projects / linux-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
md/raid5-cache: stripe reclaim only counts valid stripes
[linux-block.git] / drivers / md / raid5-cache.c
CommitLineData
f6bed0ef
SL		 1/*
2 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
b4c625c6 3 * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
f6bed0ef
SL		 4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 */
15#include <linux/kernel.h>
16#include <linux/wait.h>
17#include <linux/blkdev.h>
18#include <linux/slab.h>
19#include <linux/raid/md_p.h>
5cb2fbd6 20#include <linux/crc32c.h>
f6bed0ef 21#include <linux/random.h>
ce1ccd07 22#include <linux/kthread.h>
03b047f4 23#include <linux/types.h>
f6bed0ef
SL		 24#include "md.h"
25#include "raid5.h"
1e6d690b 26#include "bitmap.h"
f6bed0ef
SL		 27
28/*
29 * metadata/data stored in disk with 4k size unit (a block) regardless
30 * underneath hardware sector size. only works with PAGE_SIZE == 4096
31 */
32#define BLOCK_SECTORS (8)
33
0576b1c6 34/*
a39f7afd
SL		 35 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
36 *
37 * In write through mode, the reclaim runs every log->max_free_space.
38 * This can prevent the recovery scans for too long
0576b1c6
SL		 39 */
40#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
41#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
42
a39f7afd
SL		 43/* wake up reclaim thread periodically */
44#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
45/* start flush with these full stripes */
46#define R5C_FULL_STRIPE_FLUSH_BATCH 256
47/* reclaim stripes in groups */
48#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
49
c38d29b3
CH		 50/*
51 * We only need 2 bios per I/O unit to make progress, but ensure we
52 * have a few more available to not get too tight.
53 */
54#define R5L_POOL_SIZE 4
55
2ded3703
SL		 56/*
57 * r5c journal modes of the array: write-back or write-through.
58 * write-through mode has identical behavior as existing log only
59 * implementation.
60 */
61enum r5c_journal_mode {
62 R5C_JOURNAL_MODE_WRITE_THROUGH = 0,
63 R5C_JOURNAL_MODE_WRITE_BACK = 1,
64};
65
2c7da14b
SL		 66static char *r5c_journal_mode_str[] = {"write-through",
67 "write-back"};
2ded3703
SL		 68/*
69 * raid5 cache state machine
70 *
9b69173e 71 * With the RAID cache, each stripe works in two phases:
2ded3703
SL		 72 * - caching phase
73 * - writing-out phase
74 *
75 * These two phases are controlled by bit STRIPE_R5C_CACHING:
76 * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
77 * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
78 *
79 * When there is no journal, or the journal is in write-through mode,
80 * the stripe is always in writing-out phase.
81 *
82 * For write-back journal, the stripe is sent to caching phase on write
83 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
84 * the write-out phase by clearing STRIPE_R5C_CACHING.
85 *
86 * Stripes in caching phase do not write the raid disks. Instead, all
87 * writes are committed from the log device. Therefore, a stripe in
88 * caching phase handles writes as:
89 * - write to log device
90 * - return IO
91 *
92 * Stripes in writing-out phase handle writes as:
93 * - calculate parity
94 * - write pending data and parity to journal
95 * - write data and parity to raid disks
96 * - return IO for pending writes
97 */
98
f6bed0ef
SL		 99struct r5l_log {
100 struct md_rdev *rdev;
101
102 u32 uuid_checksum;
103
104 sector_t device_size; /* log device size, round to
105 * BLOCK_SECTORS */
0576b1c6
SL		 106 sector_t max_free_space; /* reclaim run if free space is at
107 * this size */
f6bed0ef
SL		 108
109 sector_t last_checkpoint; /* log tail. where recovery scan
110 * starts from */
111 u64 last_cp_seq; /* log tail sequence */
112
113 sector_t log_start; /* log head. where new data appends */
114 u64 seq; /* log head sequence */
115
17036461 116 sector_t next_checkpoint;
17036461 117
f6bed0ef
SL		 118 struct mutex io_mutex;
119 struct r5l_io_unit *current_io; /* current io_unit accepting new data */
120
121 spinlock_t io_list_lock;
122 struct list_head running_ios; /* io_units which are still running,
123 * and have not yet been completely
124 * written to the log */
125 struct list_head io_end_ios; /* io_units which have been completely
126 * written to the log but not yet written
127 * to the RAID */
a8c34f91
SL		 128 struct list_head flushing_ios; /* io_units which are waiting for log
129 * cache flush */
04732f74 130 struct list_head finished_ios; /* io_units which settle down in log disk */
a8c34f91 131 struct bio flush_bio;
f6bed0ef 132
5036c390
CH		 133 struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */
134
f6bed0ef 135 struct kmem_cache *io_kc;
5036c390 136 mempool_t *io_pool;
c38d29b3 137 struct bio_set *bs;
e8deb638 138 mempool_t *meta_pool;
f6bed0ef 139
0576b1c6
SL		 140 struct md_thread *reclaim_thread;
141 unsigned long reclaim_target; /* number of space that need to be
142 * reclaimed. if it's 0, reclaim spaces
143 * used by io_units which are in
144 * IO_UNIT_STRIPE_END state (eg, reclaim
145 * dones't wait for specific io_unit
146 * switching to IO_UNIT_STRIPE_END
147 * state) */
0fd22b45 148 wait_queue_head_t iounit_wait;
0576b1c6 149
f6bed0ef
SL		 150 struct list_head no_space_stripes; /* pending stripes, log has no space */
151 spinlock_t no_space_stripes_lock;
56fef7c6
CH		 152
153 bool need_cache_flush;
2ded3703
SL		 154
155 /* for r5c_cache */
156 enum r5c_journal_mode r5c_journal_mode;
a39f7afd
SL		 157
158 /* all stripes in r5cache, in the order of seq at sh->log_start */
159 struct list_head stripe_in_journal_list;
160
161 spinlock_t stripe_in_journal_lock;
162 atomic_t stripe_in_journal_count;
3bddb7f8
SL		 163
164 /* to submit async io_units, to fulfill ordering of flush */
165 struct work_struct deferred_io_work;
2e38a37f
SL		 166 /* to disable write back during in degraded mode */
167 struct work_struct disable_writeback_work;
03b047f4
SL		 168
169 /* to for chunk_aligned_read in writeback mode, details below */
170 spinlock_t tree_lock;
171 struct radix_tree_root big_stripe_tree;
f6bed0ef
SL		 172};
173
03b047f4
SL		 174/*
175 * Enable chunk_aligned_read() with write back cache.
176 *
177 * Each chunk may contain more than one stripe (for example, a 256kB
178 * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
179 * chunk_aligned_read, these stripes are grouped into one "big_stripe".
180 * For each big_stripe, we count how many stripes of this big_stripe
181 * are in the write back cache. These data are tracked in a radix tree
182 * (big_stripe_tree). We use radix_tree item pointer as the counter.
183 * r5c_tree_index() is used to calculate keys for the radix tree.
184 *
185 * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
186 * big_stripe of each chunk in the tree. If this big_stripe is in the
187 * tree, chunk_aligned_read() aborts. This look up is protected by
188 * rcu_read_lock().
189 *
190 * It is necessary to remember whether a stripe is counted in
191 * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
192 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
193 * two flags are set, the stripe is counted in big_stripe_tree. This
194 * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
195 * r5c_try_caching_write(); and moving clear_bit of
196 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
197 * r5c_finish_stripe_write_out().
198 */
199
200/*
201 * radix tree requests lowest 2 bits of data pointer to be 2b'00.
202 * So it is necessary to left shift the counter by 2 bits before using it
203 * as data pointer of the tree.
204 */
205#define R5C_RADIX_COUNT_SHIFT 2
206
207/*
208 * calculate key for big_stripe_tree
209 *
210 * sect: align_bi->bi_iter.bi_sector or sh->sector
211 */
212static inline sector_t r5c_tree_index(struct r5conf *conf,
213 sector_t sect)
214{
215 sector_t offset;
216
217 offset = sector_div(sect, conf->chunk_sectors);
218 return sect;
219}
220
f6bed0ef
SL		 221/*
222 * an IO range starts from a meta data block and end at the next meta data
223 * block. The io unit's the meta data block tracks data/parity followed it. io
224 * unit is written to log disk with normal write, as we always flush log disk
225 * first and then start move data to raid disks, there is no requirement to
226 * write io unit with FLUSH/FUA
227 */
228struct r5l_io_unit {
229 struct r5l_log *log;
230
231 struct page *meta_page; /* store meta block */
232 int meta_offset; /* current offset in meta_page */
233
f6bed0ef
SL		 234 struct bio *current_bio;/* current_bio accepting new data */
235
236 atomic_t pending_stripe;/* how many stripes not flushed to raid */
237 u64 seq; /* seq number of the metablock */
238 sector_t log_start; /* where the io_unit starts */
239 sector_t log_end; /* where the io_unit ends */
240 struct list_head log_sibling; /* log->running_ios */
241 struct list_head stripe_list; /* stripes added to the io_unit */
242
243 int state;
6143e2ce 244 bool need_split_bio;
3bddb7f8
SL		 245 struct bio *split_bio;
246
247 unsigned int has_flush:1; /* include flush request */
248 unsigned int has_fua:1; /* include fua request */
249 unsigned int has_null_flush:1; /* include empty flush request */
250 /*
251 * io isn't sent yet, flush/fua request can only be submitted till it's
252 * the first IO in running_ios list
253 */
254 unsigned int io_deferred:1;
255
256 struct bio_list flush_barriers; /* size == 0 flush bios */
f6bed0ef
SL		 257};
258
259/* r5l_io_unit state */
260enum r5l_io_unit_state {
261 IO_UNIT_RUNNING = 0, /* accepting new IO */
262 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
263 * don't accepting new bio */
264 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
a8c34f91 265 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
f6bed0ef
SL		 266};
267
2ded3703
SL		 268bool r5c_is_writeback(struct r5l_log *log)
269{
270 return (log != NULL &&
271 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
272}
273
f6bed0ef
SL		 274static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
275{
276 start += inc;
277 if (start >= log->device_size)
278 start = start - log->device_size;
279 return start;
280}
281
282static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
283 sector_t end)
284{
285 if (end >= start)
286 return end - start;
287 else
288 return end + log->device_size - start;
289}
290
291static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
292{
293 sector_t used_size;
294
295 used_size = r5l_ring_distance(log, log->last_checkpoint,
296 log->log_start);
297
298 return log->device_size > used_size + size;
299}
300
f6bed0ef
SL		 301static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
302 enum r5l_io_unit_state state)
303{
f6bed0ef
SL		 304 if (WARN_ON(io->state >= state))
305 return;
306 io->state = state;
f6bed0ef
SL		 307}
308
1e6d690b
SL		 309static void
310r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev,
311 struct bio_list *return_bi)
312{
313 struct bio *wbi, *wbi2;
314
315 wbi = dev->written;
316 dev->written = NULL;
317 while (wbi && wbi->bi_iter.bi_sector <
318 dev->sector + STRIPE_SECTORS) {
319 wbi2 = r5_next_bio(wbi, dev->sector);
320 if (!raid5_dec_bi_active_stripes(wbi)) {
321 md_write_end(conf->mddev);
322 bio_list_add(return_bi, wbi);
323 }
324 wbi = wbi2;
325 }
326}
327
328void r5c_handle_cached_data_endio(struct r5conf *conf,
329 struct stripe_head *sh, int disks, struct bio_list *return_bi)
330{
331 int i;
332
333 for (i = sh->disks; i--; ) {
334 if (sh->dev[i].written) {
335 set_bit(R5_UPTODATE, &sh->dev[i].flags);
336 r5c_return_dev_pending_writes(conf, &sh->dev[i],
337 return_bi);
338 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
339 STRIPE_SECTORS,
340 !test_bit(STRIPE_DEGRADED, &sh->state),
341 0);
342 }
343 }
344}
345
a39f7afd
SL		 346/* Check whether we should flush some stripes to free up stripe cache */
347void r5c_check_stripe_cache_usage(struct r5conf *conf)
348{
349 int total_cached;
350
351 if (!r5c_is_writeback(conf->log))
352 return;
353
354 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
355 atomic_read(&conf->r5c_cached_full_stripes);
356
357 /*
358 * The following condition is true for either of the following:
359 * - stripe cache pressure high:
360 * total_cached > 3/4 min_nr_stripes ||
361 * empty_inactive_list_nr > 0
362 * - stripe cache pressure moderate:
363 * total_cached > 1/2 min_nr_stripes
364 */
365 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
366 atomic_read(&conf->empty_inactive_list_nr) > 0)
367 r5l_wake_reclaim(conf->log, 0);
368}
369
370/*
371 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
372 * stripes in the cache
373 */
374void r5c_check_cached_full_stripe(struct r5conf *conf)
375{
376 if (!r5c_is_writeback(conf->log))
377 return;
378
379 /*
380 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
381 * or a full stripe (chunk size / 4k stripes).
382 */
383 if (atomic_read(&conf->r5c_cached_full_stripes) >=
384 min(R5C_FULL_STRIPE_FLUSH_BATCH,
385 conf->chunk_sectors >> STRIPE_SHIFT))
386 r5l_wake_reclaim(conf->log, 0);
387}
388
389/*
390 * Total log space (in sectors) needed to flush all data in cache
391 *
39b99586
SL		 392 * To avoid deadlock due to log space, it is necessary to reserve log
393 * space to flush critical stripes (stripes that occupying log space near
394 * last_checkpoint). This function helps check how much log space is
395 * required to flush all cached stripes.
a39f7afd 396 *
39b99586
SL		 397 * To reduce log space requirements, two mechanisms are used to give cache
398 * flush higher priorities:
399 * 1. In handle_stripe_dirtying() and schedule_reconstruction(),
400 * stripes ALREADY in journal can be flushed w/o pending writes;
401 * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
402 * can be delayed (r5l_add_no_space_stripe).
a39f7afd 403 *
39b99586
SL		 404 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
405 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
406 * pages of journal space. For stripes that has not passed 1, flushing it
407 * requires (conf->raid_disks + 1) pages of journal space. There are at
408 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
409 * required to flush all cached stripes (in pages) is:
410 *
411 * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
412 * (group_cnt + 1) * (raid_disks + 1)
413 * or
414 * (stripe_in_journal_count) * (max_degraded + 1) +
415 * (group_cnt + 1) * (raid_disks - max_degraded)
a39f7afd
SL		 416 */
417static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
418{
419 struct r5l_log *log = conf->log;
420
421 if (!r5c_is_writeback(log))
422 return 0;
423
39b99586
SL		 424 return BLOCK_SECTORS *
425 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
426 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
a39f7afd
SL		 427}
428
429/*
430 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
431 *
432 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
433 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
434 * device is less than 2x of reclaim_required_space.
435 */
436static inline void r5c_update_log_state(struct r5l_log *log)
437{
438 struct r5conf *conf = log->rdev->mddev->private;
439 sector_t free_space;
440 sector_t reclaim_space;
f687a33e 441 bool wake_reclaim = false;
a39f7afd
SL		 442
443 if (!r5c_is_writeback(log))
444 return;
445
446 free_space = r5l_ring_distance(log, log->log_start,
447 log->last_checkpoint);
448 reclaim_space = r5c_log_required_to_flush_cache(conf);
449 if (free_space < 2 * reclaim_space)
450 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
f687a33e
SL		 451 else {
452 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
453 wake_reclaim = true;
a39f7afd 454 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
f687a33e 455 }
a39f7afd
SL		 456 if (free_space < 3 * reclaim_space)
457 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
458 else
459 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
f687a33e
SL		 460
461 if (wake_reclaim)
462 r5l_wake_reclaim(log, 0);
a39f7afd
SL		 463}
464
2ded3703
SL		 465/*
466 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
467 * This function should only be called in write-back mode.
468 */
a39f7afd 469void r5c_make_stripe_write_out(struct stripe_head *sh)
2ded3703
SL		 470{
471 struct r5conf *conf = sh->raid_conf;
472 struct r5l_log *log = conf->log;
473
474 BUG_ON(!r5c_is_writeback(log));
475
476 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
477 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1e6d690b
SL		 478
479 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
480 atomic_inc(&conf->preread_active_stripes);
1e6d690b
SL		 481}
482
483static void r5c_handle_data_cached(struct stripe_head *sh)
484{
485 int i;
486
487 for (i = sh->disks; i--; )
488 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
489 set_bit(R5_InJournal, &sh->dev[i].flags);
490 clear_bit(R5_LOCKED, &sh->dev[i].flags);
491 }
492 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
493}
494
495/*
496 * this journal write must contain full parity,
497 * it may also contain some data pages
498 */
499static void r5c_handle_parity_cached(struct stripe_head *sh)
500{
501 int i;
502
503 for (i = sh->disks; i--; )
504 if (test_bit(R5_InJournal, &sh->dev[i].flags))
505 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2ded3703
SL		 506}
507
508/*
509 * Setting proper flags after writing (or flushing) data and/or parity to the
510 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
511 */
512static void r5c_finish_cache_stripe(struct stripe_head *sh)
513{
514 struct r5l_log *log = sh->raid_conf->log;
515
516 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
517 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
518 /*
519 * Set R5_InJournal for parity dev[pd_idx]. This means
520 * all data AND parity in the journal. For RAID 6, it is
521 * NOT necessary to set the flag for dev[qd_idx], as the
522 * two parities are written out together.
523 */
524 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
1e6d690b
SL		 525 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
526 r5c_handle_data_cached(sh);
527 } else {
528 r5c_handle_parity_cached(sh);
529 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
530 }
2ded3703
SL		 531}
532
d8858f43
CH		 533static void r5l_io_run_stripes(struct r5l_io_unit *io)
534{
535 struct stripe_head *sh, *next;
536
537 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
538 list_del_init(&sh->log_list);
2ded3703
SL		 539
540 r5c_finish_cache_stripe(sh);
541
d8858f43
CH		 542 set_bit(STRIPE_HANDLE, &sh->state);
543 raid5_release_stripe(sh);
544 }
545}
546
56fef7c6
CH		 547static void r5l_log_run_stripes(struct r5l_log *log)
548{
549 struct r5l_io_unit *io, *next;
550
551 assert_spin_locked(&log->io_list_lock);
552
553 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
554 /* don't change list order */
555 if (io->state < IO_UNIT_IO_END)
556 break;
557
558 list_move_tail(&io->log_sibling, &log->finished_ios);
559 r5l_io_run_stripes(io);
560 }
561}
562
3848c0bc
CH		 563static void r5l_move_to_end_ios(struct r5l_log *log)
564{
565 struct r5l_io_unit *io, *next;
566
567 assert_spin_locked(&log->io_list_lock);
568
569 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
570 /* don't change list order */
571 if (io->state < IO_UNIT_IO_END)
572 break;
573 list_move_tail(&io->log_sibling, &log->io_end_ios);
574 }
575}
576
3bddb7f8 577static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
f6bed0ef
SL		 578static void r5l_log_endio(struct bio *bio)
579{
580 struct r5l_io_unit *io = bio->bi_private;
3bddb7f8 581 struct r5l_io_unit *io_deferred;
f6bed0ef 582 struct r5l_log *log = io->log;
509ffec7 583 unsigned long flags;
f6bed0ef 584
6e74a9cf
SL		 585 if (bio->bi_error)
586 md_error(log->rdev->mddev, log->rdev);
587
f6bed0ef 588 bio_put(bio);
e8deb638 589 mempool_free(io->meta_page, log->meta_pool);
f6bed0ef 590
509ffec7
CH		 591 spin_lock_irqsave(&log->io_list_lock, flags);
592 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
56fef7c6 593 if (log->need_cache_flush)
3848c0bc 594 r5l_move_to_end_ios(log);
56fef7c6
CH		 595 else
596 r5l_log_run_stripes(log);
3bddb7f8
SL		 597 if (!list_empty(&log->running_ios)) {
598 /*
599 * FLUSH/FUA io_unit is deferred because of ordering, now we
600 * can dispatch it
601 */
602 io_deferred = list_first_entry(&log->running_ios,
603 struct r5l_io_unit, log_sibling);
604 if (io_deferred->io_deferred)
605 schedule_work(&log->deferred_io_work);
606 }
607
509ffec7
CH		 608 spin_unlock_irqrestore(&log->io_list_lock, flags);
609
56fef7c6
CH		 610 if (log->need_cache_flush)
611 md_wakeup_thread(log->rdev->mddev->thread);
3bddb7f8
SL		 612
613 if (io->has_null_flush) {
614 struct bio *bi;
615
616 WARN_ON(bio_list_empty(&io->flush_barriers));
617 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
618 bio_endio(bi);
619 atomic_dec(&io->pending_stripe);
620 }
621 if (atomic_read(&io->pending_stripe) == 0)
622 __r5l_stripe_write_finished(io);
623 }
624}
625
626static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
627{
628 unsigned long flags;
629
630 spin_lock_irqsave(&log->io_list_lock, flags);
631 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
632 spin_unlock_irqrestore(&log->io_list_lock, flags);
633
634 if (io->has_flush)
20737738 635 io->current_bio->bi_opf |= REQ_PREFLUSH;
3bddb7f8 636 if (io->has_fua)
20737738 637 io->current_bio->bi_opf |= REQ_FUA;
3bddb7f8
SL		 638 submit_bio(io->current_bio);
639
640 if (!io->split_bio)
641 return;
642
643 if (io->has_flush)
20737738 644 io->split_bio->bi_opf |= REQ_PREFLUSH;
3bddb7f8 645 if (io->has_fua)
20737738 646 io->split_bio->bi_opf |= REQ_FUA;
3bddb7f8
SL		 647 submit_bio(io->split_bio);
648}
649
650/* deferred io_unit will be dispatched here */
651static void r5l_submit_io_async(struct work_struct *work)
652{
653 struct r5l_log *log = container_of(work, struct r5l_log,
654 deferred_io_work);
655 struct r5l_io_unit *io = NULL;
656 unsigned long flags;
657
658 spin_lock_irqsave(&log->io_list_lock, flags);
659 if (!list_empty(&log->running_ios)) {
660 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
661 log_sibling);
662 if (!io->io_deferred)
663 io = NULL;
664 else
665 io->io_deferred = 0;
666 }
667 spin_unlock_irqrestore(&log->io_list_lock, flags);
668 if (io)
669 r5l_do_submit_io(log, io);
f6bed0ef
SL		 670}
671
2e38a37f
SL		 672static void r5c_disable_writeback_async(struct work_struct *work)
673{
674 struct r5l_log *log = container_of(work, struct r5l_log,
675 disable_writeback_work);
676 struct mddev *mddev = log->rdev->mddev;
677
678 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
679 return;
680 pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
681 mdname(mddev));
682 mddev_suspend(mddev);
683 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
684 mddev_resume(mddev);
685}
686
f6bed0ef
SL		 687static void r5l_submit_current_io(struct r5l_log *log)
688{
689 struct r5l_io_unit *io = log->current_io;
3bddb7f8 690 struct bio *bio;
f6bed0ef 691 struct r5l_meta_block *block;
509ffec7 692 unsigned long flags;
f6bed0ef 693 u32 crc;
3bddb7f8 694 bool do_submit = true;
f6bed0ef
SL		 695
696 if (!io)
697 return;
698
699 block = page_address(io->meta_page);
700 block->meta_size = cpu_to_le32(io->meta_offset);
5cb2fbd6 701 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
f6bed0ef 702 block->checksum = cpu_to_le32(crc);
3bddb7f8 703 bio = io->current_bio;
f6bed0ef
SL		 704
705 log->current_io = NULL;
509ffec7 706 spin_lock_irqsave(&log->io_list_lock, flags);
3bddb7f8
SL		 707 if (io->has_flush || io->has_fua) {
708 if (io != list_first_entry(&log->running_ios,
709 struct r5l_io_unit, log_sibling)) {
710 io->io_deferred = 1;
711 do_submit = false;
712 }
713 }
509ffec7 714 spin_unlock_irqrestore(&log->io_list_lock, flags);
3bddb7f8
SL		 715 if (do_submit)
716 r5l_do_submit_io(log, io);
f6bed0ef
SL		 717}
718
6143e2ce 719static struct bio *r5l_bio_alloc(struct r5l_log *log)
b349feb3 720{
c38d29b3 721 struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
b349feb3 722
796a5cf0 723 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
b349feb3 724 bio->bi_bdev = log->rdev->bdev;
1e932a37 725 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
b349feb3 726
b349feb3
CH		 727 return bio;
728}
729
c1b99198
CH		 730static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
731{
732 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
733
a39f7afd 734 r5c_update_log_state(log);
c1b99198
CH		 735 /*
736 * If we filled up the log device start from the beginning again,
737 * which will require a new bio.
738 *
739 * Note: for this to work properly the log size needs to me a multiple
740 * of BLOCK_SECTORS.
741 */
742 if (log->log_start == 0)
6143e2ce 743 io->need_split_bio = true;
c1b99198
CH		 744
745 io->log_end = log->log_start;
746}
747
f6bed0ef
SL		 748static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
749{
750 struct r5l_io_unit *io;
751 struct r5l_meta_block *block;
f6bed0ef 752
5036c390
CH		 753 io = mempool_alloc(log->io_pool, GFP_ATOMIC);
754 if (!io)
755 return NULL;
756 memset(io, 0, sizeof(*io));
757
51039cd0 758 io->log = log;
51039cd0
CH		 759 INIT_LIST_HEAD(&io->log_sibling);
760 INIT_LIST_HEAD(&io->stripe_list);
3bddb7f8 761 bio_list_init(&io->flush_barriers);
51039cd0 762 io->state = IO_UNIT_RUNNING;
f6bed0ef 763
e8deb638 764 io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
f6bed0ef 765 block = page_address(io->meta_page);
e8deb638 766 clear_page(block);
f6bed0ef
SL		 767 block->magic = cpu_to_le32(R5LOG_MAGIC);
768 block->version = R5LOG_VERSION;
769 block->seq = cpu_to_le64(log->seq);
770 block->position = cpu_to_le64(log->log_start);
771
772 io->log_start = log->log_start;
773 io->meta_offset = sizeof(struct r5l_meta_block);
2b8ef16e 774 io->seq = log->seq++;
f6bed0ef 775
6143e2ce
CH		 776 io->current_bio = r5l_bio_alloc(log);
777 io->current_bio->bi_end_io = r5l_log_endio;
778 io->current_bio->bi_private = io;
b349feb3 779 bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
f6bed0ef 780
c1b99198 781 r5_reserve_log_entry(log, io);
f6bed0ef
SL		 782
783 spin_lock_irq(&log->io_list_lock);
784 list_add_tail(&io->log_sibling, &log->running_ios);
785 spin_unlock_irq(&log->io_list_lock);
786
787 return io;
788}
789
790static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
791{
22581f58
CH		 792 if (log->current_io &&
793 log->current_io->meta_offset + payload_size > PAGE_SIZE)
f6bed0ef 794 r5l_submit_current_io(log);
f6bed0ef 795
5036c390 796 if (!log->current_io) {
22581f58 797 log->current_io = r5l_new_meta(log);
5036c390
CH		 798 if (!log->current_io)
799 return -ENOMEM;
800 }
801
f6bed0ef
SL		 802 return 0;
803}
804
805static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
806 sector_t location,
807 u32 checksum1, u32 checksum2,
808 bool checksum2_valid)
809{
810 struct r5l_io_unit *io = log->current_io;
811 struct r5l_payload_data_parity *payload;
812
813 payload = page_address(io->meta_page) + io->meta_offset;
814 payload->header.type = cpu_to_le16(type);
815 payload->header.flags = cpu_to_le16(0);
816 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
817 (PAGE_SHIFT - 9));
818 payload->location = cpu_to_le64(location);
819 payload->checksum[0] = cpu_to_le32(checksum1);
820 if (checksum2_valid)
821 payload->checksum[1] = cpu_to_le32(checksum2);
822
823 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
824 sizeof(__le32) * (1 + !!checksum2_valid);
825}
826
827static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
828{
829 struct r5l_io_unit *io = log->current_io;
830
6143e2ce 831 if (io->need_split_bio) {
3bddb7f8
SL		 832 BUG_ON(io->split_bio);
833 io->split_bio = io->current_bio;
6143e2ce 834 io->current_bio = r5l_bio_alloc(log);
3bddb7f8
SL		 835 bio_chain(io->current_bio, io->split_bio);
836 io->need_split_bio = false;
f6bed0ef 837 }
f6bed0ef 838
6143e2ce
CH		 839 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
840 BUG();
841
c1b99198 842 r5_reserve_log_entry(log, io);
f6bed0ef
SL		 843}
844
5036c390 845static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
f6bed0ef
SL		 846 int data_pages, int parity_pages)
847{
848 int i;
849 int meta_size;
5036c390 850 int ret;
f6bed0ef
SL		 851 struct r5l_io_unit *io;
852
853 meta_size =
854 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
855 * data_pages) +
856 sizeof(struct r5l_payload_data_parity) +
857 sizeof(__le32) * parity_pages;
858
5036c390
CH		 859 ret = r5l_get_meta(log, meta_size);
860 if (ret)
861 return ret;
862
f6bed0ef
SL		 863 io = log->current_io;
864
3bddb7f8
SL		 865 if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
866 io->has_flush = 1;
867
f6bed0ef 868 for (i = 0; i < sh->disks; i++) {
1e6d690b
SL		 869 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
870 test_bit(R5_InJournal, &sh->dev[i].flags))
f6bed0ef
SL		 871 continue;
872 if (i == sh->pd_idx || i == sh->qd_idx)
873 continue;
3bddb7f8
SL		 874 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
875 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
876 io->has_fua = 1;
877 /*
878 * we need to flush journal to make sure recovery can
879 * reach the data with fua flag
880 */
881 io->has_flush = 1;
882 }
f6bed0ef
SL		 883 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
884 raid5_compute_blocknr(sh, i, 0),
885 sh->dev[i].log_checksum, 0, false);
886 r5l_append_payload_page(log, sh->dev[i].page);
887 }
888
2ded3703 889 if (parity_pages == 2) {
f6bed0ef
SL		 890 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
891 sh->sector, sh->dev[sh->pd_idx].log_checksum,
892 sh->dev[sh->qd_idx].log_checksum, true);
893 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
894 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
2ded3703 895 } else if (parity_pages == 1) {
f6bed0ef
SL		 896 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
897 sh->sector, sh->dev[sh->pd_idx].log_checksum,
898 0, false);
899 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
2ded3703
SL		 900 } else /* Just writing data, not parity, in caching phase */
901 BUG_ON(parity_pages != 0);
f6bed0ef
SL		 902
903 list_add_tail(&sh->log_list, &io->stripe_list);
904 atomic_inc(&io->pending_stripe);
905 sh->log_io = io;
5036c390 906
a39f7afd
SL		 907 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
908 return 0;
909
910 if (sh->log_start == MaxSector) {
911 BUG_ON(!list_empty(&sh->r5c));
912 sh->log_start = io->log_start;
913 spin_lock_irq(&log->stripe_in_journal_lock);
914 list_add_tail(&sh->r5c,
915 &log->stripe_in_journal_list);
916 spin_unlock_irq(&log->stripe_in_journal_lock);
917 atomic_inc(&log->stripe_in_journal_count);
918 }
5036c390 919 return 0;
f6bed0ef
SL		 920}
921
a39f7afd
SL		 922/* add stripe to no_space_stripes, and then wake up reclaim */
923static inline void r5l_add_no_space_stripe(struct r5l_log *log,
924 struct stripe_head *sh)
925{
926 spin_lock(&log->no_space_stripes_lock);
927 list_add_tail(&sh->log_list, &log->no_space_stripes);
928 spin_unlock(&log->no_space_stripes_lock);
929}
930
f6bed0ef
SL		 931/*
932 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
933 * data from log to raid disks), so we shouldn't wait for reclaim here
934 */
935int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
936{
a39f7afd 937 struct r5conf *conf = sh->raid_conf;
f6bed0ef
SL		 938 int write_disks = 0;
939 int data_pages, parity_pages;
f6bed0ef
SL		 940 int reserve;
941 int i;
5036c390 942 int ret = 0;
a39f7afd 943 bool wake_reclaim = false;
f6bed0ef
SL		 944
945 if (!log)
946 return -EAGAIN;
947 /* Don't support stripe batch */
948 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
949 test_bit(STRIPE_SYNCING, &sh->state)) {
950 /* the stripe is written to log, we start writing it to raid */
951 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
952 return -EAGAIN;
953 }
954
2ded3703
SL		 955 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
956
f6bed0ef
SL		 957 for (i = 0; i < sh->disks; i++) {
958 void *addr;
959
1e6d690b
SL		 960 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
961 test_bit(R5_InJournal, &sh->dev[i].flags))
f6bed0ef 962 continue;
1e6d690b 963
f6bed0ef
SL		 964 write_disks++;
965 /* checksum is already calculated in last run */
966 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
967 continue;
968 addr = kmap_atomic(sh->dev[i].page);
5cb2fbd6
SL		 969 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
970 addr, PAGE_SIZE);
f6bed0ef
SL		 971 kunmap_atomic(addr);
972 }
973 parity_pages = 1 + !!(sh->qd_idx >= 0);
974 data_pages = write_disks - parity_pages;
975
f6bed0ef 976 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
253f9fd4
SL		 977 /*
978 * The stripe must enter state machine again to finish the write, so
979 * don't delay.
980 */
981 clear_bit(STRIPE_DELAYED, &sh->state);
f6bed0ef
SL		 982 atomic_inc(&sh->count);
983
984 mutex_lock(&log->io_mutex);
985 /* meta + data */
986 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
f6bed0ef 987
a39f7afd
SL		 988 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
989 if (!r5l_has_free_space(log, reserve)) {
990 r5l_add_no_space_stripe(log, sh);
991 wake_reclaim = true;
992 } else {
993 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
994 if (ret) {
995 spin_lock_irq(&log->io_list_lock);
996 list_add_tail(&sh->log_list,
997 &log->no_mem_stripes);
998 spin_unlock_irq(&log->io_list_lock);
999 }
1000 }
1001 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1002 /*
1003 * log space critical, do not process stripes that are
1004 * not in cache yet (sh->log_start == MaxSector).
1005 */
1006 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1007 sh->log_start == MaxSector) {
1008 r5l_add_no_space_stripe(log, sh);
1009 wake_reclaim = true;
1010 reserve = 0;
1011 } else if (!r5l_has_free_space(log, reserve)) {
1012 if (sh->log_start == log->last_checkpoint)
1013 BUG();
1014 else
1015 r5l_add_no_space_stripe(log, sh);
1016 } else {
1017 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1018 if (ret) {
1019 spin_lock_irq(&log->io_list_lock);
1020 list_add_tail(&sh->log_list,
1021 &log->no_mem_stripes);
1022 spin_unlock_irq(&log->io_list_lock);
1023 }
5036c390 1024 }
f6bed0ef 1025 }
f6bed0ef 1026
5036c390 1027 mutex_unlock(&log->io_mutex);
a39f7afd
SL		 1028 if (wake_reclaim)
1029 r5l_wake_reclaim(log, reserve);
f6bed0ef
SL		 1030 return 0;
1031}
1032
1033void r5l_write_stripe_run(struct r5l_log *log)
1034{
1035 if (!log)
1036 return;
1037 mutex_lock(&log->io_mutex);
1038 r5l_submit_current_io(log);
1039 mutex_unlock(&log->io_mutex);
1040}
1041
828cbe98
SL		 1042int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1043{
1044 if (!log)
1045 return -ENODEV;
3bddb7f8
SL		 1046
1047 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1048 /*
1049 * in write through (journal only)
1050 * we flush log disk cache first, then write stripe data to
1051 * raid disks. So if bio is finished, the log disk cache is
1052 * flushed already. The recovery guarantees we can recovery
1053 * the bio from log disk, so we don't need to flush again
1054 */
1055 if (bio->bi_iter.bi_size == 0) {
1056 bio_endio(bio);
1057 return 0;
1058 }
1059 bio->bi_opf &= ~REQ_PREFLUSH;
1060 } else {
1061 /* write back (with cache) */
1062 if (bio->bi_iter.bi_size == 0) {
1063 mutex_lock(&log->io_mutex);
1064 r5l_get_meta(log, 0);
1065 bio_list_add(&log->current_io->flush_barriers, bio);
1066 log->current_io->has_flush = 1;
1067 log->current_io->has_null_flush = 1;
1068 atomic_inc(&log->current_io->pending_stripe);
1069 r5l_submit_current_io(log);
1070 mutex_unlock(&log->io_mutex);
1071 return 0;
1072 }
828cbe98 1073 }
828cbe98
SL		 1074 return -EAGAIN;
1075}
1076
f6bed0ef
SL		 1077/* This will run after log space is reclaimed */
1078static void r5l_run_no_space_stripes(struct r5l_log *log)
1079{
1080 struct stripe_head *sh;
1081
1082 spin_lock(&log->no_space_stripes_lock);
1083 while (!list_empty(&log->no_space_stripes)) {
1084 sh = list_first_entry(&log->no_space_stripes,
1085 struct stripe_head, log_list);
1086 list_del_init(&sh->log_list);
1087 set_bit(STRIPE_HANDLE, &sh->state);
1088 raid5_release_stripe(sh);
1089 }
1090 spin_unlock(&log->no_space_stripes_lock);
1091}
1092
a39f7afd
SL		 1093/*
1094 * calculate new last_checkpoint
1095 * for write through mode, returns log->next_checkpoint
1096 * for write back, returns log_start of first sh in stripe_in_journal_list
1097 */
1098static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1099{
1100 struct stripe_head *sh;
1101 struct r5l_log *log = conf->log;
1102 sector_t new_cp;
1103 unsigned long flags;
1104
1105 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1106 return log->next_checkpoint;
1107
1108 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1109 if (list_empty(&conf->log->stripe_in_journal_list)) {
1110 /* all stripes flushed */
d3014e21 1111 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
a39f7afd
SL		 1112 return log->next_checkpoint;
1113 }
1114 sh = list_first_entry(&conf->log->stripe_in_journal_list,
1115 struct stripe_head, r5c);
1116 new_cp = sh->log_start;
1117 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1118 return new_cp;
1119}
1120
17036461
CH		 1121static sector_t r5l_reclaimable_space(struct r5l_log *log)
1122{
a39f7afd
SL		 1123 struct r5conf *conf = log->rdev->mddev->private;
1124
17036461 1125 return r5l_ring_distance(log, log->last_checkpoint,
a39f7afd 1126 r5c_calculate_new_cp(conf));
17036461
CH		 1127}
1128
5036c390
CH		 1129static void r5l_run_no_mem_stripe(struct r5l_log *log)
1130{
1131 struct stripe_head *sh;
1132
1133 assert_spin_locked(&log->io_list_lock);
1134
1135 if (!list_empty(&log->no_mem_stripes)) {
1136 sh = list_first_entry(&log->no_mem_stripes,
1137 struct stripe_head, log_list);
1138 list_del_init(&sh->log_list);
1139 set_bit(STRIPE_HANDLE, &sh->state);
1140 raid5_release_stripe(sh);
1141 }
1142}
1143
04732f74 1144static bool r5l_complete_finished_ios(struct r5l_log *log)
17036461
CH		 1145{
1146 struct r5l_io_unit *io, *next;
1147 bool found = false;
1148
1149 assert_spin_locked(&log->io_list_lock);
1150
04732f74 1151 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
17036461
CH		 1152 /* don't change list order */
1153 if (io->state < IO_UNIT_STRIPE_END)
1154 break;
1155
1156 log->next_checkpoint = io->log_start;
17036461
CH		 1157
1158 list_del(&io->log_sibling);
5036c390
CH		 1159 mempool_free(io, log->io_pool);
1160 r5l_run_no_mem_stripe(log);
17036461
CH		 1161
1162 found = true;
1163 }
1164
1165 return found;
1166}
1167
509ffec7
CH		 1168static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1169{
1170 struct r5l_log *log = io->log;
a39f7afd 1171 struct r5conf *conf = log->rdev->mddev->private;
509ffec7
CH		 1172 unsigned long flags;
1173
1174 spin_lock_irqsave(&log->io_list_lock, flags);
1175 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
17036461 1176
04732f74 1177 if (!r5l_complete_finished_ios(log)) {
85f2f9a4
SL		 1178 spin_unlock_irqrestore(&log->io_list_lock, flags);
1179 return;
1180 }
509ffec7 1181
a39f7afd
SL		 1182 if (r5l_reclaimable_space(log) > log->max_free_space ||
1183 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
509ffec7
CH		 1184 r5l_wake_reclaim(log, 0);
1185
509ffec7
CH		 1186 spin_unlock_irqrestore(&log->io_list_lock, flags);
1187 wake_up(&log->iounit_wait);
1188}
1189
0576b1c6
SL		 1190void r5l_stripe_write_finished(struct stripe_head *sh)
1191{
1192 struct r5l_io_unit *io;
1193
0576b1c6 1194 io = sh->log_io;
0576b1c6
SL		 1195 sh->log_io = NULL;
1196
509ffec7
CH		 1197 if (io && atomic_dec_and_test(&io->pending_stripe))
1198 __r5l_stripe_write_finished(io);
0576b1c6
SL		 1199}
1200
a8c34f91
SL		 1201static void r5l_log_flush_endio(struct bio *bio)
1202{
1203 struct r5l_log *log = container_of(bio, struct r5l_log,
1204 flush_bio);
1205 unsigned long flags;
1206 struct r5l_io_unit *io;
a8c34f91 1207
6e74a9cf
SL		 1208 if (bio->bi_error)
1209 md_error(log->rdev->mddev, log->rdev);
1210
a8c34f91 1211 spin_lock_irqsave(&log->io_list_lock, flags);
d8858f43
CH		 1212 list_for_each_entry(io, &log->flushing_ios, log_sibling)
1213 r5l_io_run_stripes(io);
04732f74 1214 list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
a8c34f91
SL		 1215 spin_unlock_irqrestore(&log->io_list_lock, flags);
1216}
1217
0576b1c6
SL		 1218/*
1219 * Starting dispatch IO to raid.
1220 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1221 * broken meta in the middle of a log causes recovery can't find meta at the
1222 * head of log. If operations require meta at the head persistent in log, we
1223 * must make sure meta before it persistent in log too. A case is:
1224 *
1225 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1226 * data/parity must be persistent in log before we do the write to raid disks.
1227 *
1228 * The solution is we restrictly maintain io_unit list order. In this case, we
1229 * only write stripes of an io_unit to raid disks till the io_unit is the first
1230 * one whose data/parity is in log.
1231 */
1232void r5l_flush_stripe_to_raid(struct r5l_log *log)
1233{
a8c34f91 1234 bool do_flush;
56fef7c6
CH		 1235
1236 if (!log || !log->need_cache_flush)
0576b1c6 1237 return;
0576b1c6
SL		 1238
1239 spin_lock_irq(&log->io_list_lock);
a8c34f91
SL		 1240 /* flush bio is running */
1241 if (!list_empty(&log->flushing_ios)) {
1242 spin_unlock_irq(&log->io_list_lock);
1243 return;
0576b1c6 1244 }
a8c34f91
SL		 1245 list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1246 do_flush = !list_empty(&log->flushing_ios);
0576b1c6 1247 spin_unlock_irq(&log->io_list_lock);
a8c34f91
SL		 1248
1249 if (!do_flush)
1250 return;
1251 bio_reset(&log->flush_bio);
1252 log->flush_bio.bi_bdev = log->rdev->bdev;
1253 log->flush_bio.bi_end_io = r5l_log_flush_endio;
70fd7614 1254 log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
4e49ea4a 1255 submit_bio(&log->flush_bio);
0576b1c6
SL		 1256}
1257
0576b1c6 1258static void r5l_write_super(struct r5l_log *log, sector_t cp);
4b482044
SL		 1259static void r5l_write_super_and_discard_space(struct r5l_log *log,
1260 sector_t end)
1261{
1262 struct block_device *bdev = log->rdev->bdev;
1263 struct mddev *mddev;
1264
1265 r5l_write_super(log, end);
1266
1267 if (!blk_queue_discard(bdev_get_queue(bdev)))
1268 return;
1269
1270 mddev = log->rdev->mddev;
1271 /*
8e018c21
SL		 1272 * Discard could zero data, so before discard we must make sure
1273 * superblock is updated to new log tail. Updating superblock (either
1274 * directly call md_update_sb() or depend on md thread) must hold
1275 * reconfig mutex. On the other hand, raid5_quiesce is called with
1276 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1277 * for all IO finish, hence waitting for reclaim thread, while reclaim
1278 * thread is calling this function and waitting for reconfig mutex. So
1279 * there is a deadlock. We workaround this issue with a trylock.
1280 * FIXME: we could miss discard if we can't take reconfig mutex
4b482044 1281 */
2953079c
SL		 1282 set_mask_bits(&mddev->sb_flags, 0,
1283 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
8e018c21
SL		 1284 if (!mddev_trylock(mddev))
1285 return;
1286 md_update_sb(mddev, 1);
1287 mddev_unlock(mddev);
4b482044 1288
6e74a9cf 1289 /* discard IO error really doesn't matter, ignore it */
4b482044
SL		 1290 if (log->last_checkpoint < end) {
1291 blkdev_issue_discard(bdev,
1292 log->last_checkpoint + log->rdev->data_offset,
1293 end - log->last_checkpoint, GFP_NOIO, 0);
1294 } else {
1295 blkdev_issue_discard(bdev,
1296 log->last_checkpoint + log->rdev->data_offset,
1297 log->device_size - log->last_checkpoint,
1298 GFP_NOIO, 0);
1299 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1300 GFP_NOIO, 0);
1301 }
1302}
1303
a39f7afd
SL		 1304/*
1305 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1306 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1307 *
1308 * must hold conf->device_lock
1309 */
1310static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
0576b1c6 1311{
a39f7afd
SL		 1312 BUG_ON(list_empty(&sh->lru));
1313 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1314 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
0576b1c6 1315
0576b1c6 1316 /*
a39f7afd
SL		 1317 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1318 * raid5_release_stripe() while holding conf->device_lock
0576b1c6 1319 */
a39f7afd
SL		 1320 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1321 assert_spin_locked(&conf->device_lock);
0576b1c6 1322
a39f7afd
SL		 1323 list_del_init(&sh->lru);
1324 atomic_inc(&sh->count);
17036461 1325
a39f7afd
SL		 1326 set_bit(STRIPE_HANDLE, &sh->state);
1327 atomic_inc(&conf->active_stripes);
1328 r5c_make_stripe_write_out(sh);
0576b1c6 1329
a39f7afd
SL		 1330 raid5_release_stripe(sh);
1331}
1332
1333/*
1334 * if num == 0, flush all full stripes
1335 * if num > 0, flush all full stripes. If less than num full stripes are
1336 * flushed, flush some partial stripes until totally num stripes are
1337 * flushed or there is no more cached stripes.
1338 */
1339void r5c_flush_cache(struct r5conf *conf, int num)
1340{
1341 int count;
1342 struct stripe_head *sh, *next;
1343
1344 assert_spin_locked(&conf->device_lock);
1345 if (!conf->log)
0576b1c6
SL		 1346 return;
1347
a39f7afd
SL		 1348 count = 0;
1349 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1350 r5c_flush_stripe(conf, sh);
1351 count++;
1352 }
1353
1354 if (count >= num)
1355 return;
1356 list_for_each_entry_safe(sh, next,
1357 &conf->r5c_partial_stripe_list, lru) {
1358 r5c_flush_stripe(conf, sh);
1359 if (++count >= num)
1360 break;
1361 }
1362}
1363
1364static void r5c_do_reclaim(struct r5conf *conf)
1365{
1366 struct r5l_log *log = conf->log;
1367 struct stripe_head *sh;
1368 int count = 0;
1369 unsigned long flags;
1370 int total_cached;
1371 int stripes_to_flush;
1372
1373 if (!r5c_is_writeback(log))
1374 return;
1375
1376 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1377 atomic_read(&conf->r5c_cached_full_stripes);
1378
1379 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1380 atomic_read(&conf->empty_inactive_list_nr) > 0)
1381 /*
1382 * if stripe cache pressure high, flush all full stripes and
1383 * some partial stripes
1384 */
1385 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1386 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1387 atomic_read(&conf->r5c_cached_full_stripes) >
1388 R5C_FULL_STRIPE_FLUSH_BATCH)
1389 /*
1390 * if stripe cache pressure moderate, or if there is many full
1391 * stripes,flush all full stripes
1392 */
1393 stripes_to_flush = 0;
1394 else
1395 /* no need to flush */
1396 stripes_to_flush = -1;
1397
1398 if (stripes_to_flush >= 0) {
1399 spin_lock_irqsave(&conf->device_lock, flags);
1400 r5c_flush_cache(conf, stripes_to_flush);
1401 spin_unlock_irqrestore(&conf->device_lock, flags);
1402 }
1403
1404 /* if log space is tight, flush stripes on stripe_in_journal_list */
1405 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1406 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1407 spin_lock(&conf->device_lock);
1408 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1409 /*
1410 * stripes on stripe_in_journal_list could be in any
1411 * state of the stripe_cache state machine. In this
1412 * case, we only want to flush stripe on
1413 * r5c_cached_full/partial_stripes. The following
1414 * condition makes sure the stripe is on one of the
1415 * two lists.
1416 */
1417 if (!list_empty(&sh->lru) &&
1418 !test_bit(STRIPE_HANDLE, &sh->state) &&
1419 atomic_read(&sh->count) == 0) {
1420 r5c_flush_stripe(conf, sh);
e8fd52ee
SL		 1421 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1422 break;
a39f7afd 1423 }
a39f7afd
SL		 1424 }
1425 spin_unlock(&conf->device_lock);
1426 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1427 }
f687a33e
SL		 1428
1429 if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1430 r5l_run_no_space_stripes(log);
1431
a39f7afd
SL		 1432 md_wakeup_thread(conf->mddev->thread);
1433}
1434
0576b1c6
SL		 1435static void r5l_do_reclaim(struct r5l_log *log)
1436{
a39f7afd 1437 struct r5conf *conf = log->rdev->mddev->private;
0576b1c6 1438 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
17036461
CH		 1439 sector_t reclaimable;
1440 sector_t next_checkpoint;
a39f7afd 1441 bool write_super;
0576b1c6
SL		 1442
1443 spin_lock_irq(&log->io_list_lock);
a39f7afd
SL		 1444 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1445 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
0576b1c6
SL		 1446 /*
1447 * move proper io_unit to reclaim list. We should not change the order.
1448 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1449 * shouldn't reuse space of an unreclaimable io_unit
1450 */
1451 while (1) {
17036461
CH		 1452 reclaimable = r5l_reclaimable_space(log);
1453 if (reclaimable >= reclaim_target ||
0576b1c6
SL		 1454 (list_empty(&log->running_ios) &&
1455 list_empty(&log->io_end_ios) &&
a8c34f91 1456 list_empty(&log->flushing_ios) &&
04732f74 1457 list_empty(&log->finished_ios)))
0576b1c6
SL		 1458 break;
1459
17036461
CH		 1460 md_wakeup_thread(log->rdev->mddev->thread);
1461 wait_event_lock_irq(log->iounit_wait,
1462 r5l_reclaimable_space(log) > reclaimable,
1463 log->io_list_lock);
0576b1c6 1464 }
17036461 1465
a39f7afd 1466 next_checkpoint = r5c_calculate_new_cp(conf);
0576b1c6
SL		 1467 spin_unlock_irq(&log->io_list_lock);
1468
a39f7afd 1469 if (reclaimable == 0 || !write_super)
0576b1c6
SL		 1470 return;
1471
0576b1c6
SL		 1472 /*
1473 * write_super will flush cache of each raid disk. We must write super
1474 * here, because the log area might be reused soon and we don't want to
1475 * confuse recovery
1476 */
4b482044 1477 r5l_write_super_and_discard_space(log, next_checkpoint);
0576b1c6
SL		 1478
1479 mutex_lock(&log->io_mutex);
17036461 1480 log->last_checkpoint = next_checkpoint;
a39f7afd 1481 r5c_update_log_state(log);
0576b1c6 1482 mutex_unlock(&log->io_mutex);
0576b1c6 1483
17036461 1484 r5l_run_no_space_stripes(log);
0576b1c6
SL		 1485}
1486
1487static void r5l_reclaim_thread(struct md_thread *thread)
1488{
1489 struct mddev *mddev = thread->mddev;
1490 struct r5conf *conf = mddev->private;
1491 struct r5l_log *log = conf->log;
1492
1493 if (!log)
1494 return;
a39f7afd 1495 r5c_do_reclaim(conf);
0576b1c6
SL		 1496 r5l_do_reclaim(log);
1497}
1498
a39f7afd 1499void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
f6bed0ef 1500{
0576b1c6
SL		 1501 unsigned long target;
1502 unsigned long new = (unsigned long)space; /* overflow in theory */
1503
a39f7afd
SL		 1504 if (!log)
1505 return;
0576b1c6
SL		 1506 do {
1507 target = log->reclaim_target;
1508 if (new < target)
1509 return;
1510 } while (cmpxchg(&log->reclaim_target, target, new) != target);
1511 md_wakeup_thread(log->reclaim_thread);
f6bed0ef
SL		 1512}
1513
e6c033f7
SL		 1514void r5l_quiesce(struct r5l_log *log, int state)
1515{
4b482044 1516 struct mddev *mddev;
e6c033f7
SL		 1517 if (!log || state == 2)
1518 return;
ce1ccd07
SL		 1519 if (state == 0)
1520 kthread_unpark(log->reclaim_thread->tsk);
1521 else if (state == 1) {
4b482044
SL		 1522 /* make sure r5l_write_super_and_discard_space exits */
1523 mddev = log->rdev->mddev;
1524 wake_up(&mddev->sb_wait);
ce1ccd07 1525 kthread_park(log->reclaim_thread->tsk);
a39f7afd 1526 r5l_wake_reclaim(log, MaxSector);
e6c033f7
SL		 1527 r5l_do_reclaim(log);
1528 }
1529}
1530
6e74a9cf
SL		 1531bool r5l_log_disk_error(struct r5conf *conf)
1532{
f6b6ec5c
SL		 1533 struct r5l_log *log;
1534 bool ret;
7dde2ad3 1535 /* don't allow write if journal disk is missing */
f6b6ec5c
SL		 1536 rcu_read_lock();
1537 log = rcu_dereference(conf->log);
1538
1539 if (!log)
1540 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1541 else
1542 ret = test_bit(Faulty, &log->rdev->flags);
1543 rcu_read_unlock();
1544 return ret;
6e74a9cf
SL		 1545}
1546
355810d1
SL		 1547struct r5l_recovery_ctx {
1548 struct page *meta_page; /* current meta */
1549 sector_t meta_total_blocks; /* total size of current meta and data */
1550 sector_t pos; /* recovery position */
1551 u64 seq; /* recovery position seq */
b4c625c6
SL		 1552 int data_parity_stripes; /* number of data_parity stripes */
1553 int data_only_stripes; /* number of data_only stripes */
1554 struct list_head cached_list;
355810d1
SL		 1555};
1556
9ed988f5
SL		 1557static int r5l_recovery_read_meta_block(struct r5l_log *log,
1558 struct r5l_recovery_ctx *ctx)
355810d1
SL		 1559{
1560 struct page *page = ctx->meta_page;
1561 struct r5l_meta_block *mb;
1562 u32 crc, stored_crc;
1563
796a5cf0
MC		 1564 if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, REQ_OP_READ, 0,
1565 false))
355810d1
SL		 1566 return -EIO;
1567
1568 mb = page_address(page);
1569 stored_crc = le32_to_cpu(mb->checksum);
1570 mb->checksum = 0;
1571
1572 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1573 le64_to_cpu(mb->seq) != ctx->seq ||
1574 mb->version != R5LOG_VERSION ||
1575 le64_to_cpu(mb->position) != ctx->pos)
1576 return -EINVAL;
1577
5cb2fbd6 1578 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
355810d1
SL		 1579 if (stored_crc != crc)
1580 return -EINVAL;
1581
1582 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1583 return -EINVAL;
1584
1585 ctx->meta_total_blocks = BLOCK_SECTORS;
1586
1587 return 0;
1588}
1589
9ed988f5
SL		 1590static void
1591r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1592 struct page *page,
1593 sector_t pos, u64 seq)
355810d1 1594{
355810d1 1595 struct r5l_meta_block *mb;
355810d1 1596
355810d1 1597 mb = page_address(page);
9ed988f5 1598 clear_page(mb);
355810d1
SL		 1599 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1600 mb->version = R5LOG_VERSION;
1601 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1602 mb->seq = cpu_to_le64(seq);
1603 mb->position = cpu_to_le64(pos);
9ed988f5 1604}
355810d1 1605
9ed988f5
SL		 1606static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1607 u64 seq)
1608{
1609 struct page *page;
5c88f403 1610 struct r5l_meta_block *mb;
355810d1 1611
9ed988f5
SL		 1612 page = alloc_page(GFP_KERNEL);
1613 if (!page)
1614 return -ENOMEM;
1615 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
5c88f403
SL		 1616 mb = page_address(page);
1617 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1618 mb, PAGE_SIZE));
796a5cf0 1619 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
20737738 1620 REQ_FUA, false)) {
355810d1
SL		 1621 __free_page(page);
1622 return -EIO;
1623 }
1624 __free_page(page);
1625 return 0;
1626}
355810d1 1627
b4c625c6
SL		 1628/*
1629 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1630 * to mark valid (potentially not flushed) data in the journal.
1631 *
1632 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1633 * so there should not be any mismatch here.
1634 */
1635static void r5l_recovery_load_data(struct r5l_log *log,
1636 struct stripe_head *sh,
1637 struct r5l_recovery_ctx *ctx,
1638 struct r5l_payload_data_parity *payload,
1639 sector_t log_offset)
1640{
1641 struct mddev *mddev = log->rdev->mddev;
1642 struct r5conf *conf = mddev->private;
1643 int dd_idx;
1644
1645 raid5_compute_sector(conf,
1646 le64_to_cpu(payload->location), 0,
1647 &dd_idx, sh);
1648 sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1649 sh->dev[dd_idx].page, REQ_OP_READ, 0, false);
1650 sh->dev[dd_idx].log_checksum =
1651 le32_to_cpu(payload->checksum[0]);
1652 ctx->meta_total_blocks += BLOCK_SECTORS;
1653
1654 set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1655 set_bit(STRIPE_R5C_CACHING, &sh->state);
1656}
1657
1658static void r5l_recovery_load_parity(struct r5l_log *log,
1659 struct stripe_head *sh,
1660 struct r5l_recovery_ctx *ctx,
1661 struct r5l_payload_data_parity *payload,
1662 sector_t log_offset)
1663{
1664 struct mddev *mddev = log->rdev->mddev;
1665 struct r5conf *conf = mddev->private;
1666
1667 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1668 sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1669 sh->dev[sh->pd_idx].page, REQ_OP_READ, 0, false);
1670 sh->dev[sh->pd_idx].log_checksum =
1671 le32_to_cpu(payload->checksum[0]);
1672 set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1673
1674 if (sh->qd_idx >= 0) {
1675 sync_page_io(log->rdev,
1676 r5l_ring_add(log, log_offset, BLOCK_SECTORS),
1677 PAGE_SIZE, sh->dev[sh->qd_idx].page,
1678 REQ_OP_READ, 0, false);
1679 sh->dev[sh->qd_idx].log_checksum =
1680 le32_to_cpu(payload->checksum[1]);
1681 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
355810d1 1682 }
b4c625c6
SL		 1683 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1684}
355810d1 1685
b4c625c6
SL		 1686static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1687{
1688 int i;
1689
1690 sh->state = 0;
1691 sh->log_start = MaxSector;
1692 for (i = sh->disks; i--; )
1693 sh->dev[i].flags = 0;
1694}
1695
1696static void
1697r5l_recovery_replay_one_stripe(struct r5conf *conf,
1698 struct stripe_head *sh,
1699 struct r5l_recovery_ctx *ctx)
1700{
1701 struct md_rdev *rdev, *rrdev;
1702 int disk_index;
1703 int data_count = 0;
355810d1 1704
b4c625c6 1705 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
355810d1
SL		 1706 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1707 continue;
b4c625c6
SL		 1708 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1709 continue;
1710 data_count++;
355810d1
SL		 1711 }
1712
b4c625c6
SL		 1713 /*
1714 * stripes that only have parity must have been flushed
1715 * before the crash that we are now recovering from, so
1716 * there is nothing more to recovery.
1717 */
1718 if (data_count == 0)
1719 goto out;
355810d1 1720
b4c625c6
SL		 1721 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1722 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
355810d1
SL		 1723 continue;
1724
1725 /* in case device is broken */
b4c625c6 1726 rcu_read_lock();
355810d1 1727 rdev = rcu_dereference(conf->disks[disk_index].rdev);
b4c625c6
SL		 1728 if (rdev) {
1729 atomic_inc(&rdev->nr_pending);
1730 rcu_read_unlock();
1731 sync_page_io(rdev, sh->sector, PAGE_SIZE,
796a5cf0
MC		 1732 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1733 false);
b4c625c6
SL		 1734 rdev_dec_pending(rdev, rdev->mddev);
1735 rcu_read_lock();
1736 }
355810d1 1737 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
b4c625c6
SL		 1738 if (rrdev) {
1739 atomic_inc(&rrdev->nr_pending);
1740 rcu_read_unlock();
1741 sync_page_io(rrdev, sh->sector, PAGE_SIZE,
796a5cf0
MC		 1742 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1743 false);
b4c625c6
SL		 1744 rdev_dec_pending(rrdev, rrdev->mddev);
1745 rcu_read_lock();
1746 }
1747 rcu_read_unlock();
355810d1 1748 }
b4c625c6
SL		 1749 ctx->data_parity_stripes++;
1750out:
1751 r5l_recovery_reset_stripe(sh);
1752}
1753
1754static struct stripe_head *
1755r5c_recovery_alloc_stripe(struct r5conf *conf,
3c66abba 1756 sector_t stripe_sect)
b4c625c6
SL		 1757{
1758 struct stripe_head *sh;
1759
1760 sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1761 if (!sh)
1762 return NULL; /* no more stripe available */
1763
1764 r5l_recovery_reset_stripe(sh);
b4c625c6
SL		 1765
1766 return sh;
1767}
1768
1769static struct stripe_head *
1770r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1771{
1772 struct stripe_head *sh;
1773
1774 list_for_each_entry(sh, list, lru)
1775 if (sh->sector == sect)
1776 return sh;
1777 return NULL;
1778}
1779
1780static void
1781r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1782 struct r5l_recovery_ctx *ctx)
1783{
1784 struct stripe_head *sh, *next;
1785
1786 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1787 r5l_recovery_reset_stripe(sh);
1788 list_del_init(&sh->lru);
1789 raid5_release_stripe(sh);
1790 }
1791}
1792
1793static void
1794r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1795 struct r5l_recovery_ctx *ctx)
1796{
1797 struct stripe_head *sh, *next;
1798
1799 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1800 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1801 r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1802 list_del_init(&sh->lru);
1803 raid5_release_stripe(sh);
1804 }
1805}
1806
1807/* if matches return 0; otherwise return -EINVAL */
1808static int
1809r5l_recovery_verify_data_checksum(struct r5l_log *log, struct page *page,
1810 sector_t log_offset, __le32 log_checksum)
1811{
1812 void *addr;
1813 u32 checksum;
1814
1815 sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1816 page, REQ_OP_READ, 0, false);
1817 addr = kmap_atomic(page);
1818 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1819 kunmap_atomic(addr);
1820 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1821}
1822
1823/*
1824 * before loading data to stripe cache, we need verify checksum for all data,
1825 * if there is mismatch for any data page, we drop all data in the mata block
1826 */
1827static int
1828r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1829 struct r5l_recovery_ctx *ctx)
1830{
1831 struct mddev *mddev = log->rdev->mddev;
1832 struct r5conf *conf = mddev->private;
1833 struct r5l_meta_block *mb = page_address(ctx->meta_page);
1834 sector_t mb_offset = sizeof(struct r5l_meta_block);
1835 sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1836 struct page *page;
1837 struct r5l_payload_data_parity *payload;
1838
1839 page = alloc_page(GFP_KERNEL);
1840 if (!page)
1841 return -ENOMEM;
1842
1843 while (mb_offset < le32_to_cpu(mb->meta_size)) {
1844 payload = (void *)mb + mb_offset;
1845
1846 if (payload->header.type == R5LOG_PAYLOAD_DATA) {
1847 if (r5l_recovery_verify_data_checksum(
1848 log, page, log_offset,
1849 payload->checksum[0]) < 0)
1850 goto mismatch;
1851 } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) {
1852 if (r5l_recovery_verify_data_checksum(
1853 log, page, log_offset,
1854 payload->checksum[0]) < 0)
1855 goto mismatch;
1856 if (conf->max_degraded == 2 && /* q for RAID 6 */
1857 r5l_recovery_verify_data_checksum(
1858 log, page,
1859 r5l_ring_add(log, log_offset,
1860 BLOCK_SECTORS),
1861 payload->checksum[1]) < 0)
1862 goto mismatch;
1863 } else /* not R5LOG_PAYLOAD_DATA or R5LOG_PAYLOAD_PARITY */
1864 goto mismatch;
1865
1866 log_offset = r5l_ring_add(log, log_offset,
1867 le32_to_cpu(payload->size));
1868
1869 mb_offset += sizeof(struct r5l_payload_data_parity) +
1870 sizeof(__le32) *
1871 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
1872 }
1873
1874 put_page(page);
355810d1
SL		 1875 return 0;
1876
b4c625c6
SL		 1877mismatch:
1878 put_page(page);
355810d1
SL		 1879 return -EINVAL;
1880}
1881
b4c625c6
SL		 1882/*
1883 * Analyze all data/parity pages in one meta block
1884 * Returns:
1885 * 0 for success
1886 * -EINVAL for unknown playload type
1887 * -EAGAIN for checksum mismatch of data page
1888 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
1889 */
1890static int
1891r5c_recovery_analyze_meta_block(struct r5l_log *log,
1892 struct r5l_recovery_ctx *ctx,
1893 struct list_head *cached_stripe_list)
355810d1 1894{
b4c625c6
SL		 1895 struct mddev *mddev = log->rdev->mddev;
1896 struct r5conf *conf = mddev->private;
355810d1 1897 struct r5l_meta_block *mb;
b4c625c6
SL		 1898 struct r5l_payload_data_parity *payload;
1899 int mb_offset;
355810d1 1900 sector_t log_offset;
b4c625c6
SL		 1901 sector_t stripe_sect;
1902 struct stripe_head *sh;
1903 int ret;
1904
1905 /*
1906 * for mismatch in data blocks, we will drop all data in this mb, but
1907 * we will still read next mb for other data with FLUSH flag, as
1908 * io_unit could finish out of order.
1909 */
1910 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
1911 if (ret == -EINVAL)
1912 return -EAGAIN;
1913 else if (ret)
1914 return ret; /* -ENOMEM duo to alloc_page() failed */
355810d1
SL		 1915
1916 mb = page_address(ctx->meta_page);
b4c625c6 1917 mb_offset = sizeof(struct r5l_meta_block);
355810d1
SL		 1918 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1919
b4c625c6 1920 while (mb_offset < le32_to_cpu(mb->meta_size)) {
355810d1
SL		 1921 int dd;
1922
b4c625c6
SL		 1923 payload = (void *)mb + mb_offset;
1924 stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ?
1925 raid5_compute_sector(
1926 conf, le64_to_cpu(payload->location), 0, &dd,
1927 NULL)
1928 : le64_to_cpu(payload->location);
1929
1930 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
1931 stripe_sect);
1932
1933 if (!sh) {
3c66abba 1934 sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
b4c625c6
SL		 1935 /*
1936 * cannot get stripe from raid5_get_active_stripe
1937 * try replay some stripes
1938 */
1939 if (!sh) {
1940 r5c_recovery_replay_stripes(
1941 cached_stripe_list, ctx);
1942 sh = r5c_recovery_alloc_stripe(
3c66abba 1943 conf, stripe_sect);
b4c625c6
SL		 1944 }
1945 if (!sh) {
1946 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
1947 mdname(mddev),
1948 conf->min_nr_stripes * 2);
1949 raid5_set_cache_size(mddev,
1950 conf->min_nr_stripes * 2);
3c66abba
SL		 1951 sh = r5c_recovery_alloc_stripe(conf,
1952 stripe_sect);
b4c625c6
SL		 1953 }
1954 if (!sh) {
1955 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
1956 mdname(mddev));
1957 return -ENOMEM;
1958 }
1959 list_add_tail(&sh->lru, cached_stripe_list);
1960 }
1961
1962 if (payload->header.type == R5LOG_PAYLOAD_DATA) {
f7b7bee7
ZL		 1963 if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
1964 test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
b4c625c6 1965 r5l_recovery_replay_one_stripe(conf, sh, ctx);
b4c625c6
SL		 1966 list_move_tail(&sh->lru, cached_stripe_list);
1967 }
1968 r5l_recovery_load_data(log, sh, ctx, payload,
1969 log_offset);
1970 } else if (payload->header.type == R5LOG_PAYLOAD_PARITY)
1971 r5l_recovery_load_parity(log, sh, ctx, payload,
1972 log_offset);
1973 else
355810d1 1974 return -EINVAL;
b4c625c6
SL		 1975
1976 log_offset = r5l_ring_add(log, log_offset,
1977 le32_to_cpu(payload->size));
1978
1979 mb_offset += sizeof(struct r5l_payload_data_parity) +
1980 sizeof(__le32) *
1981 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
355810d1 1982 }
b4c625c6 1983
355810d1
SL		 1984 return 0;
1985}
1986
b4c625c6
SL		 1987/*
1988 * Load the stripe into cache. The stripe will be written out later by
1989 * the stripe cache state machine.
1990 */
1991static void r5c_recovery_load_one_stripe(struct r5l_log *log,
1992 struct stripe_head *sh)
355810d1 1993{
b4c625c6
SL		 1994 struct r5dev *dev;
1995 int i;
1996
1997 for (i = sh->disks; i--; ) {
1998 dev = sh->dev + i;
1999 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2000 set_bit(R5_InJournal, &dev->flags);
2001 set_bit(R5_UPTODATE, &dev->flags);
2002 }
2003 }
b4c625c6
SL		 2004}
2005
2006/*
2007 * Scan through the log for all to-be-flushed data
2008 *
2009 * For stripes with data and parity, namely Data-Parity stripe
2010 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2011 *
2012 * For stripes with only data, namely Data-Only stripe
2013 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2014 *
2015 * For a stripe, if we see data after parity, we should discard all previous
2016 * data and parity for this stripe, as these data are already flushed to
2017 * the array.
2018 *
2019 * At the end of the scan, we return the new journal_tail, which points to
2020 * first data-only stripe on the journal device, or next invalid meta block.
2021 */
2022static int r5c_recovery_flush_log(struct r5l_log *log,
2023 struct r5l_recovery_ctx *ctx)
2024{
bc8f167f 2025 struct stripe_head *sh;
b4c625c6
SL		 2026 int ret = 0;
2027
2028 /* scan through the log */
355810d1 2029 while (1) {
b4c625c6
SL		 2030 if (r5l_recovery_read_meta_block(log, ctx))
2031 break;
2032
2033 ret = r5c_recovery_analyze_meta_block(log, ctx,
2034 &ctx->cached_list);
2035 /*
2036 * -EAGAIN means mismatch in data block, in this case, we still
2037 * try scan the next metablock
2038 */
2039 if (ret && ret != -EAGAIN)
2040 break; /* ret == -EINVAL or -ENOMEM */
355810d1
SL		 2041 ctx->seq++;
2042 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2043 }
b4c625c6
SL		 2044
2045 if (ret == -ENOMEM) {
2046 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2047 return ret;
2048 }
2049
2050 /* replay data-parity stripes */
2051 r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2052
2053 /* load data-only stripes to stripe cache */
bc8f167f 2054 list_for_each_entry(sh, &ctx->cached_list, lru) {
b4c625c6
SL		 2055 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2056 r5c_recovery_load_one_stripe(log, sh);
b4c625c6
SL		 2057 ctx->data_only_stripes++;
2058 }
2059
2060 return 0;
355810d1
SL		 2061}
2062
b4c625c6
SL		 2063/*
2064 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2065 * log will start here. but we can't let superblock point to last valid
2066 * meta block. The log might looks like:
2067 * | meta 1| meta 2| meta 3|
2068 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2069 * superblock points to meta 1, we write a new valid meta 2n. if crash
2070 * happens again, new recovery will start from meta 1. Since meta 2n is
2071 * valid now, recovery will think meta 3 is valid, which is wrong.
2072 * The solution is we create a new meta in meta2 with its seq == meta
3c6edc66
SL		 2073 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2074 * will not think meta 3 is a valid meta, because its seq doesn't match
b4c625c6
SL		 2075 */
2076
2077/*
2078 * Before recovery, the log looks like the following
2079 *
2080 * ---------------------------------------------
2081 * | valid log | invalid log |
2082 * ---------------------------------------------
2083 * ^
2084 * |- log->last_checkpoint
2085 * |- log->last_cp_seq
2086 *
2087 * Now we scan through the log until we see invalid entry
2088 *
2089 * ---------------------------------------------
2090 * | valid log | invalid log |
2091 * ---------------------------------------------
2092 * ^ ^
2093 * |- log->last_checkpoint |- ctx->pos
2094 * |- log->last_cp_seq |- ctx->seq
2095 *
2096 * From this point, we need to increase seq number by 10 to avoid
2097 * confusing next recovery.
2098 *
2099 * ---------------------------------------------
2100 * | valid log | invalid log |
2101 * ---------------------------------------------
2102 * ^ ^
2103 * |- log->last_checkpoint |- ctx->pos+1
3c6edc66 2104 * |- log->last_cp_seq |- ctx->seq+10001
b4c625c6
SL		 2105 *
2106 * However, it is not safe to start the state machine yet, because data only
2107 * parities are not yet secured in RAID. To save these data only parities, we
2108 * rewrite them from seq+11.
2109 *
2110 * -----------------------------------------------------------------
2111 * | valid log | data only stripes | invalid log |
2112 * -----------------------------------------------------------------
2113 * ^ ^
2114 * |- log->last_checkpoint |- ctx->pos+n
3c6edc66 2115 * |- log->last_cp_seq |- ctx->seq+10000+n
b4c625c6
SL		 2116 *
2117 * If failure happens again during this process, the recovery can safe start
2118 * again from log->last_checkpoint.
2119 *
2120 * Once data only stripes are rewritten to journal, we move log_tail
2121 *
2122 * -----------------------------------------------------------------
2123 * | old log | data only stripes | invalid log |
2124 * -----------------------------------------------------------------
2125 * ^ ^
2126 * |- log->last_checkpoint |- ctx->pos+n
3c6edc66 2127 * |- log->last_cp_seq |- ctx->seq+10000+n
b4c625c6
SL		 2128 *
2129 * Then we can safely start the state machine. If failure happens from this
2130 * point on, the recovery will start from new log->last_checkpoint.
2131 */
2132static int
2133r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2134 struct r5l_recovery_ctx *ctx)
355810d1 2135{
a85dd7b8 2136 struct stripe_head *sh;
b4c625c6 2137 struct mddev *mddev = log->rdev->mddev;
355810d1 2138 struct page *page;
3c66abba 2139 sector_t next_checkpoint = MaxSector;
355810d1 2140
b4c625c6
SL		 2141 page = alloc_page(GFP_KERNEL);
2142 if (!page) {
2143 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2144 mdname(mddev));
355810d1 2145 return -ENOMEM;
b4c625c6 2146 }
355810d1 2147
3c66abba
SL		 2148 WARN_ON(list_empty(&ctx->cached_list));
2149
a85dd7b8 2150 list_for_each_entry(sh, &ctx->cached_list, lru) {
b4c625c6
SL		 2151 struct r5l_meta_block *mb;
2152 int i;
2153 int offset;
2154 sector_t write_pos;
2155
2156 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2157 r5l_recovery_create_empty_meta_block(log, page,
2158 ctx->pos, ctx->seq);
2159 mb = page_address(page);
2160 offset = le32_to_cpu(mb->meta_size);
fc833c2a 2161 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
b4c625c6
SL		 2162
2163 for (i = sh->disks; i--; ) {
2164 struct r5dev *dev = &sh->dev[i];
2165 struct r5l_payload_data_parity *payload;
2166 void *addr;
2167
2168 if (test_bit(R5_InJournal, &dev->flags)) {
2169 payload = (void *)mb + offset;
2170 payload->header.type = cpu_to_le16(
2171 R5LOG_PAYLOAD_DATA);
2172 payload->size = BLOCK_SECTORS;
2173 payload->location = cpu_to_le64(
2174 raid5_compute_blocknr(sh, i, 0));
2175 addr = kmap_atomic(dev->page);
2176 payload->checksum[0] = cpu_to_le32(
2177 crc32c_le(log->uuid_checksum, addr,
2178 PAGE_SIZE));
2179 kunmap_atomic(addr);
2180 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2181 dev->page, REQ_OP_WRITE, 0, false);
2182 write_pos = r5l_ring_add(log, write_pos,
2183 BLOCK_SECTORS);
2184 offset += sizeof(__le32) +
2185 sizeof(struct r5l_payload_data_parity);
2186
2187 }
2188 }
2189 mb->meta_size = cpu_to_le32(offset);
5c88f403
SL		 2190 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2191 mb, PAGE_SIZE));
b4c625c6 2192 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
20737738 2193 REQ_OP_WRITE, REQ_FUA, false);
b4c625c6 2194 sh->log_start = ctx->pos;
3c66abba
SL		 2195 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2196 atomic_inc(&log->stripe_in_journal_count);
b4c625c6
SL		 2197 ctx->pos = write_pos;
2198 ctx->seq += 1;
3c66abba 2199 next_checkpoint = sh->log_start;
355810d1 2200 }
3c66abba 2201 log->next_checkpoint = next_checkpoint;
355810d1
SL		 2202 __free_page(page);
2203 return 0;
2204}
2205
a85dd7b8
SL		 2206static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2207 struct r5l_recovery_ctx *ctx)
2208{
2209 struct mddev *mddev = log->rdev->mddev;
2210 struct r5conf *conf = mddev->private;
2211 struct stripe_head *sh, *next;
2212
2213 if (ctx->data_only_stripes == 0)
2214 return;
2215
2216 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2217
2218 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2219 r5c_make_stripe_write_out(sh);
2220 set_bit(STRIPE_HANDLE, &sh->state);
2221 list_del_init(&sh->lru);
2222 raid5_release_stripe(sh);
2223 }
2224
2225 md_wakeup_thread(conf->mddev->thread);
2226 /* reuse conf->wait_for_quiescent in recovery */
2227 wait_event(conf->wait_for_quiescent,
2228 atomic_read(&conf->active_stripes) == 0);
2229
2230 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2231}
2232
f6bed0ef
SL		 2233static int r5l_recovery_log(struct r5l_log *log)
2234{
5aabf7c4 2235 struct mddev *mddev = log->rdev->mddev;
355810d1 2236 struct r5l_recovery_ctx ctx;
5aabf7c4 2237 int ret;
43b96748 2238 sector_t pos;
355810d1
SL		 2239
2240 ctx.pos = log->last_checkpoint;
2241 ctx.seq = log->last_cp_seq;
2242 ctx.meta_page = alloc_page(GFP_KERNEL);
b4c625c6
SL		 2243 ctx.data_only_stripes = 0;
2244 ctx.data_parity_stripes = 0;
2245 INIT_LIST_HEAD(&ctx.cached_list);
2246
355810d1
SL		 2247 if (!ctx.meta_page)
2248 return -ENOMEM;
2249
5aabf7c4 2250 ret = r5c_recovery_flush_log(log, &ctx);
355810d1
SL		 2251 __free_page(ctx.meta_page);
2252
5aabf7c4
SL		 2253 if (ret)
2254 return ret;
b4c625c6 2255
3c6edc66
SL		 2256 pos = ctx.pos;
2257 ctx.seq += 10000;
43b96748 2258
43b96748 2259
5aabf7c4
SL		 2260 if ((ctx.data_only_stripes == 0) && (ctx.data_parity_stripes == 0))
2261 pr_debug("md/raid:%s: starting from clean shutdown\n",
2262 mdname(mddev));
a85dd7b8 2263 else
99f17890 2264 pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
5aabf7c4
SL		 2265 mdname(mddev), ctx.data_only_stripes,
2266 ctx.data_parity_stripes);
2267
a85dd7b8
SL		 2268 if (ctx.data_only_stripes == 0) {
2269 log->next_checkpoint = ctx.pos;
2270 r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq++);
2271 ctx.pos = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
2272 } else if (r5c_recovery_rewrite_data_only_stripes(log, &ctx)) {
2273 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2274 mdname(mddev));
2275 return -EIO;
b4c625c6
SL		 2276 }
2277
5aabf7c4 2278 log->log_start = ctx.pos;
5aabf7c4 2279 log->seq = ctx.seq;
43b96748
J		 2280 log->last_checkpoint = pos;
2281 r5l_write_super(log, pos);
a85dd7b8
SL		 2282
2283 r5c_recovery_flush_data_only_stripes(log, &ctx);
f6bed0ef
SL		 2284 return 0;
2285}
2286
2287static void r5l_write_super(struct r5l_log *log, sector_t cp)
2288{
2289 struct mddev *mddev = log->rdev->mddev;
2290
2291 log->rdev->journal_tail = cp;
2953079c 2292 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
f6bed0ef
SL		 2293}
2294
2c7da14b
SL		 2295static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2296{
2297 struct r5conf *conf = mddev->private;
2298 int ret;
2299
2300 if (!conf->log)
2301 return 0;
2302
2303 switch (conf->log->r5c_journal_mode) {
2304 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2305 ret = snprintf(
2306 page, PAGE_SIZE, "[%s] %s\n",
2307 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2308 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2309 break;
2310 case R5C_JOURNAL_MODE_WRITE_BACK:
2311 ret = snprintf(
2312 page, PAGE_SIZE, "%s [%s]\n",
2313 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2314 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2315 break;
2316 default:
2317 ret = 0;
2318 }
2319 return ret;
2320}
2321
2322static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2323 const char *page, size_t length)
2324{
2325 struct r5conf *conf = mddev->private;
2326 struct r5l_log *log = conf->log;
2327 int val = -1, i;
2328 int len = length;
2329
2330 if (!log)
2331 return -ENODEV;
2332
2333 if (len && page[len - 1] == '\n')
2334 len -= 1;
2335 for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++)
2336 if (strlen(r5c_journal_mode_str[i]) == len &&
2337 strncmp(page, r5c_journal_mode_str[i], len) == 0) {
2338 val = i;
2339 break;
2340 }
2341 if (val < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2342 val > R5C_JOURNAL_MODE_WRITE_BACK)
2343 return -EINVAL;
2344
2e38a37f
SL		 2345 if (raid5_calc_degraded(conf) > 0 &&
2346 val == R5C_JOURNAL_MODE_WRITE_BACK)
2347 return -EINVAL;
2348
2c7da14b
SL		 2349 mddev_suspend(mddev);
2350 conf->log->r5c_journal_mode = val;
2351 mddev_resume(mddev);
2352
2353 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2354 mdname(mddev), val, r5c_journal_mode_str[val]);
2355 return length;
2356}
2357
2358struct md_sysfs_entry
2359r5c_journal_mode = __ATTR(journal_mode, 0644,
2360 r5c_journal_mode_show, r5c_journal_mode_store);
2361
2ded3703
SL		 2362/*
2363 * Try handle write operation in caching phase. This function should only
2364 * be called in write-back mode.
2365 *
2366 * If all outstanding writes can be handled in caching phase, returns 0
2367 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2368 * and returns -EAGAIN
2369 */
2370int r5c_try_caching_write(struct r5conf *conf,
2371 struct stripe_head *sh,
2372 struct stripe_head_state *s,
2373 int disks)
2374{
2375 struct r5l_log *log = conf->log;
1e6d690b
SL		 2376 int i;
2377 struct r5dev *dev;
2378 int to_cache = 0;
03b047f4
SL		 2379 void **pslot;
2380 sector_t tree_index;
2381 int ret;
2382 uintptr_t refcount;
2ded3703
SL		 2383
2384 BUG_ON(!r5c_is_writeback(log));
2385
1e6d690b
SL		 2386 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2387 /*
2388 * There are two different scenarios here:
2389 * 1. The stripe has some data cached, and it is sent to
2390 * write-out phase for reclaim
2391 * 2. The stripe is clean, and this is the first write
2392 *
2393 * For 1, return -EAGAIN, so we continue with
2394 * handle_stripe_dirtying().
2395 *
2396 * For 2, set STRIPE_R5C_CACHING and continue with caching
2397 * write.
2398 */
2399
2400 /* case 1: anything injournal or anything in written */
2401 if (s->injournal > 0 || s->written > 0)
2402 return -EAGAIN;
2403 /* case 2 */
2404 set_bit(STRIPE_R5C_CACHING, &sh->state);
2405 }
2406
2e38a37f
SL		 2407 /*
2408 * When run in degraded mode, array is set to write-through mode.
2409 * This check helps drain pending write safely in the transition to
2410 * write-through mode.
2411 */
2412 if (s->failed) {
2413 r5c_make_stripe_write_out(sh);
2414 return -EAGAIN;
2415 }
2416
1e6d690b
SL		 2417 for (i = disks; i--; ) {
2418 dev = &sh->dev[i];
2419 /* if non-overwrite, use writing-out phase */
2420 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2421 !test_bit(R5_InJournal, &dev->flags)) {
2422 r5c_make_stripe_write_out(sh);
2423 return -EAGAIN;
2424 }
2425 }
2426
03b047f4
SL		 2427 /* if the stripe is not counted in big_stripe_tree, add it now */
2428 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2429 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2430 tree_index = r5c_tree_index(conf, sh->sector);
2431 spin_lock(&log->tree_lock);
2432 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2433 tree_index);
2434 if (pslot) {
2435 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2436 pslot, &log->tree_lock) >>
2437 R5C_RADIX_COUNT_SHIFT;
2438 radix_tree_replace_slot(
2439 &log->big_stripe_tree, pslot,
2440 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2441 } else {
2442 /*
2443 * this radix_tree_insert can fail safely, so no
2444 * need to call radix_tree_preload()
2445 */
2446 ret = radix_tree_insert(
2447 &log->big_stripe_tree, tree_index,
2448 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2449 if (ret) {
2450 spin_unlock(&log->tree_lock);
2451 r5c_make_stripe_write_out(sh);
2452 return -EAGAIN;
2453 }
2454 }
2455 spin_unlock(&log->tree_lock);
2456
2457 /*
2458 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2459 * counted in the radix tree
2460 */
2461 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2462 atomic_inc(&conf->r5c_cached_partial_stripes);
2463 }
2464
1e6d690b
SL		 2465 for (i = disks; i--; ) {
2466 dev = &sh->dev[i];
2467 if (dev->towrite) {
2468 set_bit(R5_Wantwrite, &dev->flags);
2469 set_bit(R5_Wantdrain, &dev->flags);
2470 set_bit(R5_LOCKED, &dev->flags);
2471 to_cache++;
2472 }
2473 }
2474
2475 if (to_cache) {
2476 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2477 /*
2478 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2479 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2480 * r5c_handle_data_cached()
2481 */
2482 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2483 }
2484
2485 return 0;
2486}
2487
2488/*
2489 * free extra pages (orig_page) we allocated for prexor
2490 */
2491void r5c_release_extra_page(struct stripe_head *sh)
2492{
d7bd398e 2493 struct r5conf *conf = sh->raid_conf;
1e6d690b 2494 int i;
d7bd398e
SL		 2495 bool using_disk_info_extra_page;
2496
2497 using_disk_info_extra_page =
2498 sh->dev[0].orig_page == conf->disks[0].extra_page;
1e6d690b
SL		 2499
2500 for (i = sh->disks; i--; )
2501 if (sh->dev[i].page != sh->dev[i].orig_page) {
2502 struct page *p = sh->dev[i].orig_page;
2503
2504 sh->dev[i].orig_page = sh->dev[i].page;
86aa1397
SL		 2505 clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2506
d7bd398e
SL		 2507 if (!using_disk_info_extra_page)
2508 put_page(p);
1e6d690b 2509 }
d7bd398e
SL		 2510
2511 if (using_disk_info_extra_page) {
2512 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2513 md_wakeup_thread(conf->mddev->thread);
2514 }
2515}
2516
2517void r5c_use_extra_page(struct stripe_head *sh)
2518{
2519 struct r5conf *conf = sh->raid_conf;
2520 int i;
2521 struct r5dev *dev;
2522
2523 for (i = sh->disks; i--; ) {
2524 dev = &sh->dev[i];
2525 if (dev->orig_page != dev->page)
2526 put_page(dev->orig_page);
2527 dev->orig_page = conf->disks[i].extra_page;
2528 }
2ded3703
SL		 2529}
2530
2531/*
2532 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2533 * stripe is committed to RAID disks.
2534 */
2535void r5c_finish_stripe_write_out(struct r5conf *conf,
2536 struct stripe_head *sh,
2537 struct stripe_head_state *s)
2538{
03b047f4 2539 struct r5l_log *log = conf->log;
1e6d690b
SL		 2540 int i;
2541 int do_wakeup = 0;
03b047f4
SL		 2542 sector_t tree_index;
2543 void **pslot;
2544 uintptr_t refcount;
1e6d690b 2545
03b047f4 2546 if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2ded3703
SL		 2547 return;
2548
2549 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2550 clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2551
03b047f4 2552 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2ded3703 2553 return;
1e6d690b
SL		 2554
2555 for (i = sh->disks; i--; ) {
2556 clear_bit(R5_InJournal, &sh->dev[i].flags);
2557 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2558 do_wakeup = 1;
2559 }
2560
2561 /*
2562 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2563 * We updated R5_InJournal, so we also update s->injournal.
2564 */
2565 s->injournal = 0;
2566
2567 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2568 if (atomic_dec_and_test(&conf->pending_full_writes))
2569 md_wakeup_thread(conf->mddev->thread);
2570
2571 if (do_wakeup)
2572 wake_up(&conf->wait_for_overlap);
a39f7afd 2573
03b047f4 2574 spin_lock_irq(&log->stripe_in_journal_lock);
a39f7afd 2575 list_del_init(&sh->r5c);
03b047f4 2576 spin_unlock_irq(&log->stripe_in_journal_lock);
a39f7afd 2577 sh->log_start = MaxSector;
03b047f4
SL		 2578
2579 atomic_dec(&log->stripe_in_journal_count);
2580 r5c_update_log_state(log);
2581
2582 /* stop counting this stripe in big_stripe_tree */
2583 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2584 test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2585 tree_index = r5c_tree_index(conf, sh->sector);
2586 spin_lock(&log->tree_lock);
2587 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2588 tree_index);
2589 BUG_ON(pslot == NULL);
2590 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2591 pslot, &log->tree_lock) >>
2592 R5C_RADIX_COUNT_SHIFT;
2593 if (refcount == 1)
2594 radix_tree_delete(&log->big_stripe_tree, tree_index);
2595 else
2596 radix_tree_replace_slot(
2597 &log->big_stripe_tree, pslot,
2598 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2599 spin_unlock(&log->tree_lock);
2600 }
2601
2602 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2603 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2604 atomic_dec(&conf->r5c_cached_partial_stripes);
2605 }
2606
2607 if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2608 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2609 atomic_dec(&conf->r5c_cached_full_stripes);
2610 }
1e6d690b
SL		 2611}
2612
2613int
2614r5c_cache_data(struct r5l_log *log, struct stripe_head *sh,
2615 struct stripe_head_state *s)
2616{
a39f7afd 2617 struct r5conf *conf = sh->raid_conf;
1e6d690b
SL		 2618 int pages = 0;
2619 int reserve;
2620 int i;
2621 int ret = 0;
2622
2623 BUG_ON(!log);
2624
2625 for (i = 0; i < sh->disks; i++) {
2626 void *addr;
2627
2628 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2629 continue;
2630 addr = kmap_atomic(sh->dev[i].page);
2631 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2632 addr, PAGE_SIZE);
2633 kunmap_atomic(addr);
2634 pages++;
2635 }
2636 WARN_ON(pages == 0);
2637
2638 /*
2639 * The stripe must enter state machine again to call endio, so
2640 * don't delay.
2641 */
2642 clear_bit(STRIPE_DELAYED, &sh->state);
2643 atomic_inc(&sh->count);
2644
2645 mutex_lock(&log->io_mutex);
2646 /* meta + data */
2647 reserve = (1 + pages) << (PAGE_SHIFT - 9);
1e6d690b 2648
a39f7afd
SL		 2649 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2650 sh->log_start == MaxSector)
2651 r5l_add_no_space_stripe(log, sh);
2652 else if (!r5l_has_free_space(log, reserve)) {
2653 if (sh->log_start == log->last_checkpoint)
2654 BUG();
2655 else
2656 r5l_add_no_space_stripe(log, sh);
1e6d690b
SL		 2657 } else {
2658 ret = r5l_log_stripe(log, sh, pages, 0);
2659 if (ret) {
2660 spin_lock_irq(&log->io_list_lock);
2661 list_add_tail(&sh->log_list, &log->no_mem_stripes);
2662 spin_unlock_irq(&log->io_list_lock);
2663 }
2664 }
2665
2666 mutex_unlock(&log->io_mutex);
2667 return 0;
f6bed0ef
SL		 2668}
2669
03b047f4
SL		 2670/* check whether this big stripe is in write back cache. */
2671bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2672{
2673 struct r5l_log *log = conf->log;
2674 sector_t tree_index;
2675 void *slot;
2676
2677 if (!log)
2678 return false;
2679
2680 WARN_ON_ONCE(!rcu_read_lock_held());
2681 tree_index = r5c_tree_index(conf, sect);
2682 slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2683 return slot != NULL;
2684}
2685
f6bed0ef
SL		 2686static int r5l_load_log(struct r5l_log *log)
2687{
2688 struct md_rdev *rdev = log->rdev;
2689 struct page *page;
2690 struct r5l_meta_block *mb;
2691 sector_t cp = log->rdev->journal_tail;
2692 u32 stored_crc, expected_crc;
2693 bool create_super = false;
d30dfeb9 2694 int ret = 0;
f6bed0ef
SL		 2695
2696 /* Make sure it's valid */
2697 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2698 cp = 0;
2699 page = alloc_page(GFP_KERNEL);
2700 if (!page)
2701 return -ENOMEM;
2702
796a5cf0 2703 if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
f6bed0ef
SL		 2704 ret = -EIO;
2705 goto ioerr;
2706 }
2707 mb = page_address(page);
2708
2709 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2710 mb->version != R5LOG_VERSION) {
2711 create_super = true;
2712 goto create;
2713 }
2714 stored_crc = le32_to_cpu(mb->checksum);
2715 mb->checksum = 0;
5cb2fbd6 2716 expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
f6bed0ef
SL		 2717 if (stored_crc != expected_crc) {
2718 create_super = true;
2719 goto create;
2720 }
2721 if (le64_to_cpu(mb->position) != cp) {
2722 create_super = true;
2723 goto create;
2724 }
2725create:
2726 if (create_super) {
2727 log->last_cp_seq = prandom_u32();
2728 cp = 0;
56056c2e 2729 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
f6bed0ef
SL		 2730 /*
2731 * Make sure super points to correct address. Log might have
2732 * data very soon. If super hasn't correct log tail address,
2733 * recovery can't find the log
2734 */
2735 r5l_write_super(log, cp);
2736 } else
2737 log->last_cp_seq = le64_to_cpu(mb->seq);
2738
2739 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
0576b1c6
SL		 2740 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
2741 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
2742 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
f6bed0ef
SL		 2743 log->last_checkpoint = cp;
2744
2745 __free_page(page);
2746
d30dfeb9
J		 2747 if (create_super) {
2748 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
2749 log->seq = log->last_cp_seq + 1;
2750 log->next_checkpoint = cp;
2751 } else
2752 ret = r5l_recovery_log(log);
2753
3d7e7e1d
ZL		 2754 r5c_update_log_state(log);
2755 return ret;
f6bed0ef
SL		 2756ioerr:
2757 __free_page(page);
2758 return ret;
2759}
2760
2e38a37f
SL		 2761void r5c_update_on_rdev_error(struct mddev *mddev)
2762{
2763 struct r5conf *conf = mddev->private;
2764 struct r5l_log *log = conf->log;
2765
2766 if (!log)
2767 return;
2768
2769 if (raid5_calc_degraded(conf) > 0 &&
2770 conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
2771 schedule_work(&log->disable_writeback_work);
2772}
2773
f6bed0ef
SL		 2774int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
2775{
c888a8f9 2776 struct request_queue *q = bdev_get_queue(rdev->bdev);
f6bed0ef
SL		 2777 struct r5l_log *log;
2778
2779 if (PAGE_SIZE != 4096)
2780 return -EINVAL;
c757ec95
SL		 2781
2782 /*
2783 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
2784 * raid_disks r5l_payload_data_parity.
2785 *
2786 * Write journal and cache does not work for very big array
2787 * (raid_disks > 203)
2788 */
2789 if (sizeof(struct r5l_meta_block) +
2790 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
2791 conf->raid_disks) > PAGE_SIZE) {
2792 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
2793 mdname(conf->mddev), conf->raid_disks);
2794 return -EINVAL;
2795 }
2796
f6bed0ef
SL		 2797 log = kzalloc(sizeof(*log), GFP_KERNEL);
2798 if (!log)
2799 return -ENOMEM;
2800 log->rdev = rdev;
2801
c888a8f9 2802 log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
56fef7c6 2803
5cb2fbd6
SL		 2804 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
2805 sizeof(rdev->mddev->uuid));
f6bed0ef
SL		 2806
2807 mutex_init(&log->io_mutex);
2808
2809 spin_lock_init(&log->io_list_lock);
2810 INIT_LIST_HEAD(&log->running_ios);
0576b1c6 2811 INIT_LIST_HEAD(&log->io_end_ios);
a8c34f91 2812 INIT_LIST_HEAD(&log->flushing_ios);
04732f74 2813 INIT_LIST_HEAD(&log->finished_ios);
3a83f467 2814 bio_init(&log->flush_bio, NULL, 0);
f6bed0ef
SL		 2815
2816 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
2817 if (!log->io_kc)
2818 goto io_kc;
2819
5036c390
CH		 2820 log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
2821 if (!log->io_pool)
2822 goto io_pool;
2823
c38d29b3
CH		 2824 log->bs = bioset_create(R5L_POOL_SIZE, 0);
2825 if (!log->bs)
2826 goto io_bs;
2827
e8deb638
CH		 2828 log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
2829 if (!log->meta_pool)
2830 goto out_mempool;
2831
03b047f4
SL		 2832 spin_lock_init(&log->tree_lock);
2833 INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
2834
0576b1c6
SL		 2835 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
2836 log->rdev->mddev, "reclaim");
2837 if (!log->reclaim_thread)
2838 goto reclaim_thread;
a39f7afd
SL		 2839 log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
2840
0fd22b45 2841 init_waitqueue_head(&log->iounit_wait);
0576b1c6 2842
5036c390
CH		 2843 INIT_LIST_HEAD(&log->no_mem_stripes);
2844
f6bed0ef
SL		 2845 INIT_LIST_HEAD(&log->no_space_stripes);
2846 spin_lock_init(&log->no_space_stripes_lock);
2847
3bddb7f8 2848 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
2e38a37f 2849 INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3bddb7f8 2850
2ded3703 2851 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
a39f7afd
SL		 2852 INIT_LIST_HEAD(&log->stripe_in_journal_list);
2853 spin_lock_init(&log->stripe_in_journal_lock);
2854 atomic_set(&log->stripe_in_journal_count, 0);
2ded3703 2855
d2250f10
SL		 2856 rcu_assign_pointer(conf->log, log);
2857
f6bed0ef
SL		 2858 if (r5l_load_log(log))
2859 goto error;
2860
a62ab49e 2861 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
f6bed0ef 2862 return 0;
e8deb638 2863
f6bed0ef 2864error:
d2250f10 2865 rcu_assign_pointer(conf->log, NULL);
0576b1c6
SL		 2866 md_unregister_thread(&log->reclaim_thread);
2867reclaim_thread:
e8deb638
CH		 2868 mempool_destroy(log->meta_pool);
2869out_mempool:
c38d29b3
CH		 2870 bioset_free(log->bs);
2871io_bs:
5036c390
CH		 2872 mempool_destroy(log->io_pool);
2873io_pool:
f6bed0ef
SL		 2874 kmem_cache_destroy(log->io_kc);
2875io_kc:
2876 kfree(log);
2877 return -EINVAL;
2878}
2879
2880void r5l_exit_log(struct r5l_log *log)
2881{
2e38a37f 2882 flush_work(&log->disable_writeback_work);
0576b1c6 2883 md_unregister_thread(&log->reclaim_thread);
e8deb638 2884 mempool_destroy(log->meta_pool);
c38d29b3 2885 bioset_free(log->bs);
5036c390 2886 mempool_destroy(log->io_pool);
f6bed0ef
SL		 2887 kmem_cache_destroy(log->io_kc);
2888 kfree(log);
2889}
Linux 6.x block layer and io_uring tree(s)