936 return sh;
1937}
1938
1939static struct stripe_head *
1940r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1941{
1942 struct stripe_head *sh;
1943
1944 list_for_each_entry(sh, list, lru)
1945 if (sh->sector == sect)
1946 return sh;
1947 return NULL;
1948}
1949
1950static void
1951r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1952 struct r5l_recovery_ctx *ctx)
1953{
1954 struct stripe_head *sh, *next;
1955
1956 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1957 r5l_recovery_reset_stripe(sh);
1958 list_del_init(entry: &sh->lru);
1959 raid5_release_stripe(sh);
1960 }
1961}
1962
1963static void
1964r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1965 struct r5l_recovery_ctx *ctx)
1966{
1967 struct stripe_head *sh, *next;
1968
1969 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1970 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1971 r5l_recovery_replay_one_stripe(conf: sh->raid_conf, sh, ctx);
1972 list_del_init(entry: &sh->lru);
1973 raid5_release_stripe(sh);
1974 }
1975}
1976
1977/* if matches return 0; otherwise return -EINVAL */
1978static int
1979r5l_recovery_verify_data_checksum(struct r5l_log *log,
1980 struct r5l_recovery_ctx *ctx,
1981 struct page *page,
1982 sector_t log_offset, __le32 log_checksum)
1983{
1984 void *addr;
1985 u32 checksum;
1986
1987 r5l_recovery_read_page(log, ctx, page, offset: log_offset);
1988 addr = kmap_atomic(page);
1989 checksum = crc32c_le(crc: log->uuid_checksum, address: addr, PAGE_SIZE);
1990 kunmap_atomic(addr);
1991 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1992}
1993
1994/*
1995 * before loading data to stripe cache, we need verify checksum for all data,
1996 * if there is mismatch for any data page, we drop all data in the mata block
1997 */
1998static int
1999r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
2000 struct r5l_recovery_ctx *ctx)
2001{
2002 struct mddev *mddev = log->rdev->mddev;
2003 struct r5conf *conf = mddev->private;
2004 struct r5l_meta_block *mb = page_address(ctx->meta_page);
2005 sector_t mb_offset = sizeof(struct r5l_meta_block);
2006 sector_t log_offset = r5l_ring_add(log, start: ctx->pos, BLOCK_SECTORS);
2007 struct page *page;
2008 struct r5l_payload_data_parity *payload;
2009 struct r5l_payload_flush *payload_flush;
2010
2011 page = alloc_page(GFP_KERNEL);
2012 if (!page)
2013 return -ENOMEM;
2014
2015 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2016 payload = (void *)mb + mb_offset;
2017 payload_flush = (void *)mb + mb_offset;
2018
2019 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2020 if (r5l_recovery_verify_data_checksum(
2021 log, ctx, page, log_offset,
2022 log_checksum: payload->checksum[0]) < 0)
2023 goto mismatch;
2024 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2025 if (r5l_recovery_verify_data_checksum(
2026 log, ctx, page, log_offset,
2027 log_checksum: payload->checksum[0]) < 0)
2028 goto mismatch;
2029 if (conf->max_degraded == 2 && /* q for RAID 6 */
2030 r5l_recovery_verify_data_checksum(
2031 log, ctx, page,
2032 log_offset: r5l_ring_add(log, start: log_offset,
2033 BLOCK_SECTORS),
2034 log_checksum: payload->checksum[1]) < 0)
2035 goto mismatch;
2036 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2037 /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2038 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2039 goto mismatch;
2040
2041 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2042 mb_offset += sizeof(struct r5l_payload_flush) +
2043 le32_to_cpu(payload_flush->size);
2044 } else {
2045 /* DATA or PARITY payload */
2046 log_offset = r5l_ring_add(log, start: log_offset,
2047 le32_to_cpu(payload->size));
2048 mb_offset += sizeof(struct r5l_payload_data_parity) +
2049 sizeof(__le32) *
2050 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2051 }
2052
2053 }
2054
2055 put_page(page);
2056 return 0;
2057
2058mismatch:
2059 put_page(page);
2060 return -EINVAL;
2061}
2062
2063/*
2064 * Analyze all data/parity pages in one meta block
2065 * Returns:
2066 * 0 for success
2067 * -EINVAL for unknown playload type
2068 * -EAGAIN for checksum mismatch of data page
2069 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2070 */
2071static int
2072r5c_recovery_analyze_meta_block(struct r5l_log *log,
2073 struct r5l_recovery_ctx *ctx,
2074 struct list_head *cached_stripe_list)
2075{
2076 struct mddev *mddev = log->rdev->mddev;
2077 struct r5conf *conf = mddev->private;
2078 struct r5l_meta_block *mb;
2079 struct r5l_payload_data_parity *payload;
2080 struct r5l_payload_flush *payload_flush;
2081 int mb_offset;
2082 sector_t log_offset;
2083 sector_t stripe_sect;
2084 struct stripe_head *sh;
2085 int ret;
2086
2087 /*
2088 * for mismatch in data blocks, we will drop all data in this mb, but
2089 * we will still read next mb for other data with FLUSH flag, as
2090 * io_unit could finish out of order.
2091 */
2092 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2093 if (ret == -EINVAL)
2094 return -EAGAIN;
2095 else if (ret)
2096 return ret; /* -ENOMEM duo to alloc_page() failed */
2097
2098 mb = page_address(ctx->meta_page);
2099 mb_offset = sizeof(struct r5l_meta_block);
2100 log_offset = r5l_ring_add(log, start: ctx->pos, BLOCK_SECTORS);
2101
2102 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2103 int dd;
2104
2105 payload = (void *)mb + mb_offset;
2106 payload_flush = (void *)mb + mb_offset;
2107
2108 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2109 int i, count;
2110
2111 count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2112 for (i = 0; i < count; ++i) {
2113 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2114 sh = r5c_recovery_lookup_stripe(list: cached_stripe_list,
2115 sect: stripe_sect);
2116 if (sh) {
2117 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2118 r5l_recovery_reset_stripe(sh);
2119 list_del_init(entry: &sh->lru);
2120 raid5_release_stripe(sh);
2121 }
2122 }
2123
2124 mb_offset += sizeof(struct r5l_payload_flush) +
2125 le32_to_cpu(payload_flush->size);
2126 continue;
2127 }
2128
2129 /* DATA or PARITY payload */
2130 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2131 raid5_compute_sector(
2132 conf, le64_to_cpu(payload->location), previous: 0, dd_idx: &dd,
2133 NULL)
2134 : le64_to_cpu(payload->location);
2135
2136 sh = r5c_recovery_lookup_stripe(list: cached_stripe_list,
2137 sect: stripe_sect);
2138
2139 if (!sh) {
2140 sh = r5c_recovery_alloc_stripe(conf, stripe_sect, noblock: 1);
2141 /*
2142 * cannot get stripe from raid5_get_active_stripe
2143 * try replay some stripes
2144 */
2145 if (!sh) {
2146 r5c_recovery_replay_stripes(
2147 cached_stripe_list, ctx);
2148 sh = r5c_recovery_alloc_stripe(
2149 conf, stripe_sect, noblock: 1);
2150 }
2151 if (!sh) {
2152 int new_size = conf->min_nr_stripes * 2;
2153 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2154 mdname(mddev),
2155 new_size);
2156 ret = raid5_set_cache_size(mddev, size: new_size);
2157 if (conf->min_nr_stripes <= new_size / 2) {
2158 pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2159 mdname(mddev),
2160 ret,
2161 new_size,
2162 conf->min_nr_stripes,
2163 conf->max_nr_stripes);
2164 return -ENOMEM;
2165 }
2166 sh = r5c_recovery_alloc_stripe(
2167 conf, stripe_sect, noblock: 0);
2168 }
2169 if (!sh) {
2170 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2171 mdname(mddev));
2172 return -ENOMEM;
2173 }
2174 list_add_tail(new: &sh->lru, head: cached_stripe_list);
2175 }
2176
2177 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2178 if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2179 test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2180 r5l_recovery_replay_one_stripe(conf, sh, ctx);
2181 list_move_tail(list: &sh->lru, head: cached_stripe_list);
2182 }
2183 r5l_recovery_load_data(log, sh, ctx, payload,
2184 log_offset);
2185 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2186 r5l_recovery_load_parity(log, sh, ctx, payload,
2187 log_offset);
2188 else
2189 return -EINVAL;
2190
2191 log_offset = r5l_ring_add(log, start: log_offset,
2192 le32_to_cpu(payload->size));
2193
2194 mb_offset += sizeof(struct r5l_payload_data_parity) +
2195 sizeof(__le32) *
2196 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2197 }
2198
2199 return 0;
2200}
2201
2202/*
2203 * Load the stripe into cache. The stripe will be written out later by
2204 * the stripe cache state machine.
2205 */
2206static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2207 struct stripe_head *sh)
2208{
2209 struct r5dev *dev;
2210 int i;
2211
2212 for (i = sh->disks; i--; ) {
2213 dev = sh->dev + i;
2214 if (test_and_clear_bit(nr: R5_Wantwrite, addr: &dev->flags)) {
2215 set_bit(nr: R5_InJournal, addr: &dev->flags);
2216 set_bit(nr: R5_UPTODATE, addr: &dev->flags);
2217 }
2218 }
2219}
2220
2221/*
2222 * Scan through the log for all to-be-flushed data
2223 *
2224 * For stripes with data and parity, namely Data-Parity stripe
2225 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2226 *
2227 * For stripes with only data, namely Data-Only stripe
2228 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2229 *
2230 * For a stripe, if we see data after parity, we should discard all previous
2231 * data and parity for this stripe, as these data are already flushed to
2232 * the array.
2233 *
2234 * At the end of the scan, we return the new journal_tail, which points to
2235 * first data-only stripe on the journal device, or next invalid meta block.
2236 */
2237static int r5c_recovery_flush_log(struct r5l_log *log,
2238 struct r5l_recovery_ctx *ctx)
2239{
2240 struct stripe_head *sh;
2241 int ret = 0;
2242
2243 /* scan through the log */
2244 while (1) {
2245 if (r5l_recovery_read_meta_block(log, ctx))
2246 break;
2247
2248 ret = r5c_recovery_analyze_meta_block(log, ctx,
2249 cached_stripe_list: &ctx->cached_list);
2250 /*
2251 * -EAGAIN means mismatch in data block, in this case, we still
2252 * try scan the next metablock
2253 */
2254 if (ret && ret != -EAGAIN)
2255 break; /* ret == -EINVAL or -ENOMEM */
2256 ctx->seq++;
2257 ctx->pos = r5l_ring_add(log, start: ctx->pos, inc: ctx->meta_total_blocks);
2258 }
2259
2260 if (ret == -ENOMEM) {
2261 r5c_recovery_drop_stripes(cached_stripe_list: &ctx->cached_list, ctx);
2262 return ret;
2263 }
2264
2265 /* replay data-parity stripes */
2266 r5c_recovery_replay_stripes(cached_stripe_list: &ctx->cached_list, ctx);
2267
2268 /* load data-only stripes to stripe cache */
2269 list_for_each_entry(sh, &ctx->cached_list, lru) {
2270 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2271 r5c_recovery_load_one_stripe(log, sh);
2272 ctx->data_only_stripes++;
2273 }
2274
2275 return 0;
2276}
2277
2278/*
2279 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2280 * log will start here. but we can't let superblock point to last valid
2281 * meta block. The log might looks like:
2282 * | meta 1| meta 2| meta 3|
2283 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2284 * superblock points to meta 1, we write a new valid meta 2n. if crash
2285 * happens again, new recovery will start from meta 1. Since meta 2n is
2286 * valid now, recovery will think meta 3 is valid, which is wrong.
2287 * The solution is we create a new meta in meta2 with its seq == meta
2288 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2289 * will not think meta 3 is a valid meta, because its seq doesn't match
2290 */
2291
2292/*
2293 * Before recovery, the log looks like the following
2294 *
2295 * ---------------------------------------------
2296 * | valid log | invalid log |
2297 * ---------------------------------------------
2298 * ^
2299 * |- log->last_checkpoint
2300 * |- log->last_cp_seq
2301 *
2302 * Now we scan through the log until we see invalid entry
2303 *
2304 * ---------------------------------------------
2305 * | valid log | invalid log |
2306 * ---------------------------------------------
2307 * ^ ^
2308 * |- log->last_checkpoint |- ctx->pos
2309 * |- log->last_cp_seq |- ctx->seq
2310 *
2311 * From this point, we need to increase seq number by 10 to avoid
2312 * confusing next recovery.
2313 *
2314 * ---------------------------------------------
2315 * | valid log | invalid log |
2316 * ---------------------------------------------
2317 * ^ ^
2318 * |- log->last_checkpoint |- ctx->pos+1
2319 * |- log->last_cp_seq |- ctx->seq+10001
2320 *
2321 * However, it is not safe to start the state machine yet, because data only
2322 * parities are not yet secured in RAID. To save these data only parities, we
2323 * rewrite them from seq+11.
2324 *
2325 * -----------------------------------------------------------------
2326 * | valid log | data only stripes | invalid log |
2327 * -----------------------------------------------------------------
2328 * ^ ^
2329 * |- log->last_checkpoint |- ctx->pos+n
2330 * |- log->last_cp_seq |- ctx->seq+10000+n
2331 *
2332 * If failure happens again during this process, the recovery can safe start
2333 * again from log->last_checkpoint.
2334 *
2335 * Once data only stripes are rewritten to journal, we move log_tail
2336 *
2337 * -----------------------------------------------------------------
2338 * | old log | data only stripes | invalid log |
2339 * -----------------------------------------------------------------
2340 * ^ ^
2341 * |- log->last_checkpoint |- ctx->pos+n
2342 * |- log->last_cp_seq |- ctx->seq+10000+n
2343 *
2344 * Then we can safely start the state machine. If failure happens from this
2345 * point on, the recovery will start from new log->last_checkpoint.
2346 */
2347static int
2348r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2349 struct r5l_recovery_ctx *ctx)
2350{
2351 struct stripe_head *sh;
2352 struct mddev *mddev = log->rdev->mddev;
2353 struct page *page;
2354 sector_t next_checkpoint = MaxSector;
2355
2356 page = alloc_page(GFP_KERNEL);
2357 if (!page) {
2358 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2359 mdname(mddev));
2360 return -ENOMEM;
2361 }
2362
2363 WARN_ON(list_empty(&ctx->cached_list));
2364
2365 list_for_each_entry(sh, &ctx->cached_list, lru) {
2366 struct r5l_meta_block *mb;
2367 int i;
2368 int offset;
2369 sector_t write_pos;
2370
2371 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2372 r5l_recovery_create_empty_meta_block(log, page,
2373 pos: ctx->pos, seq: ctx->seq);
2374 mb = page_address(page);
2375 offset = le32_to_cpu(mb->meta_size);
2376 write_pos = r5l_ring_add(log, start: ctx->pos, BLOCK_SECTORS);
2377
2378 for (i = sh->disks; i--; ) {
2379 struct r5dev *dev = &sh->dev[i];
2380 struct r5l_payload_data_parity *payload;
2381 void *addr;
2382
2383 if (test_bit(R5_InJournal, &dev->flags)) {
2384 payload = (void *)mb + offset;
2385 payload->header.type = cpu_to_le16(
2386 R5LOG_PAYLOAD_DATA);
2387 payload->size = cpu_to_le32(BLOCK_SECTORS);
2388 payload->location = cpu_to_le64(
2389 raid5_compute_blocknr(sh, i, 0));
2390 addr = kmap_atomic(page: dev->page);
2391 payload->checksum[0] = cpu_to_le32(
2392 crc32c_le(log->uuid_checksum, addr,
2393 PAGE_SIZE));
2394 kunmap_atomic(addr);
2395 sync_page_io(rdev: log->rdev, sector: write_pos, PAGE_SIZE,
2396 page: dev->page, opf: REQ_OP_WRITE, metadata_op: false);
2397 write_pos = r5l_ring_add(log, start: write_pos,
2398 BLOCK_SECTORS);
2399 offset += sizeof(__le32) +
2400 sizeof(struct r5l_payload_data_parity);
2401
2402 }
2403 }
2404 mb->meta_size = cpu_to_le32(offset);
2405 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2406 mb, PAGE_SIZE));
2407 sync_page_io(rdev: log->rdev, sector: ctx->pos, PAGE_SIZE, page,
2408 opf: REQ_OP_WRITE | REQ_SYNC | REQ_FUA, metadata_op: false);
2409 sh->log_start = ctx->pos;
2410 list_add_tail(new: &sh->r5c, head: &log->stripe_in_journal_list);
2411 atomic_inc(v: &log->stripe_in_journal_count);
2412 ctx->pos = write_pos;
2413 ctx->seq += 1;
2414 next_checkpoint = sh->log_start;
2415 }
2416 log->next_checkpoint = next_checkpoint;
2417 __free_page(page);
2418 return 0;
2419}
2420
2421static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2422 struct r5l_recovery_ctx *ctx)
2423{
2424 struct mddev *mddev = log->rdev->mddev;
2425 struct r5conf *conf = mddev->private;
2426 struct stripe_head *sh, *next;
2427 bool cleared_pending = false;
2428
2429 if (ctx->data_only_stripes == 0)
2430 return;
2431
2432 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2433 cleared_pending = true;
2434 clear_bit(nr: MD_SB_CHANGE_PENDING, addr: &mddev->sb_flags);
2435 }
2436 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2437
2438 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2439 r5c_make_stripe_write_out(sh);
2440 set_bit(nr: STRIPE_HANDLE, addr: &sh->state);
2441 list_del_init(entry: &sh->lru);
2442 raid5_release_stripe(sh);
2443 }
2444
2445 /* reuse conf->wait_for_quiescent in recovery */
2446 wait_event(conf->wait_for_quiescent,
2447 atomic_read(&conf->active_stripes) == 0);
2448
2449 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2450 if (cleared_pending)
2451 set_bit(nr: MD_SB_CHANGE_PENDING, addr: &mddev->sb_flags);
2452}
2453
2454static int r5l_recovery_log(struct r5l_log *log)
2455{
2456 struct mddev *mddev = log->rdev->mddev;
2457 struct r5l_recovery_ctx *ctx;
2458 int ret;
2459 sector_t pos;
2460
2461 ctx = kzalloc(size: sizeof(*ctx), GFP_KERNEL);
2462 if (!ctx)
2463 return -ENOMEM;
2464
2465 ctx->pos = log->last_checkpoint;
2466 ctx->seq = log->last_cp_seq;
2467 INIT_LIST_HEAD(list: &ctx->cached_list);
2468 ctx->meta_page = alloc_page(GFP_KERNEL);
2469
2470 if (!ctx->meta_page) {
2471 ret = -ENOMEM;
2472 goto meta_page;
2473 }
2474
2475 if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2476 ret = -ENOMEM;
2477 goto ra_pool;
2478 }
2479
2480 ret = r5c_recovery_flush_log(log, ctx);
2481
2482 if (ret)
2483 goto error;
2484
2485 pos = ctx->pos;
2486 ctx->seq += 10000;
2487
2488 if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2489 pr_info("md/raid:%s: starting from clean shutdown\n",
2490 mdname(mddev));
2491 else
2492 pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2493 mdname(mddev), ctx->data_only_stripes,
2494 ctx->data_parity_stripes);
2495
2496 if (ctx->data_only_stripes == 0) {
2497 log->next_checkpoint = ctx->pos;
2498 r5l_log_write_empty_meta_block(log, pos: ctx->pos, seq: ctx->seq++);
2499 ctx->pos = r5l_ring_add(log, start: ctx->pos, BLOCK_SECTORS);
2500 } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2501 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2502 mdname(mddev));
2503 ret = -EIO;
2504 goto error;
2505 }
2506
2507 log->log_start = ctx->pos;
2508 log->seq = ctx->seq;
2509 log->last_checkpoint = pos;
2510 r5l_write_super(log, cp: pos);
2511
2512 r5c_recovery_flush_data_only_stripes(log, ctx);
2513 ret = 0;
2514error:
2515 r5l_recovery_free_ra_pool(log, ctx);
2516ra_pool:
2517 __free_page(ctx->meta_page);
2518meta_page:
2519 kfree(objp: ctx);
2520 return ret;
2521}
2522
2523static void r5l_write_super(struct r5l_log *log, sector_t cp)
2524{
2525 struct mddev *mddev = log->rdev->mddev;
2526
2527 log->rdev->journal_tail = cp;
2528 set_bit(nr: MD_SB_CHANGE_DEVS, addr: &mddev->sb_flags);
2529}
2530
2531static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2532{
2533 struct r5conf *conf;
2534 int ret;
2535
2536 ret = mddev_lock(mddev);
2537 if (ret)
2538 return ret;
2539
2540 conf = mddev->private;
2541 if (!conf || !conf->log)
2542 goto out_unlock;
2543
2544 switch (conf->log->r5c_journal_mode) {
2545 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2546 ret = snprintf(
2547 buf: page, PAGE_SIZE, fmt: "[%s] %s\n",
2548 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2549 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2550 break;
2551 case R5C_JOURNAL_MODE_WRITE_BACK:
2552 ret = snprintf(
2553 buf: page, PAGE_SIZE, fmt: "%s [%s]\n",
2554 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2555 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2556 break;
2557 default:
2558 ret = 0;
2559 }
2560
2561out_unlock:
2562 mddev_unlock(mddev);
2563 return ret;
2564}
2565
2566/*
2567 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2568 *
2569 * @mode as defined in 'enum r5c_journal_mode'.
2570 *
2571 */
2572int r5c_journal_mode_set(struct mddev *mddev, int mode)
2573{
2574 struct r5conf *conf;
2575
2576 if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2577 mode > R5C_JOURNAL_MODE_WRITE_BACK)
2578 return -EINVAL;
2579
2580 conf = mddev->private;
2581 if (!conf || !conf->log)
2582 return -ENODEV;
2583
2584 if (raid5_calc_degraded(conf) > 0 &&
2585 mode == R5C_JOURNAL_MODE_WRITE_BACK)
2586 return -EINVAL;
2587
2588 conf->log->r5c_journal_mode = mode;
2589
2590 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2591 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2592 return 0;
2593}
2594EXPORT_SYMBOL(r5c_journal_mode_set);
2595
2596static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2597 const char *page, size_t length)
2598{
2599 int mode = ARRAY_SIZE(r5c_journal_mode_str);
2600 size_t len = length;
2601 int ret;
2602
2603 if (len < 2)
2604 return -EINVAL;
2605
2606 if (page[len - 1] == '\n')
2607 len--;
2608
2609 while (mode--)
2610 if (strlen(r5c_journal_mode_str[mode]) == len &&
2611 !strncmp(page, r5c_journal_mode_str[mode], len))
2612 break;
2613 ret = mddev_suspend_and_lock(mddev);
2614 if (ret)
2615 return ret;
2616 ret = r5c_journal_mode_set(mddev, mode);
2617 mddev_unlock_and_resume(mddev);
2618 return ret ?: length;
2619}
2620
2621struct md_sysfs_entry
2622r5c_journal_mode = __ATTR(journal_mode, 0644,
2623 r5c_journal_mode_show, r5c_journal_mode_store);
2624
2625/*
2626 * Try handle write operation in caching phase. This function should only
2627 * be called in write-back mode.
2628 *
2629 * If all outstanding writes can be handled in caching phase, returns 0
2630 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2631 * and returns -EAGAIN
2632 */
2633int r5c_try_caching_write(struct r5conf *conf,
2634 struct stripe_head *sh,
2635 struct stripe_head_state *s,
2636 int disks)
2637{
2638 struct r5l_log *log = READ_ONCE(conf->log);
2639 int i;
2640 struct r5dev *dev;
2641 int to_cache = 0;
2642 void __rcu **pslot;
2643 sector_t tree_index;
2644 int ret;
2645 uintptr_t refcount;
2646
2647 BUG_ON(!r5c_is_writeback(log));
2648
2649 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2650 /*
2651 * There are two different scenarios here:
2652 * 1. The stripe has some data cached, and it is sent to
2653 * write-out phase for reclaim
2654 * 2. The stripe is clean, and this is the first write
2655 *
2656 * For 1, return -EAGAIN, so we continue with
2657 * handle_stripe_dirtying().
2658 *
2659 * For 2, set STRIPE_R5C_CACHING and continue with caching
2660 * write.
2661 */
2662
2663 /* case 1: anything injournal or anything in written */
2664 if (s->injournal > 0 || s->written > 0)
2665 return -EAGAIN;
2666 /* case 2 */
2667 set_bit(nr: STRIPE_R5C_CACHING, addr: &sh->state);
2668 }
2669
2670 /*
2671 * When run in degraded mode, array is set to write-through mode.
2672 * This check helps drain pending write safely in the transition to
2673 * write-through mode.
2674 *
2675 * When a stripe is syncing, the write is also handled in write
2676 * through mode.
2677 */
2678 if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2679 r5c_make_stripe_write_out(sh);
2680 return -EAGAIN;
2681 }
2682
2683 for (i = disks; i--; ) {
2684 dev = &sh->dev[i];
2685 /* if non-overwrite, use writing-out phase */
2686 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2687 !test_bit(R5_InJournal, &dev->flags)) {
2688 r5c_make_stripe_write_out(sh);
2689 return -EAGAIN;
2690 }
2691 }
2692
2693 /* if the stripe is not counted in big_stripe_tree, add it now */
2694 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2695 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2696 tree_index = r5c_tree_index(conf, sect: sh->sector);
2697 spin_lock(lock: &log->tree_lock);
2698 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2699 index: tree_index);
2700 if (pslot) {
2701 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2702 slot: pslot, treelock: &log->tree_lock) >>
2703 R5C_RADIX_COUNT_SHIFT;
2704 radix_tree_replace_slot(
2705 &log->big_stripe_tree, slot: pslot,
2706 entry: (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2707 } else {
2708 /*
2709 * this radix_tree_insert can fail safely, so no
2710 * need to call radix_tree_preload()
2711 */
2712 ret = radix_tree_insert(
2713 &log->big_stripe_tree, index: tree_index,
2714 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2715 if (ret) {
2716 spin_unlock(lock: &log->tree_lock);
2717 r5c_make_stripe_write_out(sh);
2718 return -EAGAIN;
2719 }
2720 }
2721 spin_unlock(lock: &log->tree_lock);
2722
2723 /*
2724 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2725 * counted in the radix tree
2726 */
2727 set_bit(nr: STRIPE_R5C_PARTIAL_STRIPE, addr: &sh->state);
2728 atomic_inc(v: &conf->r5c_cached_partial_stripes);
2729 }
2730
2731 for (i = disks; i--; ) {
2732 dev = &sh->dev[i];
2733 if (dev->towrite) {
2734 set_bit(nr: R5_Wantwrite, addr: &dev->flags);
2735 set_bit(nr: R5_Wantdrain, addr: &dev->flags);
2736 set_bit(nr: R5_LOCKED, addr: &dev->flags);
2737 to_cache++;
2738 }
2739 }
2740
2741 if (to_cache) {
2742 set_bit(nr: STRIPE_OP_BIODRAIN, addr: &s->ops_request);
2743 /*
2744 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2745 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2746 * r5c_handle_data_cached()
2747 */
2748 set_bit(nr: STRIPE_LOG_TRAPPED, addr: &sh->state);
2749 }
2750
2751 return 0;
2752}
2753
2754/*
2755 * free extra pages (orig_page) we allocated for prexor
2756 */
2757void r5c_release_extra_page(struct stripe_head *sh)
2758{
2759 struct r5conf *conf = sh->raid_conf;
2760 int i;
2761 bool using_disk_info_extra_page;
2762
2763 using_disk_info_extra_page =
2764 sh->dev[0].orig_page == conf->disks[0].extra_page;
2765
2766 for (i = sh->disks; i--; )
2767 if (sh->dev[i].page != sh->dev[i].orig_page) {
2768 struct page *p = sh->dev[i].orig_page;
2769
2770 sh->dev[i].orig_page = sh->dev[i].page;
2771 clear_bit(nr: R5_OrigPageUPTDODATE, addr: &sh->dev[i].flags);
2772
2773 if (!using_disk_info_extra_page)
2774 put_page(page: p);
2775 }
2776
2777 if (using_disk_info_extra_page) {
2778 clear_bit(nr: R5C_EXTRA_PAGE_IN_USE, addr: &conf->cache_state);
2779 md_wakeup_thread(thread: conf->mddev->thread);
2780 }
2781}
2782
2783void r5c_use_extra_page(struct stripe_head *sh)
2784{
2785 struct r5conf *conf = sh->raid_conf;
2786 int i;
2787 struct r5dev *dev;
2788
2789 for (i = sh->disks; i--; ) {
2790 dev = &sh->dev[i];
2791 if (dev->orig_page != dev->page)
2792 put_page(page: dev->orig_page);
2793 dev->orig_page = conf->disks[i].extra_page;
2794 }
2795}
2796
2797/*
2798 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2799 * stripe is committed to RAID disks.
2800 */
2801void r5c_finish_stripe_write_out(struct r5conf *conf,
2802 struct stripe_head *sh,
2803 struct stripe_head_state *s)
2804{
2805 struct r5l_log *log = READ_ONCE(conf->log);
2806 int i;
2807 int do_wakeup = 0;
2808 sector_t tree_index;
2809 void __rcu **pslot;
2810 uintptr_t refcount;
2811
2812 if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2813 return;
2814
2815 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2816 clear_bit(nr: R5_InJournal, addr: &sh->dev[sh->pd_idx].flags);
2817
2818 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2819 return;
2820
2821 for (i = sh->disks; i--; ) {
2822 clear_bit(nr: R5_InJournal, addr: &sh->dev[i].flags);
2823 if (test_and_clear_bit(nr: R5_Overlap, addr: &sh->dev[i].flags))
2824 do_wakeup = 1;
2825 }
2826
2827 /*
2828 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2829 * We updated R5_InJournal, so we also update s->injournal.
2830 */
2831 s->injournal = 0;
2832
2833 if (test_and_clear_bit(nr: STRIPE_FULL_WRITE, addr: &sh->state))
2834 if (atomic_dec_and_test(v: &conf->pending_full_writes))
2835 md_wakeup_thread(thread: conf->mddev->thread);
2836
2837 if (do_wakeup)
2838 wake_up(&conf->wait_for_overlap);
2839
2840 spin_lock_irq(lock: &log->stripe_in_journal_lock);
2841 list_del_init(entry: &sh->r5c);
2842 spin_unlock_irq(lock: &log->stripe_in_journal_lock);
2843 sh->log_start = MaxSector;
2844
2845 atomic_dec(v: &log->stripe_in_journal_count);
2846 r5c_update_log_state(log);
2847
2848 /* stop counting this stripe in big_stripe_tree */
2849 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2850 test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2851 tree_index = r5c_tree_index(conf, sect: sh->sector);
2852 spin_lock(lock: &log->tree_lock);
2853 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2854 index: tree_index);
2855 BUG_ON(pslot == NULL);
2856 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2857 slot: pslot, treelock: &log->tree_lock) >>
2858 R5C_RADIX_COUNT_SHIFT;
2859 if (refcount == 1)
2860 radix_tree_delete(&log->big_stripe_tree, tree_index);
2861 else
2862 radix_tree_replace_slot(
2863 &log->big_stripe_tree, slot: pslot,
2864 entry: (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2865 spin_unlock(lock: &log->tree_lock);
2866 }
2867
2868 if (test_and_clear_bit(nr: STRIPE_R5C_PARTIAL_STRIPE, addr: &sh->state)) {
2869 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2870 atomic_dec(v: &conf->r5c_flushing_partial_stripes);
2871 atomic_dec(v: &conf->r5c_cached_partial_stripes);
2872 }
2873
2874 if (test_and_clear_bit(nr: STRIPE_R5C_FULL_STRIPE, addr: &sh->state)) {
2875 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2876 atomic_dec(v: &conf->r5c_flushing_full_stripes);
2877 atomic_dec(v: &conf->r5c_cached_full_stripes);
2878 }
2879
2880 r5l_append_flush_payload(log, sect: sh->sector);
2881 /* stripe is flused to raid disks, we can do resync now */
2882 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2883 set_bit(nr: STRIPE_HANDLE, addr: &sh->state);
2884}
2885
2886int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2887{
2888 struct r5conf *conf = sh->raid_conf;
2889 int pages = 0;
2890 int reserve;
2891 int i;
2892 int ret = 0;
2893
2894 BUG_ON(!log);
2895
2896 for (i = 0; i < sh->disks; i++) {
2897 void *addr;
2898
2899 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2900 continue;
2901 addr = kmap_atomic(page: sh->dev[i].page);
2902 sh->dev[i].log_checksum = crc32c_le(crc: log->uuid_checksum,
2903 address: addr, PAGE_SIZE);
2904 kunmap_atomic(addr);
2905 pages++;
2906 }
2907 WARN_ON(pages == 0);
2908
2909 /*
2910 * The stripe must enter state machine again to call endio, so
2911 * don't delay.
2912 */
2913 clear_bit(nr: STRIPE_DELAYED, addr: &sh->state);
2914 atomic_inc(v: &sh->count);
2915
2916 mutex_lock(&log->io_mutex);
2917 /* meta + data */
2918 reserve = (1 + pages) << (PAGE_SHIFT - 9);
2919
2920 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2921 sh->log_start == MaxSector)
2922 r5l_add_no_space_stripe(log, sh);
2923 else if (!r5l_has_free_space(log, size: reserve)) {
2924 if (sh->log_start == log->last_checkpoint)
2925 BUG();
2926 else
2927 r5l_add_no_space_stripe(log, sh);
2928 } else {
2929 ret = r5l_log_stripe(log, sh, data_pages: pages, parity_pages: 0);
2930 if (ret) {
2931 spin_lock_irq(lock: &log->io_list_lock);
2932 list_add_tail(new: &sh->log_list, head: &log->no_mem_stripes);
2933 spin_unlock_irq(lock: &log->io_list_lock);
2934 }
2935 }
2936
2937 mutex_unlock(lock: &log->io_mutex);
2938 return 0;
2939}
2940
2941/* check whether this big stripe is in write back cache. */
2942bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2943{
2944 struct r5l_log *log = READ_ONCE(conf->log);
2945 sector_t tree_index;
2946 void *slot;
2947
2948 if (!log)
2949 return false;
2950
2951 WARN_ON_ONCE(!rcu_read_lock_held());
2952 tree_index = r5c_tree_index(conf, sect);
2953 slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2954 return slot != NULL;
2955}
2956
2957static int r5l_load_log(struct r5l_log *log)
2958{
2959 struct md_rdev *rdev = log->rdev;
2960 struct page *page;
2961 struct r5l_meta_block *mb;
2962 sector_t cp = log->rdev->journal_tail;
2963 u32 stored_crc, expected_crc;
2964 bool create_super = false;
2965 int ret = 0;
2966
2967 /* Make sure it's valid */
2968 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2969 cp = 0;
2970 page = alloc_page(GFP_KERNEL);
2971 if (!page)
2972 return -ENOMEM;
2973
2974 if (!sync_page_io(rdev, sector: cp, PAGE_SIZE, page, opf: REQ_OP_READ, metadata_op: false)) {
2975 ret = -EIO;
2976 goto ioerr;
2977 }
2978 mb = page_address(page);
2979
2980 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2981 mb->version != R5LOG_VERSION) {
2982 create_super = true;
2983 goto create;
2984 }
2985 stored_crc = le32_to_cpu(mb->checksum);
2986 mb->checksum = 0;
2987 expected_crc = crc32c_le(crc: log->uuid_checksum, address: mb, PAGE_SIZE);
2988 if (stored_crc != expected_crc) {
2989 create_super = true;
2990 goto create;
2991 }
2992 if (le64_to_cpu(mb->position) != cp) {
2993 create_super = true;
2994 goto create;
2995 }
2996create:
2997 if (create_super) {
2998 log->last_cp_seq = get_random_u32();
2999 cp = 0;
3000 r5l_log_write_empty_meta_block(log, pos: cp, seq: log->last_cp_seq);
3001 /*
3002 * Make sure super points to correct address. Log might have
3003 * data very soon. If super hasn't correct log tail address,
3004 * recovery can't find the log
3005 */
3006 r5l_write_super(log, cp);
3007 } else
3008 log->last_cp_seq = le64_to_cpu(mb->seq);
3009
3010 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3011 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3012 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3013 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3014 log->last_checkpoint = cp;
3015
3016 __free_page(page);
3017
3018 if (create_super) {
3019 log->log_start = r5l_ring_add(log, start: cp, BLOCK_SECTORS);
3020 log->seq = log->last_cp_seq + 1;
3021 log->next_checkpoint = cp;
3022 } else
3023 ret = r5l_recovery_log(log);
3024
3025 r5c_update_log_state(log);
3026 return ret;
3027ioerr:
3028 __free_page(page);
3029 return ret;
3030}
3031
3032int r5l_start(struct r5l_log *log)
3033{
3034 int ret;
3035
3036 if (!log)
3037 return 0;
3038
3039 ret = r5l_load_log(log);
3040 if (ret) {
3041 struct mddev *mddev = log->rdev->mddev;
3042 struct r5conf *conf = mddev->private;
3043
3044 r5l_exit_log(conf);
3045 }
3046 return ret;
3047}
3048
3049void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3050{
3051 struct r5conf *conf = mddev->private;
3052 struct r5l_log *log = READ_ONCE(conf->log);
3053
3054 if (!log)
3055 return;
3056
3057 if ((raid5_calc_degraded(conf) > 0 ||
3058 test_bit(Journal, &rdev->flags)) &&
3059 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3060 schedule_work(work: &log->disable_writeback_work);
3061}
3062
3063int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3064{
3065 struct r5l_log *log;
3066 struct md_thread *thread;
3067 int ret;
3068
3069 pr_debug("md/raid:%s: using device %pg as journal\n",
3070 mdname(conf->mddev), rdev->bdev);
3071
3072 if (PAGE_SIZE != 4096)
3073 return -EINVAL;
3074
3075 /*
3076 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3077 * raid_disks r5l_payload_data_parity.
3078 *
3079 * Write journal and cache does not work for very big array
3080 * (raid_disks > 203)
3081 */
3082 if (sizeof(struct r5l_meta_block) +
3083 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3084 conf->raid_disks) > PAGE_SIZE) {
3085 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3086 mdname(conf->mddev), conf->raid_disks);
3087 return -EINVAL;
3088 }
3089
3090 log = kzalloc(size: sizeof(*log), GFP_KERNEL);
3091 if (!log)
3092 return -ENOMEM;
3093 log->rdev = rdev;
3094 log->need_cache_flush = bdev_write_cache(bdev: rdev->bdev);
3095 log->uuid_checksum = crc32c_le(crc: ~0, address: rdev->mddev->uuid,
3096 length: sizeof(rdev->mddev->uuid));
3097
3098 mutex_init(&log->io_mutex);
3099
3100 spin_lock_init(&log->io_list_lock);
3101 INIT_LIST_HEAD(list: &log->running_ios);
3102 INIT_LIST_HEAD(list: &log->io_end_ios);
3103 INIT_LIST_HEAD(list: &log->flushing_ios);
3104 INIT_LIST_HEAD(list: &log->finished_ios);
3105
3106 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3107 if (!log->io_kc)
3108 goto io_kc;
3109
3110 ret = mempool_init_slab_pool(pool: &log->io_pool, R5L_POOL_SIZE, kc: log->io_kc);
3111 if (ret)
3112 goto io_pool;
3113
3114 ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, flags: BIOSET_NEED_BVECS);
3115 if (ret)
3116 goto io_bs;
3117
3118 ret = mempool_init_page_pool(pool: &log->meta_pool, R5L_POOL_SIZE, order: 0);
3119 if (ret)
3120 goto out_mempool;
3121
3122 spin_lock_init(&log->tree_lock);
3123 INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3124
3125 thread = md_register_thread(run: r5l_reclaim_thread, mddev: log->rdev->mddev,
3126 name: "reclaim");
3127 if (!thread)
3128 goto reclaim_thread;
3129
3130 thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3131 rcu_assign_pointer(log->reclaim_thread, thread);
3132
3133 init_waitqueue_head(&log->iounit_wait);
3134
3135 INIT_LIST_HEAD(list: &log->no_mem_stripes);
3136
3137 INIT_LIST_HEAD(list: &log->no_space_stripes);
3138 spin_lock_init(&log->no_space_stripes_lock);
3139
3140 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3141 INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3142
3143 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3144 INIT_LIST_HEAD(list: &log->stripe_in_journal_list);
3145 spin_lock_init(&log->stripe_in_journal_lock);
3146 atomic_set(v: &log->stripe_in_journal_count, i: 0);
3147
3148 WRITE_ONCE(conf->log, log);
3149
3150 set_bit(nr: MD_HAS_JOURNAL, addr: &conf->mddev->flags);
3151 return 0;
3152
3153reclaim_thread:
3154 mempool_exit(pool: &log->meta_pool);
3155out_mempool:
3156 bioset_exit(&log->bs);
3157io_bs:
3158 mempool_exit(pool: &log->io_pool);
3159io_pool:
3160 kmem_cache_destroy(s: log->io_kc);
3161io_kc:
3162 kfree(objp: log);
3163 return -EINVAL;
3164}
3165
3166void r5l_exit_log(struct r5conf *conf)
3167{
3168 struct r5l_log *log = conf->log;
3169
3170 md_unregister_thread(mddev: conf->mddev, threadp: &log->reclaim_thread);
3171
3172 /*
3173 * 'reconfig_mutex' is held by caller, set 'confg->log' to NULL to
3174 * ensure disable_writeback_work wakes up and exits.
3175 */
3176 WRITE_ONCE(conf->log, NULL);
3177 wake_up(&conf->mddev->sb_wait);
3178 flush_work(work: &log->disable_writeback_work);
3179
3180 mempool_exit(pool: &log->meta_pool);
3181 bioset_exit(&log->bs);
3182 mempool_exit(pool: &log->io_pool);
3183 kmem_cache_destroy(s: log->io_kc);
3184 kfree(objp: log);
3185}
3186

source code of linux/drivers/md/raid5-cache.c