Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
4 * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
5 */
6#include <linux/kernel.h>
7#include <linux/wait.h>
8#include <linux/blkdev.h>
9#include <linux/slab.h>
10#include <linux/raid/md_p.h>
11#include <linux/crc32c.h>
12#include <linux/random.h>
13#include <linux/kthread.h>
14#include <linux/types.h>
15#include "md.h"
16#include "raid5.h"
17#include "md-bitmap.h"
18#include "raid5-log.h"
19
20/*
21 * metadata/data stored in disk with 4k size unit (a block) regardless
22 * underneath hardware sector size. only works with PAGE_SIZE == 4096
23 */
24#define BLOCK_SECTORS (8)
25#define BLOCK_SECTOR_SHIFT (3)
26
27/*
28 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
29 *
30 * In write through mode, the reclaim runs every log->max_free_space.
31 * This can prevent the recovery scans for too long
32 */
33#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
34#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
35
36/* wake up reclaim thread periodically */
37#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
38/* start flush with these full stripes */
39#define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
40/* reclaim stripes in groups */
41#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
42
43/*
44 * We only need 2 bios per I/O unit to make progress, but ensure we
45 * have a few more available to not get too tight.
46 */
47#define R5L_POOL_SIZE 4
48
49static char *r5c_journal_mode_str[] = {"write-through",
50 "write-back"};
51/*
52 * raid5 cache state machine
53 *
54 * With the RAID cache, each stripe works in two phases:
55 * - caching phase
56 * - writing-out phase
57 *
58 * These two phases are controlled by bit STRIPE_R5C_CACHING:
59 * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
60 * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
61 *
62 * When there is no journal, or the journal is in write-through mode,
63 * the stripe is always in writing-out phase.
64 *
65 * For write-back journal, the stripe is sent to caching phase on write
66 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
67 * the write-out phase by clearing STRIPE_R5C_CACHING.
68 *
69 * Stripes in caching phase do not write the raid disks. Instead, all
70 * writes are committed from the log device. Therefore, a stripe in
71 * caching phase handles writes as:
72 * - write to log device
73 * - return IO
74 *
75 * Stripes in writing-out phase handle writes as:
76 * - calculate parity
77 * - write pending data and parity to journal
78 * - write data and parity to raid disks
79 * - return IO for pending writes
80 */
81
82struct r5l_log {
83 struct md_rdev *rdev;
84
85 u32 uuid_checksum;
86
87 sector_t device_size; /* log device size, round to
88 * BLOCK_SECTORS */
89 sector_t max_free_space; /* reclaim run if free space is at
90 * this size */
91
92 sector_t last_checkpoint; /* log tail. where recovery scan
93 * starts from */
94 u64 last_cp_seq; /* log tail sequence */
95
96 sector_t log_start; /* log head. where new data appends */
97 u64 seq; /* log head sequence */
98
99 sector_t next_checkpoint;
100
101 struct mutex io_mutex;
102 struct r5l_io_unit *current_io; /* current io_unit accepting new data */
103
104 spinlock_t io_list_lock;
105 struct list_head running_ios; /* io_units which are still running,
106 * and have not yet been completely
107 * written to the log */
108 struct list_head io_end_ios; /* io_units which have been completely
109 * written to the log but not yet written
110 * to the RAID */
111 struct list_head flushing_ios; /* io_units which are waiting for log
112 * cache flush */
113 struct list_head finished_ios; /* io_units which settle down in log disk */
114 struct bio flush_bio;
115
116 struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */
117
118 struct kmem_cache *io_kc;
119 mempool_t io_pool;
120 struct bio_set bs;
121 mempool_t meta_pool;
122
123 struct md_thread __rcu *reclaim_thread;
124 unsigned long reclaim_target; /* number of space that need to be
125 * reclaimed. if it's 0, reclaim spaces
126 * used by io_units which are in
127 * IO_UNIT_STRIPE_END state (eg, reclaim
128 * doesn't wait for specific io_unit
129 * switching to IO_UNIT_STRIPE_END
130 * state) */
131 wait_queue_head_t iounit_wait;
132
133 struct list_head no_space_stripes; /* pending stripes, log has no space */
134 spinlock_t no_space_stripes_lock;
135
136 bool need_cache_flush;
137
138 /* for r5c_cache */
139 enum r5c_journal_mode r5c_journal_mode;
140
141 /* all stripes in r5cache, in the order of seq at sh->log_start */
142 struct list_head stripe_in_journal_list;
143
144 spinlock_t stripe_in_journal_lock;
145 atomic_t stripe_in_journal_count;
146
147 /* to submit async io_units, to fulfill ordering of flush */
148 struct work_struct deferred_io_work;
149 /* to disable write back during in degraded mode */
150 struct work_struct disable_writeback_work;
151
152 /* to for chunk_aligned_read in writeback mode, details below */
153 spinlock_t tree_lock;
154 struct radix_tree_root big_stripe_tree;
155};
156
157/*
158 * Enable chunk_aligned_read() with write back cache.
159 *
160 * Each chunk may contain more than one stripe (for example, a 256kB
161 * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
162 * chunk_aligned_read, these stripes are grouped into one "big_stripe".
163 * For each big_stripe, we count how many stripes of this big_stripe
164 * are in the write back cache. These data are tracked in a radix tree
165 * (big_stripe_tree). We use radix_tree item pointer as the counter.
166 * r5c_tree_index() is used to calculate keys for the radix tree.
167 *
168 * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
169 * big_stripe of each chunk in the tree. If this big_stripe is in the
170 * tree, chunk_aligned_read() aborts. This look up is protected by
171 * rcu_read_lock().
172 *
173 * It is necessary to remember whether a stripe is counted in
174 * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
175 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
176 * two flags are set, the stripe is counted in big_stripe_tree. This
177 * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
178 * r5c_try_caching_write(); and moving clear_bit of
179 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
180 * r5c_finish_stripe_write_out().
181 */
182
183/*
184 * radix tree requests lowest 2 bits of data pointer to be 2b'00.
185 * So it is necessary to left shift the counter by 2 bits before using it
186 * as data pointer of the tree.
187 */
188#define R5C_RADIX_COUNT_SHIFT 2
189
190/*
191 * calculate key for big_stripe_tree
192 *
193 * sect: align_bi->bi_iter.bi_sector or sh->sector
194 */
195static inline sector_t r5c_tree_index(struct r5conf *conf,
196 sector_t sect)
197{
198 sector_div(sect, conf->chunk_sectors);
199 return sect;
200}
201
202/*
203 * an IO range starts from a meta data block and end at the next meta data
204 * block. The io unit's the meta data block tracks data/parity followed it. io
205 * unit is written to log disk with normal write, as we always flush log disk
206 * first and then start move data to raid disks, there is no requirement to
207 * write io unit with FLUSH/FUA
208 */
209struct r5l_io_unit {
210 struct r5l_log *log;
211
212 struct page *meta_page; /* store meta block */
213 int meta_offset; /* current offset in meta_page */
214
215 struct bio *current_bio;/* current_bio accepting new data */
216
217 atomic_t pending_stripe;/* how many stripes not flushed to raid */
218 u64 seq; /* seq number of the metablock */
219 sector_t log_start; /* where the io_unit starts */
220 sector_t log_end; /* where the io_unit ends */
221 struct list_head log_sibling; /* log->running_ios */
222 struct list_head stripe_list; /* stripes added to the io_unit */
223
224 int state;
225 bool need_split_bio;
226 struct bio *split_bio;
227
228 unsigned int has_flush:1; /* include flush request */
229 unsigned int has_fua:1; /* include fua request */
230 unsigned int has_null_flush:1; /* include null flush request */
231 unsigned int has_flush_payload:1; /* include flush payload */
232 /*
233 * io isn't sent yet, flush/fua request can only be submitted till it's
234 * the first IO in running_ios list
235 */
236 unsigned int io_deferred:1;
237
238 struct bio_list flush_barriers; /* size == 0 flush bios */
239};
240
241/* r5l_io_unit state */
242enum r5l_io_unit_state {
243 IO_UNIT_RUNNING = 0, /* accepting new IO */
244 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
245 * don't accepting new bio */
246 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
247 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
248};
249
250bool r5c_is_writeback(struct r5l_log *log)
251{
252 return (log != NULL &&
253 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
254}
255
256static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
257{
258 start += inc;
259 if (start >= log->device_size)
260 start = start - log->device_size;
261 return start;
262}
263
264static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
265 sector_t end)
266{
267 if (end >= start)
268 return end - start;
269 else
270 return end + log->device_size - start;
271}
272
273static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
274{
275 sector_t used_size;
276
277 used_size = r5l_ring_distance(log, log->last_checkpoint,
278 log->log_start);
279
280 return log->device_size > used_size + size;
281}
282
283static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
284 enum r5l_io_unit_state state)
285{
286 if (WARN_ON(io->state >= state))
287 return;
288 io->state = state;
289}
290
291static void
292r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
293{
294 struct bio *wbi, *wbi2;
295
296 wbi = dev->written;
297 dev->written = NULL;
298 while (wbi && wbi->bi_iter.bi_sector <
299 dev->sector + RAID5_STRIPE_SECTORS(conf)) {
300 wbi2 = r5_next_bio(conf, wbi, dev->sector);
301 md_write_end(conf->mddev);
302 bio_endio(wbi);
303 wbi = wbi2;
304 }
305}
306
307void r5c_handle_cached_data_endio(struct r5conf *conf,
308 struct stripe_head *sh, int disks)
309{
310 int i;
311
312 for (i = sh->disks; i--; ) {
313 if (sh->dev[i].written) {
314 set_bit(R5_UPTODATE, &sh->dev[i].flags);
315 r5c_return_dev_pending_writes(conf, &sh->dev[i]);
316 }
317 }
318}
319
320void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
321
322/* Check whether we should flush some stripes to free up stripe cache */
323void r5c_check_stripe_cache_usage(struct r5conf *conf)
324{
325 int total_cached;
326 struct r5l_log *log = READ_ONCE(conf->log);
327
328 if (!r5c_is_writeback(log))
329 return;
330
331 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
332 atomic_read(&conf->r5c_cached_full_stripes);
333
334 /*
335 * The following condition is true for either of the following:
336 * - stripe cache pressure high:
337 * total_cached > 3/4 min_nr_stripes ||
338 * empty_inactive_list_nr > 0
339 * - stripe cache pressure moderate:
340 * total_cached > 1/2 min_nr_stripes
341 */
342 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
343 atomic_read(&conf->empty_inactive_list_nr) > 0)
344 r5l_wake_reclaim(log, 0);
345}
346
347/*
348 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
349 * stripes in the cache
350 */
351void r5c_check_cached_full_stripe(struct r5conf *conf)
352{
353 struct r5l_log *log = READ_ONCE(conf->log);
354
355 if (!r5c_is_writeback(log))
356 return;
357
358 /*
359 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
360 * or a full stripe (chunk size / 4k stripes).
361 */
362 if (atomic_read(&conf->r5c_cached_full_stripes) >=
363 min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
364 conf->chunk_sectors >> RAID5_STRIPE_SHIFT(conf)))
365 r5l_wake_reclaim(log, 0);
366}
367
368/*
369 * Total log space (in sectors) needed to flush all data in cache
370 *
371 * To avoid deadlock due to log space, it is necessary to reserve log
372 * space to flush critical stripes (stripes that occupying log space near
373 * last_checkpoint). This function helps check how much log space is
374 * required to flush all cached stripes.
375 *
376 * To reduce log space requirements, two mechanisms are used to give cache
377 * flush higher priorities:
378 * 1. In handle_stripe_dirtying() and schedule_reconstruction(),
379 * stripes ALREADY in journal can be flushed w/o pending writes;
380 * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
381 * can be delayed (r5l_add_no_space_stripe).
382 *
383 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
384 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
385 * pages of journal space. For stripes that has not passed 1, flushing it
386 * requires (conf->raid_disks + 1) pages of journal space. There are at
387 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
388 * required to flush all cached stripes (in pages) is:
389 *
390 * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
391 * (group_cnt + 1) * (raid_disks + 1)
392 * or
393 * (stripe_in_journal_count) * (max_degraded + 1) +
394 * (group_cnt + 1) * (raid_disks - max_degraded)
395 */
396static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
397{
398 struct r5l_log *log = READ_ONCE(conf->log);
399
400 if (!r5c_is_writeback(log))
401 return 0;
402
403 return BLOCK_SECTORS *
404 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
405 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
406}
407
408/*
409 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
410 *
411 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
412 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
413 * device is less than 2x of reclaim_required_space.
414 */
415static inline void r5c_update_log_state(struct r5l_log *log)
416{
417 struct r5conf *conf = log->rdev->mddev->private;
418 sector_t free_space;
419 sector_t reclaim_space;
420 bool wake_reclaim = false;
421
422 if (!r5c_is_writeback(log))
423 return;
424
425 free_space = r5l_ring_distance(log, log->log_start,
426 log->last_checkpoint);
427 reclaim_space = r5c_log_required_to_flush_cache(conf);
428 if (free_space < 2 * reclaim_space)
429 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
430 else {
431 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
432 wake_reclaim = true;
433 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
434 }
435 if (free_space < 3 * reclaim_space)
436 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
437 else
438 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
439
440 if (wake_reclaim)
441 r5l_wake_reclaim(log, 0);
442}
443
444/*
445 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
446 * This function should only be called in write-back mode.
447 */
448void r5c_make_stripe_write_out(struct stripe_head *sh)
449{
450 struct r5conf *conf = sh->raid_conf;
451 struct r5l_log *log = READ_ONCE(conf->log);
452
453 BUG_ON(!r5c_is_writeback(log));
454
455 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
456 clear_bit(STRIPE_R5C_CACHING, &sh->state);
457
458 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
459 atomic_inc(&conf->preread_active_stripes);
460}
461
462static void r5c_handle_data_cached(struct stripe_head *sh)
463{
464 int i;
465
466 for (i = sh->disks; i--; )
467 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
468 set_bit(R5_InJournal, &sh->dev[i].flags);
469 clear_bit(R5_LOCKED, &sh->dev[i].flags);
470 }
471 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
472}
473
474/*
475 * this journal write must contain full parity,
476 * it may also contain some data pages
477 */
478static void r5c_handle_parity_cached(struct stripe_head *sh)
479{
480 int i;
481
482 for (i = sh->disks; i--; )
483 if (test_bit(R5_InJournal, &sh->dev[i].flags))
484 set_bit(R5_Wantwrite, &sh->dev[i].flags);
485}
486
487/*
488 * Setting proper flags after writing (or flushing) data and/or parity to the
489 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
490 */
491static void r5c_finish_cache_stripe(struct stripe_head *sh)
492{
493 struct r5l_log *log = READ_ONCE(sh->raid_conf->log);
494
495 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
496 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
497 /*
498 * Set R5_InJournal for parity dev[pd_idx]. This means
499 * all data AND parity in the journal. For RAID 6, it is
500 * NOT necessary to set the flag for dev[qd_idx], as the
501 * two parities are written out together.
502 */
503 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
504 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
505 r5c_handle_data_cached(sh);
506 } else {
507 r5c_handle_parity_cached(sh);
508 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
509 }
510}
511
512static void r5l_io_run_stripes(struct r5l_io_unit *io)
513{
514 struct stripe_head *sh, *next;
515
516 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
517 list_del_init(&sh->log_list);
518
519 r5c_finish_cache_stripe(sh);
520
521 set_bit(STRIPE_HANDLE, &sh->state);
522 raid5_release_stripe(sh);
523 }
524}
525
526static void r5l_log_run_stripes(struct r5l_log *log)
527{
528 struct r5l_io_unit *io, *next;
529
530 lockdep_assert_held(&log->io_list_lock);
531
532 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
533 /* don't change list order */
534 if (io->state < IO_UNIT_IO_END)
535 break;
536
537 list_move_tail(&io->log_sibling, &log->finished_ios);
538 r5l_io_run_stripes(io);
539 }
540}
541
542static void r5l_move_to_end_ios(struct r5l_log *log)
543{
544 struct r5l_io_unit *io, *next;
545
546 lockdep_assert_held(&log->io_list_lock);
547
548 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
549 /* don't change list order */
550 if (io->state < IO_UNIT_IO_END)
551 break;
552 list_move_tail(&io->log_sibling, &log->io_end_ios);
553 }
554}
555
556static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
557static void r5l_log_endio(struct bio *bio)
558{
559 struct r5l_io_unit *io = bio->bi_private;
560 struct r5l_io_unit *io_deferred;
561 struct r5l_log *log = io->log;
562 unsigned long flags;
563 bool has_null_flush;
564 bool has_flush_payload;
565
566 if (bio->bi_status)
567 md_error(log->rdev->mddev, log->rdev);
568
569 bio_put(bio);
570 mempool_free(io->meta_page, &log->meta_pool);
571
572 spin_lock_irqsave(&log->io_list_lock, flags);
573 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
574
575 /*
576 * if the io doesn't not have null_flush or flush payload,
577 * it is not safe to access it after releasing io_list_lock.
578 * Therefore, it is necessary to check the condition with
579 * the lock held.
580 */
581 has_null_flush = io->has_null_flush;
582 has_flush_payload = io->has_flush_payload;
583
584 if (log->need_cache_flush && !list_empty(&io->stripe_list))
585 r5l_move_to_end_ios(log);
586 else
587 r5l_log_run_stripes(log);
588 if (!list_empty(&log->running_ios)) {
589 /*
590 * FLUSH/FUA io_unit is deferred because of ordering, now we
591 * can dispatch it
592 */
593 io_deferred = list_first_entry(&log->running_ios,
594 struct r5l_io_unit, log_sibling);
595 if (io_deferred->io_deferred)
596 schedule_work(&log->deferred_io_work);
597 }
598
599 spin_unlock_irqrestore(&log->io_list_lock, flags);
600
601 if (log->need_cache_flush)
602 md_wakeup_thread(log->rdev->mddev->thread);
603
604 /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
605 if (has_null_flush) {
606 struct bio *bi;
607
608 WARN_ON(bio_list_empty(&io->flush_barriers));
609 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
610 bio_endio(bi);
611 if (atomic_dec_and_test(&io->pending_stripe)) {
612 __r5l_stripe_write_finished(io);
613 return;
614 }
615 }
616 }
617 /* decrease pending_stripe for flush payload */
618 if (has_flush_payload)
619 if (atomic_dec_and_test(&io->pending_stripe))
620 __r5l_stripe_write_finished(io);
621}
622
623static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
624{
625 unsigned long flags;
626
627 spin_lock_irqsave(&log->io_list_lock, flags);
628 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
629 spin_unlock_irqrestore(&log->io_list_lock, flags);
630
631 /*
632 * In case of journal device failures, submit_bio will get error
633 * and calls endio, then active stripes will continue write
634 * process. Therefore, it is not necessary to check Faulty bit
635 * of journal device here.
636 *
637 * We can't check split_bio after current_bio is submitted. If
638 * io->split_bio is null, after current_bio is submitted, current_bio
639 * might already be completed and the io_unit is freed. We submit
640 * split_bio first to avoid the issue.
641 */
642 if (io->split_bio) {
643 if (io->has_flush)
644 io->split_bio->bi_opf |= REQ_PREFLUSH;
645 if (io->has_fua)
646 io->split_bio->bi_opf |= REQ_FUA;
647 submit_bio(io->split_bio);
648 }
649
650 if (io->has_flush)
651 io->current_bio->bi_opf |= REQ_PREFLUSH;
652 if (io->has_fua)
653 io->current_bio->bi_opf |= REQ_FUA;
654 submit_bio(io->current_bio);
655}
656
657/* deferred io_unit will be dispatched here */
658static void r5l_submit_io_async(struct work_struct *work)
659{
660 struct r5l_log *log = container_of(work, struct r5l_log,
661 deferred_io_work);
662 struct r5l_io_unit *io = NULL;
663 unsigned long flags;
664
665 spin_lock_irqsave(&log->io_list_lock, flags);
666 if (!list_empty(&log->running_ios)) {
667 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
668 log_sibling);
669 if (!io->io_deferred)
670 io = NULL;
671 else
672 io->io_deferred = 0;
673 }
674 spin_unlock_irqrestore(&log->io_list_lock, flags);
675 if (io)
676 r5l_do_submit_io(log, io);
677}
678
679static void r5c_disable_writeback_async(struct work_struct *work)
680{
681 struct r5l_log *log = container_of(work, struct r5l_log,
682 disable_writeback_work);
683 struct mddev *mddev = log->rdev->mddev;
684 struct r5conf *conf = mddev->private;
685
686 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
687 return;
688 pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
689 mdname(mddev));
690
691 /* wait superblock change before suspend */
692 wait_event(mddev->sb_wait,
693 !READ_ONCE(conf->log) ||
694 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
695
696 log = READ_ONCE(conf->log);
697 if (log) {
698 mddev_suspend(mddev, false);
699 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
700 mddev_resume(mddev);
701 }
702}
703
704static void r5l_submit_current_io(struct r5l_log *log)
705{
706 struct r5l_io_unit *io = log->current_io;
707 struct r5l_meta_block *block;
708 unsigned long flags;
709 u32 crc;
710 bool do_submit = true;
711
712 if (!io)
713 return;
714
715 block = page_address(io->meta_page);
716 block->meta_size = cpu_to_le32(io->meta_offset);
717 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
718 block->checksum = cpu_to_le32(crc);
719
720 log->current_io = NULL;
721 spin_lock_irqsave(&log->io_list_lock, flags);
722 if (io->has_flush || io->has_fua) {
723 if (io != list_first_entry(&log->running_ios,
724 struct r5l_io_unit, log_sibling)) {
725 io->io_deferred = 1;
726 do_submit = false;
727 }
728 }
729 spin_unlock_irqrestore(&log->io_list_lock, flags);
730 if (do_submit)
731 r5l_do_submit_io(log, io);
732}
733
734static struct bio *r5l_bio_alloc(struct r5l_log *log)
735{
736 struct bio *bio = bio_alloc_bioset(log->rdev->bdev, BIO_MAX_VECS,
737 REQ_OP_WRITE, GFP_NOIO, &log->bs);
738
739 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
740
741 return bio;
742}
743
744static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
745{
746 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
747
748 r5c_update_log_state(log);
749 /*
750 * If we filled up the log device start from the beginning again,
751 * which will require a new bio.
752 *
753 * Note: for this to work properly the log size needs to me a multiple
754 * of BLOCK_SECTORS.
755 */
756 if (log->log_start == 0)
757 io->need_split_bio = true;
758
759 io->log_end = log->log_start;
760}
761
762static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
763{
764 struct r5l_io_unit *io;
765 struct r5l_meta_block *block;
766
767 io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
768 if (!io)
769 return NULL;
770 memset(io, 0, sizeof(*io));
771
772 io->log = log;
773 INIT_LIST_HEAD(&io->log_sibling);
774 INIT_LIST_HEAD(&io->stripe_list);
775 bio_list_init(&io->flush_barriers);
776 io->state = IO_UNIT_RUNNING;
777
778 io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
779 block = page_address(io->meta_page);
780 clear_page(block);
781 block->magic = cpu_to_le32(R5LOG_MAGIC);
782 block->version = R5LOG_VERSION;
783 block->seq = cpu_to_le64(log->seq);
784 block->position = cpu_to_le64(log->log_start);
785
786 io->log_start = log->log_start;
787 io->meta_offset = sizeof(struct r5l_meta_block);
788 io->seq = log->seq++;
789
790 io->current_bio = r5l_bio_alloc(log);
791 io->current_bio->bi_end_io = r5l_log_endio;
792 io->current_bio->bi_private = io;
793 __bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
794
795 r5_reserve_log_entry(log, io);
796
797 spin_lock_irq(&log->io_list_lock);
798 list_add_tail(&io->log_sibling, &log->running_ios);
799 spin_unlock_irq(&log->io_list_lock);
800
801 return io;
802}
803
804static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
805{
806 if (log->current_io &&
807 log->current_io->meta_offset + payload_size > PAGE_SIZE)
808 r5l_submit_current_io(log);
809
810 if (!log->current_io) {
811 log->current_io = r5l_new_meta(log);
812 if (!log->current_io)
813 return -ENOMEM;
814 }
815
816 return 0;
817}
818
819static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
820 sector_t location,
821 u32 checksum1, u32 checksum2,
822 bool checksum2_valid)
823{
824 struct r5l_io_unit *io = log->current_io;
825 struct r5l_payload_data_parity *payload;
826
827 payload = page_address(io->meta_page) + io->meta_offset;
828 payload->header.type = cpu_to_le16(type);
829 payload->header.flags = cpu_to_le16(0);
830 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
831 (PAGE_SHIFT - 9));
832 payload->location = cpu_to_le64(location);
833 payload->checksum[0] = cpu_to_le32(checksum1);
834 if (checksum2_valid)
835 payload->checksum[1] = cpu_to_le32(checksum2);
836
837 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
838 sizeof(__le32) * (1 + !!checksum2_valid);
839}
840
841static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
842{
843 struct r5l_io_unit *io = log->current_io;
844
845 if (io->need_split_bio) {
846 BUG_ON(io->split_bio);
847 io->split_bio = io->current_bio;
848 io->current_bio = r5l_bio_alloc(log);
849 bio_chain(io->current_bio, io->split_bio);
850 io->need_split_bio = false;
851 }
852
853 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
854 BUG();
855
856 r5_reserve_log_entry(log, io);
857}
858
859static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
860{
861 struct mddev *mddev = log->rdev->mddev;
862 struct r5conf *conf = mddev->private;
863 struct r5l_io_unit *io;
864 struct r5l_payload_flush *payload;
865 int meta_size;
866
867 /*
868 * payload_flush requires extra writes to the journal.
869 * To avoid handling the extra IO in quiesce, just skip
870 * flush_payload
871 */
872 if (conf->quiesce)
873 return;
874
875 mutex_lock(&log->io_mutex);
876 meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
877
878 if (r5l_get_meta(log, meta_size)) {
879 mutex_unlock(&log->io_mutex);
880 return;
881 }
882
883 /* current implementation is one stripe per flush payload */
884 io = log->current_io;
885 payload = page_address(io->meta_page) + io->meta_offset;
886 payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
887 payload->header.flags = cpu_to_le16(0);
888 payload->size = cpu_to_le32(sizeof(__le64));
889 payload->flush_stripes[0] = cpu_to_le64(sect);
890 io->meta_offset += meta_size;
891 /* multiple flush payloads count as one pending_stripe */
892 if (!io->has_flush_payload) {
893 io->has_flush_payload = 1;
894 atomic_inc(&io->pending_stripe);
895 }
896 mutex_unlock(&log->io_mutex);
897}
898
899static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
900 int data_pages, int parity_pages)
901{
902 int i;
903 int meta_size;
904 int ret;
905 struct r5l_io_unit *io;
906
907 meta_size =
908 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
909 * data_pages) +
910 sizeof(struct r5l_payload_data_parity) +
911 sizeof(__le32) * parity_pages;
912
913 ret = r5l_get_meta(log, meta_size);
914 if (ret)
915 return ret;
916
917 io = log->current_io;
918
919 if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
920 io->has_flush = 1;
921
922 for (i = 0; i < sh->disks; i++) {
923 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
924 test_bit(R5_InJournal, &sh->dev[i].flags))
925 continue;
926 if (i == sh->pd_idx || i == sh->qd_idx)
927 continue;
928 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
929 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
930 io->has_fua = 1;
931 /*
932 * we need to flush journal to make sure recovery can
933 * reach the data with fua flag
934 */
935 io->has_flush = 1;
936 }
937 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
938 raid5_compute_blocknr(sh, i, 0),
939 sh->dev[i].log_checksum, 0, false);
940 r5l_append_payload_page(log, sh->dev[i].page);
941 }
942
943 if (parity_pages == 2) {
944 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
945 sh->sector, sh->dev[sh->pd_idx].log_checksum,
946 sh->dev[sh->qd_idx].log_checksum, true);
947 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
948 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
949 } else if (parity_pages == 1) {
950 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
951 sh->sector, sh->dev[sh->pd_idx].log_checksum,
952 0, false);
953 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
954 } else /* Just writing data, not parity, in caching phase */
955 BUG_ON(parity_pages != 0);
956
957 list_add_tail(&sh->log_list, &io->stripe_list);
958 atomic_inc(&io->pending_stripe);
959 sh->log_io = io;
960
961 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
962 return 0;
963
964 if (sh->log_start == MaxSector) {
965 BUG_ON(!list_empty(&sh->r5c));
966 sh->log_start = io->log_start;
967 spin_lock_irq(&log->stripe_in_journal_lock);
968 list_add_tail(&sh->r5c,
969 &log->stripe_in_journal_list);
970 spin_unlock_irq(&log->stripe_in_journal_lock);
971 atomic_inc(&log->stripe_in_journal_count);
972 }
973 return 0;
974}
975
976/* add stripe to no_space_stripes, and then wake up reclaim */
977static inline void r5l_add_no_space_stripe(struct r5l_log *log,
978 struct stripe_head *sh)
979{
980 spin_lock(&log->no_space_stripes_lock);
981 list_add_tail(&sh->log_list, &log->no_space_stripes);
982 spin_unlock(&log->no_space_stripes_lock);
983}
984
985/*
986 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
987 * data from log to raid disks), so we shouldn't wait for reclaim here
988 */
989int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
990{
991 struct r5conf *conf = sh->raid_conf;
992 int write_disks = 0;
993 int data_pages, parity_pages;
994 int reserve;
995 int i;
996 int ret = 0;
997 bool wake_reclaim = false;
998
999 if (!log)
1000 return -EAGAIN;
1001 /* Don't support stripe batch */
1002 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1003 test_bit(STRIPE_SYNCING, &sh->state)) {
1004 /* the stripe is written to log, we start writing it to raid */
1005 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1006 return -EAGAIN;
1007 }
1008
1009 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1010
1011 for (i = 0; i < sh->disks; i++) {
1012 void *addr;
1013
1014 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1015 test_bit(R5_InJournal, &sh->dev[i].flags))
1016 continue;
1017
1018 write_disks++;
1019 /* checksum is already calculated in last run */
1020 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1021 continue;
1022 addr = kmap_atomic(sh->dev[i].page);
1023 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1024 addr, PAGE_SIZE);
1025 kunmap_atomic(addr);
1026 }
1027 parity_pages = 1 + !!(sh->qd_idx >= 0);
1028 data_pages = write_disks - parity_pages;
1029
1030 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1031 /*
1032 * The stripe must enter state machine again to finish the write, so
1033 * don't delay.
1034 */
1035 clear_bit(STRIPE_DELAYED, &sh->state);
1036 atomic_inc(&sh->count);
1037
1038 mutex_lock(&log->io_mutex);
1039 /* meta + data */
1040 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1041
1042 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1043 if (!r5l_has_free_space(log, reserve)) {
1044 r5l_add_no_space_stripe(log, sh);
1045 wake_reclaim = true;
1046 } else {
1047 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1048 if (ret) {
1049 spin_lock_irq(&log->io_list_lock);
1050 list_add_tail(&sh->log_list,
1051 &log->no_mem_stripes);
1052 spin_unlock_irq(&log->io_list_lock);
1053 }
1054 }
1055 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1056 /*
1057 * log space critical, do not process stripes that are
1058 * not in cache yet (sh->log_start == MaxSector).
1059 */
1060 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1061 sh->log_start == MaxSector) {
1062 r5l_add_no_space_stripe(log, sh);
1063 wake_reclaim = true;
1064 reserve = 0;
1065 } else if (!r5l_has_free_space(log, reserve)) {
1066 if (sh->log_start == log->last_checkpoint)
1067 BUG();
1068 else
1069 r5l_add_no_space_stripe(log, sh);
1070 } else {
1071 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1072 if (ret) {
1073 spin_lock_irq(&log->io_list_lock);
1074 list_add_tail(&sh->log_list,
1075 &log->no_mem_stripes);
1076 spin_unlock_irq(&log->io_list_lock);
1077 }
1078 }
1079 }
1080
1081 mutex_unlock(&log->io_mutex);
1082 if (wake_reclaim)
1083 r5l_wake_reclaim(log, reserve);
1084 return 0;
1085}
1086
1087void r5l_write_stripe_run(struct r5l_log *log)
1088{
1089 if (!log)
1090 return;
1091 mutex_lock(&log->io_mutex);
1092 r5l_submit_current_io(log);
1093 mutex_unlock(&log->io_mutex);
1094}
1095
1096int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1097{
1098 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1099 /*
1100 * in write through (journal only)
1101 * we flush log disk cache first, then write stripe data to
1102 * raid disks. So if bio is finished, the log disk cache is
1103 * flushed already. The recovery guarantees we can recovery
1104 * the bio from log disk, so we don't need to flush again
1105 */
1106 if (bio->bi_iter.bi_size == 0) {
1107 bio_endio(bio);
1108 return 0;
1109 }
1110 bio->bi_opf &= ~REQ_PREFLUSH;
1111 } else {
1112 /* write back (with cache) */
1113 if (bio->bi_iter.bi_size == 0) {
1114 mutex_lock(&log->io_mutex);
1115 r5l_get_meta(log, 0);
1116 bio_list_add(&log->current_io->flush_barriers, bio);
1117 log->current_io->has_flush = 1;
1118 log->current_io->has_null_flush = 1;
1119 atomic_inc(&log->current_io->pending_stripe);
1120 r5l_submit_current_io(log);
1121 mutex_unlock(&log->io_mutex);
1122 return 0;
1123 }
1124 }
1125 return -EAGAIN;
1126}
1127
1128/* This will run after log space is reclaimed */
1129static void r5l_run_no_space_stripes(struct r5l_log *log)
1130{
1131 struct stripe_head *sh;
1132
1133 spin_lock(&log->no_space_stripes_lock);
1134 while (!list_empty(&log->no_space_stripes)) {
1135 sh = list_first_entry(&log->no_space_stripes,
1136 struct stripe_head, log_list);
1137 list_del_init(&sh->log_list);
1138 set_bit(STRIPE_HANDLE, &sh->state);
1139 raid5_release_stripe(sh);
1140 }
1141 spin_unlock(&log->no_space_stripes_lock);
1142}
1143
1144/*
1145 * calculate new last_checkpoint
1146 * for write through mode, returns log->next_checkpoint
1147 * for write back, returns log_start of first sh in stripe_in_journal_list
1148 */
1149static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1150{
1151 struct stripe_head *sh;
1152 struct r5l_log *log = READ_ONCE(conf->log);
1153 sector_t new_cp;
1154 unsigned long flags;
1155
1156 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1157 return log->next_checkpoint;
1158
1159 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1160 if (list_empty(&log->stripe_in_journal_list)) {
1161 /* all stripes flushed */
1162 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1163 return log->next_checkpoint;
1164 }
1165 sh = list_first_entry(&log->stripe_in_journal_list,
1166 struct stripe_head, r5c);
1167 new_cp = sh->log_start;
1168 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1169 return new_cp;
1170}
1171
1172static sector_t r5l_reclaimable_space(struct r5l_log *log)
1173{
1174 struct r5conf *conf = log->rdev->mddev->private;
1175
1176 return r5l_ring_distance(log, log->last_checkpoint,
1177 r5c_calculate_new_cp(conf));
1178}
1179
1180static void r5l_run_no_mem_stripe(struct r5l_log *log)
1181{
1182 struct stripe_head *sh;
1183
1184 lockdep_assert_held(&log->io_list_lock);
1185
1186 if (!list_empty(&log->no_mem_stripes)) {
1187 sh = list_first_entry(&log->no_mem_stripes,
1188 struct stripe_head, log_list);
1189 list_del_init(&sh->log_list);
1190 set_bit(STRIPE_HANDLE, &sh->state);
1191 raid5_release_stripe(sh);
1192 }
1193}
1194
1195static bool r5l_complete_finished_ios(struct r5l_log *log)
1196{
1197 struct r5l_io_unit *io, *next;
1198 bool found = false;
1199
1200 lockdep_assert_held(&log->io_list_lock);
1201
1202 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1203 /* don't change list order */
1204 if (io->state < IO_UNIT_STRIPE_END)
1205 break;
1206
1207 log->next_checkpoint = io->log_start;
1208
1209 list_del(&io->log_sibling);
1210 mempool_free(io, &log->io_pool);
1211 r5l_run_no_mem_stripe(log);
1212
1213 found = true;
1214 }
1215
1216 return found;
1217}
1218
1219static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1220{
1221 struct r5l_log *log = io->log;
1222 struct r5conf *conf = log->rdev->mddev->private;
1223 unsigned long flags;
1224
1225 spin_lock_irqsave(&log->io_list_lock, flags);
1226 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1227
1228 if (!r5l_complete_finished_ios(log)) {
1229 spin_unlock_irqrestore(&log->io_list_lock, flags);
1230 return;
1231 }
1232
1233 if (r5l_reclaimable_space(log) > log->max_free_space ||
1234 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1235 r5l_wake_reclaim(log, 0);
1236
1237 spin_unlock_irqrestore(&log->io_list_lock, flags);
1238 wake_up(&log->iounit_wait);
1239}
1240
1241void r5l_stripe_write_finished(struct stripe_head *sh)
1242{
1243 struct r5l_io_unit *io;
1244
1245 io = sh->log_io;
1246 sh->log_io = NULL;
1247
1248 if (io && atomic_dec_and_test(&io->pending_stripe))
1249 __r5l_stripe_write_finished(io);
1250}
1251
1252static void r5l_log_flush_endio(struct bio *bio)
1253{
1254 struct r5l_log *log = container_of(bio, struct r5l_log,
1255 flush_bio);
1256 unsigned long flags;
1257 struct r5l_io_unit *io;
1258
1259 if (bio->bi_status)
1260 md_error(log->rdev->mddev, log->rdev);
1261 bio_uninit(bio);
1262
1263 spin_lock_irqsave(&log->io_list_lock, flags);
1264 list_for_each_entry(io, &log->flushing_ios, log_sibling)
1265 r5l_io_run_stripes(io);
1266 list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1267 spin_unlock_irqrestore(&log->io_list_lock, flags);
1268}
1269
1270/*
1271 * Starting dispatch IO to raid.
1272 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1273 * broken meta in the middle of a log causes recovery can't find meta at the
1274 * head of log. If operations require meta at the head persistent in log, we
1275 * must make sure meta before it persistent in log too. A case is:
1276 *
1277 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1278 * data/parity must be persistent in log before we do the write to raid disks.
1279 *
1280 * The solution is we restrictly maintain io_unit list order. In this case, we
1281 * only write stripes of an io_unit to raid disks till the io_unit is the first
1282 * one whose data/parity is in log.
1283 */
1284void r5l_flush_stripe_to_raid(struct r5l_log *log)
1285{
1286 bool do_flush;
1287
1288 if (!log || !log->need_cache_flush)
1289 return;
1290
1291 spin_lock_irq(&log->io_list_lock);
1292 /* flush bio is running */
1293 if (!list_empty(&log->flushing_ios)) {
1294 spin_unlock_irq(&log->io_list_lock);
1295 return;
1296 }
1297 list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1298 do_flush = !list_empty(&log->flushing_ios);
1299 spin_unlock_irq(&log->io_list_lock);
1300
1301 if (!do_flush)
1302 return;
1303 bio_init(&log->flush_bio, log->rdev->bdev, NULL, 0,
1304 REQ_OP_WRITE | REQ_PREFLUSH);
1305 log->flush_bio.bi_end_io = r5l_log_flush_endio;
1306 submit_bio(&log->flush_bio);
1307}
1308
1309static void r5l_write_super(struct r5l_log *log, sector_t cp);
1310static void r5l_write_super_and_discard_space(struct r5l_log *log,
1311 sector_t end)
1312{
1313 struct block_device *bdev = log->rdev->bdev;
1314 struct mddev *mddev;
1315
1316 r5l_write_super(log, end);
1317
1318 if (!bdev_max_discard_sectors(bdev))
1319 return;
1320
1321 mddev = log->rdev->mddev;
1322 /*
1323 * Discard could zero data, so before discard we must make sure
1324 * superblock is updated to new log tail. Updating superblock (either
1325 * directly call md_update_sb() or depend on md thread) must hold
1326 * reconfig mutex. On the other hand, raid5_quiesce is called with
1327 * reconfig_mutex hold. The first step of raid5_quiesce() is waiting
1328 * for all IO finish, hence waiting for reclaim thread, while reclaim
1329 * thread is calling this function and waiting for reconfig mutex. So
1330 * there is a deadlock. We workaround this issue with a trylock.
1331 * FIXME: we could miss discard if we can't take reconfig mutex
1332 */
1333 set_mask_bits(&mddev->sb_flags, 0,
1334 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1335 if (!mddev_trylock(mddev))
1336 return;
1337 md_update_sb(mddev, 1);
1338 mddev_unlock(mddev);
1339
1340 /* discard IO error really doesn't matter, ignore it */
1341 if (log->last_checkpoint < end) {
1342 blkdev_issue_discard(bdev,
1343 log->last_checkpoint + log->rdev->data_offset,
1344 end - log->last_checkpoint, GFP_NOIO);
1345 } else {
1346 blkdev_issue_discard(bdev,
1347 log->last_checkpoint + log->rdev->data_offset,
1348 log->device_size - log->last_checkpoint,
1349 GFP_NOIO);
1350 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1351 GFP_NOIO);
1352 }
1353}
1354
1355/*
1356 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1357 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1358 *
1359 * must hold conf->device_lock
1360 */
1361static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1362{
1363 BUG_ON(list_empty(&sh->lru));
1364 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1365 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1366
1367 /*
1368 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1369 * raid5_release_stripe() while holding conf->device_lock
1370 */
1371 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1372 lockdep_assert_held(&conf->device_lock);
1373
1374 list_del_init(&sh->lru);
1375 atomic_inc(&sh->count);
1376
1377 set_bit(STRIPE_HANDLE, &sh->state);
1378 atomic_inc(&conf->active_stripes);
1379 r5c_make_stripe_write_out(sh);
1380
1381 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1382 atomic_inc(&conf->r5c_flushing_partial_stripes);
1383 else
1384 atomic_inc(&conf->r5c_flushing_full_stripes);
1385 raid5_release_stripe(sh);
1386}
1387
1388/*
1389 * if num == 0, flush all full stripes
1390 * if num > 0, flush all full stripes. If less than num full stripes are
1391 * flushed, flush some partial stripes until totally num stripes are
1392 * flushed or there is no more cached stripes.
1393 */
1394void r5c_flush_cache(struct r5conf *conf, int num)
1395{
1396 int count;
1397 struct stripe_head *sh, *next;
1398
1399 lockdep_assert_held(&conf->device_lock);
1400 if (!READ_ONCE(conf->log))
1401 return;
1402
1403 count = 0;
1404 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1405 r5c_flush_stripe(conf, sh);
1406 count++;
1407 }
1408
1409 if (count >= num)
1410 return;
1411 list_for_each_entry_safe(sh, next,
1412 &conf->r5c_partial_stripe_list, lru) {
1413 r5c_flush_stripe(conf, sh);
1414 if (++count >= num)
1415 break;
1416 }
1417}
1418
1419static void r5c_do_reclaim(struct r5conf *conf)
1420{
1421 struct r5l_log *log = READ_ONCE(conf->log);
1422 struct stripe_head *sh;
1423 int count = 0;
1424 unsigned long flags;
1425 int total_cached;
1426 int stripes_to_flush;
1427 int flushing_partial, flushing_full;
1428
1429 if (!r5c_is_writeback(log))
1430 return;
1431
1432 flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1433 flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1434 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1435 atomic_read(&conf->r5c_cached_full_stripes) -
1436 flushing_full - flushing_partial;
1437
1438 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1439 atomic_read(&conf->empty_inactive_list_nr) > 0)
1440 /*
1441 * if stripe cache pressure high, flush all full stripes and
1442 * some partial stripes
1443 */
1444 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1445 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1446 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1447 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1448 /*
1449 * if stripe cache pressure moderate, or if there is many full
1450 * stripes,flush all full stripes
1451 */
1452 stripes_to_flush = 0;
1453 else
1454 /* no need to flush */
1455 stripes_to_flush = -1;
1456
1457 if (stripes_to_flush >= 0) {
1458 spin_lock_irqsave(&conf->device_lock, flags);
1459 r5c_flush_cache(conf, stripes_to_flush);
1460 spin_unlock_irqrestore(&conf->device_lock, flags);
1461 }
1462
1463 /* if log space is tight, flush stripes on stripe_in_journal_list */
1464 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1465 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1466 spin_lock(&conf->device_lock);
1467 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1468 /*
1469 * stripes on stripe_in_journal_list could be in any
1470 * state of the stripe_cache state machine. In this
1471 * case, we only want to flush stripe on
1472 * r5c_cached_full/partial_stripes. The following
1473 * condition makes sure the stripe is on one of the
1474 * two lists.
1475 */
1476 if (!list_empty(&sh->lru) &&
1477 !test_bit(STRIPE_HANDLE, &sh->state) &&
1478 atomic_read(&sh->count) == 0) {
1479 r5c_flush_stripe(conf, sh);
1480 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1481 break;
1482 }
1483 }
1484 spin_unlock(&conf->device_lock);
1485 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1486 }
1487
1488 if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1489 r5l_run_no_space_stripes(log);
1490
1491 md_wakeup_thread(conf->mddev->thread);
1492}
1493
1494static void r5l_do_reclaim(struct r5l_log *log)
1495{
1496 struct r5conf *conf = log->rdev->mddev->private;
1497 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1498 sector_t reclaimable;
1499 sector_t next_checkpoint;
1500 bool write_super;
1501
1502 spin_lock_irq(&log->io_list_lock);
1503 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1504 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1505 /*
1506 * move proper io_unit to reclaim list. We should not change the order.
1507 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1508 * shouldn't reuse space of an unreclaimable io_unit
1509 */
1510 while (1) {
1511 reclaimable = r5l_reclaimable_space(log);
1512 if (reclaimable >= reclaim_target ||
1513 (list_empty(&log->running_ios) &&
1514 list_empty(&log->io_end_ios) &&
1515 list_empty(&log->flushing_ios) &&
1516 list_empty(&log->finished_ios)))
1517 break;
1518
1519 md_wakeup_thread(log->rdev->mddev->thread);
1520 wait_event_lock_irq(log->iounit_wait,
1521 r5l_reclaimable_space(log) > reclaimable,
1522 log->io_list_lock);
1523 }
1524
1525 next_checkpoint = r5c_calculate_new_cp(conf);
1526 spin_unlock_irq(&log->io_list_lock);
1527
1528 if (reclaimable == 0 || !write_super)
1529 return;
1530
1531 /*
1532 * write_super will flush cache of each raid disk. We must write super
1533 * here, because the log area might be reused soon and we don't want to
1534 * confuse recovery
1535 */
1536 r5l_write_super_and_discard_space(log, next_checkpoint);
1537
1538 mutex_lock(&log->io_mutex);
1539 log->last_checkpoint = next_checkpoint;
1540 r5c_update_log_state(log);
1541 mutex_unlock(&log->io_mutex);
1542
1543 r5l_run_no_space_stripes(log);
1544}
1545
1546static void r5l_reclaim_thread(struct md_thread *thread)
1547{
1548 struct mddev *mddev = thread->mddev;
1549 struct r5conf *conf = mddev->private;
1550 struct r5l_log *log = READ_ONCE(conf->log);
1551
1552 if (!log)
1553 return;
1554 r5c_do_reclaim(conf);
1555 r5l_do_reclaim(log);
1556}
1557
1558void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1559{
1560 unsigned long target;
1561 unsigned long new = (unsigned long)space; /* overflow in theory */
1562
1563 if (!log)
1564 return;
1565
1566 target = READ_ONCE(log->reclaim_target);
1567 do {
1568 if (new < target)
1569 return;
1570 } while (!try_cmpxchg(&log->reclaim_target, &target, new));
1571 md_wakeup_thread(log->reclaim_thread);
1572}
1573
1574void r5l_quiesce(struct r5l_log *log, int quiesce)
1575{
1576 struct mddev *mddev = log->rdev->mddev;
1577 struct md_thread *thread = rcu_dereference_protected(
1578 log->reclaim_thread, lockdep_is_held(&mddev->reconfig_mutex));
1579
1580 if (quiesce) {
1581 /* make sure r5l_write_super_and_discard_space exits */
1582 wake_up(&mddev->sb_wait);
1583 kthread_park(thread->tsk);
1584 r5l_wake_reclaim(log, MaxSector);
1585 r5l_do_reclaim(log);
1586 } else
1587 kthread_unpark(thread->tsk);
1588}
1589
1590bool r5l_log_disk_error(struct r5conf *conf)
1591{
1592 struct r5l_log *log = READ_ONCE(conf->log);
1593
1594 /* don't allow write if journal disk is missing */
1595 if (!log)
1596 return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1597 else
1598 return test_bit(Faulty, &log->rdev->flags);
1599}
1600
1601#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1602
1603struct r5l_recovery_ctx {
1604 struct page *meta_page; /* current meta */
1605 sector_t meta_total_blocks; /* total size of current meta and data */
1606 sector_t pos; /* recovery position */
1607 u64 seq; /* recovery position seq */
1608 int data_parity_stripes; /* number of data_parity stripes */
1609 int data_only_stripes; /* number of data_only stripes */
1610 struct list_head cached_list;
1611
1612 /*
1613 * read ahead page pool (ra_pool)
1614 * in recovery, log is read sequentially. It is not efficient to
1615 * read every page with sync_page_io(). The read ahead page pool
1616 * reads multiple pages with one IO, so further log read can
1617 * just copy data from the pool.
1618 */
1619 struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1620 struct bio_vec ra_bvec[R5L_RECOVERY_PAGE_POOL_SIZE];
1621 sector_t pool_offset; /* offset of first page in the pool */
1622 int total_pages; /* total allocated pages */
1623 int valid_pages; /* pages with valid data */
1624};
1625
1626static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1627 struct r5l_recovery_ctx *ctx)
1628{
1629 struct page *page;
1630
1631 ctx->valid_pages = 0;
1632 ctx->total_pages = 0;
1633 while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1634 page = alloc_page(GFP_KERNEL);
1635
1636 if (!page)
1637 break;
1638 ctx->ra_pool[ctx->total_pages] = page;
1639 ctx->total_pages += 1;
1640 }
1641
1642 if (ctx->total_pages == 0)
1643 return -ENOMEM;
1644
1645 ctx->pool_offset = 0;
1646 return 0;
1647}
1648
1649static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1650 struct r5l_recovery_ctx *ctx)
1651{
1652 int i;
1653
1654 for (i = 0; i < ctx->total_pages; ++i)
1655 put_page(ctx->ra_pool[i]);
1656}
1657
1658/*
1659 * fetch ctx->valid_pages pages from offset
1660 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1661 * However, if the offset is close to the end of the journal device,
1662 * ctx->valid_pages could be smaller than ctx->total_pages
1663 */
1664static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1665 struct r5l_recovery_ctx *ctx,
1666 sector_t offset)
1667{
1668 struct bio bio;
1669 int ret;
1670
1671 bio_init(&bio, log->rdev->bdev, ctx->ra_bvec,
1672 R5L_RECOVERY_PAGE_POOL_SIZE, REQ_OP_READ);
1673 bio.bi_iter.bi_sector = log->rdev->data_offset + offset;
1674
1675 ctx->valid_pages = 0;
1676 ctx->pool_offset = offset;
1677
1678 while (ctx->valid_pages < ctx->total_pages) {
1679 __bio_add_page(&bio, ctx->ra_pool[ctx->valid_pages], PAGE_SIZE,
1680 0);
1681 ctx->valid_pages += 1;
1682
1683 offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1684
1685 if (offset == 0) /* reached end of the device */
1686 break;
1687 }
1688
1689 ret = submit_bio_wait(&bio);
1690 bio_uninit(&bio);
1691 return ret;
1692}
1693
1694/*
1695 * try read a page from the read ahead page pool, if the page is not in the
1696 * pool, call r5l_recovery_fetch_ra_pool
1697 */
1698static int r5l_recovery_read_page(struct r5l_log *log,
1699 struct r5l_recovery_ctx *ctx,
1700 struct page *page,
1701 sector_t offset)
1702{
1703 int ret;
1704
1705 if (offset < ctx->pool_offset ||
1706 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1707 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1708 if (ret)
1709 return ret;
1710 }
1711
1712 BUG_ON(offset < ctx->pool_offset ||
1713 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1714
1715 memcpy(page_address(page),
1716 page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1717 BLOCK_SECTOR_SHIFT]),
1718 PAGE_SIZE);
1719 return 0;
1720}
1721
1722static int r5l_recovery_read_meta_block(struct r5l_log *log,
1723 struct r5l_recovery_ctx *ctx)
1724{
1725 struct page *page = ctx->meta_page;
1726 struct r5l_meta_block *mb;
1727 u32 crc, stored_crc;
1728 int ret;
1729
1730 ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1731 if (ret != 0)
1732 return ret;
1733
1734 mb = page_address(page);
1735 stored_crc = le32_to_cpu(mb->checksum);
1736 mb->checksum = 0;
1737
1738 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1739 le64_to_cpu(mb->seq) != ctx->seq ||
1740 mb->version != R5LOG_VERSION ||
1741 le64_to_cpu(mb->position) != ctx->pos)
1742 return -EINVAL;
1743
1744 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1745 if (stored_crc != crc)
1746 return -EINVAL;
1747
1748 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1749 return -EINVAL;
1750
1751 ctx->meta_total_blocks = BLOCK_SECTORS;
1752
1753 return 0;
1754}
1755
1756static void
1757r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1758 struct page *page,
1759 sector_t pos, u64 seq)
1760{
1761 struct r5l_meta_block *mb;
1762
1763 mb = page_address(page);
1764 clear_page(mb);
1765 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1766 mb->version = R5LOG_VERSION;
1767 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1768 mb->seq = cpu_to_le64(seq);
1769 mb->position = cpu_to_le64(pos);
1770}
1771
1772static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1773 u64 seq)
1774{
1775 struct page *page;
1776 struct r5l_meta_block *mb;
1777
1778 page = alloc_page(GFP_KERNEL);
1779 if (!page)
1780 return -ENOMEM;
1781 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1782 mb = page_address(page);
1783 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1784 mb, PAGE_SIZE));
1785 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE |
1786 REQ_SYNC | REQ_FUA, false)) {
1787 __free_page(page);
1788 return -EIO;
1789 }
1790 __free_page(page);
1791 return 0;
1792}
1793
1794/*
1795 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1796 * to mark valid (potentially not flushed) data in the journal.
1797 *
1798 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1799 * so there should not be any mismatch here.
1800 */
1801static void r5l_recovery_load_data(struct r5l_log *log,
1802 struct stripe_head *sh,
1803 struct r5l_recovery_ctx *ctx,
1804 struct r5l_payload_data_parity *payload,
1805 sector_t log_offset)
1806{
1807 struct mddev *mddev = log->rdev->mddev;
1808 struct r5conf *conf = mddev->private;
1809 int dd_idx;
1810
1811 raid5_compute_sector(conf,
1812 le64_to_cpu(payload->location), 0,
1813 &dd_idx, sh);
1814 r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1815 sh->dev[dd_idx].log_checksum =
1816 le32_to_cpu(payload->checksum[0]);
1817 ctx->meta_total_blocks += BLOCK_SECTORS;
1818
1819 set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1820 set_bit(STRIPE_R5C_CACHING, &sh->state);
1821}
1822
1823static void r5l_recovery_load_parity(struct r5l_log *log,
1824 struct stripe_head *sh,
1825 struct r5l_recovery_ctx *ctx,
1826 struct r5l_payload_data_parity *payload,
1827 sector_t log_offset)
1828{
1829 struct mddev *mddev = log->rdev->mddev;
1830 struct r5conf *conf = mddev->private;
1831
1832 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1833 r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1834 sh->dev[sh->pd_idx].log_checksum =
1835 le32_to_cpu(payload->checksum[0]);
1836 set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1837
1838 if (sh->qd_idx >= 0) {
1839 r5l_recovery_read_page(
1840 log, ctx, sh->dev[sh->qd_idx].page,
1841 r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1842 sh->dev[sh->qd_idx].log_checksum =
1843 le32_to_cpu(payload->checksum[1]);
1844 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1845 }
1846 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1847}
1848
1849static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1850{
1851 int i;
1852
1853 sh->state = 0;
1854 sh->log_start = MaxSector;
1855 for (i = sh->disks; i--; )
1856 sh->dev[i].flags = 0;
1857}
1858
1859static void
1860r5l_recovery_replay_one_stripe(struct r5conf *conf,
1861 struct stripe_head *sh,
1862 struct r5l_recovery_ctx *ctx)
1863{
1864 struct md_rdev *rdev, *rrdev;
1865 int disk_index;
1866 int data_count = 0;
1867
1868 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1869 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1870 continue;
1871 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1872 continue;
1873 data_count++;
1874 }
1875
1876 /*
1877 * stripes that only have parity must have been flushed
1878 * before the crash that we are now recovering from, so
1879 * there is nothing more to recovery.
1880 */
1881 if (data_count == 0)
1882 goto out;
1883
1884 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1885 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1886 continue;
1887
1888 /* in case device is broken */
1889 rdev = conf->disks[disk_index].rdev;
1890 if (rdev) {
1891 atomic_inc(&rdev->nr_pending);
1892 sync_page_io(rdev, sh->sector, PAGE_SIZE,
1893 sh->dev[disk_index].page, REQ_OP_WRITE,
1894 false);
1895 rdev_dec_pending(rdev, rdev->mddev);
1896 }
1897 rrdev = conf->disks[disk_index].replacement;
1898 if (rrdev) {
1899 atomic_inc(&rrdev->nr_pending);
1900 sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1901 sh->dev[disk_index].page, REQ_OP_WRITE,
1902 false);
1903 rdev_dec_pending(rrdev, rrdev->mddev);
1904 }
1905 }
1906 ctx->data_parity_stripes++;
1907out:
1908 r5l_recovery_reset_stripe(sh);
1909}
1910
1911static struct stripe_head *
1912r5c_recovery_alloc_stripe(
1913 struct r5conf *conf,
1914 sector_t stripe_sect,
1915 int noblock)
1916{
1917 struct stripe_head *sh;
1918
1919 sh = raid5_get_active_stripe(conf, NULL, stripe_sect,
1920 noblock ? R5_GAS_NOBLOCK : 0);
1921 if (!sh)
1922 return NULL; /* no more stripe available */
1923
1924 r5l_recovery_reset_stripe(sh);
1925
1926 return sh;
1927}
1928
1929static struct stripe_head *
1930r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1931{
1932 struct stripe_head *sh;
1933
1934 list_for_each_entry(sh, list, lru)
1935 if (sh->sector == sect)
1936 return sh;
1937 return NULL;
1938}
1939
1940static void
1941r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1942 struct r5l_recovery_ctx *ctx)
1943{
1944 struct stripe_head *sh, *next;
1945
1946 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1947 r5l_recovery_reset_stripe(sh);
1948 list_del_init(&sh->lru);
1949 raid5_release_stripe(sh);
1950 }
1951}
1952
1953static void
1954r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1955 struct r5l_recovery_ctx *ctx)
1956{
1957 struct stripe_head *sh, *next;
1958
1959 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1960 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1961 r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1962 list_del_init(&sh->lru);
1963 raid5_release_stripe(sh);
1964 }
1965}
1966
1967/* if matches return 0; otherwise return -EINVAL */
1968static int
1969r5l_recovery_verify_data_checksum(struct r5l_log *log,
1970 struct r5l_recovery_ctx *ctx,
1971 struct page *page,
1972 sector_t log_offset, __le32 log_checksum)
1973{
1974 void *addr;
1975 u32 checksum;
1976
1977 r5l_recovery_read_page(log, ctx, page, log_offset);
1978 addr = kmap_atomic(page);
1979 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1980 kunmap_atomic(addr);
1981 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1982}
1983
1984/*
1985 * before loading data to stripe cache, we need verify checksum for all data,
1986 * if there is mismatch for any data page, we drop all data in the mata block
1987 */
1988static int
1989r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1990 struct r5l_recovery_ctx *ctx)
1991{
1992 struct mddev *mddev = log->rdev->mddev;
1993 struct r5conf *conf = mddev->private;
1994 struct r5l_meta_block *mb = page_address(ctx->meta_page);
1995 sector_t mb_offset = sizeof(struct r5l_meta_block);
1996 sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1997 struct page *page;
1998 struct r5l_payload_data_parity *payload;
1999 struct r5l_payload_flush *payload_flush;
2000
2001 page = alloc_page(GFP_KERNEL);
2002 if (!page)
2003 return -ENOMEM;
2004
2005 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2006 payload = (void *)mb + mb_offset;
2007 payload_flush = (void *)mb + mb_offset;
2008
2009 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2010 if (r5l_recovery_verify_data_checksum(
2011 log, ctx, page, log_offset,
2012 payload->checksum[0]) < 0)
2013 goto mismatch;
2014 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2015 if (r5l_recovery_verify_data_checksum(
2016 log, ctx, page, log_offset,
2017 payload->checksum[0]) < 0)
2018 goto mismatch;
2019 if (conf->max_degraded == 2 && /* q for RAID 6 */
2020 r5l_recovery_verify_data_checksum(
2021 log, ctx, page,
2022 r5l_ring_add(log, log_offset,
2023 BLOCK_SECTORS),
2024 payload->checksum[1]) < 0)
2025 goto mismatch;
2026 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2027 /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2028 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2029 goto mismatch;
2030
2031 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2032 mb_offset += sizeof(struct r5l_payload_flush) +
2033 le32_to_cpu(payload_flush->size);
2034 } else {
2035 /* DATA or PARITY payload */
2036 log_offset = r5l_ring_add(log, log_offset,
2037 le32_to_cpu(payload->size));
2038 mb_offset += sizeof(struct r5l_payload_data_parity) +
2039 sizeof(__le32) *
2040 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2041 }
2042
2043 }
2044
2045 put_page(page);
2046 return 0;
2047
2048mismatch:
2049 put_page(page);
2050 return -EINVAL;
2051}
2052
2053/*
2054 * Analyze all data/parity pages in one meta block
2055 * Returns:
2056 * 0 for success
2057 * -EINVAL for unknown playload type
2058 * -EAGAIN for checksum mismatch of data page
2059 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2060 */
2061static int
2062r5c_recovery_analyze_meta_block(struct r5l_log *log,
2063 struct r5l_recovery_ctx *ctx,
2064 struct list_head *cached_stripe_list)
2065{
2066 struct mddev *mddev = log->rdev->mddev;
2067 struct r5conf *conf = mddev->private;
2068 struct r5l_meta_block *mb;
2069 struct r5l_payload_data_parity *payload;
2070 struct r5l_payload_flush *payload_flush;
2071 int mb_offset;
2072 sector_t log_offset;
2073 sector_t stripe_sect;
2074 struct stripe_head *sh;
2075 int ret;
2076
2077 /*
2078 * for mismatch in data blocks, we will drop all data in this mb, but
2079 * we will still read next mb for other data with FLUSH flag, as
2080 * io_unit could finish out of order.
2081 */
2082 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2083 if (ret == -EINVAL)
2084 return -EAGAIN;
2085 else if (ret)
2086 return ret; /* -ENOMEM duo to alloc_page() failed */
2087
2088 mb = page_address(ctx->meta_page);
2089 mb_offset = sizeof(struct r5l_meta_block);
2090 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2091
2092 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2093 int dd;
2094
2095 payload = (void *)mb + mb_offset;
2096 payload_flush = (void *)mb + mb_offset;
2097
2098 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2099 int i, count;
2100
2101 count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2102 for (i = 0; i < count; ++i) {
2103 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2104 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2105 stripe_sect);
2106 if (sh) {
2107 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2108 r5l_recovery_reset_stripe(sh);
2109 list_del_init(&sh->lru);
2110 raid5_release_stripe(sh);
2111 }
2112 }
2113
2114 mb_offset += sizeof(struct r5l_payload_flush) +
2115 le32_to_cpu(payload_flush->size);
2116 continue;
2117 }
2118
2119 /* DATA or PARITY payload */
2120 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2121 raid5_compute_sector(
2122 conf, le64_to_cpu(payload->location), 0, &dd,
2123 NULL)
2124 : le64_to_cpu(payload->location);
2125
2126 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2127 stripe_sect);
2128
2129 if (!sh) {
2130 sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2131 /*
2132 * cannot get stripe from raid5_get_active_stripe
2133 * try replay some stripes
2134 */
2135 if (!sh) {
2136 r5c_recovery_replay_stripes(
2137 cached_stripe_list, ctx);
2138 sh = r5c_recovery_alloc_stripe(
2139 conf, stripe_sect, 1);
2140 }
2141 if (!sh) {
2142 int new_size = conf->min_nr_stripes * 2;
2143 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2144 mdname(mddev),
2145 new_size);
2146 ret = raid5_set_cache_size(mddev, new_size);
2147 if (conf->min_nr_stripes <= new_size / 2) {
2148 pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2149 mdname(mddev),
2150 ret,
2151 new_size,
2152 conf->min_nr_stripes,
2153 conf->max_nr_stripes);
2154 return -ENOMEM;
2155 }
2156 sh = r5c_recovery_alloc_stripe(
2157 conf, stripe_sect, 0);
2158 }
2159 if (!sh) {
2160 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2161 mdname(mddev));
2162 return -ENOMEM;
2163 }
2164 list_add_tail(&sh->lru, cached_stripe_list);
2165 }
2166
2167 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2168 if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2169 test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2170 r5l_recovery_replay_one_stripe(conf, sh, ctx);
2171 list_move_tail(&sh->lru, cached_stripe_list);
2172 }
2173 r5l_recovery_load_data(log, sh, ctx, payload,
2174 log_offset);
2175 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2176 r5l_recovery_load_parity(log, sh, ctx, payload,
2177 log_offset);
2178 else
2179 return -EINVAL;
2180
2181 log_offset = r5l_ring_add(log, log_offset,
2182 le32_to_cpu(payload->size));
2183
2184 mb_offset += sizeof(struct r5l_payload_data_parity) +
2185 sizeof(__le32) *
2186 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2187 }
2188
2189 return 0;
2190}
2191
2192/*
2193 * Load the stripe into cache. The stripe will be written out later by
2194 * the stripe cache state machine.
2195 */
2196static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2197 struct stripe_head *sh)
2198{
2199 struct r5dev *dev;
2200 int i;
2201
2202 for (i = sh->disks; i--; ) {
2203 dev = sh->dev + i;
2204 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2205 set_bit(R5_InJournal, &dev->flags);
2206 set_bit(R5_UPTODATE, &dev->flags);
2207 }
2208 }
2209}
2210
2211/*
2212 * Scan through the log for all to-be-flushed data
2213 *
2214 * For stripes with data and parity, namely Data-Parity stripe
2215 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2216 *
2217 * For stripes with only data, namely Data-Only stripe
2218 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2219 *
2220 * For a stripe, if we see data after parity, we should discard all previous
2221 * data and parity for this stripe, as these data are already flushed to
2222 * the array.
2223 *
2224 * At the end of the scan, we return the new journal_tail, which points to
2225 * first data-only stripe on the journal device, or next invalid meta block.
2226 */
2227static int r5c_recovery_flush_log(struct r5l_log *log,
2228 struct r5l_recovery_ctx *ctx)
2229{
2230 struct stripe_head *sh;
2231 int ret = 0;
2232
2233 /* scan through the log */
2234 while (1) {
2235 if (r5l_recovery_read_meta_block(log, ctx))
2236 break;
2237
2238 ret = r5c_recovery_analyze_meta_block(log, ctx,
2239 &ctx->cached_list);
2240 /*
2241 * -EAGAIN means mismatch in data block, in this case, we still
2242 * try scan the next metablock
2243 */
2244 if (ret && ret != -EAGAIN)
2245 break; /* ret == -EINVAL or -ENOMEM */
2246 ctx->seq++;
2247 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2248 }
2249
2250 if (ret == -ENOMEM) {
2251 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2252 return ret;
2253 }
2254
2255 /* replay data-parity stripes */
2256 r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2257
2258 /* load data-only stripes to stripe cache */
2259 list_for_each_entry(sh, &ctx->cached_list, lru) {
2260 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2261 r5c_recovery_load_one_stripe(log, sh);
2262 ctx->data_only_stripes++;
2263 }
2264
2265 return 0;
2266}
2267
2268/*
2269 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2270 * log will start here. but we can't let superblock point to last valid
2271 * meta block. The log might looks like:
2272 * | meta 1| meta 2| meta 3|
2273 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2274 * superblock points to meta 1, we write a new valid meta 2n. if crash
2275 * happens again, new recovery will start from meta 1. Since meta 2n is
2276 * valid now, recovery will think meta 3 is valid, which is wrong.
2277 * The solution is we create a new meta in meta2 with its seq == meta
2278 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2279 * will not think meta 3 is a valid meta, because its seq doesn't match
2280 */
2281
2282/*
2283 * Before recovery, the log looks like the following
2284 *
2285 * ---------------------------------------------
2286 * | valid log | invalid log |
2287 * ---------------------------------------------
2288 * ^
2289 * |- log->last_checkpoint
2290 * |- log->last_cp_seq
2291 *
2292 * Now we scan through the log until we see invalid entry
2293 *
2294 * ---------------------------------------------
2295 * | valid log | invalid log |
2296 * ---------------------------------------------
2297 * ^ ^
2298 * |- log->last_checkpoint |- ctx->pos
2299 * |- log->last_cp_seq |- ctx->seq
2300 *
2301 * From this point, we need to increase seq number by 10 to avoid
2302 * confusing next recovery.
2303 *
2304 * ---------------------------------------------
2305 * | valid log | invalid log |
2306 * ---------------------------------------------
2307 * ^ ^
2308 * |- log->last_checkpoint |- ctx->pos+1
2309 * |- log->last_cp_seq |- ctx->seq+10001
2310 *
2311 * However, it is not safe to start the state machine yet, because data only
2312 * parities are not yet secured in RAID. To save these data only parities, we
2313 * rewrite them from seq+11.
2314 *
2315 * -----------------------------------------------------------------
2316 * | valid log | data only stripes | invalid log |
2317 * -----------------------------------------------------------------
2318 * ^ ^
2319 * |- log->last_checkpoint |- ctx->pos+n
2320 * |- log->last_cp_seq |- ctx->seq+10000+n
2321 *
2322 * If failure happens again during this process, the recovery can safe start
2323 * again from log->last_checkpoint.
2324 *
2325 * Once data only stripes are rewritten to journal, we move log_tail
2326 *
2327 * -----------------------------------------------------------------
2328 * | old log | data only stripes | invalid log |
2329 * -----------------------------------------------------------------
2330 * ^ ^
2331 * |- log->last_checkpoint |- ctx->pos+n
2332 * |- log->last_cp_seq |- ctx->seq+10000+n
2333 *
2334 * Then we can safely start the state machine. If failure happens from this
2335 * point on, the recovery will start from new log->last_checkpoint.
2336 */
2337static int
2338r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2339 struct r5l_recovery_ctx *ctx)
2340{
2341 struct stripe_head *sh;
2342 struct mddev *mddev = log->rdev->mddev;
2343 struct page *page;
2344 sector_t next_checkpoint = MaxSector;
2345
2346 page = alloc_page(GFP_KERNEL);
2347 if (!page) {
2348 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2349 mdname(mddev));
2350 return -ENOMEM;
2351 }
2352
2353 WARN_ON(list_empty(&ctx->cached_list));
2354
2355 list_for_each_entry(sh, &ctx->cached_list, lru) {
2356 struct r5l_meta_block *mb;
2357 int i;
2358 int offset;
2359 sector_t write_pos;
2360
2361 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2362 r5l_recovery_create_empty_meta_block(log, page,
2363 ctx->pos, ctx->seq);
2364 mb = page_address(page);
2365 offset = le32_to_cpu(mb->meta_size);
2366 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2367
2368 for (i = sh->disks; i--; ) {
2369 struct r5dev *dev = &sh->dev[i];
2370 struct r5l_payload_data_parity *payload;
2371 void *addr;
2372
2373 if (test_bit(R5_InJournal, &dev->flags)) {
2374 payload = (void *)mb + offset;
2375 payload->header.type = cpu_to_le16(
2376 R5LOG_PAYLOAD_DATA);
2377 payload->size = cpu_to_le32(BLOCK_SECTORS);
2378 payload->location = cpu_to_le64(
2379 raid5_compute_blocknr(sh, i, 0));
2380 addr = kmap_atomic(dev->page);
2381 payload->checksum[0] = cpu_to_le32(
2382 crc32c_le(log->uuid_checksum, addr,
2383 PAGE_SIZE));
2384 kunmap_atomic(addr);
2385 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2386 dev->page, REQ_OP_WRITE, false);
2387 write_pos = r5l_ring_add(log, write_pos,
2388 BLOCK_SECTORS);
2389 offset += sizeof(__le32) +
2390 sizeof(struct r5l_payload_data_parity);
2391
2392 }
2393 }
2394 mb->meta_size = cpu_to_le32(offset);
2395 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2396 mb, PAGE_SIZE));
2397 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2398 REQ_OP_WRITE | REQ_SYNC | REQ_FUA, false);
2399 sh->log_start = ctx->pos;
2400 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2401 atomic_inc(&log->stripe_in_journal_count);
2402 ctx->pos = write_pos;
2403 ctx->seq += 1;
2404 next_checkpoint = sh->log_start;
2405 }
2406 log->next_checkpoint = next_checkpoint;
2407 __free_page(page);
2408 return 0;
2409}
2410
2411static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2412 struct r5l_recovery_ctx *ctx)
2413{
2414 struct mddev *mddev = log->rdev->mddev;
2415 struct r5conf *conf = mddev->private;
2416 struct stripe_head *sh, *next;
2417 bool cleared_pending = false;
2418
2419 if (ctx->data_only_stripes == 0)
2420 return;
2421
2422 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2423 cleared_pending = true;
2424 clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2425 }
2426 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2427
2428 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2429 r5c_make_stripe_write_out(sh);
2430 set_bit(STRIPE_HANDLE, &sh->state);
2431 list_del_init(&sh->lru);
2432 raid5_release_stripe(sh);
2433 }
2434
2435 /* reuse conf->wait_for_quiescent in recovery */
2436 wait_event(conf->wait_for_quiescent,
2437 atomic_read(&conf->active_stripes) == 0);
2438
2439 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2440 if (cleared_pending)
2441 set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2442}
2443
2444static int r5l_recovery_log(struct r5l_log *log)
2445{
2446 struct mddev *mddev = log->rdev->mddev;
2447 struct r5l_recovery_ctx *ctx;
2448 int ret;
2449 sector_t pos;
2450
2451 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2452 if (!ctx)
2453 return -ENOMEM;
2454
2455 ctx->pos = log->last_checkpoint;
2456 ctx->seq = log->last_cp_seq;
2457 INIT_LIST_HEAD(&ctx->cached_list);
2458 ctx->meta_page = alloc_page(GFP_KERNEL);
2459
2460 if (!ctx->meta_page) {
2461 ret = -ENOMEM;
2462 goto meta_page;
2463 }
2464
2465 if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2466 ret = -ENOMEM;
2467 goto ra_pool;
2468 }
2469
2470 ret = r5c_recovery_flush_log(log, ctx);
2471
2472 if (ret)
2473 goto error;
2474
2475 pos = ctx->pos;
2476 ctx->seq += 10000;
2477
2478 if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2479 pr_info("md/raid:%s: starting from clean shutdown\n",
2480 mdname(mddev));
2481 else
2482 pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2483 mdname(mddev), ctx->data_only_stripes,
2484 ctx->data_parity_stripes);
2485
2486 if (ctx->data_only_stripes == 0) {
2487 log->next_checkpoint = ctx->pos;
2488 r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2489 ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2490 } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2491 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2492 mdname(mddev));
2493 ret = -EIO;
2494 goto error;
2495 }
2496
2497 log->log_start = ctx->pos;
2498 log->seq = ctx->seq;
2499 log->last_checkpoint = pos;
2500 r5l_write_super(log, pos);
2501
2502 r5c_recovery_flush_data_only_stripes(log, ctx);
2503 ret = 0;
2504error:
2505 r5l_recovery_free_ra_pool(log, ctx);
2506ra_pool:
2507 __free_page(ctx->meta_page);
2508meta_page:
2509 kfree(ctx);
2510 return ret;
2511}
2512
2513static void r5l_write_super(struct r5l_log *log, sector_t cp)
2514{
2515 struct mddev *mddev = log->rdev->mddev;
2516
2517 log->rdev->journal_tail = cp;
2518 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2519}
2520
2521static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2522{
2523 struct r5conf *conf;
2524 int ret;
2525
2526 ret = mddev_lock(mddev);
2527 if (ret)
2528 return ret;
2529
2530 conf = mddev->private;
2531 if (!conf || !conf->log)
2532 goto out_unlock;
2533
2534 switch (conf->log->r5c_journal_mode) {
2535 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2536 ret = snprintf(
2537 page, PAGE_SIZE, "[%s] %s\n",
2538 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2539 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2540 break;
2541 case R5C_JOURNAL_MODE_WRITE_BACK:
2542 ret = snprintf(
2543 page, PAGE_SIZE, "%s [%s]\n",
2544 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2545 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2546 break;
2547 default:
2548 ret = 0;
2549 }
2550
2551out_unlock:
2552 mddev_unlock(mddev);
2553 return ret;
2554}
2555
2556/*
2557 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2558 *
2559 * @mode as defined in 'enum r5c_journal_mode'.
2560 *
2561 */
2562int r5c_journal_mode_set(struct mddev *mddev, int mode)
2563{
2564 struct r5conf *conf;
2565
2566 if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2567 mode > R5C_JOURNAL_MODE_WRITE_BACK)
2568 return -EINVAL;
2569
2570 conf = mddev->private;
2571 if (!conf || !conf->log)
2572 return -ENODEV;
2573
2574 if (raid5_calc_degraded(conf) > 0 &&
2575 mode == R5C_JOURNAL_MODE_WRITE_BACK)
2576 return -EINVAL;
2577
2578 conf->log->r5c_journal_mode = mode;
2579
2580 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2581 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2582 return 0;
2583}
2584EXPORT_SYMBOL(r5c_journal_mode_set);
2585
2586static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2587 const char *page, size_t length)
2588{
2589 int mode = ARRAY_SIZE(r5c_journal_mode_str);
2590 size_t len = length;
2591 int ret;
2592
2593 if (len < 2)
2594 return -EINVAL;
2595
2596 if (page[len - 1] == '\n')
2597 len--;
2598
2599 while (mode--)
2600 if (strlen(r5c_journal_mode_str[mode]) == len &&
2601 !strncmp(page, r5c_journal_mode_str[mode], len))
2602 break;
2603 ret = mddev_suspend_and_lock(mddev);
2604 if (ret)
2605 return ret;
2606 ret = r5c_journal_mode_set(mddev, mode);
2607 mddev_unlock_and_resume(mddev);
2608 return ret ?: length;
2609}
2610
2611struct md_sysfs_entry
2612r5c_journal_mode = __ATTR(journal_mode, 0644,
2613 r5c_journal_mode_show, r5c_journal_mode_store);
2614
2615/*
2616 * Try handle write operation in caching phase. This function should only
2617 * be called in write-back mode.
2618 *
2619 * If all outstanding writes can be handled in caching phase, returns 0
2620 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2621 * and returns -EAGAIN
2622 */
2623int r5c_try_caching_write(struct r5conf *conf,
2624 struct stripe_head *sh,
2625 struct stripe_head_state *s,
2626 int disks)
2627{
2628 struct r5l_log *log = READ_ONCE(conf->log);
2629 int i;
2630 struct r5dev *dev;
2631 int to_cache = 0;
2632 void __rcu **pslot;
2633 sector_t tree_index;
2634 int ret;
2635 uintptr_t refcount;
2636
2637 BUG_ON(!r5c_is_writeback(log));
2638
2639 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2640 /*
2641 * There are two different scenarios here:
2642 * 1. The stripe has some data cached, and it is sent to
2643 * write-out phase for reclaim
2644 * 2. The stripe is clean, and this is the first write
2645 *
2646 * For 1, return -EAGAIN, so we continue with
2647 * handle_stripe_dirtying().
2648 *
2649 * For 2, set STRIPE_R5C_CACHING and continue with caching
2650 * write.
2651 */
2652
2653 /* case 1: anything injournal or anything in written */
2654 if (s->injournal > 0 || s->written > 0)
2655 return -EAGAIN;
2656 /* case 2 */
2657 set_bit(STRIPE_R5C_CACHING, &sh->state);
2658 }
2659
2660 /*
2661 * When run in degraded mode, array is set to write-through mode.
2662 * This check helps drain pending write safely in the transition to
2663 * write-through mode.
2664 *
2665 * When a stripe is syncing, the write is also handled in write
2666 * through mode.
2667 */
2668 if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2669 r5c_make_stripe_write_out(sh);
2670 return -EAGAIN;
2671 }
2672
2673 for (i = disks; i--; ) {
2674 dev = &sh->dev[i];
2675 /* if non-overwrite, use writing-out phase */
2676 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2677 !test_bit(R5_InJournal, &dev->flags)) {
2678 r5c_make_stripe_write_out(sh);
2679 return -EAGAIN;
2680 }
2681 }
2682
2683 /* if the stripe is not counted in big_stripe_tree, add it now */
2684 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2685 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2686 tree_index = r5c_tree_index(conf, sh->sector);
2687 spin_lock(&log->tree_lock);
2688 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2689 tree_index);
2690 if (pslot) {
2691 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2692 pslot, &log->tree_lock) >>
2693 R5C_RADIX_COUNT_SHIFT;
2694 radix_tree_replace_slot(
2695 &log->big_stripe_tree, pslot,
2696 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2697 } else {
2698 /*
2699 * this radix_tree_insert can fail safely, so no
2700 * need to call radix_tree_preload()
2701 */
2702 ret = radix_tree_insert(
2703 &log->big_stripe_tree, tree_index,
2704 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2705 if (ret) {
2706 spin_unlock(&log->tree_lock);
2707 r5c_make_stripe_write_out(sh);
2708 return -EAGAIN;
2709 }
2710 }
2711 spin_unlock(&log->tree_lock);
2712
2713 /*
2714 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2715 * counted in the radix tree
2716 */
2717 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2718 atomic_inc(&conf->r5c_cached_partial_stripes);
2719 }
2720
2721 for (i = disks; i--; ) {
2722 dev = &sh->dev[i];
2723 if (dev->towrite) {
2724 set_bit(R5_Wantwrite, &dev->flags);
2725 set_bit(R5_Wantdrain, &dev->flags);
2726 set_bit(R5_LOCKED, &dev->flags);
2727 to_cache++;
2728 }
2729 }
2730
2731 if (to_cache) {
2732 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2733 /*
2734 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2735 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2736 * r5c_handle_data_cached()
2737 */
2738 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2739 }
2740
2741 return 0;
2742}
2743
2744/*
2745 * free extra pages (orig_page) we allocated for prexor
2746 */
2747void r5c_release_extra_page(struct stripe_head *sh)
2748{
2749 struct r5conf *conf = sh->raid_conf;
2750 int i;
2751 bool using_disk_info_extra_page;
2752
2753 using_disk_info_extra_page =
2754 sh->dev[0].orig_page == conf->disks[0].extra_page;
2755
2756 for (i = sh->disks; i--; )
2757 if (sh->dev[i].page != sh->dev[i].orig_page) {
2758 struct page *p = sh->dev[i].orig_page;
2759
2760 sh->dev[i].orig_page = sh->dev[i].page;
2761 clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2762
2763 if (!using_disk_info_extra_page)
2764 put_page(p);
2765 }
2766
2767 if (using_disk_info_extra_page) {
2768 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2769 md_wakeup_thread(conf->mddev->thread);
2770 }
2771}
2772
2773void r5c_use_extra_page(struct stripe_head *sh)
2774{
2775 struct r5conf *conf = sh->raid_conf;
2776 int i;
2777 struct r5dev *dev;
2778
2779 for (i = sh->disks; i--; ) {
2780 dev = &sh->dev[i];
2781 if (dev->orig_page != dev->page)
2782 put_page(dev->orig_page);
2783 dev->orig_page = conf->disks[i].extra_page;
2784 }
2785}
2786
2787/*
2788 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2789 * stripe is committed to RAID disks.
2790 */
2791void r5c_finish_stripe_write_out(struct r5conf *conf,
2792 struct stripe_head *sh,
2793 struct stripe_head_state *s)
2794{
2795 struct r5l_log *log = READ_ONCE(conf->log);
2796 int i;
2797 sector_t tree_index;
2798 void __rcu **pslot;
2799 uintptr_t refcount;
2800
2801 if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2802 return;
2803
2804 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2805 clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2806
2807 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2808 return;
2809
2810 for (i = sh->disks; i--; ) {
2811 clear_bit(R5_InJournal, &sh->dev[i].flags);
2812 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2813 wake_up_bit(&sh->dev[i].flags, R5_Overlap);
2814 }
2815
2816 /*
2817 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2818 * We updated R5_InJournal, so we also update s->injournal.
2819 */
2820 s->injournal = 0;
2821
2822 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2823 if (atomic_dec_and_test(&conf->pending_full_writes))
2824 md_wakeup_thread(conf->mddev->thread);
2825
2826 spin_lock_irq(&log->stripe_in_journal_lock);
2827 list_del_init(&sh->r5c);
2828 spin_unlock_irq(&log->stripe_in_journal_lock);
2829 sh->log_start = MaxSector;
2830
2831 atomic_dec(&log->stripe_in_journal_count);
2832 r5c_update_log_state(log);
2833
2834 /* stop counting this stripe in big_stripe_tree */
2835 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2836 test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2837 tree_index = r5c_tree_index(conf, sh->sector);
2838 spin_lock(&log->tree_lock);
2839 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2840 tree_index);
2841 BUG_ON(pslot == NULL);
2842 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2843 pslot, &log->tree_lock) >>
2844 R5C_RADIX_COUNT_SHIFT;
2845 if (refcount == 1)
2846 radix_tree_delete(&log->big_stripe_tree, tree_index);
2847 else
2848 radix_tree_replace_slot(
2849 &log->big_stripe_tree, pslot,
2850 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2851 spin_unlock(&log->tree_lock);
2852 }
2853
2854 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2855 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2856 atomic_dec(&conf->r5c_flushing_partial_stripes);
2857 atomic_dec(&conf->r5c_cached_partial_stripes);
2858 }
2859
2860 if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2861 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2862 atomic_dec(&conf->r5c_flushing_full_stripes);
2863 atomic_dec(&conf->r5c_cached_full_stripes);
2864 }
2865
2866 r5l_append_flush_payload(log, sh->sector);
2867 /* stripe is flused to raid disks, we can do resync now */
2868 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2869 set_bit(STRIPE_HANDLE, &sh->state);
2870}
2871
2872int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2873{
2874 struct r5conf *conf = sh->raid_conf;
2875 int pages = 0;
2876 int reserve;
2877 int i;
2878 int ret = 0;
2879
2880 BUG_ON(!log);
2881
2882 for (i = 0; i < sh->disks; i++) {
2883 void *addr;
2884
2885 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2886 continue;
2887 addr = kmap_atomic(sh->dev[i].page);
2888 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2889 addr, PAGE_SIZE);
2890 kunmap_atomic(addr);
2891 pages++;
2892 }
2893 WARN_ON(pages == 0);
2894
2895 /*
2896 * The stripe must enter state machine again to call endio, so
2897 * don't delay.
2898 */
2899 clear_bit(STRIPE_DELAYED, &sh->state);
2900 atomic_inc(&sh->count);
2901
2902 mutex_lock(&log->io_mutex);
2903 /* meta + data */
2904 reserve = (1 + pages) << (PAGE_SHIFT - 9);
2905
2906 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2907 sh->log_start == MaxSector)
2908 r5l_add_no_space_stripe(log, sh);
2909 else if (!r5l_has_free_space(log, reserve)) {
2910 if (sh->log_start == log->last_checkpoint)
2911 BUG();
2912 else
2913 r5l_add_no_space_stripe(log, sh);
2914 } else {
2915 ret = r5l_log_stripe(log, sh, pages, 0);
2916 if (ret) {
2917 spin_lock_irq(&log->io_list_lock);
2918 list_add_tail(&sh->log_list, &log->no_mem_stripes);
2919 spin_unlock_irq(&log->io_list_lock);
2920 }
2921 }
2922
2923 mutex_unlock(&log->io_mutex);
2924 return 0;
2925}
2926
2927/* check whether this big stripe is in write back cache. */
2928bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2929{
2930 struct r5l_log *log = READ_ONCE(conf->log);
2931 sector_t tree_index;
2932 void *slot;
2933
2934 if (!log)
2935 return false;
2936
2937 tree_index = r5c_tree_index(conf, sect);
2938 slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2939 return slot != NULL;
2940}
2941
2942static int r5l_load_log(struct r5l_log *log)
2943{
2944 struct md_rdev *rdev = log->rdev;
2945 struct page *page;
2946 struct r5l_meta_block *mb;
2947 sector_t cp = log->rdev->journal_tail;
2948 u32 stored_crc, expected_crc;
2949 bool create_super = false;
2950 int ret = 0;
2951
2952 /* Make sure it's valid */
2953 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2954 cp = 0;
2955 page = alloc_page(GFP_KERNEL);
2956 if (!page)
2957 return -ENOMEM;
2958
2959 if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, false)) {
2960 ret = -EIO;
2961 goto ioerr;
2962 }
2963 mb = page_address(page);
2964
2965 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2966 mb->version != R5LOG_VERSION) {
2967 create_super = true;
2968 goto create;
2969 }
2970 stored_crc = le32_to_cpu(mb->checksum);
2971 mb->checksum = 0;
2972 expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2973 if (stored_crc != expected_crc) {
2974 create_super = true;
2975 goto create;
2976 }
2977 if (le64_to_cpu(mb->position) != cp) {
2978 create_super = true;
2979 goto create;
2980 }
2981create:
2982 if (create_super) {
2983 log->last_cp_seq = get_random_u32();
2984 cp = 0;
2985 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2986 /*
2987 * Make sure super points to correct address. Log might have
2988 * data very soon. If super hasn't correct log tail address,
2989 * recovery can't find the log
2990 */
2991 r5l_write_super(log, cp);
2992 } else
2993 log->last_cp_seq = le64_to_cpu(mb->seq);
2994
2995 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
2996 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
2997 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
2998 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
2999 log->last_checkpoint = cp;
3000
3001 __free_page(page);
3002
3003 if (create_super) {
3004 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3005 log->seq = log->last_cp_seq + 1;
3006 log->next_checkpoint = cp;
3007 } else
3008 ret = r5l_recovery_log(log);
3009
3010 r5c_update_log_state(log);
3011 return ret;
3012ioerr:
3013 __free_page(page);
3014 return ret;
3015}
3016
3017int r5l_start(struct r5l_log *log)
3018{
3019 int ret;
3020
3021 if (!log)
3022 return 0;
3023
3024 ret = r5l_load_log(log);
3025 if (ret) {
3026 struct mddev *mddev = log->rdev->mddev;
3027 struct r5conf *conf = mddev->private;
3028
3029 r5l_exit_log(conf);
3030 }
3031 return ret;
3032}
3033
3034void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3035{
3036 struct r5conf *conf = mddev->private;
3037 struct r5l_log *log = READ_ONCE(conf->log);
3038
3039 if (!log)
3040 return;
3041
3042 if ((raid5_calc_degraded(conf) > 0 ||
3043 test_bit(Journal, &rdev->flags)) &&
3044 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3045 schedule_work(&log->disable_writeback_work);
3046}
3047
3048int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3049{
3050 struct r5l_log *log;
3051 struct md_thread *thread;
3052 int ret;
3053
3054 pr_debug("md/raid:%s: using device %pg as journal\n",
3055 mdname(conf->mddev), rdev->bdev);
3056
3057 if (PAGE_SIZE != 4096)
3058 return -EINVAL;
3059
3060 /*
3061 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3062 * raid_disks r5l_payload_data_parity.
3063 *
3064 * Write journal and cache does not work for very big array
3065 * (raid_disks > 203)
3066 */
3067 if (sizeof(struct r5l_meta_block) +
3068 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3069 conf->raid_disks) > PAGE_SIZE) {
3070 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3071 mdname(conf->mddev), conf->raid_disks);
3072 return -EINVAL;
3073 }
3074
3075 log = kzalloc(sizeof(*log), GFP_KERNEL);
3076 if (!log)
3077 return -ENOMEM;
3078 log->rdev = rdev;
3079 log->need_cache_flush = bdev_write_cache(rdev->bdev);
3080 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3081 sizeof(rdev->mddev->uuid));
3082
3083 mutex_init(&log->io_mutex);
3084
3085 spin_lock_init(&log->io_list_lock);
3086 INIT_LIST_HEAD(&log->running_ios);
3087 INIT_LIST_HEAD(&log->io_end_ios);
3088 INIT_LIST_HEAD(&log->flushing_ios);
3089 INIT_LIST_HEAD(&log->finished_ios);
3090
3091 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3092 if (!log->io_kc)
3093 goto io_kc;
3094
3095 ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3096 if (ret)
3097 goto io_pool;
3098
3099 ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3100 if (ret)
3101 goto io_bs;
3102
3103 ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3104 if (ret)
3105 goto out_mempool;
3106
3107 spin_lock_init(&log->tree_lock);
3108 INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3109
3110 thread = md_register_thread(r5l_reclaim_thread, log->rdev->mddev,
3111 "reclaim");
3112 if (!thread)
3113 goto reclaim_thread;
3114
3115 thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3116 rcu_assign_pointer(log->reclaim_thread, thread);
3117
3118 init_waitqueue_head(&log->iounit_wait);
3119
3120 INIT_LIST_HEAD(&log->no_mem_stripes);
3121
3122 INIT_LIST_HEAD(&log->no_space_stripes);
3123 spin_lock_init(&log->no_space_stripes_lock);
3124
3125 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3126 INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3127
3128 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3129 INIT_LIST_HEAD(&log->stripe_in_journal_list);
3130 spin_lock_init(&log->stripe_in_journal_lock);
3131 atomic_set(&log->stripe_in_journal_count, 0);
3132
3133 WRITE_ONCE(conf->log, log);
3134
3135 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3136 return 0;
3137
3138reclaim_thread:
3139 mempool_exit(&log->meta_pool);
3140out_mempool:
3141 bioset_exit(&log->bs);
3142io_bs:
3143 mempool_exit(&log->io_pool);
3144io_pool:
3145 kmem_cache_destroy(log->io_kc);
3146io_kc:
3147 kfree(log);
3148 return -EINVAL;
3149}
3150
3151void r5l_exit_log(struct r5conf *conf)
3152{
3153 struct r5l_log *log = conf->log;
3154
3155 md_unregister_thread(conf->mddev, &log->reclaim_thread);
3156
3157 /*
3158 * 'reconfig_mutex' is held by caller, set 'confg->log' to NULL to
3159 * ensure disable_writeback_work wakes up and exits.
3160 */
3161 WRITE_ONCE(conf->log, NULL);
3162 wake_up(&conf->mddev->sb_wait);
3163 flush_work(&log->disable_writeback_work);
3164
3165 mempool_exit(&log->meta_pool);
3166 bioset_exit(&log->bs);
3167 mempool_exit(&log->io_pool);
3168 kmem_cache_destroy(log->io_kc);
3169 kfree(log);
3170}
1/*
2 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
3 *
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
11 * more details.
12 *
13 */
14#include <linux/kernel.h>
15#include <linux/wait.h>
16#include <linux/blkdev.h>
17#include <linux/slab.h>
18#include <linux/raid/md_p.h>
19#include <linux/crc32c.h>
20#include <linux/random.h>
21#include "md.h"
22#include "raid5.h"
23
24/*
25 * metadata/data stored in disk with 4k size unit (a block) regardless
26 * underneath hardware sector size. only works with PAGE_SIZE == 4096
27 */
28#define BLOCK_SECTORS (8)
29
30/*
31 * reclaim runs every 1/4 disk size or 10G reclaimable space. This can prevent
32 * recovery scans a very long log
33 */
34#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
35#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
36
37/*
38 * We only need 2 bios per I/O unit to make progress, but ensure we
39 * have a few more available to not get too tight.
40 */
41#define R5L_POOL_SIZE 4
42
43struct r5l_log {
44 struct md_rdev *rdev;
45
46 u32 uuid_checksum;
47
48 sector_t device_size; /* log device size, round to
49 * BLOCK_SECTORS */
50 sector_t max_free_space; /* reclaim run if free space is at
51 * this size */
52
53 sector_t last_checkpoint; /* log tail. where recovery scan
54 * starts from */
55 u64 last_cp_seq; /* log tail sequence */
56
57 sector_t log_start; /* log head. where new data appends */
58 u64 seq; /* log head sequence */
59
60 sector_t next_checkpoint;
61 u64 next_cp_seq;
62
63 struct mutex io_mutex;
64 struct r5l_io_unit *current_io; /* current io_unit accepting new data */
65
66 spinlock_t io_list_lock;
67 struct list_head running_ios; /* io_units which are still running,
68 * and have not yet been completely
69 * written to the log */
70 struct list_head io_end_ios; /* io_units which have been completely
71 * written to the log but not yet written
72 * to the RAID */
73 struct list_head flushing_ios; /* io_units which are waiting for log
74 * cache flush */
75 struct list_head finished_ios; /* io_units which settle down in log disk */
76 struct bio flush_bio;
77
78 struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */
79
80 struct kmem_cache *io_kc;
81 mempool_t *io_pool;
82 struct bio_set *bs;
83 mempool_t *meta_pool;
84
85 struct md_thread *reclaim_thread;
86 unsigned long reclaim_target; /* number of space that need to be
87 * reclaimed. if it's 0, reclaim spaces
88 * used by io_units which are in
89 * IO_UNIT_STRIPE_END state (eg, reclaim
90 * dones't wait for specific io_unit
91 * switching to IO_UNIT_STRIPE_END
92 * state) */
93 wait_queue_head_t iounit_wait;
94
95 struct list_head no_space_stripes; /* pending stripes, log has no space */
96 spinlock_t no_space_stripes_lock;
97
98 bool need_cache_flush;
99 bool in_teardown;
100};
101
102/*
103 * an IO range starts from a meta data block and end at the next meta data
104 * block. The io unit's the meta data block tracks data/parity followed it. io
105 * unit is written to log disk with normal write, as we always flush log disk
106 * first and then start move data to raid disks, there is no requirement to
107 * write io unit with FLUSH/FUA
108 */
109struct r5l_io_unit {
110 struct r5l_log *log;
111
112 struct page *meta_page; /* store meta block */
113 int meta_offset; /* current offset in meta_page */
114
115 struct bio *current_bio;/* current_bio accepting new data */
116
117 atomic_t pending_stripe;/* how many stripes not flushed to raid */
118 u64 seq; /* seq number of the metablock */
119 sector_t log_start; /* where the io_unit starts */
120 sector_t log_end; /* where the io_unit ends */
121 struct list_head log_sibling; /* log->running_ios */
122 struct list_head stripe_list; /* stripes added to the io_unit */
123
124 int state;
125 bool need_split_bio;
126};
127
128/* r5l_io_unit state */
129enum r5l_io_unit_state {
130 IO_UNIT_RUNNING = 0, /* accepting new IO */
131 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
132 * don't accepting new bio */
133 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
134 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
135};
136
137static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
138{
139 start += inc;
140 if (start >= log->device_size)
141 start = start - log->device_size;
142 return start;
143}
144
145static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
146 sector_t end)
147{
148 if (end >= start)
149 return end - start;
150 else
151 return end + log->device_size - start;
152}
153
154static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
155{
156 sector_t used_size;
157
158 used_size = r5l_ring_distance(log, log->last_checkpoint,
159 log->log_start);
160
161 return log->device_size > used_size + size;
162}
163
164static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
165 enum r5l_io_unit_state state)
166{
167 if (WARN_ON(io->state >= state))
168 return;
169 io->state = state;
170}
171
172static void r5l_io_run_stripes(struct r5l_io_unit *io)
173{
174 struct stripe_head *sh, *next;
175
176 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
177 list_del_init(&sh->log_list);
178 set_bit(STRIPE_HANDLE, &sh->state);
179 raid5_release_stripe(sh);
180 }
181}
182
183static void r5l_log_run_stripes(struct r5l_log *log)
184{
185 struct r5l_io_unit *io, *next;
186
187 assert_spin_locked(&log->io_list_lock);
188
189 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
190 /* don't change list order */
191 if (io->state < IO_UNIT_IO_END)
192 break;
193
194 list_move_tail(&io->log_sibling, &log->finished_ios);
195 r5l_io_run_stripes(io);
196 }
197}
198
199static void r5l_move_to_end_ios(struct r5l_log *log)
200{
201 struct r5l_io_unit *io, *next;
202
203 assert_spin_locked(&log->io_list_lock);
204
205 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
206 /* don't change list order */
207 if (io->state < IO_UNIT_IO_END)
208 break;
209 list_move_tail(&io->log_sibling, &log->io_end_ios);
210 }
211}
212
213static void r5l_log_endio(struct bio *bio)
214{
215 struct r5l_io_unit *io = bio->bi_private;
216 struct r5l_log *log = io->log;
217 unsigned long flags;
218
219 if (bio->bi_error)
220 md_error(log->rdev->mddev, log->rdev);
221
222 bio_put(bio);
223 mempool_free(io->meta_page, log->meta_pool);
224
225 spin_lock_irqsave(&log->io_list_lock, flags);
226 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
227 if (log->need_cache_flush)
228 r5l_move_to_end_ios(log);
229 else
230 r5l_log_run_stripes(log);
231 spin_unlock_irqrestore(&log->io_list_lock, flags);
232
233 if (log->need_cache_flush)
234 md_wakeup_thread(log->rdev->mddev->thread);
235}
236
237static void r5l_submit_current_io(struct r5l_log *log)
238{
239 struct r5l_io_unit *io = log->current_io;
240 struct r5l_meta_block *block;
241 unsigned long flags;
242 u32 crc;
243
244 if (!io)
245 return;
246
247 block = page_address(io->meta_page);
248 block->meta_size = cpu_to_le32(io->meta_offset);
249 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
250 block->checksum = cpu_to_le32(crc);
251
252 log->current_io = NULL;
253 spin_lock_irqsave(&log->io_list_lock, flags);
254 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
255 spin_unlock_irqrestore(&log->io_list_lock, flags);
256
257 submit_bio(WRITE, io->current_bio);
258}
259
260static struct bio *r5l_bio_alloc(struct r5l_log *log)
261{
262 struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
263
264 bio->bi_rw = WRITE;
265 bio->bi_bdev = log->rdev->bdev;
266 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
267
268 return bio;
269}
270
271static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
272{
273 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
274
275 /*
276 * If we filled up the log device start from the beginning again,
277 * which will require a new bio.
278 *
279 * Note: for this to work properly the log size needs to me a multiple
280 * of BLOCK_SECTORS.
281 */
282 if (log->log_start == 0)
283 io->need_split_bio = true;
284
285 io->log_end = log->log_start;
286}
287
288static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
289{
290 struct r5l_io_unit *io;
291 struct r5l_meta_block *block;
292
293 io = mempool_alloc(log->io_pool, GFP_ATOMIC);
294 if (!io)
295 return NULL;
296 memset(io, 0, sizeof(*io));
297
298 io->log = log;
299 INIT_LIST_HEAD(&io->log_sibling);
300 INIT_LIST_HEAD(&io->stripe_list);
301 io->state = IO_UNIT_RUNNING;
302
303 io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
304 block = page_address(io->meta_page);
305 clear_page(block);
306 block->magic = cpu_to_le32(R5LOG_MAGIC);
307 block->version = R5LOG_VERSION;
308 block->seq = cpu_to_le64(log->seq);
309 block->position = cpu_to_le64(log->log_start);
310
311 io->log_start = log->log_start;
312 io->meta_offset = sizeof(struct r5l_meta_block);
313 io->seq = log->seq++;
314
315 io->current_bio = r5l_bio_alloc(log);
316 io->current_bio->bi_end_io = r5l_log_endio;
317 io->current_bio->bi_private = io;
318 bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
319
320 r5_reserve_log_entry(log, io);
321
322 spin_lock_irq(&log->io_list_lock);
323 list_add_tail(&io->log_sibling, &log->running_ios);
324 spin_unlock_irq(&log->io_list_lock);
325
326 return io;
327}
328
329static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
330{
331 if (log->current_io &&
332 log->current_io->meta_offset + payload_size > PAGE_SIZE)
333 r5l_submit_current_io(log);
334
335 if (!log->current_io) {
336 log->current_io = r5l_new_meta(log);
337 if (!log->current_io)
338 return -ENOMEM;
339 }
340
341 return 0;
342}
343
344static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
345 sector_t location,
346 u32 checksum1, u32 checksum2,
347 bool checksum2_valid)
348{
349 struct r5l_io_unit *io = log->current_io;
350 struct r5l_payload_data_parity *payload;
351
352 payload = page_address(io->meta_page) + io->meta_offset;
353 payload->header.type = cpu_to_le16(type);
354 payload->header.flags = cpu_to_le16(0);
355 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
356 (PAGE_SHIFT - 9));
357 payload->location = cpu_to_le64(location);
358 payload->checksum[0] = cpu_to_le32(checksum1);
359 if (checksum2_valid)
360 payload->checksum[1] = cpu_to_le32(checksum2);
361
362 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
363 sizeof(__le32) * (1 + !!checksum2_valid);
364}
365
366static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
367{
368 struct r5l_io_unit *io = log->current_io;
369
370 if (io->need_split_bio) {
371 struct bio *prev = io->current_bio;
372
373 io->current_bio = r5l_bio_alloc(log);
374 bio_chain(io->current_bio, prev);
375
376 submit_bio(WRITE, prev);
377 }
378
379 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
380 BUG();
381
382 r5_reserve_log_entry(log, io);
383}
384
385static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
386 int data_pages, int parity_pages)
387{
388 int i;
389 int meta_size;
390 int ret;
391 struct r5l_io_unit *io;
392
393 meta_size =
394 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
395 * data_pages) +
396 sizeof(struct r5l_payload_data_parity) +
397 sizeof(__le32) * parity_pages;
398
399 ret = r5l_get_meta(log, meta_size);
400 if (ret)
401 return ret;
402
403 io = log->current_io;
404
405 for (i = 0; i < sh->disks; i++) {
406 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
407 continue;
408 if (i == sh->pd_idx || i == sh->qd_idx)
409 continue;
410 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
411 raid5_compute_blocknr(sh, i, 0),
412 sh->dev[i].log_checksum, 0, false);
413 r5l_append_payload_page(log, sh->dev[i].page);
414 }
415
416 if (sh->qd_idx >= 0) {
417 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
418 sh->sector, sh->dev[sh->pd_idx].log_checksum,
419 sh->dev[sh->qd_idx].log_checksum, true);
420 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
421 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
422 } else {
423 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
424 sh->sector, sh->dev[sh->pd_idx].log_checksum,
425 0, false);
426 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
427 }
428
429 list_add_tail(&sh->log_list, &io->stripe_list);
430 atomic_inc(&io->pending_stripe);
431 sh->log_io = io;
432
433 return 0;
434}
435
436static void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
437/*
438 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
439 * data from log to raid disks), so we shouldn't wait for reclaim here
440 */
441int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
442{
443 int write_disks = 0;
444 int data_pages, parity_pages;
445 int meta_size;
446 int reserve;
447 int i;
448 int ret = 0;
449
450 if (!log)
451 return -EAGAIN;
452 /* Don't support stripe batch */
453 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
454 test_bit(STRIPE_SYNCING, &sh->state)) {
455 /* the stripe is written to log, we start writing it to raid */
456 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
457 return -EAGAIN;
458 }
459
460 for (i = 0; i < sh->disks; i++) {
461 void *addr;
462
463 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
464 continue;
465 write_disks++;
466 /* checksum is already calculated in last run */
467 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
468 continue;
469 addr = kmap_atomic(sh->dev[i].page);
470 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
471 addr, PAGE_SIZE);
472 kunmap_atomic(addr);
473 }
474 parity_pages = 1 + !!(sh->qd_idx >= 0);
475 data_pages = write_disks - parity_pages;
476
477 meta_size =
478 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
479 * data_pages) +
480 sizeof(struct r5l_payload_data_parity) +
481 sizeof(__le32) * parity_pages;
482 /* Doesn't work with very big raid array */
483 if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE)
484 return -EINVAL;
485
486 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
487 /*
488 * The stripe must enter state machine again to finish the write, so
489 * don't delay.
490 */
491 clear_bit(STRIPE_DELAYED, &sh->state);
492 atomic_inc(&sh->count);
493
494 mutex_lock(&log->io_mutex);
495 /* meta + data */
496 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
497 if (!r5l_has_free_space(log, reserve)) {
498 spin_lock(&log->no_space_stripes_lock);
499 list_add_tail(&sh->log_list, &log->no_space_stripes);
500 spin_unlock(&log->no_space_stripes_lock);
501
502 r5l_wake_reclaim(log, reserve);
503 } else {
504 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
505 if (ret) {
506 spin_lock_irq(&log->io_list_lock);
507 list_add_tail(&sh->log_list, &log->no_mem_stripes);
508 spin_unlock_irq(&log->io_list_lock);
509 }
510 }
511
512 mutex_unlock(&log->io_mutex);
513 return 0;
514}
515
516void r5l_write_stripe_run(struct r5l_log *log)
517{
518 if (!log)
519 return;
520 mutex_lock(&log->io_mutex);
521 r5l_submit_current_io(log);
522 mutex_unlock(&log->io_mutex);
523}
524
525int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
526{
527 if (!log)
528 return -ENODEV;
529 /*
530 * we flush log disk cache first, then write stripe data to raid disks.
531 * So if bio is finished, the log disk cache is flushed already. The
532 * recovery guarantees we can recovery the bio from log disk, so we
533 * don't need to flush again
534 */
535 if (bio->bi_iter.bi_size == 0) {
536 bio_endio(bio);
537 return 0;
538 }
539 bio->bi_rw &= ~REQ_FLUSH;
540 return -EAGAIN;
541}
542
543/* This will run after log space is reclaimed */
544static void r5l_run_no_space_stripes(struct r5l_log *log)
545{
546 struct stripe_head *sh;
547
548 spin_lock(&log->no_space_stripes_lock);
549 while (!list_empty(&log->no_space_stripes)) {
550 sh = list_first_entry(&log->no_space_stripes,
551 struct stripe_head, log_list);
552 list_del_init(&sh->log_list);
553 set_bit(STRIPE_HANDLE, &sh->state);
554 raid5_release_stripe(sh);
555 }
556 spin_unlock(&log->no_space_stripes_lock);
557}
558
559static sector_t r5l_reclaimable_space(struct r5l_log *log)
560{
561 return r5l_ring_distance(log, log->last_checkpoint,
562 log->next_checkpoint);
563}
564
565static void r5l_run_no_mem_stripe(struct r5l_log *log)
566{
567 struct stripe_head *sh;
568
569 assert_spin_locked(&log->io_list_lock);
570
571 if (!list_empty(&log->no_mem_stripes)) {
572 sh = list_first_entry(&log->no_mem_stripes,
573 struct stripe_head, log_list);
574 list_del_init(&sh->log_list);
575 set_bit(STRIPE_HANDLE, &sh->state);
576 raid5_release_stripe(sh);
577 }
578}
579
580static bool r5l_complete_finished_ios(struct r5l_log *log)
581{
582 struct r5l_io_unit *io, *next;
583 bool found = false;
584
585 assert_spin_locked(&log->io_list_lock);
586
587 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
588 /* don't change list order */
589 if (io->state < IO_UNIT_STRIPE_END)
590 break;
591
592 log->next_checkpoint = io->log_start;
593 log->next_cp_seq = io->seq;
594
595 list_del(&io->log_sibling);
596 mempool_free(io, log->io_pool);
597 r5l_run_no_mem_stripe(log);
598
599 found = true;
600 }
601
602 return found;
603}
604
605static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
606{
607 struct r5l_log *log = io->log;
608 unsigned long flags;
609
610 spin_lock_irqsave(&log->io_list_lock, flags);
611 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
612
613 if (!r5l_complete_finished_ios(log)) {
614 spin_unlock_irqrestore(&log->io_list_lock, flags);
615 return;
616 }
617
618 if (r5l_reclaimable_space(log) > log->max_free_space)
619 r5l_wake_reclaim(log, 0);
620
621 spin_unlock_irqrestore(&log->io_list_lock, flags);
622 wake_up(&log->iounit_wait);
623}
624
625void r5l_stripe_write_finished(struct stripe_head *sh)
626{
627 struct r5l_io_unit *io;
628
629 io = sh->log_io;
630 sh->log_io = NULL;
631
632 if (io && atomic_dec_and_test(&io->pending_stripe))
633 __r5l_stripe_write_finished(io);
634}
635
636static void r5l_log_flush_endio(struct bio *bio)
637{
638 struct r5l_log *log = container_of(bio, struct r5l_log,
639 flush_bio);
640 unsigned long flags;
641 struct r5l_io_unit *io;
642
643 if (bio->bi_error)
644 md_error(log->rdev->mddev, log->rdev);
645
646 spin_lock_irqsave(&log->io_list_lock, flags);
647 list_for_each_entry(io, &log->flushing_ios, log_sibling)
648 r5l_io_run_stripes(io);
649 list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
650 spin_unlock_irqrestore(&log->io_list_lock, flags);
651}
652
653/*
654 * Starting dispatch IO to raid.
655 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
656 * broken meta in the middle of a log causes recovery can't find meta at the
657 * head of log. If operations require meta at the head persistent in log, we
658 * must make sure meta before it persistent in log too. A case is:
659 *
660 * stripe data/parity is in log, we start write stripe to raid disks. stripe
661 * data/parity must be persistent in log before we do the write to raid disks.
662 *
663 * The solution is we restrictly maintain io_unit list order. In this case, we
664 * only write stripes of an io_unit to raid disks till the io_unit is the first
665 * one whose data/parity is in log.
666 */
667void r5l_flush_stripe_to_raid(struct r5l_log *log)
668{
669 bool do_flush;
670
671 if (!log || !log->need_cache_flush)
672 return;
673
674 spin_lock_irq(&log->io_list_lock);
675 /* flush bio is running */
676 if (!list_empty(&log->flushing_ios)) {
677 spin_unlock_irq(&log->io_list_lock);
678 return;
679 }
680 list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
681 do_flush = !list_empty(&log->flushing_ios);
682 spin_unlock_irq(&log->io_list_lock);
683
684 if (!do_flush)
685 return;
686 bio_reset(&log->flush_bio);
687 log->flush_bio.bi_bdev = log->rdev->bdev;
688 log->flush_bio.bi_end_io = r5l_log_flush_endio;
689 submit_bio(WRITE_FLUSH, &log->flush_bio);
690}
691
692static void r5l_write_super(struct r5l_log *log, sector_t cp);
693static void r5l_write_super_and_discard_space(struct r5l_log *log,
694 sector_t end)
695{
696 struct block_device *bdev = log->rdev->bdev;
697 struct mddev *mddev;
698
699 r5l_write_super(log, end);
700
701 if (!blk_queue_discard(bdev_get_queue(bdev)))
702 return;
703
704 mddev = log->rdev->mddev;
705 /*
706 * This is to avoid a deadlock. r5l_quiesce holds reconfig_mutex and
707 * wait for this thread to finish. This thread waits for
708 * MD_CHANGE_PENDING clear, which is supposed to be done in
709 * md_check_recovery(). md_check_recovery() tries to get
710 * reconfig_mutex. Since r5l_quiesce already holds the mutex,
711 * md_check_recovery() fails, so the PENDING never get cleared. The
712 * in_teardown check workaround this issue.
713 */
714 if (!log->in_teardown) {
715 set_bit(MD_CHANGE_DEVS, &mddev->flags);
716 set_bit(MD_CHANGE_PENDING, &mddev->flags);
717 md_wakeup_thread(mddev->thread);
718 wait_event(mddev->sb_wait,
719 !test_bit(MD_CHANGE_PENDING, &mddev->flags) ||
720 log->in_teardown);
721 /*
722 * r5l_quiesce could run after in_teardown check and hold
723 * mutex first. Superblock might get updated twice.
724 */
725 if (log->in_teardown)
726 md_update_sb(mddev, 1);
727 } else {
728 WARN_ON(!mddev_is_locked(mddev));
729 md_update_sb(mddev, 1);
730 }
731
732 /* discard IO error really doesn't matter, ignore it */
733 if (log->last_checkpoint < end) {
734 blkdev_issue_discard(bdev,
735 log->last_checkpoint + log->rdev->data_offset,
736 end - log->last_checkpoint, GFP_NOIO, 0);
737 } else {
738 blkdev_issue_discard(bdev,
739 log->last_checkpoint + log->rdev->data_offset,
740 log->device_size - log->last_checkpoint,
741 GFP_NOIO, 0);
742 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
743 GFP_NOIO, 0);
744 }
745}
746
747
748static void r5l_do_reclaim(struct r5l_log *log)
749{
750 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
751 sector_t reclaimable;
752 sector_t next_checkpoint;
753 u64 next_cp_seq;
754
755 spin_lock_irq(&log->io_list_lock);
756 /*
757 * move proper io_unit to reclaim list. We should not change the order.
758 * reclaimable/unreclaimable io_unit can be mixed in the list, we
759 * shouldn't reuse space of an unreclaimable io_unit
760 */
761 while (1) {
762 reclaimable = r5l_reclaimable_space(log);
763 if (reclaimable >= reclaim_target ||
764 (list_empty(&log->running_ios) &&
765 list_empty(&log->io_end_ios) &&
766 list_empty(&log->flushing_ios) &&
767 list_empty(&log->finished_ios)))
768 break;
769
770 md_wakeup_thread(log->rdev->mddev->thread);
771 wait_event_lock_irq(log->iounit_wait,
772 r5l_reclaimable_space(log) > reclaimable,
773 log->io_list_lock);
774 }
775
776 next_checkpoint = log->next_checkpoint;
777 next_cp_seq = log->next_cp_seq;
778 spin_unlock_irq(&log->io_list_lock);
779
780 BUG_ON(reclaimable < 0);
781 if (reclaimable == 0)
782 return;
783
784 /*
785 * write_super will flush cache of each raid disk. We must write super
786 * here, because the log area might be reused soon and we don't want to
787 * confuse recovery
788 */
789 r5l_write_super_and_discard_space(log, next_checkpoint);
790
791 mutex_lock(&log->io_mutex);
792 log->last_checkpoint = next_checkpoint;
793 log->last_cp_seq = next_cp_seq;
794 mutex_unlock(&log->io_mutex);
795
796 r5l_run_no_space_stripes(log);
797}
798
799static void r5l_reclaim_thread(struct md_thread *thread)
800{
801 struct mddev *mddev = thread->mddev;
802 struct r5conf *conf = mddev->private;
803 struct r5l_log *log = conf->log;
804
805 if (!log)
806 return;
807 r5l_do_reclaim(log);
808}
809
810static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
811{
812 unsigned long target;
813 unsigned long new = (unsigned long)space; /* overflow in theory */
814
815 do {
816 target = log->reclaim_target;
817 if (new < target)
818 return;
819 } while (cmpxchg(&log->reclaim_target, target, new) != target);
820 md_wakeup_thread(log->reclaim_thread);
821}
822
823void r5l_quiesce(struct r5l_log *log, int state)
824{
825 struct mddev *mddev;
826 if (!log || state == 2)
827 return;
828 if (state == 0) {
829 log->in_teardown = 0;
830 /*
831 * This is a special case for hotadd. In suspend, the array has
832 * no journal. In resume, journal is initialized as well as the
833 * reclaim thread.
834 */
835 if (log->reclaim_thread)
836 return;
837 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
838 log->rdev->mddev, "reclaim");
839 } else if (state == 1) {
840 /*
841 * at this point all stripes are finished, so io_unit is at
842 * least in STRIPE_END state
843 */
844 log->in_teardown = 1;
845 /* make sure r5l_write_super_and_discard_space exits */
846 mddev = log->rdev->mddev;
847 wake_up(&mddev->sb_wait);
848 r5l_wake_reclaim(log, -1L);
849 md_unregister_thread(&log->reclaim_thread);
850 r5l_do_reclaim(log);
851 }
852}
853
854bool r5l_log_disk_error(struct r5conf *conf)
855{
856 struct r5l_log *log;
857 bool ret;
858 /* don't allow write if journal disk is missing */
859 rcu_read_lock();
860 log = rcu_dereference(conf->log);
861
862 if (!log)
863 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
864 else
865 ret = test_bit(Faulty, &log->rdev->flags);
866 rcu_read_unlock();
867 return ret;
868}
869
870struct r5l_recovery_ctx {
871 struct page *meta_page; /* current meta */
872 sector_t meta_total_blocks; /* total size of current meta and data */
873 sector_t pos; /* recovery position */
874 u64 seq; /* recovery position seq */
875};
876
877static int r5l_read_meta_block(struct r5l_log *log,
878 struct r5l_recovery_ctx *ctx)
879{
880 struct page *page = ctx->meta_page;
881 struct r5l_meta_block *mb;
882 u32 crc, stored_crc;
883
884 if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false))
885 return -EIO;
886
887 mb = page_address(page);
888 stored_crc = le32_to_cpu(mb->checksum);
889 mb->checksum = 0;
890
891 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
892 le64_to_cpu(mb->seq) != ctx->seq ||
893 mb->version != R5LOG_VERSION ||
894 le64_to_cpu(mb->position) != ctx->pos)
895 return -EINVAL;
896
897 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
898 if (stored_crc != crc)
899 return -EINVAL;
900
901 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
902 return -EINVAL;
903
904 ctx->meta_total_blocks = BLOCK_SECTORS;
905
906 return 0;
907}
908
909static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
910 struct r5l_recovery_ctx *ctx,
911 sector_t stripe_sect,
912 int *offset, sector_t *log_offset)
913{
914 struct r5conf *conf = log->rdev->mddev->private;
915 struct stripe_head *sh;
916 struct r5l_payload_data_parity *payload;
917 int disk_index;
918
919 sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
920 while (1) {
921 payload = page_address(ctx->meta_page) + *offset;
922
923 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
924 raid5_compute_sector(conf,
925 le64_to_cpu(payload->location), 0,
926 &disk_index, sh);
927
928 sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
929 sh->dev[disk_index].page, READ, false);
930 sh->dev[disk_index].log_checksum =
931 le32_to_cpu(payload->checksum[0]);
932 set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
933 ctx->meta_total_blocks += BLOCK_SECTORS;
934 } else {
935 disk_index = sh->pd_idx;
936 sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
937 sh->dev[disk_index].page, READ, false);
938 sh->dev[disk_index].log_checksum =
939 le32_to_cpu(payload->checksum[0]);
940 set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
941
942 if (sh->qd_idx >= 0) {
943 disk_index = sh->qd_idx;
944 sync_page_io(log->rdev,
945 r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
946 PAGE_SIZE, sh->dev[disk_index].page,
947 READ, false);
948 sh->dev[disk_index].log_checksum =
949 le32_to_cpu(payload->checksum[1]);
950 set_bit(R5_Wantwrite,
951 &sh->dev[disk_index].flags);
952 }
953 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
954 }
955
956 *log_offset = r5l_ring_add(log, *log_offset,
957 le32_to_cpu(payload->size));
958 *offset += sizeof(struct r5l_payload_data_parity) +
959 sizeof(__le32) *
960 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
961 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
962 break;
963 }
964
965 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
966 void *addr;
967 u32 checksum;
968
969 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
970 continue;
971 addr = kmap_atomic(sh->dev[disk_index].page);
972 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
973 kunmap_atomic(addr);
974 if (checksum != sh->dev[disk_index].log_checksum)
975 goto error;
976 }
977
978 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
979 struct md_rdev *rdev, *rrdev;
980
981 if (!test_and_clear_bit(R5_Wantwrite,
982 &sh->dev[disk_index].flags))
983 continue;
984
985 /* in case device is broken */
986 rdev = rcu_dereference(conf->disks[disk_index].rdev);
987 if (rdev)
988 sync_page_io(rdev, stripe_sect, PAGE_SIZE,
989 sh->dev[disk_index].page, WRITE, false);
990 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
991 if (rrdev)
992 sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
993 sh->dev[disk_index].page, WRITE, false);
994 }
995 raid5_release_stripe(sh);
996 return 0;
997
998error:
999 for (disk_index = 0; disk_index < sh->disks; disk_index++)
1000 sh->dev[disk_index].flags = 0;
1001 raid5_release_stripe(sh);
1002 return -EINVAL;
1003}
1004
1005static int r5l_recovery_flush_one_meta(struct r5l_log *log,
1006 struct r5l_recovery_ctx *ctx)
1007{
1008 struct r5conf *conf = log->rdev->mddev->private;
1009 struct r5l_payload_data_parity *payload;
1010 struct r5l_meta_block *mb;
1011 int offset;
1012 sector_t log_offset;
1013 sector_t stripe_sector;
1014
1015 mb = page_address(ctx->meta_page);
1016 offset = sizeof(struct r5l_meta_block);
1017 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1018
1019 while (offset < le32_to_cpu(mb->meta_size)) {
1020 int dd;
1021
1022 payload = (void *)mb + offset;
1023 stripe_sector = raid5_compute_sector(conf,
1024 le64_to_cpu(payload->location), 0, &dd, NULL);
1025 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
1026 &offset, &log_offset))
1027 return -EINVAL;
1028 }
1029 return 0;
1030}
1031
1032/* copy data/parity from log to raid disks */
1033static void r5l_recovery_flush_log(struct r5l_log *log,
1034 struct r5l_recovery_ctx *ctx)
1035{
1036 while (1) {
1037 if (r5l_read_meta_block(log, ctx))
1038 return;
1039 if (r5l_recovery_flush_one_meta(log, ctx))
1040 return;
1041 ctx->seq++;
1042 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
1043 }
1044}
1045
1046static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1047 u64 seq)
1048{
1049 struct page *page;
1050 struct r5l_meta_block *mb;
1051 u32 crc;
1052
1053 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1054 if (!page)
1055 return -ENOMEM;
1056 mb = page_address(page);
1057 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1058 mb->version = R5LOG_VERSION;
1059 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1060 mb->seq = cpu_to_le64(seq);
1061 mb->position = cpu_to_le64(pos);
1062 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1063 mb->checksum = cpu_to_le32(crc);
1064
1065 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) {
1066 __free_page(page);
1067 return -EIO;
1068 }
1069 __free_page(page);
1070 return 0;
1071}
1072
1073static int r5l_recovery_log(struct r5l_log *log)
1074{
1075 struct r5l_recovery_ctx ctx;
1076
1077 ctx.pos = log->last_checkpoint;
1078 ctx.seq = log->last_cp_seq;
1079 ctx.meta_page = alloc_page(GFP_KERNEL);
1080 if (!ctx.meta_page)
1081 return -ENOMEM;
1082
1083 r5l_recovery_flush_log(log, &ctx);
1084 __free_page(ctx.meta_page);
1085
1086 /*
1087 * we did a recovery. Now ctx.pos points to an invalid meta block. New
1088 * log will start here. but we can't let superblock point to last valid
1089 * meta block. The log might looks like:
1090 * | meta 1| meta 2| meta 3|
1091 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
1092 * superblock points to meta 1, we write a new valid meta 2n. if crash
1093 * happens again, new recovery will start from meta 1. Since meta 2n is
1094 * valid now, recovery will think meta 3 is valid, which is wrong.
1095 * The solution is we create a new meta in meta2 with its seq == meta
1096 * 1's seq + 10 and let superblock points to meta2. The same recovery will
1097 * not think meta 3 is a valid meta, because its seq doesn't match
1098 */
1099 if (ctx.seq > log->last_cp_seq + 1) {
1100 int ret;
1101
1102 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
1103 if (ret)
1104 return ret;
1105 log->seq = ctx.seq + 11;
1106 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
1107 r5l_write_super(log, ctx.pos);
1108 } else {
1109 log->log_start = ctx.pos;
1110 log->seq = ctx.seq;
1111 }
1112 return 0;
1113}
1114
1115static void r5l_write_super(struct r5l_log *log, sector_t cp)
1116{
1117 struct mddev *mddev = log->rdev->mddev;
1118
1119 log->rdev->journal_tail = cp;
1120 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1121}
1122
1123static int r5l_load_log(struct r5l_log *log)
1124{
1125 struct md_rdev *rdev = log->rdev;
1126 struct page *page;
1127 struct r5l_meta_block *mb;
1128 sector_t cp = log->rdev->journal_tail;
1129 u32 stored_crc, expected_crc;
1130 bool create_super = false;
1131 int ret;
1132
1133 /* Make sure it's valid */
1134 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
1135 cp = 0;
1136 page = alloc_page(GFP_KERNEL);
1137 if (!page)
1138 return -ENOMEM;
1139
1140 if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) {
1141 ret = -EIO;
1142 goto ioerr;
1143 }
1144 mb = page_address(page);
1145
1146 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1147 mb->version != R5LOG_VERSION) {
1148 create_super = true;
1149 goto create;
1150 }
1151 stored_crc = le32_to_cpu(mb->checksum);
1152 mb->checksum = 0;
1153 expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1154 if (stored_crc != expected_crc) {
1155 create_super = true;
1156 goto create;
1157 }
1158 if (le64_to_cpu(mb->position) != cp) {
1159 create_super = true;
1160 goto create;
1161 }
1162create:
1163 if (create_super) {
1164 log->last_cp_seq = prandom_u32();
1165 cp = 0;
1166 /*
1167 * Make sure super points to correct address. Log might have
1168 * data very soon. If super hasn't correct log tail address,
1169 * recovery can't find the log
1170 */
1171 r5l_write_super(log, cp);
1172 } else
1173 log->last_cp_seq = le64_to_cpu(mb->seq);
1174
1175 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1176 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1177 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1178 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1179 log->last_checkpoint = cp;
1180
1181 __free_page(page);
1182
1183 return r5l_recovery_log(log);
1184ioerr:
1185 __free_page(page);
1186 return ret;
1187}
1188
1189int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1190{
1191 struct r5l_log *log;
1192
1193 if (PAGE_SIZE != 4096)
1194 return -EINVAL;
1195 log = kzalloc(sizeof(*log), GFP_KERNEL);
1196 if (!log)
1197 return -ENOMEM;
1198 log->rdev = rdev;
1199
1200 log->need_cache_flush = (rdev->bdev->bd_disk->queue->flush_flags != 0);
1201
1202 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1203 sizeof(rdev->mddev->uuid));
1204
1205 mutex_init(&log->io_mutex);
1206
1207 spin_lock_init(&log->io_list_lock);
1208 INIT_LIST_HEAD(&log->running_ios);
1209 INIT_LIST_HEAD(&log->io_end_ios);
1210 INIT_LIST_HEAD(&log->flushing_ios);
1211 INIT_LIST_HEAD(&log->finished_ios);
1212 bio_init(&log->flush_bio);
1213
1214 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1215 if (!log->io_kc)
1216 goto io_kc;
1217
1218 log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
1219 if (!log->io_pool)
1220 goto io_pool;
1221
1222 log->bs = bioset_create(R5L_POOL_SIZE, 0);
1223 if (!log->bs)
1224 goto io_bs;
1225
1226 log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
1227 if (!log->meta_pool)
1228 goto out_mempool;
1229
1230 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1231 log->rdev->mddev, "reclaim");
1232 if (!log->reclaim_thread)
1233 goto reclaim_thread;
1234 init_waitqueue_head(&log->iounit_wait);
1235
1236 INIT_LIST_HEAD(&log->no_mem_stripes);
1237
1238 INIT_LIST_HEAD(&log->no_space_stripes);
1239 spin_lock_init(&log->no_space_stripes_lock);
1240
1241 if (r5l_load_log(log))
1242 goto error;
1243
1244 rcu_assign_pointer(conf->log, log);
1245 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1246 return 0;
1247
1248error:
1249 md_unregister_thread(&log->reclaim_thread);
1250reclaim_thread:
1251 mempool_destroy(log->meta_pool);
1252out_mempool:
1253 bioset_free(log->bs);
1254io_bs:
1255 mempool_destroy(log->io_pool);
1256io_pool:
1257 kmem_cache_destroy(log->io_kc);
1258io_kc:
1259 kfree(log);
1260 return -EINVAL;
1261}
1262
1263void r5l_exit_log(struct r5l_log *log)
1264{
1265 md_unregister_thread(&log->reclaim_thread);
1266 mempool_destroy(log->meta_pool);
1267 bioset_free(log->bs);
1268 mempool_destroy(log->io_pool);
1269 kmem_cache_destroy(log->io_kc);
1270 kfree(log);
1271}