projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
md/md0: optimize raid0 discard handling
[linux-2.6-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)
effe6ee7 33#define BLOCK_SECTOR_SHIFT (3)
f6bed0ef 34
0576b1c6 35/*
a39f7afd
SL		 36 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
37 *
38 * In write through mode, the reclaim runs every log->max_free_space.
39 * This can prevent the recovery scans for too long
0576b1c6
SL		 40 */
41#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
42#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
43
a39f7afd
SL		 44/* wake up reclaim thread periodically */
45#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
46/* start flush with these full stripes */
84890c03 47#define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
a39f7afd
SL		 48/* reclaim stripes in groups */
49#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
50
c38d29b3
CH		 51/*
52 * We only need 2 bios per I/O unit to make progress, but ensure we
53 * have a few more available to not get too tight.
54 */
55#define R5L_POOL_SIZE 4
56
2c7da14b
SL		 57static char *r5c_journal_mode_str[] = {"write-through",
58 "write-back"};
2ded3703
SL		 59/*
60 * raid5 cache state machine
61 *
9b69173e 62 * With the RAID cache, each stripe works in two phases:
2ded3703
SL		 63 * - caching phase
64 * - writing-out phase
65 *
66 * These two phases are controlled by bit STRIPE_R5C_CACHING:
67 * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
68 * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
69 *
70 * When there is no journal, or the journal is in write-through mode,
71 * the stripe is always in writing-out phase.
72 *
73 * For write-back journal, the stripe is sent to caching phase on write
74 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
75 * the write-out phase by clearing STRIPE_R5C_CACHING.
76 *
77 * Stripes in caching phase do not write the raid disks. Instead, all
78 * writes are committed from the log device. Therefore, a stripe in
79 * caching phase handles writes as:
80 * - write to log device
81 * - return IO
82 *
83 * Stripes in writing-out phase handle writes as:
84 * - calculate parity
85 * - write pending data and parity to journal
86 * - write data and parity to raid disks
87 * - return IO for pending writes
88 */
89
f6bed0ef
SL		 90struct r5l_log {
91 struct md_rdev *rdev;
92
93 u32 uuid_checksum;
94
95 sector_t device_size; /* log device size, round to
96 * BLOCK_SECTORS */
0576b1c6
SL		 97 sector_t max_free_space; /* reclaim run if free space is at
98 * this size */
f6bed0ef
SL		 99
100 sector_t last_checkpoint; /* log tail. where recovery scan
101 * starts from */
102 u64 last_cp_seq; /* log tail sequence */
103
104 sector_t log_start; /* log head. where new data appends */
105 u64 seq; /* log head sequence */
106
17036461 107 sector_t next_checkpoint;
17036461 108
f6bed0ef
SL		 109 struct mutex io_mutex;
110 struct r5l_io_unit *current_io; /* current io_unit accepting new data */
111
112 spinlock_t io_list_lock;
113 struct list_head running_ios; /* io_units which are still running,
114 * and have not yet been completely
115 * written to the log */
116 struct list_head io_end_ios; /* io_units which have been completely
117 * written to the log but not yet written
118 * to the RAID */
a8c34f91
SL		 119 struct list_head flushing_ios; /* io_units which are waiting for log
120 * cache flush */
04732f74 121 struct list_head finished_ios; /* io_units which settle down in log disk */
a8c34f91 122 struct bio flush_bio;
f6bed0ef 123
5036c390
CH		 124 struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */
125
f6bed0ef 126 struct kmem_cache *io_kc;
5036c390 127 mempool_t *io_pool;
c38d29b3 128 struct bio_set *bs;
e8deb638 129 mempool_t *meta_pool;
f6bed0ef 130
0576b1c6
SL		 131 struct md_thread *reclaim_thread;
132 unsigned long reclaim_target; /* number of space that need to be
133 * reclaimed. if it's 0, reclaim spaces
134 * used by io_units which are in
135 * IO_UNIT_STRIPE_END state (eg, reclaim
136 * dones't wait for specific io_unit
137 * switching to IO_UNIT_STRIPE_END
138 * state) */
0fd22b45 139 wait_queue_head_t iounit_wait;
0576b1c6 140
f6bed0ef
SL		 141 struct list_head no_space_stripes; /* pending stripes, log has no space */
142 spinlock_t no_space_stripes_lock;
56fef7c6
CH		 143
144 bool need_cache_flush;
2ded3703
SL		 145
146 /* for r5c_cache */
147 enum r5c_journal_mode r5c_journal_mode;
a39f7afd
SL		 148
149 /* all stripes in r5cache, in the order of seq at sh->log_start */
150 struct list_head stripe_in_journal_list;
151
152 spinlock_t stripe_in_journal_lock;
153 atomic_t stripe_in_journal_count;
3bddb7f8
SL		 154
155 /* to submit async io_units, to fulfill ordering of flush */
156 struct work_struct deferred_io_work;
2e38a37f
SL		 157 /* to disable write back during in degraded mode */
158 struct work_struct disable_writeback_work;
03b047f4
SL		 159
160 /* to for chunk_aligned_read in writeback mode, details below */
161 spinlock_t tree_lock;
162 struct radix_tree_root big_stripe_tree;
f6bed0ef
SL		 163};
164
03b047f4
SL		 165/*
166 * Enable chunk_aligned_read() with write back cache.
167 *
168 * Each chunk may contain more than one stripe (for example, a 256kB
169 * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
170 * chunk_aligned_read, these stripes are grouped into one "big_stripe".
171 * For each big_stripe, we count how many stripes of this big_stripe
172 * are in the write back cache. These data are tracked in a radix tree
173 * (big_stripe_tree). We use radix_tree item pointer as the counter.
174 * r5c_tree_index() is used to calculate keys for the radix tree.
175 *
176 * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
177 * big_stripe of each chunk in the tree. If this big_stripe is in the
178 * tree, chunk_aligned_read() aborts. This look up is protected by
179 * rcu_read_lock().
180 *
181 * It is necessary to remember whether a stripe is counted in
182 * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
183 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
184 * two flags are set, the stripe is counted in big_stripe_tree. This
185 * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
186 * r5c_try_caching_write(); and moving clear_bit of
187 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
188 * r5c_finish_stripe_write_out().
189 */
190
191/*
192 * radix tree requests lowest 2 bits of data pointer to be 2b'00.
193 * So it is necessary to left shift the counter by 2 bits before using it
194 * as data pointer of the tree.
195 */
196#define R5C_RADIX_COUNT_SHIFT 2
197
198/*
199 * calculate key for big_stripe_tree
200 *
201 * sect: align_bi->bi_iter.bi_sector or sh->sector
202 */
203static inline sector_t r5c_tree_index(struct r5conf *conf,
204 sector_t sect)
205{
206 sector_t offset;
207
208 offset = sector_div(sect, conf->chunk_sectors);
209 return sect;
210}
211
f6bed0ef
SL		 212/*
213 * an IO range starts from a meta data block and end at the next meta data
214 * block. The io unit's the meta data block tracks data/parity followed it. io
215 * unit is written to log disk with normal write, as we always flush log disk
216 * first and then start move data to raid disks, there is no requirement to
217 * write io unit with FLUSH/FUA
218 */
219struct r5l_io_unit {
220 struct r5l_log *log;
221
222 struct page *meta_page; /* store meta block */
223 int meta_offset; /* current offset in meta_page */
224
f6bed0ef
SL		 225 struct bio *current_bio;/* current_bio accepting new data */
226
227 atomic_t pending_stripe;/* how many stripes not flushed to raid */
228 u64 seq; /* seq number of the metablock */
229 sector_t log_start; /* where the io_unit starts */
230 sector_t log_end; /* where the io_unit ends */
231 struct list_head log_sibling; /* log->running_ios */
232 struct list_head stripe_list; /* stripes added to the io_unit */
233
234 int state;
6143e2ce 235 bool need_split_bio;
3bddb7f8
SL		 236 struct bio *split_bio;
237
238 unsigned int has_flush:1; /* include flush request */
239 unsigned int has_fua:1; /* include fua request */
240 unsigned int has_null_flush:1; /* include empty flush request */
241 /*
242 * io isn't sent yet, flush/fua request can only be submitted till it's
243 * the first IO in running_ios list
244 */
245 unsigned int io_deferred:1;
246
247 struct bio_list flush_barriers; /* size == 0 flush bios */
f6bed0ef
SL		 248};
249
250/* r5l_io_unit state */
251enum r5l_io_unit_state {
252 IO_UNIT_RUNNING = 0, /* accepting new IO */
253 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
254 * don't accepting new bio */
255 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
a8c34f91 256 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
f6bed0ef
SL		 257};
258
2ded3703
SL		 259bool r5c_is_writeback(struct r5l_log *log)
260{
261 return (log != NULL &&
262 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
263}
264
f6bed0ef
SL		 265static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
266{
267 start += inc;
268 if (start >= log->device_size)
269 start = start - log->device_size;
270 return start;
271}
272
273static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
274 sector_t end)
275{
276 if (end >= start)
277 return end - start;
278 else
279 return end + log->device_size - start;
280}
281
282static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
283{
284 sector_t used_size;
285
286 used_size = r5l_ring_distance(log, log->last_checkpoint,
287 log->log_start);
288
289 return log->device_size > used_size + size;
290}
291
f6bed0ef
SL		 292static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
293 enum r5l_io_unit_state state)
294{
f6bed0ef
SL		 295 if (WARN_ON(io->state >= state))
296 return;
297 io->state = state;
f6bed0ef
SL		 298}
299
1e6d690b 300static void
bd83d0a2 301r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
1e6d690b
SL		 302{
303 struct bio *wbi, *wbi2;
304
305 wbi = dev->written;
306 dev->written = NULL;
307 while (wbi && wbi->bi_iter.bi_sector <
308 dev->sector + STRIPE_SECTORS) {
309 wbi2 = r5_next_bio(wbi, dev->sector);
49728050 310 md_write_end(conf->mddev);
016c76ac 311 bio_endio(wbi);
1e6d690b
SL		 312 wbi = wbi2;
313 }
314}
315
316void r5c_handle_cached_data_endio(struct r5conf *conf,
bd83d0a2 317 struct stripe_head *sh, int disks)
1e6d690b
SL		 318{
319 int i;
320
321 for (i = sh->disks; i--; ) {
322 if (sh->dev[i].written) {
323 set_bit(R5_UPTODATE, &sh->dev[i].flags);
bd83d0a2 324 r5c_return_dev_pending_writes(conf, &sh->dev[i]);
1e6d690b
SL		 325 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
326 STRIPE_SECTORS,
327 !test_bit(STRIPE_DEGRADED, &sh->state),
328 0);
329 }
330 }
331}
332
ff875738
AP		 333void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
334
a39f7afd
SL		 335/* Check whether we should flush some stripes to free up stripe cache */
336void r5c_check_stripe_cache_usage(struct r5conf *conf)
337{
338 int total_cached;
339
340 if (!r5c_is_writeback(conf->log))
341 return;
342
343 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
344 atomic_read(&conf->r5c_cached_full_stripes);
345
346 /*
347 * The following condition is true for either of the following:
348 * - stripe cache pressure high:
349 * total_cached > 3/4 min_nr_stripes ||
350 * empty_inactive_list_nr > 0
351 * - stripe cache pressure moderate:
352 * total_cached > 1/2 min_nr_stripes
353 */
354 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
355 atomic_read(&conf->empty_inactive_list_nr) > 0)
356 r5l_wake_reclaim(conf->log, 0);
357}
358
359/*
360 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
361 * stripes in the cache
362 */
363void r5c_check_cached_full_stripe(struct r5conf *conf)
364{
365 if (!r5c_is_writeback(conf->log))
366 return;
367
368 /*
369 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
370 * or a full stripe (chunk size / 4k stripes).
371 */
372 if (atomic_read(&conf->r5c_cached_full_stripes) >=
84890c03 373 min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
a39f7afd
SL		 374 conf->chunk_sectors >> STRIPE_SHIFT))
375 r5l_wake_reclaim(conf->log, 0);
376}
377
378/*
379 * Total log space (in sectors) needed to flush all data in cache
380 *
39b99586
SL		 381 * To avoid deadlock due to log space, it is necessary to reserve log
382 * space to flush critical stripes (stripes that occupying log space near
383 * last_checkpoint). This function helps check how much log space is
384 * required to flush all cached stripes.
a39f7afd 385 *
39b99586
SL		 386 * To reduce log space requirements, two mechanisms are used to give cache
387 * flush higher priorities:
388 * 1. In handle_stripe_dirtying() and schedule_reconstruction(),
389 * stripes ALREADY in journal can be flushed w/o pending writes;
390 * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
391 * can be delayed (r5l_add_no_space_stripe).
a39f7afd 392 *
39b99586
SL		 393 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
394 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
395 * pages of journal space. For stripes that has not passed 1, flushing it
396 * requires (conf->raid_disks + 1) pages of journal space. There are at
397 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
398 * required to flush all cached stripes (in pages) is:
399 *
400 * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
401 * (group_cnt + 1) * (raid_disks + 1)
402 * or
403 * (stripe_in_journal_count) * (max_degraded + 1) +
404 * (group_cnt + 1) * (raid_disks - max_degraded)
a39f7afd
SL		 405 */
406static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
407{
408 struct r5l_log *log = conf->log;
409
410 if (!r5c_is_writeback(log))
411 return 0;
412
39b99586
SL		 413 return BLOCK_SECTORS *
414 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
415 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
a39f7afd
SL		 416}
417
418/*
419 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
420 *
421 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
422 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
423 * device is less than 2x of reclaim_required_space.
424 */
425static inline void r5c_update_log_state(struct r5l_log *log)
426{
427 struct r5conf *conf = log->rdev->mddev->private;
428 sector_t free_space;
429 sector_t reclaim_space;
f687a33e 430 bool wake_reclaim = false;
a39f7afd
SL		 431
432 if (!r5c_is_writeback(log))
433 return;
434
435 free_space = r5l_ring_distance(log, log->log_start,
436 log->last_checkpoint);
437 reclaim_space = r5c_log_required_to_flush_cache(conf);
438 if (free_space < 2 * reclaim_space)
439 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
f687a33e
SL		 440 else {
441 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
442 wake_reclaim = true;
a39f7afd 443 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
f687a33e 444 }
a39f7afd
SL		 445 if (free_space < 3 * reclaim_space)
446 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
447 else
448 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
f687a33e
SL		 449
450 if (wake_reclaim)
451 r5l_wake_reclaim(log, 0);
a39f7afd
SL		 452}
453
2ded3703
SL		 454/*
455 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
456 * This function should only be called in write-back mode.
457 */
a39f7afd 458void r5c_make_stripe_write_out(struct stripe_head *sh)
2ded3703
SL		 459{
460 struct r5conf *conf = sh->raid_conf;
461 struct r5l_log *log = conf->log;
462
463 BUG_ON(!r5c_is_writeback(log));
464
465 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
466 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1e6d690b
SL		 467
468 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
469 atomic_inc(&conf->preread_active_stripes);
1e6d690b
SL		 470}
471
472static void r5c_handle_data_cached(struct stripe_head *sh)
473{
474 int i;
475
476 for (i = sh->disks; i--; )
477 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
478 set_bit(R5_InJournal, &sh->dev[i].flags);
479 clear_bit(R5_LOCKED, &sh->dev[i].flags);
480 }
481 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
482}
483
484/*
485 * this journal write must contain full parity,
486 * it may also contain some data pages
487 */
488static void r5c_handle_parity_cached(struct stripe_head *sh)
489{
490 int i;
491
492 for (i = sh->disks; i--; )
493 if (test_bit(R5_InJournal, &sh->dev[i].flags))
494 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2ded3703
SL		 495}
496
497/*
498 * Setting proper flags after writing (or flushing) data and/or parity to the
499 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
500 */
501static void r5c_finish_cache_stripe(struct stripe_head *sh)
502{
503 struct r5l_log *log = sh->raid_conf->log;
504
505 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
506 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
507 /*
508 * Set R5_InJournal for parity dev[pd_idx]. This means
509 * all data AND parity in the journal. For RAID 6, it is
510 * NOT necessary to set the flag for dev[qd_idx], as the
511 * two parities are written out together.
512 */
513 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
1e6d690b
SL		 514 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
515 r5c_handle_data_cached(sh);
516 } else {
517 r5c_handle_parity_cached(sh);
518 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
519 }
2ded3703
SL		 520}
521
d8858f43
CH		 522static void r5l_io_run_stripes(struct r5l_io_unit *io)
523{
524 struct stripe_head *sh, *next;
525
526 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
527 list_del_init(&sh->log_list);
2ded3703
SL		 528
529 r5c_finish_cache_stripe(sh);
530
d8858f43
CH		 531 set_bit(STRIPE_HANDLE, &sh->state);
532 raid5_release_stripe(sh);
533 }
534}
535
56fef7c6
CH		 536static void r5l_log_run_stripes(struct r5l_log *log)
537{
538 struct r5l_io_unit *io, *next;
539
540 assert_spin_locked(&log->io_list_lock);
541
542 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
543 /* don't change list order */
544 if (io->state < IO_UNIT_IO_END)
545 break;
546
547 list_move_tail(&io->log_sibling, &log->finished_ios);
548 r5l_io_run_stripes(io);
549 }
550}
551
3848c0bc
CH		 552static void r5l_move_to_end_ios(struct r5l_log *log)
553{
554 struct r5l_io_unit *io, *next;
555
556 assert_spin_locked(&log->io_list_lock);
557
558 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
559 /* don't change list order */
560 if (io->state < IO_UNIT_IO_END)
561 break;
562 list_move_tail(&io->log_sibling, &log->io_end_ios);
563 }
564}
565
3bddb7f8 566static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
f6bed0ef
SL		 567static void r5l_log_endio(struct bio *bio)
568{
569 struct r5l_io_unit *io = bio->bi_private;
3bddb7f8 570 struct r5l_io_unit *io_deferred;
f6bed0ef 571 struct r5l_log *log = io->log;
509ffec7 572 unsigned long flags;
f6bed0ef 573
6e74a9cf
SL		 574 if (bio->bi_error)
575 md_error(log->rdev->mddev, log->rdev);
576
f6bed0ef 577 bio_put(bio);
e8deb638 578 mempool_free(io->meta_page, log->meta_pool);
f6bed0ef 579
509ffec7
CH		 580 spin_lock_irqsave(&log->io_list_lock, flags);
581 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
ea17481f 582 if (log->need_cache_flush && !list_empty(&io->stripe_list))
3848c0bc 583 r5l_move_to_end_ios(log);
56fef7c6
CH		 584 else
585 r5l_log_run_stripes(log);
3bddb7f8
SL		 586 if (!list_empty(&log->running_ios)) {
587 /*
588 * FLUSH/FUA io_unit is deferred because of ordering, now we
589 * can dispatch it
590 */
591 io_deferred = list_first_entry(&log->running_ios,
592 struct r5l_io_unit, log_sibling);
593 if (io_deferred->io_deferred)
594 schedule_work(&log->deferred_io_work);
595 }
596
509ffec7
CH		 597 spin_unlock_irqrestore(&log->io_list_lock, flags);
598
56fef7c6
CH		 599 if (log->need_cache_flush)
600 md_wakeup_thread(log->rdev->mddev->thread);
3bddb7f8
SL		 601
602 if (io->has_null_flush) {
603 struct bio *bi;
604
605 WARN_ON(bio_list_empty(&io->flush_barriers));
606 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
607 bio_endio(bi);
608 atomic_dec(&io->pending_stripe);
609 }
3bddb7f8 610 }
ea17481f
SL		 611
612 /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
613 if (atomic_read(&io->pending_stripe) == 0)
614 __r5l_stripe_write_finished(io);
3bddb7f8
SL		 615}
616
617static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
618{
619 unsigned long flags;
620
621 spin_lock_irqsave(&log->io_list_lock, flags);
622 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
623 spin_unlock_irqrestore(&log->io_list_lock, flags);
624
625 if (io->has_flush)
20737738 626 io->current_bio->bi_opf |= REQ_PREFLUSH;
3bddb7f8 627 if (io->has_fua)
20737738 628 io->current_bio->bi_opf |= REQ_FUA;
3bddb7f8
SL		 629 submit_bio(io->current_bio);
630
631 if (!io->split_bio)
632 return;
633
634 if (io->has_flush)
20737738 635 io->split_bio->bi_opf |= REQ_PREFLUSH;
3bddb7f8 636 if (io->has_fua)
20737738 637 io->split_bio->bi_opf |= REQ_FUA;
3bddb7f8
SL		 638 submit_bio(io->split_bio);
639}
640
641/* deferred io_unit will be dispatched here */
642static void r5l_submit_io_async(struct work_struct *work)
643{
644 struct r5l_log *log = container_of(work, struct r5l_log,
645 deferred_io_work);
646 struct r5l_io_unit *io = NULL;
647 unsigned long flags;
648
649 spin_lock_irqsave(&log->io_list_lock, flags);
650 if (!list_empty(&log->running_ios)) {
651 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
652 log_sibling);
653 if (!io->io_deferred)
654 io = NULL;
655 else
656 io->io_deferred = 0;
657 }
658 spin_unlock_irqrestore(&log->io_list_lock, flags);
659 if (io)
660 r5l_do_submit_io(log, io);
f6bed0ef
SL		 661}
662
2e38a37f
SL		 663static void r5c_disable_writeback_async(struct work_struct *work)
664{
665 struct r5l_log *log = container_of(work, struct r5l_log,
666 disable_writeback_work);
667 struct mddev *mddev = log->rdev->mddev;
668
669 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
670 return;
671 pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
672 mdname(mddev));
673 mddev_suspend(mddev);
674 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
675 mddev_resume(mddev);
676}
677
f6bed0ef
SL		 678static void r5l_submit_current_io(struct r5l_log *log)
679{
680 struct r5l_io_unit *io = log->current_io;
3bddb7f8 681 struct bio *bio;
f6bed0ef 682 struct r5l_meta_block *block;
509ffec7 683 unsigned long flags;
f6bed0ef 684 u32 crc;
3bddb7f8 685 bool do_submit = true;
f6bed0ef
SL		 686
687 if (!io)
688 return;
689
690 block = page_address(io->meta_page);
691 block->meta_size = cpu_to_le32(io->meta_offset);
5cb2fbd6 692 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
f6bed0ef 693 block->checksum = cpu_to_le32(crc);
3bddb7f8 694 bio = io->current_bio;
f6bed0ef
SL		 695
696 log->current_io = NULL;
509ffec7 697 spin_lock_irqsave(&log->io_list_lock, flags);
3bddb7f8
SL		 698 if (io->has_flush || io->has_fua) {
699 if (io != list_first_entry(&log->running_ios,
700 struct r5l_io_unit, log_sibling)) {
701 io->io_deferred = 1;
702 do_submit = false;
703 }
704 }
509ffec7 705 spin_unlock_irqrestore(&log->io_list_lock, flags);
3bddb7f8
SL		 706 if (do_submit)
707 r5l_do_submit_io(log, io);
f6bed0ef
SL		 708}
709
6143e2ce 710static struct bio *r5l_bio_alloc(struct r5l_log *log)
b349feb3 711{
c38d29b3 712 struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
b349feb3 713
796a5cf0 714 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
b349feb3 715 bio->bi_bdev = log->rdev->bdev;
1e932a37 716 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
b349feb3 717
b349feb3
CH		 718 return bio;
719}
720
c1b99198
CH		 721static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
722{
723 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
724
a39f7afd 725 r5c_update_log_state(log);
c1b99198
CH		 726 /*
727 * If we filled up the log device start from the beginning again,
728 * which will require a new bio.
729 *
730 * Note: for this to work properly the log size needs to me a multiple
731 * of BLOCK_SECTORS.
732 */
733 if (log->log_start == 0)
6143e2ce 734 io->need_split_bio = true;
c1b99198
CH		 735
736 io->log_end = log->log_start;
737}
738
f6bed0ef
SL		 739static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
740{
741 struct r5l_io_unit *io;
742 struct r5l_meta_block *block;
f6bed0ef 743
5036c390
CH		 744 io = mempool_alloc(log->io_pool, GFP_ATOMIC);
745 if (!io)
746 return NULL;
747 memset(io, 0, sizeof(*io));
748
51039cd0 749 io->log = log;
51039cd0
CH		 750 INIT_LIST_HEAD(&io->log_sibling);
751 INIT_LIST_HEAD(&io->stripe_list);
3bddb7f8 752 bio_list_init(&io->flush_barriers);
51039cd0 753 io->state = IO_UNIT_RUNNING;
f6bed0ef 754
e8deb638 755 io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
f6bed0ef 756 block = page_address(io->meta_page);
e8deb638 757 clear_page(block);
f6bed0ef
SL		 758 block->magic = cpu_to_le32(R5LOG_MAGIC);
759 block->version = R5LOG_VERSION;
760 block->seq = cpu_to_le64(log->seq);
761 block->position = cpu_to_le64(log->log_start);
762
763 io->log_start = log->log_start;
764 io->meta_offset = sizeof(struct r5l_meta_block);
2b8ef16e 765 io->seq = log->seq++;
f6bed0ef 766
6143e2ce
CH		 767 io->current_bio = r5l_bio_alloc(log);
768 io->current_bio->bi_end_io = r5l_log_endio;
769 io->current_bio->bi_private = io;
b349feb3 770 bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
f6bed0ef 771
c1b99198 772 r5_reserve_log_entry(log, io);
f6bed0ef
SL		 773
774 spin_lock_irq(&log->io_list_lock);
775 list_add_tail(&io->log_sibling, &log->running_ios);
776 spin_unlock_irq(&log->io_list_lock);
777
778 return io;
779}
780
781static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
782{
22581f58
CH		 783 if (log->current_io &&
784 log->current_io->meta_offset + payload_size > PAGE_SIZE)
f6bed0ef 785 r5l_submit_current_io(log);
f6bed0ef 786
5036c390 787 if (!log->current_io) {
22581f58 788 log->current_io = r5l_new_meta(log);
5036c390
CH		 789 if (!log->current_io)
790 return -ENOMEM;
791 }
792
f6bed0ef
SL		 793 return 0;
794}
795
796static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
797 sector_t location,
798 u32 checksum1, u32 checksum2,
799 bool checksum2_valid)
800{
801 struct r5l_io_unit *io = log->current_io;
802 struct r5l_payload_data_parity *payload;
803
804 payload = page_address(io->meta_page) + io->meta_offset;
805 payload->header.type = cpu_to_le16(type);
806 payload->header.flags = cpu_to_le16(0);
807 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
808 (PAGE_SHIFT - 9));
809 payload->location = cpu_to_le64(location);
810 payload->checksum[0] = cpu_to_le32(checksum1);
811 if (checksum2_valid)
812 payload->checksum[1] = cpu_to_le32(checksum2);
813
814 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
815 sizeof(__le32) * (1 + !!checksum2_valid);
816}
817
818static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
819{
820 struct r5l_io_unit *io = log->current_io;
821
6143e2ce 822 if (io->need_split_bio) {
3bddb7f8
SL		 823 BUG_ON(io->split_bio);
824 io->split_bio = io->current_bio;
6143e2ce 825 io->current_bio = r5l_bio_alloc(log);
3bddb7f8
SL		 826 bio_chain(io->current_bio, io->split_bio);
827 io->need_split_bio = false;
f6bed0ef 828 }
f6bed0ef 829
6143e2ce
CH		 830 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
831 BUG();
832
c1b99198 833 r5_reserve_log_entry(log, io);
f6bed0ef
SL		 834}
835
ea17481f
SL		 836static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
837{
838 struct mddev *mddev = log->rdev->mddev;
839 struct r5conf *conf = mddev->private;
840 struct r5l_io_unit *io;
841 struct r5l_payload_flush *payload;
842 int meta_size;
843
844 /*
845 * payload_flush requires extra writes to the journal.
846 * To avoid handling the extra IO in quiesce, just skip
847 * flush_payload
848 */
849 if (conf->quiesce)
850 return;
851
852 mutex_lock(&log->io_mutex);
853 meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
854
855 if (r5l_get_meta(log, meta_size)) {
856 mutex_unlock(&log->io_mutex);
857 return;
858 }
859
860 /* current implementation is one stripe per flush payload */
861 io = log->current_io;
862 payload = page_address(io->meta_page) + io->meta_offset;
863 payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
864 payload->header.flags = cpu_to_le16(0);
865 payload->size = cpu_to_le32(sizeof(__le64));
866 payload->flush_stripes[0] = cpu_to_le64(sect);
867 io->meta_offset += meta_size;
868 mutex_unlock(&log->io_mutex);
869}
870
5036c390 871static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
f6bed0ef
SL		 872 int data_pages, int parity_pages)
873{
874 int i;
875 int meta_size;
5036c390 876 int ret;
f6bed0ef
SL		 877 struct r5l_io_unit *io;
878
879 meta_size =
880 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
881 * data_pages) +
882 sizeof(struct r5l_payload_data_parity) +
883 sizeof(__le32) * parity_pages;
884
5036c390
CH		 885 ret = r5l_get_meta(log, meta_size);
886 if (ret)
887 return ret;
888
f6bed0ef
SL		 889 io = log->current_io;
890
3bddb7f8
SL		 891 if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
892 io->has_flush = 1;
893
f6bed0ef 894 for (i = 0; i < sh->disks; i++) {
1e6d690b
SL		 895 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
896 test_bit(R5_InJournal, &sh->dev[i].flags))
f6bed0ef
SL		 897 continue;
898 if (i == sh->pd_idx || i == sh->qd_idx)
899 continue;
3bddb7f8
SL		 900 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
901 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
902 io->has_fua = 1;
903 /*
904 * we need to flush journal to make sure recovery can
905 * reach the data with fua flag
906 */
907 io->has_flush = 1;
908 }
f6bed0ef
SL		 909 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
910 raid5_compute_blocknr(sh, i, 0),
911 sh->dev[i].log_checksum, 0, false);
912 r5l_append_payload_page(log, sh->dev[i].page);
913 }
914
2ded3703 915 if (parity_pages == 2) {
f6bed0ef
SL		 916 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
917 sh->sector, sh->dev[sh->pd_idx].log_checksum,
918 sh->dev[sh->qd_idx].log_checksum, true);
919 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
920 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
2ded3703 921 } else if (parity_pages == 1) {
f6bed0ef
SL		 922 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
923 sh->sector, sh->dev[sh->pd_idx].log_checksum,
924 0, false);
925 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
2ded3703
SL		 926 } else /* Just writing data, not parity, in caching phase */
927 BUG_ON(parity_pages != 0);
f6bed0ef
SL		 928
929 list_add_tail(&sh->log_list, &io->stripe_list);
930 atomic_inc(&io->pending_stripe);
931 sh->log_io = io;
5036c390 932
a39f7afd
SL		 933 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
934 return 0;
935
936 if (sh->log_start == MaxSector) {
937 BUG_ON(!list_empty(&sh->r5c));
938 sh->log_start = io->log_start;
939 spin_lock_irq(&log->stripe_in_journal_lock);
940 list_add_tail(&sh->r5c,
941 &log->stripe_in_journal_list);
942 spin_unlock_irq(&log->stripe_in_journal_lock);
943 atomic_inc(&log->stripe_in_journal_count);
944 }
5036c390 945 return 0;
f6bed0ef
SL		 946}
947
a39f7afd
SL		 948/* add stripe to no_space_stripes, and then wake up reclaim */
949static inline void r5l_add_no_space_stripe(struct r5l_log *log,
950 struct stripe_head *sh)
951{
952 spin_lock(&log->no_space_stripes_lock);
953 list_add_tail(&sh->log_list, &log->no_space_stripes);
954 spin_unlock(&log->no_space_stripes_lock);
955}
956
f6bed0ef
SL		 957/*
958 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
959 * data from log to raid disks), so we shouldn't wait for reclaim here
960 */
961int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
962{
a39f7afd 963 struct r5conf *conf = sh->raid_conf;
f6bed0ef
SL		 964 int write_disks = 0;
965 int data_pages, parity_pages;
f6bed0ef
SL		 966 int reserve;
967 int i;
5036c390 968 int ret = 0;
a39f7afd 969 bool wake_reclaim = false;
f6bed0ef
SL		 970
971 if (!log)
972 return -EAGAIN;
973 /* Don't support stripe batch */
974 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
975 test_bit(STRIPE_SYNCING, &sh->state)) {
976 /* the stripe is written to log, we start writing it to raid */
977 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
978 return -EAGAIN;
979 }
980
2ded3703
SL		 981 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
982
f6bed0ef
SL		 983 for (i = 0; i < sh->disks; i++) {
984 void *addr;
985
1e6d690b
SL		 986 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
987 test_bit(R5_InJournal, &sh->dev[i].flags))
f6bed0ef 988 continue;
1e6d690b 989
f6bed0ef
SL		 990 write_disks++;
991 /* checksum is already calculated in last run */
992 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
993 continue;
994 addr = kmap_atomic(sh->dev[i].page);
5cb2fbd6
SL		 995 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
996 addr, PAGE_SIZE);
f6bed0ef
SL		 997 kunmap_atomic(addr);
998 }
999 parity_pages = 1 + !!(sh->qd_idx >= 0);
1000 data_pages = write_disks - parity_pages;
1001
f6bed0ef 1002 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
253f9fd4
SL		 1003 /*
1004 * The stripe must enter state machine again to finish the write, so
1005 * don't delay.
1006 */
1007 clear_bit(STRIPE_DELAYED, &sh->state);
f6bed0ef
SL		 1008 atomic_inc(&sh->count);
1009
1010 mutex_lock(&log->io_mutex);
1011 /* meta + data */
1012 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
f6bed0ef 1013
a39f7afd
SL		 1014 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1015 if (!r5l_has_free_space(log, reserve)) {
1016 r5l_add_no_space_stripe(log, sh);
1017 wake_reclaim = true;
1018 } else {
1019 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1020 if (ret) {
1021 spin_lock_irq(&log->io_list_lock);
1022 list_add_tail(&sh->log_list,
1023 &log->no_mem_stripes);
1024 spin_unlock_irq(&log->io_list_lock);
1025 }
1026 }
1027 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1028 /*
1029 * log space critical, do not process stripes that are
1030 * not in cache yet (sh->log_start == MaxSector).
1031 */
1032 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1033 sh->log_start == MaxSector) {
1034 r5l_add_no_space_stripe(log, sh);
1035 wake_reclaim = true;
1036 reserve = 0;
1037 } else if (!r5l_has_free_space(log, reserve)) {
1038 if (sh->log_start == log->last_checkpoint)
1039 BUG();
1040 else
1041 r5l_add_no_space_stripe(log, sh);
1042 } else {
1043 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1044 if (ret) {
1045 spin_lock_irq(&log->io_list_lock);
1046 list_add_tail(&sh->log_list,
1047 &log->no_mem_stripes);
1048 spin_unlock_irq(&log->io_list_lock);
1049 }
5036c390 1050 }
f6bed0ef 1051 }
f6bed0ef 1052
5036c390 1053 mutex_unlock(&log->io_mutex);
a39f7afd
SL		 1054 if (wake_reclaim)
1055 r5l_wake_reclaim(log, reserve);
f6bed0ef
SL		 1056 return 0;
1057}
1058
1059void r5l_write_stripe_run(struct r5l_log *log)
1060{
1061 if (!log)
1062 return;
1063 mutex_lock(&log->io_mutex);
1064 r5l_submit_current_io(log);
1065 mutex_unlock(&log->io_mutex);
1066}
1067
828cbe98
SL		 1068int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1069{
1070 if (!log)
1071 return -ENODEV;
3bddb7f8
SL		 1072
1073 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1074 /*
1075 * in write through (journal only)
1076 * we flush log disk cache first, then write stripe data to
1077 * raid disks. So if bio is finished, the log disk cache is
1078 * flushed already. The recovery guarantees we can recovery
1079 * the bio from log disk, so we don't need to flush again
1080 */
1081 if (bio->bi_iter.bi_size == 0) {
1082 bio_endio(bio);
1083 return 0;
1084 }
1085 bio->bi_opf &= ~REQ_PREFLUSH;
1086 } else {
1087 /* write back (with cache) */
1088 if (bio->bi_iter.bi_size == 0) {
1089 mutex_lock(&log->io_mutex);
1090 r5l_get_meta(log, 0);
1091 bio_list_add(&log->current_io->flush_barriers, bio);
1092 log->current_io->has_flush = 1;
1093 log->current_io->has_null_flush = 1;
1094 atomic_inc(&log->current_io->pending_stripe);
1095 r5l_submit_current_io(log);
1096 mutex_unlock(&log->io_mutex);
1097 return 0;
1098 }
828cbe98 1099 }
828cbe98
SL		 1100 return -EAGAIN;
1101}
1102
f6bed0ef
SL		 1103/* This will run after log space is reclaimed */
1104static void r5l_run_no_space_stripes(struct r5l_log *log)
1105{
1106 struct stripe_head *sh;
1107
1108 spin_lock(&log->no_space_stripes_lock);
1109 while (!list_empty(&log->no_space_stripes)) {
1110 sh = list_first_entry(&log->no_space_stripes,
1111 struct stripe_head, log_list);
1112 list_del_init(&sh->log_list);
1113 set_bit(STRIPE_HANDLE, &sh->state);
1114 raid5_release_stripe(sh);
1115 }
1116 spin_unlock(&log->no_space_stripes_lock);
1117}
1118
a39f7afd
SL		 1119/*
1120 * calculate new last_checkpoint
1121 * for write through mode, returns log->next_checkpoint
1122 * for write back, returns log_start of first sh in stripe_in_journal_list
1123 */
1124static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1125{
1126 struct stripe_head *sh;
1127 struct r5l_log *log = conf->log;
1128 sector_t new_cp;
1129 unsigned long flags;
1130
1131 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1132 return log->next_checkpoint;
1133
1134 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1135 if (list_empty(&conf->log->stripe_in_journal_list)) {
1136 /* all stripes flushed */
d3014e21 1137 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
a39f7afd
SL		 1138 return log->next_checkpoint;
1139 }
1140 sh = list_first_entry(&conf->log->stripe_in_journal_list,
1141 struct stripe_head, r5c);
1142 new_cp = sh->log_start;
1143 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1144 return new_cp;
1145}
1146
17036461
CH		 1147static sector_t r5l_reclaimable_space(struct r5l_log *log)
1148{
a39f7afd
SL		 1149 struct r5conf *conf = log->rdev->mddev->private;
1150
17036461 1151 return r5l_ring_distance(log, log->last_checkpoint,
a39f7afd 1152 r5c_calculate_new_cp(conf));
17036461
CH		 1153}
1154
5036c390
CH		 1155static void r5l_run_no_mem_stripe(struct r5l_log *log)
1156{
1157 struct stripe_head *sh;
1158
1159 assert_spin_locked(&log->io_list_lock);
1160
1161 if (!list_empty(&log->no_mem_stripes)) {
1162 sh = list_first_entry(&log->no_mem_stripes,
1163 struct stripe_head, log_list);
1164 list_del_init(&sh->log_list);
1165 set_bit(STRIPE_HANDLE, &sh->state);
1166 raid5_release_stripe(sh);
1167 }
1168}
1169
04732f74 1170static bool r5l_complete_finished_ios(struct r5l_log *log)
17036461
CH		 1171{
1172 struct r5l_io_unit *io, *next;
1173 bool found = false;
1174
1175 assert_spin_locked(&log->io_list_lock);
1176
04732f74 1177 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
17036461
CH		 1178 /* don't change list order */
1179 if (io->state < IO_UNIT_STRIPE_END)
1180 break;
1181
1182 log->next_checkpoint = io->log_start;
17036461
CH		 1183
1184 list_del(&io->log_sibling);
5036c390
CH		 1185 mempool_free(io, log->io_pool);
1186 r5l_run_no_mem_stripe(log);
17036461
CH		 1187
1188 found = true;
1189 }
1190
1191 return found;
1192}
1193
509ffec7
CH		 1194static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1195{
1196 struct r5l_log *log = io->log;
a39f7afd 1197 struct r5conf *conf = log->rdev->mddev->private;
509ffec7
CH		 1198 unsigned long flags;
1199
1200 spin_lock_irqsave(&log->io_list_lock, flags);
1201 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
17036461 1202
04732f74 1203 if (!r5l_complete_finished_ios(log)) {
85f2f9a4
SL		 1204 spin_unlock_irqrestore(&log->io_list_lock, flags);
1205 return;
1206 }
509ffec7 1207
a39f7afd
SL		 1208 if (r5l_reclaimable_space(log) > log->max_free_space ||
1209 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
509ffec7
CH		 1210 r5l_wake_reclaim(log, 0);
1211
509ffec7
CH		 1212 spin_unlock_irqrestore(&log->io_list_lock, flags);
1213 wake_up(&log->iounit_wait);
1214}
1215
0576b1c6
SL		 1216void r5l_stripe_write_finished(struct stripe_head *sh)
1217{
1218 struct r5l_io_unit *io;
1219
0576b1c6 1220 io = sh->log_io;
0576b1c6
SL		 1221 sh->log_io = NULL;
1222
509ffec7
CH		 1223 if (io && atomic_dec_and_test(&io->pending_stripe))
1224 __r5l_stripe_write_finished(io);
0576b1c6
SL		 1225}
1226
a8c34f91
SL		 1227static void r5l_log_flush_endio(struct bio *bio)
1228{
1229 struct r5l_log *log = container_of(bio, struct r5l_log,
1230 flush_bio);
1231 unsigned long flags;
1232 struct r5l_io_unit *io;
a8c34f91 1233
6e74a9cf
SL		 1234 if (bio->bi_error)
1235 md_error(log->rdev->mddev, log->rdev);
1236
a8c34f91 1237 spin_lock_irqsave(&log->io_list_lock, flags);
d8858f43
CH		 1238 list_for_each_entry(io, &log->flushing_ios, log_sibling)
1239 r5l_io_run_stripes(io);
04732f74 1240 list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
a8c34f91
SL		 1241 spin_unlock_irqrestore(&log->io_list_lock, flags);
1242}
1243
0576b1c6
SL		 1244/*
1245 * Starting dispatch IO to raid.
1246 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1247 * broken meta in the middle of a log causes recovery can't find meta at the
1248 * head of log. If operations require meta at the head persistent in log, we
1249 * must make sure meta before it persistent in log too. A case is:
1250 *
1251 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1252 * data/parity must be persistent in log before we do the write to raid disks.
1253 *
1254 * The solution is we restrictly maintain io_unit list order. In this case, we
1255 * only write stripes of an io_unit to raid disks till the io_unit is the first
1256 * one whose data/parity is in log.
1257 */
1258void r5l_flush_stripe_to_raid(struct r5l_log *log)
1259{
a8c34f91 1260 bool do_flush;
56fef7c6
CH		 1261
1262 if (!log || !log->need_cache_flush)
0576b1c6 1263 return;
0576b1c6
SL		 1264
1265 spin_lock_irq(&log->io_list_lock);
a8c34f91
SL		 1266 /* flush bio is running */
1267 if (!list_empty(&log->flushing_ios)) {
1268 spin_unlock_irq(&log->io_list_lock);
1269 return;
0576b1c6 1270 }
a8c34f91
SL		 1271 list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1272 do_flush = !list_empty(&log->flushing_ios);
0576b1c6 1273 spin_unlock_irq(&log->io_list_lock);
a8c34f91
SL		 1274
1275 if (!do_flush)
1276 return;
1277 bio_reset(&log->flush_bio);
1278 log->flush_bio.bi_bdev = log->rdev->bdev;
1279 log->flush_bio.bi_end_io = r5l_log_flush_endio;
70fd7614 1280 log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
4e49ea4a 1281 submit_bio(&log->flush_bio);
0576b1c6
SL		 1282}
1283
0576b1c6 1284static void r5l_write_super(struct r5l_log *log, sector_t cp);
4b482044
SL		 1285static void r5l_write_super_and_discard_space(struct r5l_log *log,
1286 sector_t end)
1287{
1288 struct block_device *bdev = log->rdev->bdev;
1289 struct mddev *mddev;
1290
1291 r5l_write_super(log, end);
1292
1293 if (!blk_queue_discard(bdev_get_queue(bdev)))
1294 return;
1295
1296 mddev = log->rdev->mddev;
1297 /*
8e018c21
SL		 1298 * Discard could zero data, so before discard we must make sure
1299 * superblock is updated to new log tail. Updating superblock (either
1300 * directly call md_update_sb() or depend on md thread) must hold
1301 * reconfig mutex. On the other hand, raid5_quiesce is called with
1302 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1303 * for all IO finish, hence waitting for reclaim thread, while reclaim
1304 * thread is calling this function and waitting for reconfig mutex. So
1305 * there is a deadlock. We workaround this issue with a trylock.
1306 * FIXME: we could miss discard if we can't take reconfig mutex
4b482044 1307 */
2953079c
SL		 1308 set_mask_bits(&mddev->sb_flags, 0,
1309 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
8e018c21
SL		 1310 if (!mddev_trylock(mddev))
1311 return;
1312 md_update_sb(mddev, 1);
1313 mddev_unlock(mddev);
4b482044 1314
6e74a9cf 1315 /* discard IO error really doesn't matter, ignore it */
4b482044
SL		 1316 if (log->last_checkpoint < end) {
1317 blkdev_issue_discard(bdev,
1318 log->last_checkpoint + log->rdev->data_offset,
1319 end - log->last_checkpoint, GFP_NOIO, 0);
1320 } else {
1321 blkdev_issue_discard(bdev,
1322 log->last_checkpoint + log->rdev->data_offset,
1323 log->device_size - log->last_checkpoint,
1324 GFP_NOIO, 0);
1325 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1326 GFP_NOIO, 0);
1327 }
1328}
1329
a39f7afd
SL		 1330/*
1331 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1332 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1333 *
1334 * must hold conf->device_lock
1335 */
1336static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
0576b1c6 1337{
a39f7afd
SL		 1338 BUG_ON(list_empty(&sh->lru));
1339 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1340 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
0576b1c6 1341
0576b1c6 1342 /*
a39f7afd
SL		 1343 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1344 * raid5_release_stripe() while holding conf->device_lock
0576b1c6 1345 */
a39f7afd
SL		 1346 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1347 assert_spin_locked(&conf->device_lock);
0576b1c6 1348
a39f7afd
SL		 1349 list_del_init(&sh->lru);
1350 atomic_inc(&sh->count);
17036461 1351
a39f7afd
SL		 1352 set_bit(STRIPE_HANDLE, &sh->state);
1353 atomic_inc(&conf->active_stripes);
1354 r5c_make_stripe_write_out(sh);
0576b1c6 1355
e33fbb9c
SL		 1356 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1357 atomic_inc(&conf->r5c_flushing_partial_stripes);
1358 else
1359 atomic_inc(&conf->r5c_flushing_full_stripes);
a39f7afd
SL		 1360 raid5_release_stripe(sh);
1361}
1362
1363/*
1364 * if num == 0, flush all full stripes
1365 * if num > 0, flush all full stripes. If less than num full stripes are
1366 * flushed, flush some partial stripes until totally num stripes are
1367 * flushed or there is no more cached stripes.
1368 */
1369void r5c_flush_cache(struct r5conf *conf, int num)
1370{
1371 int count;
1372 struct stripe_head *sh, *next;
1373
1374 assert_spin_locked(&conf->device_lock);
1375 if (!conf->log)
0576b1c6
SL		 1376 return;
1377
a39f7afd
SL		 1378 count = 0;
1379 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1380 r5c_flush_stripe(conf, sh);
1381 count++;
1382 }
1383
1384 if (count >= num)
1385 return;
1386 list_for_each_entry_safe(sh, next,
1387 &conf->r5c_partial_stripe_list, lru) {
1388 r5c_flush_stripe(conf, sh);
1389 if (++count >= num)
1390 break;
1391 }
1392}
1393
1394static void r5c_do_reclaim(struct r5conf *conf)
1395{
1396 struct r5l_log *log = conf->log;
1397 struct stripe_head *sh;
1398 int count = 0;
1399 unsigned long flags;
1400 int total_cached;
1401 int stripes_to_flush;
e33fbb9c 1402 int flushing_partial, flushing_full;
a39f7afd
SL		 1403
1404 if (!r5c_is_writeback(log))
1405 return;
1406
e33fbb9c
SL		 1407 flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1408 flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
a39f7afd 1409 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
e33fbb9c
SL		 1410 atomic_read(&conf->r5c_cached_full_stripes) -
1411 flushing_full - flushing_partial;
a39f7afd
SL		 1412
1413 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1414 atomic_read(&conf->empty_inactive_list_nr) > 0)
1415 /*
1416 * if stripe cache pressure high, flush all full stripes and
1417 * some partial stripes
1418 */
1419 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1420 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
e33fbb9c 1421 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
84890c03 1422 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
a39f7afd
SL		 1423 /*
1424 * if stripe cache pressure moderate, or if there is many full
1425 * stripes,flush all full stripes
1426 */
1427 stripes_to_flush = 0;
1428 else
1429 /* no need to flush */
1430 stripes_to_flush = -1;
1431
1432 if (stripes_to_flush >= 0) {
1433 spin_lock_irqsave(&conf->device_lock, flags);
1434 r5c_flush_cache(conf, stripes_to_flush);
1435 spin_unlock_irqrestore(&conf->device_lock, flags);
1436 }
1437
1438 /* if log space is tight, flush stripes on stripe_in_journal_list */
1439 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1440 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1441 spin_lock(&conf->device_lock);
1442 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1443 /*
1444 * stripes on stripe_in_journal_list could be in any
1445 * state of the stripe_cache state machine. In this
1446 * case, we only want to flush stripe on
1447 * r5c_cached_full/partial_stripes. The following
1448 * condition makes sure the stripe is on one of the
1449 * two lists.
1450 */
1451 if (!list_empty(&sh->lru) &&
1452 !test_bit(STRIPE_HANDLE, &sh->state) &&
1453 atomic_read(&sh->count) == 0) {
1454 r5c_flush_stripe(conf, sh);
e8fd52ee
SL		 1455 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1456 break;
a39f7afd 1457 }
a39f7afd
SL		 1458 }
1459 spin_unlock(&conf->device_lock);
1460 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1461 }
f687a33e
SL		 1462
1463 if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1464 r5l_run_no_space_stripes(log);
1465
a39f7afd
SL		 1466 md_wakeup_thread(conf->mddev->thread);
1467}
1468
0576b1c6
SL		 1469static void r5l_do_reclaim(struct r5l_log *log)
1470{
a39f7afd 1471 struct r5conf *conf = log->rdev->mddev->private;
0576b1c6 1472 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
17036461
CH		 1473 sector_t reclaimable;
1474 sector_t next_checkpoint;
a39f7afd 1475 bool write_super;
0576b1c6
SL		 1476
1477 spin_lock_irq(&log->io_list_lock);
a39f7afd
SL		 1478 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1479 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
0576b1c6
SL		 1480 /*
1481 * move proper io_unit to reclaim list. We should not change the order.
1482 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1483 * shouldn't reuse space of an unreclaimable io_unit
1484 */
1485 while (1) {
17036461
CH		 1486 reclaimable = r5l_reclaimable_space(log);
1487 if (reclaimable >= reclaim_target ||
0576b1c6
SL		 1488 (list_empty(&log->running_ios) &&
1489 list_empty(&log->io_end_ios) &&
a8c34f91 1490 list_empty(&log->flushing_ios) &&
04732f74 1491 list_empty(&log->finished_ios)))
0576b1c6
SL		 1492 break;
1493
17036461
CH		 1494 md_wakeup_thread(log->rdev->mddev->thread);
1495 wait_event_lock_irq(log->iounit_wait,
1496 r5l_reclaimable_space(log) > reclaimable,
1497 log->io_list_lock);
0576b1c6 1498 }
17036461 1499
a39f7afd 1500 next_checkpoint = r5c_calculate_new_cp(conf);
0576b1c6
SL		 1501 spin_unlock_irq(&log->io_list_lock);
1502
a39f7afd 1503 if (reclaimable == 0 || !write_super)
0576b1c6
SL		 1504 return;
1505
0576b1c6
SL		 1506 /*
1507 * write_super will flush cache of each raid disk. We must write super
1508 * here, because the log area might be reused soon and we don't want to
1509 * confuse recovery
1510 */
4b482044 1511 r5l_write_super_and_discard_space(log, next_checkpoint);
0576b1c6
SL		 1512
1513 mutex_lock(&log->io_mutex);
17036461 1514 log->last_checkpoint = next_checkpoint;
a39f7afd 1515 r5c_update_log_state(log);
0576b1c6 1516 mutex_unlock(&log->io_mutex);
0576b1c6 1517
17036461 1518 r5l_run_no_space_stripes(log);
0576b1c6
SL		 1519}
1520
1521static void r5l_reclaim_thread(struct md_thread *thread)
1522{
1523 struct mddev *mddev = thread->mddev;
1524 struct r5conf *conf = mddev->private;
1525 struct r5l_log *log = conf->log;
1526
1527 if (!log)
1528 return;
a39f7afd 1529 r5c_do_reclaim(conf);
0576b1c6
SL		 1530 r5l_do_reclaim(log);
1531}
1532
a39f7afd 1533void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
f6bed0ef 1534{
0576b1c6
SL		 1535 unsigned long target;
1536 unsigned long new = (unsigned long)space; /* overflow in theory */
1537
a39f7afd
SL		 1538 if (!log)
1539 return;
0576b1c6
SL		 1540 do {
1541 target = log->reclaim_target;
1542 if (new < target)
1543 return;
1544 } while (cmpxchg(&log->reclaim_target, target, new) != target);
1545 md_wakeup_thread(log->reclaim_thread);
f6bed0ef
SL		 1546}
1547
e6c033f7
SL		 1548void r5l_quiesce(struct r5l_log *log, int state)
1549{
4b482044 1550 struct mddev *mddev;
e6c033f7
SL		 1551 if (!log || state == 2)
1552 return;
ce1ccd07
SL		 1553 if (state == 0)
1554 kthread_unpark(log->reclaim_thread->tsk);
1555 else if (state == 1) {
4b482044
SL		 1556 /* make sure r5l_write_super_and_discard_space exits */
1557 mddev = log->rdev->mddev;
1558 wake_up(&mddev->sb_wait);
ce1ccd07 1559 kthread_park(log->reclaim_thread->tsk);
a39f7afd 1560 r5l_wake_reclaim(log, MaxSector);
e6c033f7
SL		 1561 r5l_do_reclaim(log);
1562 }
1563}
1564
6e74a9cf
SL		 1565bool r5l_log_disk_error(struct r5conf *conf)
1566{
f6b6ec5c
SL		 1567 struct r5l_log *log;
1568 bool ret;
7dde2ad3 1569 /* don't allow write if journal disk is missing */
f6b6ec5c
SL		 1570 rcu_read_lock();
1571 log = rcu_dereference(conf->log);
1572
1573 if (!log)
1574 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1575 else
1576 ret = test_bit(Faulty, &log->rdev->flags);
1577 rcu_read_unlock();
1578 return ret;
6e74a9cf
SL		 1579}
1580
effe6ee7
SL		 1581#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1582
355810d1
SL		 1583struct r5l_recovery_ctx {
1584 struct page *meta_page; /* current meta */
1585 sector_t meta_total_blocks; /* total size of current meta and data */
1586 sector_t pos; /* recovery position */
1587 u64 seq; /* recovery position seq */
b4c625c6
SL		 1588 int data_parity_stripes; /* number of data_parity stripes */
1589 int data_only_stripes; /* number of data_only stripes */
1590 struct list_head cached_list;
effe6ee7
SL		 1591
1592 /*
1593 * read ahead page pool (ra_pool)
1594 * in recovery, log is read sequentially. It is not efficient to
1595 * read every page with sync_page_io(). The read ahead page pool
1596 * reads multiple pages with one IO, so further log read can
1597 * just copy data from the pool.
1598 */
1599 struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1600 sector_t pool_offset; /* offset of first page in the pool */
1601 int total_pages; /* total allocated pages */
1602 int valid_pages; /* pages with valid data */
1603 struct bio *ra_bio; /* bio to do the read ahead */
355810d1
SL		 1604};
1605
effe6ee7
SL		 1606static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1607 struct r5l_recovery_ctx *ctx)
1608{
1609 struct page *page;
1610
1611 ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs);
1612 if (!ctx->ra_bio)
1613 return -ENOMEM;
1614
1615 ctx->valid_pages = 0;
1616 ctx->total_pages = 0;
1617 while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1618 page = alloc_page(GFP_KERNEL);
1619
1620 if (!page)
1621 break;
1622 ctx->ra_pool[ctx->total_pages] = page;
1623 ctx->total_pages += 1;
1624 }
1625
1626 if (ctx->total_pages == 0) {
1627 bio_put(ctx->ra_bio);
1628 return -ENOMEM;
1629 }
1630
1631 ctx->pool_offset = 0;
1632 return 0;
1633}
1634
1635static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1636 struct r5l_recovery_ctx *ctx)
1637{
1638 int i;
1639
1640 for (i = 0; i < ctx->total_pages; ++i)
1641 put_page(ctx->ra_pool[i]);
1642 bio_put(ctx->ra_bio);
1643}
1644
1645/*
1646 * fetch ctx->valid_pages pages from offset
1647 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1648 * However, if the offset is close to the end of the journal device,
1649 * ctx->valid_pages could be smaller than ctx->total_pages
1650 */
1651static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1652 struct r5l_recovery_ctx *ctx,
1653 sector_t offset)
1654{
1655 bio_reset(ctx->ra_bio);
1656 ctx->ra_bio->bi_bdev = log->rdev->bdev;
1657 bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1658 ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1659
1660 ctx->valid_pages = 0;
1661 ctx->pool_offset = offset;
1662
1663 while (ctx->valid_pages < ctx->total_pages) {
1664 bio_add_page(ctx->ra_bio,
1665 ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1666 ctx->valid_pages += 1;
1667
1668 offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1669
1670 if (offset == 0) /* reached end of the device */
1671 break;
1672 }
1673
1674 return submit_bio_wait(ctx->ra_bio);
1675}
1676
1677/*
1678 * try read a page from the read ahead page pool, if the page is not in the
1679 * pool, call r5l_recovery_fetch_ra_pool
1680 */
1681static int r5l_recovery_read_page(struct r5l_log *log,
1682 struct r5l_recovery_ctx *ctx,
1683 struct page *page,
1684 sector_t offset)
1685{
1686 int ret;
1687
1688 if (offset < ctx->pool_offset ||
1689 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1690 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1691 if (ret)
1692 return ret;
1693 }
1694
1695 BUG_ON(offset < ctx->pool_offset ||
1696 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1697
1698 memcpy(page_address(page),
1699 page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1700 BLOCK_SECTOR_SHIFT]),
1701 PAGE_SIZE);
1702 return 0;
1703}
1704
9ed988f5
SL		 1705static int r5l_recovery_read_meta_block(struct r5l_log *log,
1706 struct r5l_recovery_ctx *ctx)
355810d1
SL		 1707{
1708 struct page *page = ctx->meta_page;
1709 struct r5l_meta_block *mb;
1710 u32 crc, stored_crc;
effe6ee7 1711 int ret;
355810d1 1712
effe6ee7
SL		 1713 ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1714 if (ret != 0)
1715 return ret;
355810d1
SL		 1716
1717 mb = page_address(page);
1718 stored_crc = le32_to_cpu(mb->checksum);
1719 mb->checksum = 0;
1720
1721 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1722 le64_to_cpu(mb->seq) != ctx->seq ||
1723 mb->version != R5LOG_VERSION ||
1724 le64_to_cpu(mb->position) != ctx->pos)
1725 return -EINVAL;
1726
5cb2fbd6 1727 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
355810d1
SL		 1728 if (stored_crc != crc)
1729 return -EINVAL;
1730
1731 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1732 return -EINVAL;
1733
1734 ctx->meta_total_blocks = BLOCK_SECTORS;
1735
1736 return 0;
1737}
1738
9ed988f5
SL		 1739static void
1740r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1741 struct page *page,
1742 sector_t pos, u64 seq)
355810d1 1743{
355810d1 1744 struct r5l_meta_block *mb;
355810d1 1745
355810d1 1746 mb = page_address(page);
9ed988f5 1747 clear_page(mb);
355810d1
SL		 1748 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1749 mb->version = R5LOG_VERSION;
1750 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1751 mb->seq = cpu_to_le64(seq);
1752 mb->position = cpu_to_le64(pos);
9ed988f5 1753}
355810d1 1754
9ed988f5
SL		 1755static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1756 u64 seq)
1757{
1758 struct page *page;
5c88f403 1759 struct r5l_meta_block *mb;
355810d1 1760
9ed988f5
SL		 1761 page = alloc_page(GFP_KERNEL);
1762 if (!page)
1763 return -ENOMEM;
1764 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
5c88f403
SL		 1765 mb = page_address(page);
1766 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1767 mb, PAGE_SIZE));
796a5cf0 1768 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
20737738 1769 REQ_FUA, false)) {
355810d1
SL		 1770 __free_page(page);
1771 return -EIO;
1772 }
1773 __free_page(page);
1774 return 0;
1775}
355810d1 1776
b4c625c6
SL		 1777/*
1778 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1779 * to mark valid (potentially not flushed) data in the journal.
1780 *
1781 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1782 * so there should not be any mismatch here.
1783 */
1784static void r5l_recovery_load_data(struct r5l_log *log,
1785 struct stripe_head *sh,
1786 struct r5l_recovery_ctx *ctx,
1787 struct r5l_payload_data_parity *payload,
1788 sector_t log_offset)
1789{
1790 struct mddev *mddev = log->rdev->mddev;
1791 struct r5conf *conf = mddev->private;
1792 int dd_idx;
1793
1794 raid5_compute_sector(conf,
1795 le64_to_cpu(payload->location), 0,
1796 &dd_idx, sh);
effe6ee7 1797 r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
b4c625c6
SL		 1798 sh->dev[dd_idx].log_checksum =
1799 le32_to_cpu(payload->checksum[0]);
1800 ctx->meta_total_blocks += BLOCK_SECTORS;
1801
1802 set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1803 set_bit(STRIPE_R5C_CACHING, &sh->state);
1804}
1805
1806static void r5l_recovery_load_parity(struct r5l_log *log,
1807 struct stripe_head *sh,
1808 struct r5l_recovery_ctx *ctx,
1809 struct r5l_payload_data_parity *payload,
1810 sector_t log_offset)
1811{
1812 struct mddev *mddev = log->rdev->mddev;
1813 struct r5conf *conf = mddev->private;
1814
1815 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
effe6ee7 1816 r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
b4c625c6
SL		 1817 sh->dev[sh->pd_idx].log_checksum =
1818 le32_to_cpu(payload->checksum[0]);
1819 set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1820
1821 if (sh->qd_idx >= 0) {
effe6ee7
SL		 1822 r5l_recovery_read_page(
1823 log, ctx, sh->dev[sh->qd_idx].page,
1824 r5l_ring_add(log, log_offset, BLOCK_SECTORS));
b4c625c6
SL		 1825 sh->dev[sh->qd_idx].log_checksum =
1826 le32_to_cpu(payload->checksum[1]);
1827 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
355810d1 1828 }
b4c625c6
SL		 1829 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1830}
355810d1 1831
b4c625c6
SL		 1832static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1833{
1834 int i;
1835
1836 sh->state = 0;
1837 sh->log_start = MaxSector;
1838 for (i = sh->disks; i--; )
1839 sh->dev[i].flags = 0;
1840}
1841
1842static void
1843r5l_recovery_replay_one_stripe(struct r5conf *conf,
1844 struct stripe_head *sh,
1845 struct r5l_recovery_ctx *ctx)
1846{
1847 struct md_rdev *rdev, *rrdev;
1848 int disk_index;
1849 int data_count = 0;
355810d1 1850
b4c625c6 1851 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
355810d1
SL		 1852 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1853 continue;
b4c625c6
SL		 1854 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1855 continue;
1856 data_count++;
355810d1
SL		 1857 }
1858
b4c625c6
SL		 1859 /*
1860 * stripes that only have parity must have been flushed
1861 * before the crash that we are now recovering from, so
1862 * there is nothing more to recovery.
1863 */
1864 if (data_count == 0)
1865 goto out;
355810d1 1866
b4c625c6
SL		 1867 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1868 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
355810d1
SL		 1869 continue;
1870
1871 /* in case device is broken */
b4c625c6 1872 rcu_read_lock();
355810d1 1873 rdev = rcu_dereference(conf->disks[disk_index].rdev);
b4c625c6
SL		 1874 if (rdev) {
1875 atomic_inc(&rdev->nr_pending);
1876 rcu_read_unlock();
1877 sync_page_io(rdev, sh->sector, PAGE_SIZE,
796a5cf0
MC		 1878 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1879 false);
b4c625c6
SL		 1880 rdev_dec_pending(rdev, rdev->mddev);
1881 rcu_read_lock();
1882 }
355810d1 1883 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
b4c625c6
SL		 1884 if (rrdev) {
1885 atomic_inc(&rrdev->nr_pending);
1886 rcu_read_unlock();
1887 sync_page_io(rrdev, sh->sector, PAGE_SIZE,
796a5cf0
MC		 1888 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1889 false);
b4c625c6
SL		 1890 rdev_dec_pending(rrdev, rrdev->mddev);
1891 rcu_read_lock();
1892 }
1893 rcu_read_unlock();
355810d1 1894 }
b4c625c6
SL		 1895 ctx->data_parity_stripes++;
1896out:
1897 r5l_recovery_reset_stripe(sh);
1898}
1899
1900static struct stripe_head *
1901r5c_recovery_alloc_stripe(struct r5conf *conf,
3c66abba 1902 sector_t stripe_sect)
b4c625c6
SL		 1903{
1904 struct stripe_head *sh;
1905
1906 sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1907 if (!sh)
1908 return NULL; /* no more stripe available */
1909
1910 r5l_recovery_reset_stripe(sh);
b4c625c6
SL		 1911
1912 return sh;
1913}
1914
1915static struct stripe_head *
1916r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1917{
1918 struct stripe_head *sh;
1919
1920 list_for_each_entry(sh, list, lru)
1921 if (sh->sector == sect)
1922 return sh;
1923 return NULL;
1924}
1925
1926static void
1927r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1928 struct r5l_recovery_ctx *ctx)
1929{
1930 struct stripe_head *sh, *next;
1931
1932 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1933 r5l_recovery_reset_stripe(sh);
1934 list_del_init(&sh->lru);
1935 raid5_release_stripe(sh);
1936 }
1937}
1938
1939static void
1940r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1941 struct r5l_recovery_ctx *ctx)
1942{
1943 struct stripe_head *sh, *next;
1944
1945 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1946 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1947 r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1948 list_del_init(&sh->lru);
1949 raid5_release_stripe(sh);
1950 }
1951}
1952
1953/* if matches return 0; otherwise return -EINVAL */
1954static int
effe6ee7
SL		 1955r5l_recovery_verify_data_checksum(struct r5l_log *log,
1956 struct r5l_recovery_ctx *ctx,
1957 struct page *page,
b4c625c6
SL		 1958 sector_t log_offset, __le32 log_checksum)
1959{
1960 void *addr;
1961 u32 checksum;
1962
effe6ee7 1963 r5l_recovery_read_page(log, ctx, page, log_offset);
b4c625c6
SL		 1964 addr = kmap_atomic(page);
1965 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1966 kunmap_atomic(addr);
1967 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1968}
1969
1970/*
1971 * before loading data to stripe cache, we need verify checksum for all data,
1972 * if there is mismatch for any data page, we drop all data in the mata block
1973 */
1974static int
1975r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1976 struct r5l_recovery_ctx *ctx)
1977{
1978 struct mddev *mddev = log->rdev->mddev;
1979 struct r5conf *conf = mddev->private;
1980 struct r5l_meta_block *mb = page_address(ctx->meta_page);
1981 sector_t mb_offset = sizeof(struct r5l_meta_block);
1982 sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1983 struct page *page;
1984 struct r5l_payload_data_parity *payload;
2d4f4687 1985 struct r5l_payload_flush *payload_flush;
b4c625c6
SL		 1986
1987 page = alloc_page(GFP_KERNEL);
1988 if (!page)
1989 return -ENOMEM;
1990
1991 while (mb_offset < le32_to_cpu(mb->meta_size)) {
1992 payload = (void *)mb + mb_offset;
2d4f4687 1993 payload_flush = (void *)mb + mb_offset;
b4c625c6 1994
1ad45a9b 1995 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
b4c625c6 1996 if (r5l_recovery_verify_data_checksum(
effe6ee7 1997 log, ctx, page, log_offset,
b4c625c6
SL		 1998 payload->checksum[0]) < 0)
1999 goto mismatch;
1ad45a9b 2000 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
b4c625c6 2001 if (r5l_recovery_verify_data_checksum(
effe6ee7 2002 log, ctx, page, log_offset,
b4c625c6
SL		 2003 payload->checksum[0]) < 0)
2004 goto mismatch;
2005 if (conf->max_degraded == 2 && /* q for RAID 6 */
2006 r5l_recovery_verify_data_checksum(
effe6ee7 2007 log, ctx, page,
b4c625c6
SL		 2008 r5l_ring_add(log, log_offset,
2009 BLOCK_SECTORS),
2010 payload->checksum[1]) < 0)
2011 goto mismatch;
1ad45a9b 2012 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2d4f4687
SL		 2013 /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2014 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
b4c625c6
SL		 2015 goto mismatch;
2016
1ad45a9b 2017 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2d4f4687
SL		 2018 mb_offset += sizeof(struct r5l_payload_flush) +
2019 le32_to_cpu(payload_flush->size);
2020 } else {
2021 /* DATA or PARITY payload */
2022 log_offset = r5l_ring_add(log, log_offset,
2023 le32_to_cpu(payload->size));
2024 mb_offset += sizeof(struct r5l_payload_data_parity) +
2025 sizeof(__le32) *
2026 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2027 }
b4c625c6 2028
b4c625c6
SL		 2029 }
2030
2031 put_page(page);
355810d1
SL		 2032 return 0;
2033
b4c625c6
SL		 2034mismatch:
2035 put_page(page);
355810d1
SL		 2036 return -EINVAL;
2037}
2038
b4c625c6
SL		 2039/*
2040 * Analyze all data/parity pages in one meta block
2041 * Returns:
2042 * 0 for success
2043 * -EINVAL for unknown playload type
2044 * -EAGAIN for checksum mismatch of data page
2045 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2046 */
2047static int
2048r5c_recovery_analyze_meta_block(struct r5l_log *log,
2049 struct r5l_recovery_ctx *ctx,
2050 struct list_head *cached_stripe_list)
355810d1 2051{
b4c625c6
SL		 2052 struct mddev *mddev = log->rdev->mddev;
2053 struct r5conf *conf = mddev->private;
355810d1 2054 struct r5l_meta_block *mb;
b4c625c6 2055 struct r5l_payload_data_parity *payload;
2d4f4687 2056 struct r5l_payload_flush *payload_flush;
b4c625c6 2057 int mb_offset;
355810d1 2058 sector_t log_offset;
b4c625c6
SL		 2059 sector_t stripe_sect;
2060 struct stripe_head *sh;
2061 int ret;
2062
2063 /*
2064 * for mismatch in data blocks, we will drop all data in this mb, but
2065 * we will still read next mb for other data with FLUSH flag, as
2066 * io_unit could finish out of order.
2067 */
2068 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2069 if (ret == -EINVAL)
2070 return -EAGAIN;
2071 else if (ret)
2072 return ret; /* -ENOMEM duo to alloc_page() failed */
355810d1
SL		 2073
2074 mb = page_address(ctx->meta_page);
b4c625c6 2075 mb_offset = sizeof(struct r5l_meta_block);
355810d1
SL		 2076 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2077
b4c625c6 2078 while (mb_offset < le32_to_cpu(mb->meta_size)) {
355810d1
SL		 2079 int dd;
2080
b4c625c6 2081 payload = (void *)mb + mb_offset;
2d4f4687
SL		 2082 payload_flush = (void *)mb + mb_offset;
2083
1ad45a9b 2084 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2d4f4687
SL		 2085 int i, count;
2086
2087 count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2088 for (i = 0; i < count; ++i) {
2089 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2090 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2091 stripe_sect);
2092 if (sh) {
2093 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2094 r5l_recovery_reset_stripe(sh);
2095 list_del_init(&sh->lru);
2096 raid5_release_stripe(sh);
2097 }
2098 }
2099
2100 mb_offset += sizeof(struct r5l_payload_flush) +
2101 le32_to_cpu(payload_flush->size);
2102 continue;
2103 }
2104
2105 /* DATA or PARITY payload */
1ad45a9b 2106 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
b4c625c6
SL		 2107 raid5_compute_sector(
2108 conf, le64_to_cpu(payload->location), 0, &dd,
2109 NULL)
2110 : le64_to_cpu(payload->location);
2111
2112 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2113 stripe_sect);
2114
2115 if (!sh) {
3c66abba 2116 sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
b4c625c6
SL		 2117 /*
2118 * cannot get stripe from raid5_get_active_stripe
2119 * try replay some stripes
2120 */
2121 if (!sh) {
2122 r5c_recovery_replay_stripes(
2123 cached_stripe_list, ctx);
2124 sh = r5c_recovery_alloc_stripe(
3c66abba 2125 conf, stripe_sect);
b4c625c6
SL		 2126 }
2127 if (!sh) {
2128 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2129 mdname(mddev),
2130 conf->min_nr_stripes * 2);
2131 raid5_set_cache_size(mddev,
2132 conf->min_nr_stripes * 2);
3c66abba
SL		 2133 sh = r5c_recovery_alloc_stripe(conf,
2134 stripe_sect);
b4c625c6
SL		 2135 }
2136 if (!sh) {
2137 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2138 mdname(mddev));
2139 return -ENOMEM;
2140 }
2141 list_add_tail(&sh->lru, cached_stripe_list);
2142 }
2143
1ad45a9b 2144 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
f7b7bee7
ZL		 2145 if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2146 test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
b4c625c6 2147 r5l_recovery_replay_one_stripe(conf, sh, ctx);
b4c625c6
SL		 2148 list_move_tail(&sh->lru, cached_stripe_list);
2149 }
2150 r5l_recovery_load_data(log, sh, ctx, payload,
2151 log_offset);
1ad45a9b 2152 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
b4c625c6
SL		 2153 r5l_recovery_load_parity(log, sh, ctx, payload,
2154 log_offset);
2155 else
355810d1 2156 return -EINVAL;
b4c625c6
SL		 2157
2158 log_offset = r5l_ring_add(log, log_offset,
2159 le32_to_cpu(payload->size));
2160
2161 mb_offset += sizeof(struct r5l_payload_data_parity) +
2162 sizeof(__le32) *
2163 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
355810d1 2164 }
b4c625c6 2165
355810d1
SL		 2166 return 0;
2167}
2168
b4c625c6
SL		 2169/*
2170 * Load the stripe into cache. The stripe will be written out later by
2171 * the stripe cache state machine.
2172 */
2173static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2174 struct stripe_head *sh)
355810d1 2175{
b4c625c6
SL		 2176 struct r5dev *dev;
2177 int i;
2178
2179 for (i = sh->disks; i--; ) {
2180 dev = sh->dev + i;
2181 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2182 set_bit(R5_InJournal, &dev->flags);
2183 set_bit(R5_UPTODATE, &dev->flags);
2184 }
2185 }
b4c625c6
SL		 2186}
2187
2188/*
2189 * Scan through the log for all to-be-flushed data
2190 *
2191 * For stripes with data and parity, namely Data-Parity stripe
2192 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2193 *
2194 * For stripes with only data, namely Data-Only stripe
2195 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2196 *
2197 * For a stripe, if we see data after parity, we should discard all previous
2198 * data and parity for this stripe, as these data are already flushed to
2199 * the array.
2200 *
2201 * At the end of the scan, we return the new journal_tail, which points to
2202 * first data-only stripe on the journal device, or next invalid meta block.
2203 */
2204static int r5c_recovery_flush_log(struct r5l_log *log,
2205 struct r5l_recovery_ctx *ctx)
2206{
bc8f167f 2207 struct stripe_head *sh;
b4c625c6
SL		 2208 int ret = 0;
2209
2210 /* scan through the log */
355810d1 2211 while (1) {
b4c625c6
SL		 2212 if (r5l_recovery_read_meta_block(log, ctx))
2213 break;
2214
2215 ret = r5c_recovery_analyze_meta_block(log, ctx,
2216 &ctx->cached_list);
2217 /*
2218 * -EAGAIN means mismatch in data block, in this case, we still
2219 * try scan the next metablock
2220 */
2221 if (ret && ret != -EAGAIN)
2222 break; /* ret == -EINVAL or -ENOMEM */
355810d1
SL		 2223 ctx->seq++;
2224 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2225 }
b4c625c6
SL		 2226
2227 if (ret == -ENOMEM) {
2228 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2229 return ret;
2230 }
2231
2232 /* replay data-parity stripes */
2233 r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2234
2235 /* load data-only stripes to stripe cache */
bc8f167f 2236 list_for_each_entry(sh, &ctx->cached_list, lru) {
b4c625c6
SL		 2237 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2238 r5c_recovery_load_one_stripe(log, sh);
b4c625c6
SL		 2239 ctx->data_only_stripes++;
2240 }
2241
2242 return 0;
355810d1
SL		 2243}
2244
b4c625c6
SL		 2245/*
2246 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2247 * log will start here. but we can't let superblock point to last valid
2248 * meta block. The log might looks like:
2249 * | meta 1| meta 2| meta 3|
2250 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2251 * superblock points to meta 1, we write a new valid meta 2n. if crash
2252 * happens again, new recovery will start from meta 1. Since meta 2n is
2253 * valid now, recovery will think meta 3 is valid, which is wrong.
2254 * The solution is we create a new meta in meta2 with its seq == meta
3c6edc66
SL		 2255 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2256 * will not think meta 3 is a valid meta, because its seq doesn't match
b4c625c6
SL		 2257 */
2258
2259/*
2260 * Before recovery, the log looks like the following
2261 *
2262 * ---------------------------------------------
2263 * | valid log | invalid log |
2264 * ---------------------------------------------
2265 * ^
2266 * |- log->last_checkpoint
2267 * |- log->last_cp_seq
2268 *
2269 * Now we scan through the log until we see invalid entry
2270 *
2271 * ---------------------------------------------
2272 * | valid log | invalid log |
2273 * ---------------------------------------------
2274 * ^ ^
2275 * |- log->last_checkpoint |- ctx->pos
2276 * |- log->last_cp_seq |- ctx->seq
2277 *
2278 * From this point, we need to increase seq number by 10 to avoid
2279 * confusing next recovery.
2280 *
2281 * ---------------------------------------------
2282 * | valid log | invalid log |
2283 * ---------------------------------------------
2284 * ^ ^
2285 * |- log->last_checkpoint |- ctx->pos+1
3c6edc66 2286 * |- log->last_cp_seq |- ctx->seq+10001
b4c625c6
SL		 2287 *
2288 * However, it is not safe to start the state machine yet, because data only
2289 * parities are not yet secured in RAID. To save these data only parities, we
2290 * rewrite them from seq+11.
2291 *
2292 * -----------------------------------------------------------------
2293 * | valid log | data only stripes | invalid log |
2294 * -----------------------------------------------------------------
2295 * ^ ^
2296 * |- log->last_checkpoint |- ctx->pos+n
3c6edc66 2297 * |- log->last_cp_seq |- ctx->seq+10000+n
b4c625c6
SL		 2298 *
2299 * If failure happens again during this process, the recovery can safe start
2300 * again from log->last_checkpoint.
2301 *
2302 * Once data only stripes are rewritten to journal, we move log_tail
2303 *
2304 * -----------------------------------------------------------------
2305 * | old log | data only stripes | invalid log |
2306 * -----------------------------------------------------------------
2307 * ^ ^
2308 * |- log->last_checkpoint |- ctx->pos+n
3c6edc66 2309 * |- log->last_cp_seq |- ctx->seq+10000+n
b4c625c6
SL		 2310 *
2311 * Then we can safely start the state machine. If failure happens from this
2312 * point on, the recovery will start from new log->last_checkpoint.
2313 */
2314static int
2315r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2316 struct r5l_recovery_ctx *ctx)
355810d1 2317{
a85dd7b8 2318 struct stripe_head *sh;
b4c625c6 2319 struct mddev *mddev = log->rdev->mddev;
355810d1 2320 struct page *page;
3c66abba 2321 sector_t next_checkpoint = MaxSector;
355810d1 2322
b4c625c6
SL		 2323 page = alloc_page(GFP_KERNEL);
2324 if (!page) {
2325 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2326 mdname(mddev));
355810d1 2327 return -ENOMEM;
b4c625c6 2328 }
355810d1 2329
3c66abba
SL		 2330 WARN_ON(list_empty(&ctx->cached_list));
2331
a85dd7b8 2332 list_for_each_entry(sh, &ctx->cached_list, lru) {
b4c625c6
SL		 2333 struct r5l_meta_block *mb;
2334 int i;
2335 int offset;
2336 sector_t write_pos;
2337
2338 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2339 r5l_recovery_create_empty_meta_block(log, page,
2340 ctx->pos, ctx->seq);
2341 mb = page_address(page);
2342 offset = le32_to_cpu(mb->meta_size);
fc833c2a 2343 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
b4c625c6
SL		 2344
2345 for (i = sh->disks; i--; ) {
2346 struct r5dev *dev = &sh->dev[i];
2347 struct r5l_payload_data_parity *payload;
2348 void *addr;
2349
2350 if (test_bit(R5_InJournal, &dev->flags)) {
2351 payload = (void *)mb + offset;
2352 payload->header.type = cpu_to_le16(
2353 R5LOG_PAYLOAD_DATA);
1ad45a9b 2354 payload->size = cpu_to_le32(BLOCK_SECTORS);
b4c625c6
SL		 2355 payload->location = cpu_to_le64(
2356 raid5_compute_blocknr(sh, i, 0));
2357 addr = kmap_atomic(dev->page);
2358 payload->checksum[0] = cpu_to_le32(
2359 crc32c_le(log->uuid_checksum, addr,
2360 PAGE_SIZE));
2361 kunmap_atomic(addr);
2362 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2363 dev->page, REQ_OP_WRITE, 0, false);
2364 write_pos = r5l_ring_add(log, write_pos,
2365 BLOCK_SECTORS);
2366 offset += sizeof(__le32) +
2367 sizeof(struct r5l_payload_data_parity);
2368
2369 }
2370 }
2371 mb->meta_size = cpu_to_le32(offset);
5c88f403
SL		 2372 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2373 mb, PAGE_SIZE));
b4c625c6 2374 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
20737738 2375 REQ_OP_WRITE, REQ_FUA, false);
b4c625c6 2376 sh->log_start = ctx->pos;
3c66abba
SL		 2377 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2378 atomic_inc(&log->stripe_in_journal_count);
b4c625c6
SL		 2379 ctx->pos = write_pos;
2380 ctx->seq += 1;
3c66abba 2381 next_checkpoint = sh->log_start;
355810d1 2382 }
3c66abba 2383 log->next_checkpoint = next_checkpoint;
355810d1
SL		 2384 __free_page(page);
2385 return 0;
2386}
2387
a85dd7b8
SL		 2388static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2389 struct r5l_recovery_ctx *ctx)
2390{
2391 struct mddev *mddev = log->rdev->mddev;
2392 struct r5conf *conf = mddev->private;
2393 struct stripe_head *sh, *next;
2394
2395 if (ctx->data_only_stripes == 0)
2396 return;
2397
2398 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2399
2400 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2401 r5c_make_stripe_write_out(sh);
2402 set_bit(STRIPE_HANDLE, &sh->state);
2403 list_del_init(&sh->lru);
2404 raid5_release_stripe(sh);
2405 }
2406
2407 md_wakeup_thread(conf->mddev->thread);
2408 /* reuse conf->wait_for_quiescent in recovery */
2409 wait_event(conf->wait_for_quiescent,
2410 atomic_read(&conf->active_stripes) == 0);
2411
2412 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2413}
2414
f6bed0ef
SL		 2415static int r5l_recovery_log(struct r5l_log *log)
2416{
5aabf7c4 2417 struct mddev *mddev = log->rdev->mddev;
effe6ee7 2418 struct r5l_recovery_ctx *ctx;
5aabf7c4 2419 int ret;
43b96748 2420 sector_t pos;
355810d1 2421
effe6ee7
SL		 2422 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2423 if (!ctx)
355810d1
SL		 2424 return -ENOMEM;
2425
effe6ee7
SL		 2426 ctx->pos = log->last_checkpoint;
2427 ctx->seq = log->last_cp_seq;
2428 INIT_LIST_HEAD(&ctx->cached_list);
2429 ctx->meta_page = alloc_page(GFP_KERNEL);
355810d1 2430
effe6ee7
SL		 2431 if (!ctx->meta_page) {
2432 ret = -ENOMEM;
2433 goto meta_page;
2434 }
b4c625c6 2435
effe6ee7
SL		 2436 if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2437 ret = -ENOMEM;
2438 goto ra_pool;
2439 }
2440
2441 ret = r5c_recovery_flush_log(log, ctx);
2442
2443 if (ret)
2444 goto error;
43b96748 2445
effe6ee7
SL		 2446 pos = ctx->pos;
2447 ctx->seq += 10000;
43b96748 2448
effe6ee7 2449 if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
5aabf7c4
SL		 2450 pr_debug("md/raid:%s: starting from clean shutdown\n",
2451 mdname(mddev));
a85dd7b8 2452 else
99f17890 2453 pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
effe6ee7
SL		 2454 mdname(mddev), ctx->data_only_stripes,
2455 ctx->data_parity_stripes);
2456
2457 if (ctx->data_only_stripes == 0) {
2458 log->next_checkpoint = ctx->pos;
2459 r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2460 ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2461 } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
a85dd7b8
SL		 2462 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2463 mdname(mddev));
effe6ee7
SL		 2464 ret = -EIO;
2465 goto error;
b4c625c6
SL		 2466 }
2467
effe6ee7
SL		 2468 log->log_start = ctx->pos;
2469 log->seq = ctx->seq;
43b96748
J		 2470 log->last_checkpoint = pos;
2471 r5l_write_super(log, pos);
a85dd7b8 2472
effe6ee7
SL		 2473 r5c_recovery_flush_data_only_stripes(log, ctx);
2474 ret = 0;
2475error:
2476 r5l_recovery_free_ra_pool(log, ctx);
2477ra_pool:
2478 __free_page(ctx->meta_page);
2479meta_page:
2480 kfree(ctx);
2481 return ret;
f6bed0ef
SL		 2482}
2483
2484static void r5l_write_super(struct r5l_log *log, sector_t cp)
2485{
2486 struct mddev *mddev = log->rdev->mddev;
2487
2488 log->rdev->journal_tail = cp;
2953079c 2489 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
f6bed0ef
SL		 2490}
2491
2c7da14b
SL		 2492static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2493{
2494 struct r5conf *conf = mddev->private;
2495 int ret;
2496
2497 if (!conf->log)
2498 return 0;
2499
2500 switch (conf->log->r5c_journal_mode) {
2501 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2502 ret = snprintf(
2503 page, PAGE_SIZE, "[%s] %s\n",
2504 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2505 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2506 break;
2507 case R5C_JOURNAL_MODE_WRITE_BACK:
2508 ret = snprintf(
2509 page, PAGE_SIZE, "%s [%s]\n",
2510 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2511 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2512 break;
2513 default:
2514 ret = 0;
2515 }
2516 return ret;
2517}
2518
78e470c2
HM		 2519/*
2520 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2521 *
2522 * @mode as defined in 'enum r5c_journal_mode'.
2523 *
2524 */
2525int r5c_journal_mode_set(struct mddev *mddev, int mode)
2c7da14b
SL		 2526{
2527 struct r5conf *conf = mddev->private;
2528 struct r5l_log *log = conf->log;
2c7da14b
SL		 2529
2530 if (!log)
2531 return -ENODEV;
2532
78e470c2
HM		 2533 if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2534 mode > R5C_JOURNAL_MODE_WRITE_BACK)
2c7da14b
SL		 2535 return -EINVAL;
2536
2e38a37f 2537 if (raid5_calc_degraded(conf) > 0 &&
78e470c2 2538 mode == R5C_JOURNAL_MODE_WRITE_BACK)
2e38a37f
SL		 2539 return -EINVAL;
2540
2c7da14b 2541 mddev_suspend(mddev);
78e470c2 2542 conf->log->r5c_journal_mode = mode;
2c7da14b
SL		 2543 mddev_resume(mddev);
2544
2545 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
78e470c2
HM		 2546 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2547 return 0;
2548}
2549EXPORT_SYMBOL(r5c_journal_mode_set);
2550
2551static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2552 const char *page, size_t length)
2553{
2554 int mode = ARRAY_SIZE(r5c_journal_mode_str);
2555 size_t len = length;
2556
2557 if (len < 2)
2558 return -EINVAL;
2559
2560 if (page[len - 1] == '\n')
2561 len--;
2562
2563 while (mode--)
2564 if (strlen(r5c_journal_mode_str[mode]) == len &&
2565 !strncmp(page, r5c_journal_mode_str[mode], len))
2566 break;
2567
2568 return r5c_journal_mode_set(mddev, mode) ?: length;
2c7da14b
SL		 2569}
2570
2571struct md_sysfs_entry
2572r5c_journal_mode = __ATTR(journal_mode, 0644,
2573 r5c_journal_mode_show, r5c_journal_mode_store);
2574
2ded3703
SL		 2575/*
2576 * Try handle write operation in caching phase. This function should only
2577 * be called in write-back mode.
2578 *
2579 * If all outstanding writes can be handled in caching phase, returns 0
2580 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2581 * and returns -EAGAIN
2582 */
2583int r5c_try_caching_write(struct r5conf *conf,
2584 struct stripe_head *sh,
2585 struct stripe_head_state *s,
2586 int disks)
2587{
2588 struct r5l_log *log = conf->log;
1e6d690b
SL		 2589 int i;
2590 struct r5dev *dev;
2591 int to_cache = 0;
03b047f4
SL		 2592 void **pslot;
2593 sector_t tree_index;
2594 int ret;
2595 uintptr_t refcount;
2ded3703
SL		 2596
2597 BUG_ON(!r5c_is_writeback(log));
2598
1e6d690b
SL		 2599 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2600 /*
2601 * There are two different scenarios here:
2602 * 1. The stripe has some data cached, and it is sent to
2603 * write-out phase for reclaim
2604 * 2. The stripe is clean, and this is the first write
2605 *
2606 * For 1, return -EAGAIN, so we continue with
2607 * handle_stripe_dirtying().
2608 *
2609 * For 2, set STRIPE_R5C_CACHING and continue with caching
2610 * write.
2611 */
2612
2613 /* case 1: anything injournal or anything in written */
2614 if (s->injournal > 0 || s->written > 0)
2615 return -EAGAIN;
2616 /* case 2 */
2617 set_bit(STRIPE_R5C_CACHING, &sh->state);
2618 }
2619
2e38a37f
SL		 2620 /*
2621 * When run in degraded mode, array is set to write-through mode.
2622 * This check helps drain pending write safely in the transition to
2623 * write-through mode.
2624 */
2625 if (s->failed) {
2626 r5c_make_stripe_write_out(sh);
2627 return -EAGAIN;
2628 }
2629
1e6d690b
SL		 2630 for (i = disks; i--; ) {
2631 dev = &sh->dev[i];
2632 /* if non-overwrite, use writing-out phase */
2633 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2634 !test_bit(R5_InJournal, &dev->flags)) {
2635 r5c_make_stripe_write_out(sh);
2636 return -EAGAIN;
2637 }
2638 }
2639
03b047f4
SL		 2640 /* if the stripe is not counted in big_stripe_tree, add it now */
2641 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2642 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2643 tree_index = r5c_tree_index(conf, sh->sector);
2644 spin_lock(&log->tree_lock);
2645 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2646 tree_index);
2647 if (pslot) {
2648 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2649 pslot, &log->tree_lock) >>
2650 R5C_RADIX_COUNT_SHIFT;
2651 radix_tree_replace_slot(
2652 &log->big_stripe_tree, pslot,
2653 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2654 } else {
2655 /*
2656 * this radix_tree_insert can fail safely, so no
2657 * need to call radix_tree_preload()
2658 */
2659 ret = radix_tree_insert(
2660 &log->big_stripe_tree, tree_index,
2661 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2662 if (ret) {
2663 spin_unlock(&log->tree_lock);
2664 r5c_make_stripe_write_out(sh);
2665 return -EAGAIN;
2666 }
2667 }
2668 spin_unlock(&log->tree_lock);
2669
2670 /*
2671 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2672 * counted in the radix tree
2673 */
2674 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2675 atomic_inc(&conf->r5c_cached_partial_stripes);
2676 }
2677
1e6d690b
SL		 2678 for (i = disks; i--; ) {
2679 dev = &sh->dev[i];
2680 if (dev->towrite) {
2681 set_bit(R5_Wantwrite, &dev->flags);
2682 set_bit(R5_Wantdrain, &dev->flags);
2683 set_bit(R5_LOCKED, &dev->flags);
2684 to_cache++;
2685 }
2686 }
2687
2688 if (to_cache) {
2689 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2690 /*
2691 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2692 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2693 * r5c_handle_data_cached()
2694 */
2695 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2696 }
2697
2698 return 0;
2699}
2700
2701/*
2702 * free extra pages (orig_page) we allocated for prexor
2703 */
2704void r5c_release_extra_page(struct stripe_head *sh)
2705{
d7bd398e 2706 struct r5conf *conf = sh->raid_conf;
1e6d690b 2707 int i;
d7bd398e
SL		 2708 bool using_disk_info_extra_page;
2709
2710 using_disk_info_extra_page =
2711 sh->dev[0].orig_page == conf->disks[0].extra_page;
1e6d690b
SL		 2712
2713 for (i = sh->disks; i--; )
2714 if (sh->dev[i].page != sh->dev[i].orig_page) {
2715 struct page *p = sh->dev[i].orig_page;
2716
2717 sh->dev[i].orig_page = sh->dev[i].page;
86aa1397
SL		 2718 clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2719
d7bd398e
SL		 2720 if (!using_disk_info_extra_page)
2721 put_page(p);
1e6d690b 2722 }
d7bd398e
SL		 2723
2724 if (using_disk_info_extra_page) {
2725 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2726 md_wakeup_thread(conf->mddev->thread);
2727 }
2728}
2729
2730void r5c_use_extra_page(struct stripe_head *sh)
2731{
2732 struct r5conf *conf = sh->raid_conf;
2733 int i;
2734 struct r5dev *dev;
2735
2736 for (i = sh->disks; i--; ) {
2737 dev = &sh->dev[i];
2738 if (dev->orig_page != dev->page)
2739 put_page(dev->orig_page);
2740 dev->orig_page = conf->disks[i].extra_page;
2741 }
2ded3703
SL		 2742}
2743
2744/*
2745 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2746 * stripe is committed to RAID disks.
2747 */
2748void r5c_finish_stripe_write_out(struct r5conf *conf,
2749 struct stripe_head *sh,
2750 struct stripe_head_state *s)
2751{
03b047f4 2752 struct r5l_log *log = conf->log;
1e6d690b
SL		 2753 int i;
2754 int do_wakeup = 0;
03b047f4
SL		 2755 sector_t tree_index;
2756 void **pslot;
2757 uintptr_t refcount;
1e6d690b 2758
03b047f4 2759 if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2ded3703
SL		 2760 return;
2761
2762 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2763 clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2764
03b047f4 2765 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2ded3703 2766 return;
1e6d690b
SL		 2767
2768 for (i = sh->disks; i--; ) {
2769 clear_bit(R5_InJournal, &sh->dev[i].flags);
2770 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2771 do_wakeup = 1;
2772 }
2773
2774 /*
2775 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2776 * We updated R5_InJournal, so we also update s->injournal.
2777 */
2778 s->injournal = 0;
2779
2780 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2781 if (atomic_dec_and_test(&conf->pending_full_writes))
2782 md_wakeup_thread(conf->mddev->thread);
2783
2784 if (do_wakeup)
2785 wake_up(&conf->wait_for_overlap);
a39f7afd 2786
03b047f4 2787 spin_lock_irq(&log->stripe_in_journal_lock);
a39f7afd 2788 list_del_init(&sh->r5c);
03b047f4 2789 spin_unlock_irq(&log->stripe_in_journal_lock);
a39f7afd 2790 sh->log_start = MaxSector;
03b047f4
SL		 2791
2792 atomic_dec(&log->stripe_in_journal_count);
2793 r5c_update_log_state(log);
2794
2795 /* stop counting this stripe in big_stripe_tree */
2796 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2797 test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2798 tree_index = r5c_tree_index(conf, sh->sector);
2799 spin_lock(&log->tree_lock);
2800 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2801 tree_index);
2802 BUG_ON(pslot == NULL);
2803 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2804 pslot, &log->tree_lock) >>
2805 R5C_RADIX_COUNT_SHIFT;
2806 if (refcount == 1)
2807 radix_tree_delete(&log->big_stripe_tree, tree_index);
2808 else
2809 radix_tree_replace_slot(
2810 &log->big_stripe_tree, pslot,
2811 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2812 spin_unlock(&log->tree_lock);
2813 }
2814
2815 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2816 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
e33fbb9c 2817 atomic_dec(&conf->r5c_flushing_partial_stripes);
03b047f4
SL		 2818 atomic_dec(&conf->r5c_cached_partial_stripes);
2819 }
2820
2821 if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2822 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
e33fbb9c 2823 atomic_dec(&conf->r5c_flushing_full_stripes);
03b047f4
SL		 2824 atomic_dec(&conf->r5c_cached_full_stripes);
2825 }
ea17481f
SL		 2826
2827 r5l_append_flush_payload(log, sh->sector);
1e6d690b
SL		 2828}
2829
ff875738 2830int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
1e6d690b 2831{
a39f7afd 2832 struct r5conf *conf = sh->raid_conf;
1e6d690b
SL		 2833 int pages = 0;
2834 int reserve;
2835 int i;
2836 int ret = 0;
2837
2838 BUG_ON(!log);
2839
2840 for (i = 0; i < sh->disks; i++) {
2841 void *addr;
2842
2843 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2844 continue;
2845 addr = kmap_atomic(sh->dev[i].page);
2846 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2847 addr, PAGE_SIZE);
2848 kunmap_atomic(addr);
2849 pages++;
2850 }
2851 WARN_ON(pages == 0);
2852
2853 /*
2854 * The stripe must enter state machine again to call endio, so
2855 * don't delay.
2856 */
2857 clear_bit(STRIPE_DELAYED, &sh->state);
2858 atomic_inc(&sh->count);
2859
2860 mutex_lock(&log->io_mutex);
2861 /* meta + data */
2862 reserve = (1 + pages) << (PAGE_SHIFT - 9);
1e6d690b 2863
a39f7afd
SL		 2864 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2865 sh->log_start == MaxSector)
2866 r5l_add_no_space_stripe(log, sh);
2867 else if (!r5l_has_free_space(log, reserve)) {
2868 if (sh->log_start == log->last_checkpoint)
2869 BUG();
2870 else
2871 r5l_add_no_space_stripe(log, sh);
1e6d690b
SL		 2872 } else {
2873 ret = r5l_log_stripe(log, sh, pages, 0);
2874 if (ret) {
2875 spin_lock_irq(&log->io_list_lock);
2876 list_add_tail(&sh->log_list, &log->no_mem_stripes);
2877 spin_unlock_irq(&log->io_list_lock);
2878 }
2879 }
2880
2881 mutex_unlock(&log->io_mutex);
2882 return 0;
f6bed0ef
SL		 2883}
2884
03b047f4
SL		 2885/* check whether this big stripe is in write back cache. */
2886bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2887{
2888 struct r5l_log *log = conf->log;
2889 sector_t tree_index;
2890 void *slot;
2891
2892 if (!log)
2893 return false;
2894
2895 WARN_ON_ONCE(!rcu_read_lock_held());
2896 tree_index = r5c_tree_index(conf, sect);
2897 slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2898 return slot != NULL;
2899}
2900
f6bed0ef
SL		 2901static int r5l_load_log(struct r5l_log *log)
2902{
2903 struct md_rdev *rdev = log->rdev;
2904 struct page *page;
2905 struct r5l_meta_block *mb;
2906 sector_t cp = log->rdev->journal_tail;
2907 u32 stored_crc, expected_crc;
2908 bool create_super = false;
d30dfeb9 2909 int ret = 0;
f6bed0ef
SL		 2910
2911 /* Make sure it's valid */
2912 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2913 cp = 0;
2914 page = alloc_page(GFP_KERNEL);
2915 if (!page)
2916 return -ENOMEM;
2917
796a5cf0 2918 if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
f6bed0ef
SL		 2919 ret = -EIO;
2920 goto ioerr;
2921 }
2922 mb = page_address(page);
2923
2924 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2925 mb->version != R5LOG_VERSION) {
2926 create_super = true;
2927 goto create;
2928 }
2929 stored_crc = le32_to_cpu(mb->checksum);
2930 mb->checksum = 0;
5cb2fbd6 2931 expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
f6bed0ef
SL		 2932 if (stored_crc != expected_crc) {
2933 create_super = true;
2934 goto create;
2935 }
2936 if (le64_to_cpu(mb->position) != cp) {
2937 create_super = true;
2938 goto create;
2939 }
2940create:
2941 if (create_super) {
2942 log->last_cp_seq = prandom_u32();
2943 cp = 0;
56056c2e 2944 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
f6bed0ef
SL		 2945 /*
2946 * Make sure super points to correct address. Log might have
2947 * data very soon. If super hasn't correct log tail address,
2948 * recovery can't find the log
2949 */
2950 r5l_write_super(log, cp);
2951 } else
2952 log->last_cp_seq = le64_to_cpu(mb->seq);
2953
2954 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
0576b1c6
SL		 2955 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
2956 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
2957 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
f6bed0ef
SL		 2958 log->last_checkpoint = cp;
2959
2960 __free_page(page);
2961
d30dfeb9
J		 2962 if (create_super) {
2963 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
2964 log->seq = log->last_cp_seq + 1;
2965 log->next_checkpoint = cp;
2966 } else
2967 ret = r5l_recovery_log(log);
2968
3d7e7e1d
ZL		 2969 r5c_update_log_state(log);
2970 return ret;
f6bed0ef
SL		 2971ioerr:
2972 __free_page(page);
2973 return ret;
2974}
2975
2e38a37f
SL		 2976void r5c_update_on_rdev_error(struct mddev *mddev)
2977{
2978 struct r5conf *conf = mddev->private;
2979 struct r5l_log *log = conf->log;
2980
2981 if (!log)
2982 return;
2983
2984 if (raid5_calc_degraded(conf) > 0 &&
2985 conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
2986 schedule_work(&log->disable_writeback_work);
2987}
2988
f6bed0ef
SL		 2989int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
2990{
c888a8f9 2991 struct request_queue *q = bdev_get_queue(rdev->bdev);
f6bed0ef 2992 struct r5l_log *log;
ff875738
AP		 2993 char b[BDEVNAME_SIZE];
2994
2995 pr_debug("md/raid:%s: using device %s as journal\n",
2996 mdname(conf->mddev), bdevname(rdev->bdev, b));
f6bed0ef
SL		 2997
2998 if (PAGE_SIZE != 4096)
2999 return -EINVAL;
c757ec95
SL		 3000
3001 /*
3002 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3003 * raid_disks r5l_payload_data_parity.
3004 *
3005 * Write journal and cache does not work for very big array
3006 * (raid_disks > 203)
3007 */
3008 if (sizeof(struct r5l_meta_block) +
3009 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3010 conf->raid_disks) > PAGE_SIZE) {
3011 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3012 mdname(conf->mddev), conf->raid_disks);
3013 return -EINVAL;
3014 }
3015
f6bed0ef
SL		 3016 log = kzalloc(sizeof(*log), GFP_KERNEL);
3017 if (!log)
3018 return -ENOMEM;
3019 log->rdev = rdev;
3020
c888a8f9 3021 log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
56fef7c6 3022
5cb2fbd6
SL		 3023 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3024 sizeof(rdev->mddev->uuid));
f6bed0ef
SL		 3025
3026 mutex_init(&log->io_mutex);
3027
3028 spin_lock_init(&log->io_list_lock);
3029 INIT_LIST_HEAD(&log->running_ios);
0576b1c6 3030 INIT_LIST_HEAD(&log->io_end_ios);
a8c34f91 3031 INIT_LIST_HEAD(&log->flushing_ios);
04732f74 3032 INIT_LIST_HEAD(&log->finished_ios);
3a83f467 3033 bio_init(&log->flush_bio, NULL, 0);
f6bed0ef
SL		 3034
3035 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3036 if (!log->io_kc)
3037 goto io_kc;
3038
5036c390
CH		 3039 log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
3040 if (!log->io_pool)
3041 goto io_pool;
3042
c38d29b3
CH		 3043 log->bs = bioset_create(R5L_POOL_SIZE, 0);
3044 if (!log->bs)
3045 goto io_bs;
3046
e8deb638
CH		 3047 log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
3048 if (!log->meta_pool)
3049 goto out_mempool;
3050
03b047f4
SL		 3051 spin_lock_init(&log->tree_lock);
3052 INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3053
0576b1c6
SL		 3054 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3055 log->rdev->mddev, "reclaim");
3056 if (!log->reclaim_thread)
3057 goto reclaim_thread;
a39f7afd
SL		 3058 log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3059
0fd22b45 3060 init_waitqueue_head(&log->iounit_wait);
0576b1c6 3061
5036c390
CH		 3062 INIT_LIST_HEAD(&log->no_mem_stripes);
3063
f6bed0ef
SL		 3064 INIT_LIST_HEAD(&log->no_space_stripes);
3065 spin_lock_init(&log->no_space_stripes_lock);
3066
3bddb7f8 3067 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
2e38a37f 3068 INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3bddb7f8 3069
2ded3703 3070 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
a39f7afd
SL		 3071 INIT_LIST_HEAD(&log->stripe_in_journal_list);
3072 spin_lock_init(&log->stripe_in_journal_lock);
3073 atomic_set(&log->stripe_in_journal_count, 0);
2ded3703 3074
d2250f10
SL		 3075 rcu_assign_pointer(conf->log, log);
3076
f6bed0ef
SL		 3077 if (r5l_load_log(log))
3078 goto error;
3079
a62ab49e 3080 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
f6bed0ef 3081 return 0;
e8deb638 3082
f6bed0ef 3083error:
d2250f10 3084 rcu_assign_pointer(conf->log, NULL);
0576b1c6
SL		 3085 md_unregister_thread(&log->reclaim_thread);
3086reclaim_thread:
e8deb638
CH		 3087 mempool_destroy(log->meta_pool);
3088out_mempool:
c38d29b3
CH		 3089 bioset_free(log->bs);
3090io_bs:
5036c390
CH		 3091 mempool_destroy(log->io_pool);
3092io_pool:
f6bed0ef
SL		 3093 kmem_cache_destroy(log->io_kc);
3094io_kc:
3095 kfree(log);
3096 return -EINVAL;
3097}
3098
ff875738 3099void r5l_exit_log(struct r5conf *conf)
f6bed0ef 3100{
ff875738
AP		 3101 struct r5l_log *log = conf->log;
3102
3103 conf->log = NULL;
3104 synchronize_rcu();
3105
2e38a37f 3106 flush_work(&log->disable_writeback_work);
0576b1c6 3107 md_unregister_thread(&log->reclaim_thread);
e8deb638 3108 mempool_destroy(log->meta_pool);
c38d29b3 3109 bioset_free(log->bs);
5036c390 3110 mempool_destroy(log->io_pool);
f6bed0ef
SL		 3111 kmem_cache_destroy(log->io_kc);
3112 kfree(log);
3113}
Linux 6.x block layer and io_uring tree(s)