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 *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 * dones'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 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
317 RAID5_STRIPE_SECTORS(conf),
318 !test_bit(STRIPE_DEGRADED, &sh->state),
319 0);
320 }
321 }
322}
323
324void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
325
326/* Check whether we should flush some stripes to free up stripe cache */
327void r5c_check_stripe_cache_usage(struct r5conf *conf)
328{
329 int total_cached;
330
331 if (!r5c_is_writeback(conf->log))
332 return;
333
334 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
335 atomic_read(&conf->r5c_cached_full_stripes);
336
337 /*
338 * The following condition is true for either of the following:
339 * - stripe cache pressure high:
340 * total_cached > 3/4 min_nr_stripes ||
341 * empty_inactive_list_nr > 0
342 * - stripe cache pressure moderate:
343 * total_cached > 1/2 min_nr_stripes
344 */
345 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
346 atomic_read(&conf->empty_inactive_list_nr) > 0)
347 r5l_wake_reclaim(conf->log, 0);
348}
349
350/*
351 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
352 * stripes in the cache
353 */
354void r5c_check_cached_full_stripe(struct r5conf *conf)
355{
356 if (!r5c_is_writeback(conf->log))
357 return;
358
359 /*
360 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
361 * or a full stripe (chunk size / 4k stripes).
362 */
363 if (atomic_read(&conf->r5c_cached_full_stripes) >=
364 min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
365 conf->chunk_sectors >> RAID5_STRIPE_SHIFT(conf)))
366 r5l_wake_reclaim(conf->log, 0);
367}
368
369/*
370 * Total log space (in sectors) needed to flush all data in cache
371 *
372 * To avoid deadlock due to log space, it is necessary to reserve log
373 * space to flush critical stripes (stripes that occupying log space near
374 * last_checkpoint). This function helps check how much log space is
375 * required to flush all cached stripes.
376 *
377 * To reduce log space requirements, two mechanisms are used to give cache
378 * flush higher priorities:
379 * 1. In handle_stripe_dirtying() and schedule_reconstruction(),
380 * stripes ALREADY in journal can be flushed w/o pending writes;
381 * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
382 * can be delayed (r5l_add_no_space_stripe).
383 *
384 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
385 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
386 * pages of journal space. For stripes that has not passed 1, flushing it
387 * requires (conf->raid_disks + 1) pages of journal space. There are at
388 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
389 * required to flush all cached stripes (in pages) is:
390 *
391 * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
392 * (group_cnt + 1) * (raid_disks + 1)
393 * or
394 * (stripe_in_journal_count) * (max_degraded + 1) +
395 * (group_cnt + 1) * (raid_disks - max_degraded)
396 */
397static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
398{
399 struct r5l_log *log = conf->log;
400
401 if (!r5c_is_writeback(log))
402 return 0;
403
404 return BLOCK_SECTORS *
405 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
406 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
407}
408
409/*
410 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
411 *
412 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
413 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
414 * device is less than 2x of reclaim_required_space.
415 */
416static inline void r5c_update_log_state(struct r5l_log *log)
417{
418 struct r5conf *conf = log->rdev->mddev->private;
419 sector_t free_space;
420 sector_t reclaim_space;
421 bool wake_reclaim = false;
422
423 if (!r5c_is_writeback(log))
424 return;
425
426 free_space = r5l_ring_distance(log, log->log_start,
427 log->last_checkpoint);
428 reclaim_space = r5c_log_required_to_flush_cache(conf);
429 if (free_space < 2 * reclaim_space)
430 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
431 else {
432 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
433 wake_reclaim = true;
434 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
435 }
436 if (free_space < 3 * reclaim_space)
437 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
438 else
439 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
440
441 if (wake_reclaim)
442 r5l_wake_reclaim(log, 0);
443}
444
445/*
446 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
447 * This function should only be called in write-back mode.
448 */
449void r5c_make_stripe_write_out(struct stripe_head *sh)
450{
451 struct r5conf *conf = sh->raid_conf;
452 struct r5l_log *log = conf->log;
453
454 BUG_ON(!r5c_is_writeback(log));
455
456 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
457 clear_bit(STRIPE_R5C_CACHING, &sh->state);
458
459 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
460 atomic_inc(&conf->preread_active_stripes);
461}
462
463static void r5c_handle_data_cached(struct stripe_head *sh)
464{
465 int i;
466
467 for (i = sh->disks; i--; )
468 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
469 set_bit(R5_InJournal, &sh->dev[i].flags);
470 clear_bit(R5_LOCKED, &sh->dev[i].flags);
471 }
472 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
473}
474
475/*
476 * this journal write must contain full parity,
477 * it may also contain some data pages
478 */
479static void r5c_handle_parity_cached(struct stripe_head *sh)
480{
481 int i;
482
483 for (i = sh->disks; i--; )
484 if (test_bit(R5_InJournal, &sh->dev[i].flags))
485 set_bit(R5_Wantwrite, &sh->dev[i].flags);
486}
487
488/*
489 * Setting proper flags after writing (or flushing) data and/or parity to the
490 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
491 */
492static void r5c_finish_cache_stripe(struct stripe_head *sh)
493{
494 struct r5l_log *log = sh->raid_conf->log;
495
496 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
497 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
498 /*
499 * Set R5_InJournal for parity dev[pd_idx]. This means
500 * all data AND parity in the journal. For RAID 6, it is
501 * NOT necessary to set the flag for dev[qd_idx], as the
502 * two parities are written out together.
503 */
504 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
505 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
506 r5c_handle_data_cached(sh);
507 } else {
508 r5c_handle_parity_cached(sh);
509 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
510 }
511}
512
513static void r5l_io_run_stripes(struct r5l_io_unit *io)
514{
515 struct stripe_head *sh, *next;
516
517 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
518 list_del_init(&sh->log_list);
519
520 r5c_finish_cache_stripe(sh);
521
522 set_bit(STRIPE_HANDLE, &sh->state);
523 raid5_release_stripe(sh);
524 }
525}
526
527static void r5l_log_run_stripes(struct r5l_log *log)
528{
529 struct r5l_io_unit *io, *next;
530
531 lockdep_assert_held(&log->io_list_lock);
532
533 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
534 /* don't change list order */
535 if (io->state < IO_UNIT_IO_END)
536 break;
537
538 list_move_tail(&io->log_sibling, &log->finished_ios);
539 r5l_io_run_stripes(io);
540 }
541}
542
543static void r5l_move_to_end_ios(struct r5l_log *log)
544{
545 struct r5l_io_unit *io, *next;
546
547 lockdep_assert_held(&log->io_list_lock);
548
549 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
550 /* don't change list order */
551 if (io->state < IO_UNIT_IO_END)
552 break;
553 list_move_tail(&io->log_sibling, &log->io_end_ios);
554 }
555}
556
557static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
558static void r5l_log_endio(struct bio *bio)
559{
560 struct r5l_io_unit *io = bio->bi_private;
561 struct r5l_io_unit *io_deferred;
562 struct r5l_log *log = io->log;
563 unsigned long flags;
564 bool has_null_flush;
565 bool has_flush_payload;
566
567 if (bio->bi_status)
568 md_error(log->rdev->mddev, log->rdev);
569
570 bio_put(bio);
571 mempool_free(io->meta_page, &log->meta_pool);
572
573 spin_lock_irqsave(&log->io_list_lock, flags);
574 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
575
576 /*
577 * if the io doesn't not have null_flush or flush payload,
578 * it is not safe to access it after releasing io_list_lock.
579 * Therefore, it is necessary to check the condition with
580 * the lock held.
581 */
582 has_null_flush = io->has_null_flush;
583 has_flush_payload = io->has_flush_payload;
584
585 if (log->need_cache_flush && !list_empty(&io->stripe_list))
586 r5l_move_to_end_ios(log);
587 else
588 r5l_log_run_stripes(log);
589 if (!list_empty(&log->running_ios)) {
590 /*
591 * FLUSH/FUA io_unit is deferred because of ordering, now we
592 * can dispatch it
593 */
594 io_deferred = list_first_entry(&log->running_ios,
595 struct r5l_io_unit, log_sibling);
596 if (io_deferred->io_deferred)
597 schedule_work(&log->deferred_io_work);
598 }
599
600 spin_unlock_irqrestore(&log->io_list_lock, flags);
601
602 if (log->need_cache_flush)
603 md_wakeup_thread(log->rdev->mddev->thread);
604
605 /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
606 if (has_null_flush) {
607 struct bio *bi;
608
609 WARN_ON(bio_list_empty(&io->flush_barriers));
610 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
611 bio_endio(bi);
612 if (atomic_dec_and_test(&io->pending_stripe)) {
613 __r5l_stripe_write_finished(io);
614 return;
615 }
616 }
617 }
618 /* decrease pending_stripe for flush payload */
619 if (has_flush_payload)
620 if (atomic_dec_and_test(&io->pending_stripe))
621 __r5l_stripe_write_finished(io);
622}
623
624static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
625{
626 unsigned long flags;
627
628 spin_lock_irqsave(&log->io_list_lock, flags);
629 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
630 spin_unlock_irqrestore(&log->io_list_lock, flags);
631
632 /*
633 * In case of journal device failures, submit_bio will get error
634 * and calls endio, then active stripes will continue write
635 * process. Therefore, it is not necessary to check Faulty bit
636 * of journal device here.
637 *
638 * We can't check split_bio after current_bio is submitted. If
639 * io->split_bio is null, after current_bio is submitted, current_bio
640 * might already be completed and the io_unit is freed. We submit
641 * split_bio first to avoid the issue.
642 */
643 if (io->split_bio) {
644 if (io->has_flush)
645 io->split_bio->bi_opf |= REQ_PREFLUSH;
646 if (io->has_fua)
647 io->split_bio->bi_opf |= REQ_FUA;
648 submit_bio(io->split_bio);
649 }
650
651 if (io->has_flush)
652 io->current_bio->bi_opf |= REQ_PREFLUSH;
653 if (io->has_fua)
654 io->current_bio->bi_opf |= REQ_FUA;
655 submit_bio(io->current_bio);
656}
657
658/* deferred io_unit will be dispatched here */
659static void r5l_submit_io_async(struct work_struct *work)
660{
661 struct r5l_log *log = container_of(work, struct r5l_log,
662 deferred_io_work);
663 struct r5l_io_unit *io = NULL;
664 unsigned long flags;
665
666 spin_lock_irqsave(&log->io_list_lock, flags);
667 if (!list_empty(&log->running_ios)) {
668 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
669 log_sibling);
670 if (!io->io_deferred)
671 io = NULL;
672 else
673 io->io_deferred = 0;
674 }
675 spin_unlock_irqrestore(&log->io_list_lock, flags);
676 if (io)
677 r5l_do_submit_io(log, io);
678}
679
680static void r5c_disable_writeback_async(struct work_struct *work)
681{
682 struct r5l_log *log = container_of(work, struct r5l_log,
683 disable_writeback_work);
684 struct mddev *mddev = log->rdev->mddev;
685 struct r5conf *conf = mddev->private;
686 int locked = 0;
687
688 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
689 return;
690 pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
691 mdname(mddev));
692
693 /* wait superblock change before suspend */
694 wait_event(mddev->sb_wait,
695 conf->log == NULL ||
696 (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) &&
697 (locked = mddev_trylock(mddev))));
698 if (locked) {
699 mddev_suspend(mddev);
700 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
701 mddev_resume(mddev);
702 mddev_unlock(mddev);
703 }
704}
705
706static void r5l_submit_current_io(struct r5l_log *log)
707{
708 struct r5l_io_unit *io = log->current_io;
709 struct r5l_meta_block *block;
710 unsigned long flags;
711 u32 crc;
712 bool do_submit = true;
713
714 if (!io)
715 return;
716
717 block = page_address(io->meta_page);
718 block->meta_size = cpu_to_le32(io->meta_offset);
719 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
720 block->checksum = cpu_to_le32(crc);
721
722 log->current_io = NULL;
723 spin_lock_irqsave(&log->io_list_lock, flags);
724 if (io->has_flush || io->has_fua) {
725 if (io != list_first_entry(&log->running_ios,
726 struct r5l_io_unit, log_sibling)) {
727 io->io_deferred = 1;
728 do_submit = false;
729 }
730 }
731 spin_unlock_irqrestore(&log->io_list_lock, flags);
732 if (do_submit)
733 r5l_do_submit_io(log, io);
734}
735
736static struct bio *r5l_bio_alloc(struct r5l_log *log)
737{
738 struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, &log->bs);
739
740 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
741 bio_set_dev(bio, log->rdev->bdev);
742 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
743
744 return bio;
745}
746
747static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
748{
749 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
750
751 r5c_update_log_state(log);
752 /*
753 * If we filled up the log device start from the beginning again,
754 * which will require a new bio.
755 *
756 * Note: for this to work properly the log size needs to me a multiple
757 * of BLOCK_SECTORS.
758 */
759 if (log->log_start == 0)
760 io->need_split_bio = true;
761
762 io->log_end = log->log_start;
763}
764
765static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
766{
767 struct r5l_io_unit *io;
768 struct r5l_meta_block *block;
769
770 io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
771 if (!io)
772 return NULL;
773 memset(io, 0, sizeof(*io));
774
775 io->log = log;
776 INIT_LIST_HEAD(&io->log_sibling);
777 INIT_LIST_HEAD(&io->stripe_list);
778 bio_list_init(&io->flush_barriers);
779 io->state = IO_UNIT_RUNNING;
780
781 io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
782 block = page_address(io->meta_page);
783 clear_page(block);
784 block->magic = cpu_to_le32(R5LOG_MAGIC);
785 block->version = R5LOG_VERSION;
786 block->seq = cpu_to_le64(log->seq);
787 block->position = cpu_to_le64(log->log_start);
788
789 io->log_start = log->log_start;
790 io->meta_offset = sizeof(struct r5l_meta_block);
791 io->seq = log->seq++;
792
793 io->current_bio = r5l_bio_alloc(log);
794 io->current_bio->bi_end_io = r5l_log_endio;
795 io->current_bio->bi_private = io;
796 bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
797
798 r5_reserve_log_entry(log, io);
799
800 spin_lock_irq(&log->io_list_lock);
801 list_add_tail(&io->log_sibling, &log->running_ios);
802 spin_unlock_irq(&log->io_list_lock);
803
804 return io;
805}
806
807static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
808{
809 if (log->current_io &&
810 log->current_io->meta_offset + payload_size > PAGE_SIZE)
811 r5l_submit_current_io(log);
812
813 if (!log->current_io) {
814 log->current_io = r5l_new_meta(log);
815 if (!log->current_io)
816 return -ENOMEM;
817 }
818
819 return 0;
820}
821
822static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
823 sector_t location,
824 u32 checksum1, u32 checksum2,
825 bool checksum2_valid)
826{
827 struct r5l_io_unit *io = log->current_io;
828 struct r5l_payload_data_parity *payload;
829
830 payload = page_address(io->meta_page) + io->meta_offset;
831 payload->header.type = cpu_to_le16(type);
832 payload->header.flags = cpu_to_le16(0);
833 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
834 (PAGE_SHIFT - 9));
835 payload->location = cpu_to_le64(location);
836 payload->checksum[0] = cpu_to_le32(checksum1);
837 if (checksum2_valid)
838 payload->checksum[1] = cpu_to_le32(checksum2);
839
840 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
841 sizeof(__le32) * (1 + !!checksum2_valid);
842}
843
844static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
845{
846 struct r5l_io_unit *io = log->current_io;
847
848 if (io->need_split_bio) {
849 BUG_ON(io->split_bio);
850 io->split_bio = io->current_bio;
851 io->current_bio = r5l_bio_alloc(log);
852 bio_chain(io->current_bio, io->split_bio);
853 io->need_split_bio = false;
854 }
855
856 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
857 BUG();
858
859 r5_reserve_log_entry(log, io);
860}
861
862static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
863{
864 struct mddev *mddev = log->rdev->mddev;
865 struct r5conf *conf = mddev->private;
866 struct r5l_io_unit *io;
867 struct r5l_payload_flush *payload;
868 int meta_size;
869
870 /*
871 * payload_flush requires extra writes to the journal.
872 * To avoid handling the extra IO in quiesce, just skip
873 * flush_payload
874 */
875 if (conf->quiesce)
876 return;
877
878 mutex_lock(&log->io_mutex);
879 meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
880
881 if (r5l_get_meta(log, meta_size)) {
882 mutex_unlock(&log->io_mutex);
883 return;
884 }
885
886 /* current implementation is one stripe per flush payload */
887 io = log->current_io;
888 payload = page_address(io->meta_page) + io->meta_offset;
889 payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
890 payload->header.flags = cpu_to_le16(0);
891 payload->size = cpu_to_le32(sizeof(__le64));
892 payload->flush_stripes[0] = cpu_to_le64(sect);
893 io->meta_offset += meta_size;
894 /* multiple flush payloads count as one pending_stripe */
895 if (!io->has_flush_payload) {
896 io->has_flush_payload = 1;
897 atomic_inc(&io->pending_stripe);
898 }
899 mutex_unlock(&log->io_mutex);
900}
901
902static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
903 int data_pages, int parity_pages)
904{
905 int i;
906 int meta_size;
907 int ret;
908 struct r5l_io_unit *io;
909
910 meta_size =
911 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
912 * data_pages) +
913 sizeof(struct r5l_payload_data_parity) +
914 sizeof(__le32) * parity_pages;
915
916 ret = r5l_get_meta(log, meta_size);
917 if (ret)
918 return ret;
919
920 io = log->current_io;
921
922 if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
923 io->has_flush = 1;
924
925 for (i = 0; i < sh->disks; i++) {
926 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
927 test_bit(R5_InJournal, &sh->dev[i].flags))
928 continue;
929 if (i == sh->pd_idx || i == sh->qd_idx)
930 continue;
931 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
932 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
933 io->has_fua = 1;
934 /*
935 * we need to flush journal to make sure recovery can
936 * reach the data with fua flag
937 */
938 io->has_flush = 1;
939 }
940 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
941 raid5_compute_blocknr(sh, i, 0),
942 sh->dev[i].log_checksum, 0, false);
943 r5l_append_payload_page(log, sh->dev[i].page);
944 }
945
946 if (parity_pages == 2) {
947 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
948 sh->sector, sh->dev[sh->pd_idx].log_checksum,
949 sh->dev[sh->qd_idx].log_checksum, true);
950 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
951 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
952 } else if (parity_pages == 1) {
953 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
954 sh->sector, sh->dev[sh->pd_idx].log_checksum,
955 0, false);
956 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
957 } else /* Just writing data, not parity, in caching phase */
958 BUG_ON(parity_pages != 0);
959
960 list_add_tail(&sh->log_list, &io->stripe_list);
961 atomic_inc(&io->pending_stripe);
962 sh->log_io = io;
963
964 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
965 return 0;
966
967 if (sh->log_start == MaxSector) {
968 BUG_ON(!list_empty(&sh->r5c));
969 sh->log_start = io->log_start;
970 spin_lock_irq(&log->stripe_in_journal_lock);
971 list_add_tail(&sh->r5c,
972 &log->stripe_in_journal_list);
973 spin_unlock_irq(&log->stripe_in_journal_lock);
974 atomic_inc(&log->stripe_in_journal_count);
975 }
976 return 0;
977}
978
979/* add stripe to no_space_stripes, and then wake up reclaim */
980static inline void r5l_add_no_space_stripe(struct r5l_log *log,
981 struct stripe_head *sh)
982{
983 spin_lock(&log->no_space_stripes_lock);
984 list_add_tail(&sh->log_list, &log->no_space_stripes);
985 spin_unlock(&log->no_space_stripes_lock);
986}
987
988/*
989 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
990 * data from log to raid disks), so we shouldn't wait for reclaim here
991 */
992int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
993{
994 struct r5conf *conf = sh->raid_conf;
995 int write_disks = 0;
996 int data_pages, parity_pages;
997 int reserve;
998 int i;
999 int ret = 0;
1000 bool wake_reclaim = false;
1001
1002 if (!log)
1003 return -EAGAIN;
1004 /* Don't support stripe batch */
1005 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1006 test_bit(STRIPE_SYNCING, &sh->state)) {
1007 /* the stripe is written to log, we start writing it to raid */
1008 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1009 return -EAGAIN;
1010 }
1011
1012 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1013
1014 for (i = 0; i < sh->disks; i++) {
1015 void *addr;
1016
1017 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1018 test_bit(R5_InJournal, &sh->dev[i].flags))
1019 continue;
1020
1021 write_disks++;
1022 /* checksum is already calculated in last run */
1023 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1024 continue;
1025 addr = kmap_atomic(sh->dev[i].page);
1026 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1027 addr, PAGE_SIZE);
1028 kunmap_atomic(addr);
1029 }
1030 parity_pages = 1 + !!(sh->qd_idx >= 0);
1031 data_pages = write_disks - parity_pages;
1032
1033 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1034 /*
1035 * The stripe must enter state machine again to finish the write, so
1036 * don't delay.
1037 */
1038 clear_bit(STRIPE_DELAYED, &sh->state);
1039 atomic_inc(&sh->count);
1040
1041 mutex_lock(&log->io_mutex);
1042 /* meta + data */
1043 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1044
1045 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1046 if (!r5l_has_free_space(log, reserve)) {
1047 r5l_add_no_space_stripe(log, sh);
1048 wake_reclaim = true;
1049 } else {
1050 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1051 if (ret) {
1052 spin_lock_irq(&log->io_list_lock);
1053 list_add_tail(&sh->log_list,
1054 &log->no_mem_stripes);
1055 spin_unlock_irq(&log->io_list_lock);
1056 }
1057 }
1058 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1059 /*
1060 * log space critical, do not process stripes that are
1061 * not in cache yet (sh->log_start == MaxSector).
1062 */
1063 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1064 sh->log_start == MaxSector) {
1065 r5l_add_no_space_stripe(log, sh);
1066 wake_reclaim = true;
1067 reserve = 0;
1068 } else if (!r5l_has_free_space(log, reserve)) {
1069 if (sh->log_start == log->last_checkpoint)
1070 BUG();
1071 else
1072 r5l_add_no_space_stripe(log, sh);
1073 } else {
1074 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1075 if (ret) {
1076 spin_lock_irq(&log->io_list_lock);
1077 list_add_tail(&sh->log_list,
1078 &log->no_mem_stripes);
1079 spin_unlock_irq(&log->io_list_lock);
1080 }
1081 }
1082 }
1083
1084 mutex_unlock(&log->io_mutex);
1085 if (wake_reclaim)
1086 r5l_wake_reclaim(log, reserve);
1087 return 0;
1088}
1089
1090void r5l_write_stripe_run(struct r5l_log *log)
1091{
1092 if (!log)
1093 return;
1094 mutex_lock(&log->io_mutex);
1095 r5l_submit_current_io(log);
1096 mutex_unlock(&log->io_mutex);
1097}
1098
1099int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1100{
1101 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1102 /*
1103 * in write through (journal only)
1104 * we flush log disk cache first, then write stripe data to
1105 * raid disks. So if bio is finished, the log disk cache is
1106 * flushed already. The recovery guarantees we can recovery
1107 * the bio from log disk, so we don't need to flush again
1108 */
1109 if (bio->bi_iter.bi_size == 0) {
1110 bio_endio(bio);
1111 return 0;
1112 }
1113 bio->bi_opf &= ~REQ_PREFLUSH;
1114 } else {
1115 /* write back (with cache) */
1116 if (bio->bi_iter.bi_size == 0) {
1117 mutex_lock(&log->io_mutex);
1118 r5l_get_meta(log, 0);
1119 bio_list_add(&log->current_io->flush_barriers, bio);
1120 log->current_io->has_flush = 1;
1121 log->current_io->has_null_flush = 1;
1122 atomic_inc(&log->current_io->pending_stripe);
1123 r5l_submit_current_io(log);
1124 mutex_unlock(&log->io_mutex);
1125 return 0;
1126 }
1127 }
1128 return -EAGAIN;
1129}
1130
1131/* This will run after log space is reclaimed */
1132static void r5l_run_no_space_stripes(struct r5l_log *log)
1133{
1134 struct stripe_head *sh;
1135
1136 spin_lock(&log->no_space_stripes_lock);
1137 while (!list_empty(&log->no_space_stripes)) {
1138 sh = list_first_entry(&log->no_space_stripes,
1139 struct stripe_head, log_list);
1140 list_del_init(&sh->log_list);
1141 set_bit(STRIPE_HANDLE, &sh->state);
1142 raid5_release_stripe(sh);
1143 }
1144 spin_unlock(&log->no_space_stripes_lock);
1145}
1146
1147/*
1148 * calculate new last_checkpoint
1149 * for write through mode, returns log->next_checkpoint
1150 * for write back, returns log_start of first sh in stripe_in_journal_list
1151 */
1152static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1153{
1154 struct stripe_head *sh;
1155 struct r5l_log *log = conf->log;
1156 sector_t new_cp;
1157 unsigned long flags;
1158
1159 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1160 return log->next_checkpoint;
1161
1162 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1163 if (list_empty(&conf->log->stripe_in_journal_list)) {
1164 /* all stripes flushed */
1165 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1166 return log->next_checkpoint;
1167 }
1168 sh = list_first_entry(&conf->log->stripe_in_journal_list,
1169 struct stripe_head, r5c);
1170 new_cp = sh->log_start;
1171 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1172 return new_cp;
1173}
1174
1175static sector_t r5l_reclaimable_space(struct r5l_log *log)
1176{
1177 struct r5conf *conf = log->rdev->mddev->private;
1178
1179 return r5l_ring_distance(log, log->last_checkpoint,
1180 r5c_calculate_new_cp(conf));
1181}
1182
1183static void r5l_run_no_mem_stripe(struct r5l_log *log)
1184{
1185 struct stripe_head *sh;
1186
1187 lockdep_assert_held(&log->io_list_lock);
1188
1189 if (!list_empty(&log->no_mem_stripes)) {
1190 sh = list_first_entry(&log->no_mem_stripes,
1191 struct stripe_head, log_list);
1192 list_del_init(&sh->log_list);
1193 set_bit(STRIPE_HANDLE, &sh->state);
1194 raid5_release_stripe(sh);
1195 }
1196}
1197
1198static bool r5l_complete_finished_ios(struct r5l_log *log)
1199{
1200 struct r5l_io_unit *io, *next;
1201 bool found = false;
1202
1203 lockdep_assert_held(&log->io_list_lock);
1204
1205 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1206 /* don't change list order */
1207 if (io->state < IO_UNIT_STRIPE_END)
1208 break;
1209
1210 log->next_checkpoint = io->log_start;
1211
1212 list_del(&io->log_sibling);
1213 mempool_free(io, &log->io_pool);
1214 r5l_run_no_mem_stripe(log);
1215
1216 found = true;
1217 }
1218
1219 return found;
1220}
1221
1222static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1223{
1224 struct r5l_log *log = io->log;
1225 struct r5conf *conf = log->rdev->mddev->private;
1226 unsigned long flags;
1227
1228 spin_lock_irqsave(&log->io_list_lock, flags);
1229 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1230
1231 if (!r5l_complete_finished_ios(log)) {
1232 spin_unlock_irqrestore(&log->io_list_lock, flags);
1233 return;
1234 }
1235
1236 if (r5l_reclaimable_space(log) > log->max_free_space ||
1237 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1238 r5l_wake_reclaim(log, 0);
1239
1240 spin_unlock_irqrestore(&log->io_list_lock, flags);
1241 wake_up(&log->iounit_wait);
1242}
1243
1244void r5l_stripe_write_finished(struct stripe_head *sh)
1245{
1246 struct r5l_io_unit *io;
1247
1248 io = sh->log_io;
1249 sh->log_io = NULL;
1250
1251 if (io && atomic_dec_and_test(&io->pending_stripe))
1252 __r5l_stripe_write_finished(io);
1253}
1254
1255static void r5l_log_flush_endio(struct bio *bio)
1256{
1257 struct r5l_log *log = container_of(bio, struct r5l_log,
1258 flush_bio);
1259 unsigned long flags;
1260 struct r5l_io_unit *io;
1261
1262 if (bio->bi_status)
1263 md_error(log->rdev->mddev, log->rdev);
1264
1265 spin_lock_irqsave(&log->io_list_lock, flags);
1266 list_for_each_entry(io, &log->flushing_ios, log_sibling)
1267 r5l_io_run_stripes(io);
1268 list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1269 spin_unlock_irqrestore(&log->io_list_lock, flags);
1270}
1271
1272/*
1273 * Starting dispatch IO to raid.
1274 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1275 * broken meta in the middle of a log causes recovery can't find meta at the
1276 * head of log. If operations require meta at the head persistent in log, we
1277 * must make sure meta before it persistent in log too. A case is:
1278 *
1279 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1280 * data/parity must be persistent in log before we do the write to raid disks.
1281 *
1282 * The solution is we restrictly maintain io_unit list order. In this case, we
1283 * only write stripes of an io_unit to raid disks till the io_unit is the first
1284 * one whose data/parity is in log.
1285 */
1286void r5l_flush_stripe_to_raid(struct r5l_log *log)
1287{
1288 bool do_flush;
1289
1290 if (!log || !log->need_cache_flush)
1291 return;
1292
1293 spin_lock_irq(&log->io_list_lock);
1294 /* flush bio is running */
1295 if (!list_empty(&log->flushing_ios)) {
1296 spin_unlock_irq(&log->io_list_lock);
1297 return;
1298 }
1299 list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1300 do_flush = !list_empty(&log->flushing_ios);
1301 spin_unlock_irq(&log->io_list_lock);
1302
1303 if (!do_flush)
1304 return;
1305 bio_reset(&log->flush_bio);
1306 bio_set_dev(&log->flush_bio, log->rdev->bdev);
1307 log->flush_bio.bi_end_io = r5l_log_flush_endio;
1308 log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1309 submit_bio(&log->flush_bio);
1310}
1311
1312static void r5l_write_super(struct r5l_log *log, sector_t cp);
1313static void r5l_write_super_and_discard_space(struct r5l_log *log,
1314 sector_t end)
1315{
1316 struct block_device *bdev = log->rdev->bdev;
1317 struct mddev *mddev;
1318
1319 r5l_write_super(log, end);
1320
1321 if (!blk_queue_discard(bdev_get_queue(bdev)))
1322 return;
1323
1324 mddev = log->rdev->mddev;
1325 /*
1326 * Discard could zero data, so before discard we must make sure
1327 * superblock is updated to new log tail. Updating superblock (either
1328 * directly call md_update_sb() or depend on md thread) must hold
1329 * reconfig mutex. On the other hand, raid5_quiesce is called with
1330 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1331 * for all IO finish, hence waitting for reclaim thread, while reclaim
1332 * thread is calling this function and waitting for reconfig mutex. So
1333 * there is a deadlock. We workaround this issue with a trylock.
1334 * FIXME: we could miss discard if we can't take reconfig mutex
1335 */
1336 set_mask_bits(&mddev->sb_flags, 0,
1337 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1338 if (!mddev_trylock(mddev))
1339 return;
1340 md_update_sb(mddev, 1);
1341 mddev_unlock(mddev);
1342
1343 /* discard IO error really doesn't matter, ignore it */
1344 if (log->last_checkpoint < end) {
1345 blkdev_issue_discard(bdev,
1346 log->last_checkpoint + log->rdev->data_offset,
1347 end - log->last_checkpoint, GFP_NOIO, 0);
1348 } else {
1349 blkdev_issue_discard(bdev,
1350 log->last_checkpoint + log->rdev->data_offset,
1351 log->device_size - log->last_checkpoint,
1352 GFP_NOIO, 0);
1353 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1354 GFP_NOIO, 0);
1355 }
1356}
1357
1358/*
1359 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1360 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1361 *
1362 * must hold conf->device_lock
1363 */
1364static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1365{
1366 BUG_ON(list_empty(&sh->lru));
1367 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1368 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1369
1370 /*
1371 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1372 * raid5_release_stripe() while holding conf->device_lock
1373 */
1374 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1375 lockdep_assert_held(&conf->device_lock);
1376
1377 list_del_init(&sh->lru);
1378 atomic_inc(&sh->count);
1379
1380 set_bit(STRIPE_HANDLE, &sh->state);
1381 atomic_inc(&conf->active_stripes);
1382 r5c_make_stripe_write_out(sh);
1383
1384 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1385 atomic_inc(&conf->r5c_flushing_partial_stripes);
1386 else
1387 atomic_inc(&conf->r5c_flushing_full_stripes);
1388 raid5_release_stripe(sh);
1389}
1390
1391/*
1392 * if num == 0, flush all full stripes
1393 * if num > 0, flush all full stripes. If less than num full stripes are
1394 * flushed, flush some partial stripes until totally num stripes are
1395 * flushed or there is no more cached stripes.
1396 */
1397void r5c_flush_cache(struct r5conf *conf, int num)
1398{
1399 int count;
1400 struct stripe_head *sh, *next;
1401
1402 lockdep_assert_held(&conf->device_lock);
1403 if (!conf->log)
1404 return;
1405
1406 count = 0;
1407 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1408 r5c_flush_stripe(conf, sh);
1409 count++;
1410 }
1411
1412 if (count >= num)
1413 return;
1414 list_for_each_entry_safe(sh, next,
1415 &conf->r5c_partial_stripe_list, lru) {
1416 r5c_flush_stripe(conf, sh);
1417 if (++count >= num)
1418 break;
1419 }
1420}
1421
1422static void r5c_do_reclaim(struct r5conf *conf)
1423{
1424 struct r5l_log *log = conf->log;
1425 struct stripe_head *sh;
1426 int count = 0;
1427 unsigned long flags;
1428 int total_cached;
1429 int stripes_to_flush;
1430 int flushing_partial, flushing_full;
1431
1432 if (!r5c_is_writeback(log))
1433 return;
1434
1435 flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1436 flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1437 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1438 atomic_read(&conf->r5c_cached_full_stripes) -
1439 flushing_full - flushing_partial;
1440
1441 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1442 atomic_read(&conf->empty_inactive_list_nr) > 0)
1443 /*
1444 * if stripe cache pressure high, flush all full stripes and
1445 * some partial stripes
1446 */
1447 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1448 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1449 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1450 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1451 /*
1452 * if stripe cache pressure moderate, or if there is many full
1453 * stripes,flush all full stripes
1454 */
1455 stripes_to_flush = 0;
1456 else
1457 /* no need to flush */
1458 stripes_to_flush = -1;
1459
1460 if (stripes_to_flush >= 0) {
1461 spin_lock_irqsave(&conf->device_lock, flags);
1462 r5c_flush_cache(conf, stripes_to_flush);
1463 spin_unlock_irqrestore(&conf->device_lock, flags);
1464 }
1465
1466 /* if log space is tight, flush stripes on stripe_in_journal_list */
1467 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1468 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1469 spin_lock(&conf->device_lock);
1470 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1471 /*
1472 * stripes on stripe_in_journal_list could be in any
1473 * state of the stripe_cache state machine. In this
1474 * case, we only want to flush stripe on
1475 * r5c_cached_full/partial_stripes. The following
1476 * condition makes sure the stripe is on one of the
1477 * two lists.
1478 */
1479 if (!list_empty(&sh->lru) &&
1480 !test_bit(STRIPE_HANDLE, &sh->state) &&
1481 atomic_read(&sh->count) == 0) {
1482 r5c_flush_stripe(conf, sh);
1483 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1484 break;
1485 }
1486 }
1487 spin_unlock(&conf->device_lock);
1488 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1489 }
1490
1491 if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1492 r5l_run_no_space_stripes(log);
1493
1494 md_wakeup_thread(conf->mddev->thread);
1495}
1496
1497static void r5l_do_reclaim(struct r5l_log *log)
1498{
1499 struct r5conf *conf = log->rdev->mddev->private;
1500 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1501 sector_t reclaimable;
1502 sector_t next_checkpoint;
1503 bool write_super;
1504
1505 spin_lock_irq(&log->io_list_lock);
1506 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1507 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1508 /*
1509 * move proper io_unit to reclaim list. We should not change the order.
1510 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1511 * shouldn't reuse space of an unreclaimable io_unit
1512 */
1513 while (1) {
1514 reclaimable = r5l_reclaimable_space(log);
1515 if (reclaimable >= reclaim_target ||
1516 (list_empty(&log->running_ios) &&
1517 list_empty(&log->io_end_ios) &&
1518 list_empty(&log->flushing_ios) &&
1519 list_empty(&log->finished_ios)))
1520 break;
1521
1522 md_wakeup_thread(log->rdev->mddev->thread);
1523 wait_event_lock_irq(log->iounit_wait,
1524 r5l_reclaimable_space(log) > reclaimable,
1525 log->io_list_lock);
1526 }
1527
1528 next_checkpoint = r5c_calculate_new_cp(conf);
1529 spin_unlock_irq(&log->io_list_lock);
1530
1531 if (reclaimable == 0 || !write_super)
1532 return;
1533
1534 /*
1535 * write_super will flush cache of each raid disk. We must write super
1536 * here, because the log area might be reused soon and we don't want to
1537 * confuse recovery
1538 */
1539 r5l_write_super_and_discard_space(log, next_checkpoint);
1540
1541 mutex_lock(&log->io_mutex);
1542 log->last_checkpoint = next_checkpoint;
1543 r5c_update_log_state(log);
1544 mutex_unlock(&log->io_mutex);
1545
1546 r5l_run_no_space_stripes(log);
1547}
1548
1549static void r5l_reclaim_thread(struct md_thread *thread)
1550{
1551 struct mddev *mddev = thread->mddev;
1552 struct r5conf *conf = mddev->private;
1553 struct r5l_log *log = conf->log;
1554
1555 if (!log)
1556 return;
1557 r5c_do_reclaim(conf);
1558 r5l_do_reclaim(log);
1559}
1560
1561void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1562{
1563 unsigned long target;
1564 unsigned long new = (unsigned long)space; /* overflow in theory */
1565
1566 if (!log)
1567 return;
1568 do {
1569 target = log->reclaim_target;
1570 if (new < target)
1571 return;
1572 } while (cmpxchg(&log->reclaim_target, target, new) != target);
1573 md_wakeup_thread(log->reclaim_thread);
1574}
1575
1576void r5l_quiesce(struct r5l_log *log, int quiesce)
1577{
1578 struct mddev *mddev;
1579
1580 if (quiesce) {
1581 /* make sure r5l_write_super_and_discard_space exits */
1582 mddev = log->rdev->mddev;
1583 wake_up(&mddev->sb_wait);
1584 kthread_park(log->reclaim_thread->tsk);
1585 r5l_wake_reclaim(log, MaxSector);
1586 r5l_do_reclaim(log);
1587 } else
1588 kthread_unpark(log->reclaim_thread->tsk);
1589}
1590
1591bool r5l_log_disk_error(struct r5conf *conf)
1592{
1593 struct r5l_log *log;
1594 bool ret;
1595 /* don't allow write if journal disk is missing */
1596 rcu_read_lock();
1597 log = rcu_dereference(conf->log);
1598
1599 if (!log)
1600 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1601 else
1602 ret = test_bit(Faulty, &log->rdev->flags);
1603 rcu_read_unlock();
1604 return ret;
1605}
1606
1607#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1608
1609struct r5l_recovery_ctx {
1610 struct page *meta_page; /* current meta */
1611 sector_t meta_total_blocks; /* total size of current meta and data */
1612 sector_t pos; /* recovery position */
1613 u64 seq; /* recovery position seq */
1614 int data_parity_stripes; /* number of data_parity stripes */
1615 int data_only_stripes; /* number of data_only stripes */
1616 struct list_head cached_list;
1617
1618 /*
1619 * read ahead page pool (ra_pool)
1620 * in recovery, log is read sequentially. It is not efficient to
1621 * read every page with sync_page_io(). The read ahead page pool
1622 * reads multiple pages with one IO, so further log read can
1623 * just copy data from the pool.
1624 */
1625 struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1626 sector_t pool_offset; /* offset of first page in the pool */
1627 int total_pages; /* total allocated pages */
1628 int valid_pages; /* pages with valid data */
1629 struct bio *ra_bio; /* bio to do the read ahead */
1630};
1631
1632static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1633 struct r5l_recovery_ctx *ctx)
1634{
1635 struct page *page;
1636
1637 ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, &log->bs);
1638 if (!ctx->ra_bio)
1639 return -ENOMEM;
1640
1641 ctx->valid_pages = 0;
1642 ctx->total_pages = 0;
1643 while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1644 page = alloc_page(GFP_KERNEL);
1645
1646 if (!page)
1647 break;
1648 ctx->ra_pool[ctx->total_pages] = page;
1649 ctx->total_pages += 1;
1650 }
1651
1652 if (ctx->total_pages == 0) {
1653 bio_put(ctx->ra_bio);
1654 return -ENOMEM;
1655 }
1656
1657 ctx->pool_offset = 0;
1658 return 0;
1659}
1660
1661static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1662 struct r5l_recovery_ctx *ctx)
1663{
1664 int i;
1665
1666 for (i = 0; i < ctx->total_pages; ++i)
1667 put_page(ctx->ra_pool[i]);
1668 bio_put(ctx->ra_bio);
1669}
1670
1671/*
1672 * fetch ctx->valid_pages pages from offset
1673 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1674 * However, if the offset is close to the end of the journal device,
1675 * ctx->valid_pages could be smaller than ctx->total_pages
1676 */
1677static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1678 struct r5l_recovery_ctx *ctx,
1679 sector_t offset)
1680{
1681 bio_reset(ctx->ra_bio);
1682 bio_set_dev(ctx->ra_bio, log->rdev->bdev);
1683 bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1684 ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1685
1686 ctx->valid_pages = 0;
1687 ctx->pool_offset = offset;
1688
1689 while (ctx->valid_pages < ctx->total_pages) {
1690 bio_add_page(ctx->ra_bio,
1691 ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1692 ctx->valid_pages += 1;
1693
1694 offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1695
1696 if (offset == 0) /* reached end of the device */
1697 break;
1698 }
1699
1700 return submit_bio_wait(ctx->ra_bio);
1701}
1702
1703/*
1704 * try read a page from the read ahead page pool, if the page is not in the
1705 * pool, call r5l_recovery_fetch_ra_pool
1706 */
1707static int r5l_recovery_read_page(struct r5l_log *log,
1708 struct r5l_recovery_ctx *ctx,
1709 struct page *page,
1710 sector_t offset)
1711{
1712 int ret;
1713
1714 if (offset < ctx->pool_offset ||
1715 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1716 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1717 if (ret)
1718 return ret;
1719 }
1720
1721 BUG_ON(offset < ctx->pool_offset ||
1722 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1723
1724 memcpy(page_address(page),
1725 page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1726 BLOCK_SECTOR_SHIFT]),
1727 PAGE_SIZE);
1728 return 0;
1729}
1730
1731static int r5l_recovery_read_meta_block(struct r5l_log *log,
1732 struct r5l_recovery_ctx *ctx)
1733{
1734 struct page *page = ctx->meta_page;
1735 struct r5l_meta_block *mb;
1736 u32 crc, stored_crc;
1737 int ret;
1738
1739 ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1740 if (ret != 0)
1741 return ret;
1742
1743 mb = page_address(page);
1744 stored_crc = le32_to_cpu(mb->checksum);
1745 mb->checksum = 0;
1746
1747 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1748 le64_to_cpu(mb->seq) != ctx->seq ||
1749 mb->version != R5LOG_VERSION ||
1750 le64_to_cpu(mb->position) != ctx->pos)
1751 return -EINVAL;
1752
1753 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1754 if (stored_crc != crc)
1755 return -EINVAL;
1756
1757 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1758 return -EINVAL;
1759
1760 ctx->meta_total_blocks = BLOCK_SECTORS;
1761
1762 return 0;
1763}
1764
1765static void
1766r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1767 struct page *page,
1768 sector_t pos, u64 seq)
1769{
1770 struct r5l_meta_block *mb;
1771
1772 mb = page_address(page);
1773 clear_page(mb);
1774 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1775 mb->version = R5LOG_VERSION;
1776 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1777 mb->seq = cpu_to_le64(seq);
1778 mb->position = cpu_to_le64(pos);
1779}
1780
1781static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1782 u64 seq)
1783{
1784 struct page *page;
1785 struct r5l_meta_block *mb;
1786
1787 page = alloc_page(GFP_KERNEL);
1788 if (!page)
1789 return -ENOMEM;
1790 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1791 mb = page_address(page);
1792 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1793 mb, PAGE_SIZE));
1794 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1795 REQ_SYNC | REQ_FUA, false)) {
1796 __free_page(page);
1797 return -EIO;
1798 }
1799 __free_page(page);
1800 return 0;
1801}
1802
1803/*
1804 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1805 * to mark valid (potentially not flushed) data in the journal.
1806 *
1807 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1808 * so there should not be any mismatch here.
1809 */
1810static void r5l_recovery_load_data(struct r5l_log *log,
1811 struct stripe_head *sh,
1812 struct r5l_recovery_ctx *ctx,
1813 struct r5l_payload_data_parity *payload,
1814 sector_t log_offset)
1815{
1816 struct mddev *mddev = log->rdev->mddev;
1817 struct r5conf *conf = mddev->private;
1818 int dd_idx;
1819
1820 raid5_compute_sector(conf,
1821 le64_to_cpu(payload->location), 0,
1822 &dd_idx, sh);
1823 r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1824 sh->dev[dd_idx].log_checksum =
1825 le32_to_cpu(payload->checksum[0]);
1826 ctx->meta_total_blocks += BLOCK_SECTORS;
1827
1828 set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1829 set_bit(STRIPE_R5C_CACHING, &sh->state);
1830}
1831
1832static void r5l_recovery_load_parity(struct r5l_log *log,
1833 struct stripe_head *sh,
1834 struct r5l_recovery_ctx *ctx,
1835 struct r5l_payload_data_parity *payload,
1836 sector_t log_offset)
1837{
1838 struct mddev *mddev = log->rdev->mddev;
1839 struct r5conf *conf = mddev->private;
1840
1841 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1842 r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1843 sh->dev[sh->pd_idx].log_checksum =
1844 le32_to_cpu(payload->checksum[0]);
1845 set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1846
1847 if (sh->qd_idx >= 0) {
1848 r5l_recovery_read_page(
1849 log, ctx, sh->dev[sh->qd_idx].page,
1850 r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1851 sh->dev[sh->qd_idx].log_checksum =
1852 le32_to_cpu(payload->checksum[1]);
1853 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1854 }
1855 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1856}
1857
1858static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1859{
1860 int i;
1861
1862 sh->state = 0;
1863 sh->log_start = MaxSector;
1864 for (i = sh->disks; i--; )
1865 sh->dev[i].flags = 0;
1866}
1867
1868static void
1869r5l_recovery_replay_one_stripe(struct r5conf *conf,
1870 struct stripe_head *sh,
1871 struct r5l_recovery_ctx *ctx)
1872{
1873 struct md_rdev *rdev, *rrdev;
1874 int disk_index;
1875 int data_count = 0;
1876
1877 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1878 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1879 continue;
1880 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1881 continue;
1882 data_count++;
1883 }
1884
1885 /*
1886 * stripes that only have parity must have been flushed
1887 * before the crash that we are now recovering from, so
1888 * there is nothing more to recovery.
1889 */
1890 if (data_count == 0)
1891 goto out;
1892
1893 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1894 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1895 continue;
1896
1897 /* in case device is broken */
1898 rcu_read_lock();
1899 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1900 if (rdev) {
1901 atomic_inc(&rdev->nr_pending);
1902 rcu_read_unlock();
1903 sync_page_io(rdev, sh->sector, PAGE_SIZE,
1904 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1905 false);
1906 rdev_dec_pending(rdev, rdev->mddev);
1907 rcu_read_lock();
1908 }
1909 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1910 if (rrdev) {
1911 atomic_inc(&rrdev->nr_pending);
1912 rcu_read_unlock();
1913 sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1914 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1915 false);
1916 rdev_dec_pending(rrdev, rrdev->mddev);
1917 rcu_read_lock();
1918 }
1919 rcu_read_unlock();
1920 }
1921 ctx->data_parity_stripes++;
1922out:
1923 r5l_recovery_reset_stripe(sh);
1924}
1925
1926static struct stripe_head *
1927r5c_recovery_alloc_stripe(
1928 struct r5conf *conf,
1929 sector_t stripe_sect,
1930 int noblock)
1931{
1932 struct stripe_head *sh;
1933
1934 sh = raid5_get_active_stripe(conf, stripe_sect, 0, noblock, 0);
1935 if (!sh)
1936 return NULL; /* no more stripe available */
1937
1938 r5l_recovery_reset_stripe(sh);
1939
1940 return sh;
1941}
1942
1943static struct stripe_head *
1944r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1945{
1946 struct stripe_head *sh;
1947
1948 list_for_each_entry(sh, list, lru)
1949 if (sh->sector == sect)
1950 return sh;
1951 return NULL;
1952}
1953
1954static void
1955r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1956 struct r5l_recovery_ctx *ctx)
1957{
1958 struct stripe_head *sh, *next;
1959
1960 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1961 r5l_recovery_reset_stripe(sh);
1962 list_del_init(&sh->lru);
1963 raid5_release_stripe(sh);
1964 }
1965}
1966
1967static void
1968r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1969 struct r5l_recovery_ctx *ctx)
1970{
1971 struct stripe_head *sh, *next;
1972
1973 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1974 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1975 r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1976 list_del_init(&sh->lru);
1977 raid5_release_stripe(sh);
1978 }
1979}
1980
1981/* if matches return 0; otherwise return -EINVAL */
1982static int
1983r5l_recovery_verify_data_checksum(struct r5l_log *log,
1984 struct r5l_recovery_ctx *ctx,
1985 struct page *page,
1986 sector_t log_offset, __le32 log_checksum)
1987{
1988 void *addr;
1989 u32 checksum;
1990
1991 r5l_recovery_read_page(log, ctx, page, log_offset);
1992 addr = kmap_atomic(page);
1993 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1994 kunmap_atomic(addr);
1995 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1996}
1997
1998/*
1999 * before loading data to stripe cache, we need verify checksum for all data,
2000 * if there is mismatch for any data page, we drop all data in the mata block
2001 */
2002static int
2003r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
2004 struct r5l_recovery_ctx *ctx)
2005{
2006 struct mddev *mddev = log->rdev->mddev;
2007 struct r5conf *conf = mddev->private;
2008 struct r5l_meta_block *mb = page_address(ctx->meta_page);
2009 sector_t mb_offset = sizeof(struct r5l_meta_block);
2010 sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2011 struct page *page;
2012 struct r5l_payload_data_parity *payload;
2013 struct r5l_payload_flush *payload_flush;
2014
2015 page = alloc_page(GFP_KERNEL);
2016 if (!page)
2017 return -ENOMEM;
2018
2019 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2020 payload = (void *)mb + mb_offset;
2021 payload_flush = (void *)mb + mb_offset;
2022
2023 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2024 if (r5l_recovery_verify_data_checksum(
2025 log, ctx, page, log_offset,
2026 payload->checksum[0]) < 0)
2027 goto mismatch;
2028 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2029 if (r5l_recovery_verify_data_checksum(
2030 log, ctx, page, log_offset,
2031 payload->checksum[0]) < 0)
2032 goto mismatch;
2033 if (conf->max_degraded == 2 && /* q for RAID 6 */
2034 r5l_recovery_verify_data_checksum(
2035 log, ctx, page,
2036 r5l_ring_add(log, log_offset,
2037 BLOCK_SECTORS),
2038 payload->checksum[1]) < 0)
2039 goto mismatch;
2040 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2041 /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2042 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2043 goto mismatch;
2044
2045 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2046 mb_offset += sizeof(struct r5l_payload_flush) +
2047 le32_to_cpu(payload_flush->size);
2048 } else {
2049 /* DATA or PARITY payload */
2050 log_offset = r5l_ring_add(log, log_offset,
2051 le32_to_cpu(payload->size));
2052 mb_offset += sizeof(struct r5l_payload_data_parity) +
2053 sizeof(__le32) *
2054 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2055 }
2056
2057 }
2058
2059 put_page(page);
2060 return 0;
2061
2062mismatch:
2063 put_page(page);
2064 return -EINVAL;
2065}
2066
2067/*
2068 * Analyze all data/parity pages in one meta block
2069 * Returns:
2070 * 0 for success
2071 * -EINVAL for unknown playload type
2072 * -EAGAIN for checksum mismatch of data page
2073 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2074 */
2075static int
2076r5c_recovery_analyze_meta_block(struct r5l_log *log,
2077 struct r5l_recovery_ctx *ctx,
2078 struct list_head *cached_stripe_list)
2079{
2080 struct mddev *mddev = log->rdev->mddev;
2081 struct r5conf *conf = mddev->private;
2082 struct r5l_meta_block *mb;
2083 struct r5l_payload_data_parity *payload;
2084 struct r5l_payload_flush *payload_flush;
2085 int mb_offset;
2086 sector_t log_offset;
2087 sector_t stripe_sect;
2088 struct stripe_head *sh;
2089 int ret;
2090
2091 /*
2092 * for mismatch in data blocks, we will drop all data in this mb, but
2093 * we will still read next mb for other data with FLUSH flag, as
2094 * io_unit could finish out of order.
2095 */
2096 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2097 if (ret == -EINVAL)
2098 return -EAGAIN;
2099 else if (ret)
2100 return ret; /* -ENOMEM duo to alloc_page() failed */
2101
2102 mb = page_address(ctx->meta_page);
2103 mb_offset = sizeof(struct r5l_meta_block);
2104 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2105
2106 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2107 int dd;
2108
2109 payload = (void *)mb + mb_offset;
2110 payload_flush = (void *)mb + mb_offset;
2111
2112 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2113 int i, count;
2114
2115 count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2116 for (i = 0; i < count; ++i) {
2117 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2118 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2119 stripe_sect);
2120 if (sh) {
2121 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2122 r5l_recovery_reset_stripe(sh);
2123 list_del_init(&sh->lru);
2124 raid5_release_stripe(sh);
2125 }
2126 }
2127
2128 mb_offset += sizeof(struct r5l_payload_flush) +
2129 le32_to_cpu(payload_flush->size);
2130 continue;
2131 }
2132
2133 /* DATA or PARITY payload */
2134 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2135 raid5_compute_sector(
2136 conf, le64_to_cpu(payload->location), 0, &dd,
2137 NULL)
2138 : le64_to_cpu(payload->location);
2139
2140 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2141 stripe_sect);
2142
2143 if (!sh) {
2144 sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2145 /*
2146 * cannot get stripe from raid5_get_active_stripe
2147 * try replay some stripes
2148 */
2149 if (!sh) {
2150 r5c_recovery_replay_stripes(
2151 cached_stripe_list, ctx);
2152 sh = r5c_recovery_alloc_stripe(
2153 conf, stripe_sect, 1);
2154 }
2155 if (!sh) {
2156 int new_size = conf->min_nr_stripes * 2;
2157 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2158 mdname(mddev),
2159 new_size);
2160 ret = raid5_set_cache_size(mddev, new_size);
2161 if (conf->min_nr_stripes <= new_size / 2) {
2162 pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2163 mdname(mddev),
2164 ret,
2165 new_size,
2166 conf->min_nr_stripes,
2167 conf->max_nr_stripes);
2168 return -ENOMEM;
2169 }
2170 sh = r5c_recovery_alloc_stripe(
2171 conf, stripe_sect, 0);
2172 }
2173 if (!sh) {
2174 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2175 mdname(mddev));
2176 return -ENOMEM;
2177 }
2178 list_add_tail(&sh->lru, cached_stripe_list);
2179 }
2180
2181 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2182 if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2183 test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2184 r5l_recovery_replay_one_stripe(conf, sh, ctx);
2185 list_move_tail(&sh->lru, cached_stripe_list);
2186 }
2187 r5l_recovery_load_data(log, sh, ctx, payload,
2188 log_offset);
2189 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2190 r5l_recovery_load_parity(log, sh, ctx, payload,
2191 log_offset);
2192 else
2193 return -EINVAL;
2194
2195 log_offset = r5l_ring_add(log, log_offset,
2196 le32_to_cpu(payload->size));
2197
2198 mb_offset += sizeof(struct r5l_payload_data_parity) +
2199 sizeof(__le32) *
2200 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2201 }
2202
2203 return 0;
2204}
2205
2206/*
2207 * Load the stripe into cache. The stripe will be written out later by
2208 * the stripe cache state machine.
2209 */
2210static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2211 struct stripe_head *sh)
2212{
2213 struct r5dev *dev;
2214 int i;
2215
2216 for (i = sh->disks; i--; ) {
2217 dev = sh->dev + i;
2218 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2219 set_bit(R5_InJournal, &dev->flags);
2220 set_bit(R5_UPTODATE, &dev->flags);
2221 }
2222 }
2223}
2224
2225/*
2226 * Scan through the log for all to-be-flushed data
2227 *
2228 * For stripes with data and parity, namely Data-Parity stripe
2229 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2230 *
2231 * For stripes with only data, namely Data-Only stripe
2232 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2233 *
2234 * For a stripe, if we see data after parity, we should discard all previous
2235 * data and parity for this stripe, as these data are already flushed to
2236 * the array.
2237 *
2238 * At the end of the scan, we return the new journal_tail, which points to
2239 * first data-only stripe on the journal device, or next invalid meta block.
2240 */
2241static int r5c_recovery_flush_log(struct r5l_log *log,
2242 struct r5l_recovery_ctx *ctx)
2243{
2244 struct stripe_head *sh;
2245 int ret = 0;
2246
2247 /* scan through the log */
2248 while (1) {
2249 if (r5l_recovery_read_meta_block(log, ctx))
2250 break;
2251
2252 ret = r5c_recovery_analyze_meta_block(log, ctx,
2253 &ctx->cached_list);
2254 /*
2255 * -EAGAIN means mismatch in data block, in this case, we still
2256 * try scan the next metablock
2257 */
2258 if (ret && ret != -EAGAIN)
2259 break; /* ret == -EINVAL or -ENOMEM */
2260 ctx->seq++;
2261 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2262 }
2263
2264 if (ret == -ENOMEM) {
2265 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2266 return ret;
2267 }
2268
2269 /* replay data-parity stripes */
2270 r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2271
2272 /* load data-only stripes to stripe cache */
2273 list_for_each_entry(sh, &ctx->cached_list, lru) {
2274 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2275 r5c_recovery_load_one_stripe(log, sh);
2276 ctx->data_only_stripes++;
2277 }
2278
2279 return 0;
2280}
2281
2282/*
2283 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2284 * log will start here. but we can't let superblock point to last valid
2285 * meta block. The log might looks like:
2286 * | meta 1| meta 2| meta 3|
2287 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2288 * superblock points to meta 1, we write a new valid meta 2n. if crash
2289 * happens again, new recovery will start from meta 1. Since meta 2n is
2290 * valid now, recovery will think meta 3 is valid, which is wrong.
2291 * The solution is we create a new meta in meta2 with its seq == meta
2292 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2293 * will not think meta 3 is a valid meta, because its seq doesn't match
2294 */
2295
2296/*
2297 * Before recovery, the log looks like the following
2298 *
2299 * ---------------------------------------------
2300 * | valid log | invalid log |
2301 * ---------------------------------------------
2302 * ^
2303 * |- log->last_checkpoint
2304 * |- log->last_cp_seq
2305 *
2306 * Now we scan through the log until we see invalid entry
2307 *
2308 * ---------------------------------------------
2309 * | valid log | invalid log |
2310 * ---------------------------------------------
2311 * ^ ^
2312 * |- log->last_checkpoint |- ctx->pos
2313 * |- log->last_cp_seq |- ctx->seq
2314 *
2315 * From this point, we need to increase seq number by 10 to avoid
2316 * confusing next recovery.
2317 *
2318 * ---------------------------------------------
2319 * | valid log | invalid log |
2320 * ---------------------------------------------
2321 * ^ ^
2322 * |- log->last_checkpoint |- ctx->pos+1
2323 * |- log->last_cp_seq |- ctx->seq+10001
2324 *
2325 * However, it is not safe to start the state machine yet, because data only
2326 * parities are not yet secured in RAID. To save these data only parities, we
2327 * rewrite them from seq+11.
2328 *
2329 * -----------------------------------------------------------------
2330 * | valid log | data only stripes | invalid log |
2331 * -----------------------------------------------------------------
2332 * ^ ^
2333 * |- log->last_checkpoint |- ctx->pos+n
2334 * |- log->last_cp_seq |- ctx->seq+10000+n
2335 *
2336 * If failure happens again during this process, the recovery can safe start
2337 * again from log->last_checkpoint.
2338 *
2339 * Once data only stripes are rewritten to journal, we move log_tail
2340 *
2341 * -----------------------------------------------------------------
2342 * | old log | data only stripes | invalid log |
2343 * -----------------------------------------------------------------
2344 * ^ ^
2345 * |- log->last_checkpoint |- ctx->pos+n
2346 * |- log->last_cp_seq |- ctx->seq+10000+n
2347 *
2348 * Then we can safely start the state machine. If failure happens from this
2349 * point on, the recovery will start from new log->last_checkpoint.
2350 */
2351static int
2352r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2353 struct r5l_recovery_ctx *ctx)
2354{
2355 struct stripe_head *sh;
2356 struct mddev *mddev = log->rdev->mddev;
2357 struct page *page;
2358 sector_t next_checkpoint = MaxSector;
2359
2360 page = alloc_page(GFP_KERNEL);
2361 if (!page) {
2362 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2363 mdname(mddev));
2364 return -ENOMEM;
2365 }
2366
2367 WARN_ON(list_empty(&ctx->cached_list));
2368
2369 list_for_each_entry(sh, &ctx->cached_list, lru) {
2370 struct r5l_meta_block *mb;
2371 int i;
2372 int offset;
2373 sector_t write_pos;
2374
2375 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2376 r5l_recovery_create_empty_meta_block(log, page,
2377 ctx->pos, ctx->seq);
2378 mb = page_address(page);
2379 offset = le32_to_cpu(mb->meta_size);
2380 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2381
2382 for (i = sh->disks; i--; ) {
2383 struct r5dev *dev = &sh->dev[i];
2384 struct r5l_payload_data_parity *payload;
2385 void *addr;
2386
2387 if (test_bit(R5_InJournal, &dev->flags)) {
2388 payload = (void *)mb + offset;
2389 payload->header.type = cpu_to_le16(
2390 R5LOG_PAYLOAD_DATA);
2391 payload->size = cpu_to_le32(BLOCK_SECTORS);
2392 payload->location = cpu_to_le64(
2393 raid5_compute_blocknr(sh, i, 0));
2394 addr = kmap_atomic(dev->page);
2395 payload->checksum[0] = cpu_to_le32(
2396 crc32c_le(log->uuid_checksum, addr,
2397 PAGE_SIZE));
2398 kunmap_atomic(addr);
2399 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2400 dev->page, REQ_OP_WRITE, 0, false);
2401 write_pos = r5l_ring_add(log, write_pos,
2402 BLOCK_SECTORS);
2403 offset += sizeof(__le32) +
2404 sizeof(struct r5l_payload_data_parity);
2405
2406 }
2407 }
2408 mb->meta_size = cpu_to_le32(offset);
2409 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2410 mb, PAGE_SIZE));
2411 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2412 REQ_OP_WRITE, REQ_SYNC | REQ_FUA, false);
2413 sh->log_start = ctx->pos;
2414 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2415 atomic_inc(&log->stripe_in_journal_count);
2416 ctx->pos = write_pos;
2417 ctx->seq += 1;
2418 next_checkpoint = sh->log_start;
2419 }
2420 log->next_checkpoint = next_checkpoint;
2421 __free_page(page);
2422 return 0;
2423}
2424
2425static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2426 struct r5l_recovery_ctx *ctx)
2427{
2428 struct mddev *mddev = log->rdev->mddev;
2429 struct r5conf *conf = mddev->private;
2430 struct stripe_head *sh, *next;
2431 bool cleared_pending = false;
2432
2433 if (ctx->data_only_stripes == 0)
2434 return;
2435
2436 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2437 cleared_pending = true;
2438 clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2439 }
2440 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2441
2442 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2443 r5c_make_stripe_write_out(sh);
2444 set_bit(STRIPE_HANDLE, &sh->state);
2445 list_del_init(&sh->lru);
2446 raid5_release_stripe(sh);
2447 }
2448
2449 /* reuse conf->wait_for_quiescent in recovery */
2450 wait_event(conf->wait_for_quiescent,
2451 atomic_read(&conf->active_stripes) == 0);
2452
2453 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2454 if (cleared_pending)
2455 set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2456}
2457
2458static int r5l_recovery_log(struct r5l_log *log)
2459{
2460 struct mddev *mddev = log->rdev->mddev;
2461 struct r5l_recovery_ctx *ctx;
2462 int ret;
2463 sector_t pos;
2464
2465 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2466 if (!ctx)
2467 return -ENOMEM;
2468
2469 ctx->pos = log->last_checkpoint;
2470 ctx->seq = log->last_cp_seq;
2471 INIT_LIST_HEAD(&ctx->cached_list);
2472 ctx->meta_page = alloc_page(GFP_KERNEL);
2473
2474 if (!ctx->meta_page) {
2475 ret = -ENOMEM;
2476 goto meta_page;
2477 }
2478
2479 if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2480 ret = -ENOMEM;
2481 goto ra_pool;
2482 }
2483
2484 ret = r5c_recovery_flush_log(log, ctx);
2485
2486 if (ret)
2487 goto error;
2488
2489 pos = ctx->pos;
2490 ctx->seq += 10000;
2491
2492 if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2493 pr_info("md/raid:%s: starting from clean shutdown\n",
2494 mdname(mddev));
2495 else
2496 pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2497 mdname(mddev), ctx->data_only_stripes,
2498 ctx->data_parity_stripes);
2499
2500 if (ctx->data_only_stripes == 0) {
2501 log->next_checkpoint = ctx->pos;
2502 r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2503 ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2504 } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2505 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2506 mdname(mddev));
2507 ret = -EIO;
2508 goto error;
2509 }
2510
2511 log->log_start = ctx->pos;
2512 log->seq = ctx->seq;
2513 log->last_checkpoint = pos;
2514 r5l_write_super(log, pos);
2515
2516 r5c_recovery_flush_data_only_stripes(log, ctx);
2517 ret = 0;
2518error:
2519 r5l_recovery_free_ra_pool(log, ctx);
2520ra_pool:
2521 __free_page(ctx->meta_page);
2522meta_page:
2523 kfree(ctx);
2524 return ret;
2525}
2526
2527static void r5l_write_super(struct r5l_log *log, sector_t cp)
2528{
2529 struct mddev *mddev = log->rdev->mddev;
2530
2531 log->rdev->journal_tail = cp;
2532 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2533}
2534
2535static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2536{
2537 struct r5conf *conf;
2538 int ret;
2539
2540 spin_lock(&mddev->lock);
2541 conf = mddev->private;
2542 if (!conf || !conf->log) {
2543 spin_unlock(&mddev->lock);
2544 return 0;
2545 }
2546
2547 switch (conf->log->r5c_journal_mode) {
2548 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2549 ret = snprintf(
2550 page, PAGE_SIZE, "[%s] %s\n",
2551 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2552 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2553 break;
2554 case R5C_JOURNAL_MODE_WRITE_BACK:
2555 ret = snprintf(
2556 page, PAGE_SIZE, "%s [%s]\n",
2557 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2558 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2559 break;
2560 default:
2561 ret = 0;
2562 }
2563 spin_unlock(&mddev->lock);
2564 return ret;
2565}
2566
2567/*
2568 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2569 *
2570 * @mode as defined in 'enum r5c_journal_mode'.
2571 *
2572 */
2573int r5c_journal_mode_set(struct mddev *mddev, int mode)
2574{
2575 struct r5conf *conf;
2576
2577 if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2578 mode > R5C_JOURNAL_MODE_WRITE_BACK)
2579 return -EINVAL;
2580
2581 conf = mddev->private;
2582 if (!conf || !conf->log)
2583 return -ENODEV;
2584
2585 if (raid5_calc_degraded(conf) > 0 &&
2586 mode == R5C_JOURNAL_MODE_WRITE_BACK)
2587 return -EINVAL;
2588
2589 mddev_suspend(mddev);
2590 conf->log->r5c_journal_mode = mode;
2591 mddev_resume(mddev);
2592
2593 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2594 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2595 return 0;
2596}
2597EXPORT_SYMBOL(r5c_journal_mode_set);
2598
2599static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2600 const char *page, size_t length)
2601{
2602 int mode = ARRAY_SIZE(r5c_journal_mode_str);
2603 size_t len = length;
2604 int ret;
2605
2606 if (len < 2)
2607 return -EINVAL;
2608
2609 if (page[len - 1] == '\n')
2610 len--;
2611
2612 while (mode--)
2613 if (strlen(r5c_journal_mode_str[mode]) == len &&
2614 !strncmp(page, r5c_journal_mode_str[mode], len))
2615 break;
2616 ret = mddev_lock(mddev);
2617 if (ret)
2618 return ret;
2619 ret = r5c_journal_mode_set(mddev, mode);
2620 mddev_unlock(mddev);
2621 return ret ?: length;
2622}
2623
2624struct md_sysfs_entry
2625r5c_journal_mode = __ATTR(journal_mode, 0644,
2626 r5c_journal_mode_show, r5c_journal_mode_store);
2627
2628/*
2629 * Try handle write operation in caching phase. This function should only
2630 * be called in write-back mode.
2631 *
2632 * If all outstanding writes can be handled in caching phase, returns 0
2633 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2634 * and returns -EAGAIN
2635 */
2636int r5c_try_caching_write(struct r5conf *conf,
2637 struct stripe_head *sh,
2638 struct stripe_head_state *s,
2639 int disks)
2640{
2641 struct r5l_log *log = conf->log;
2642 int i;
2643 struct r5dev *dev;
2644 int to_cache = 0;
2645 void **pslot;
2646 sector_t tree_index;
2647 int ret;
2648 uintptr_t refcount;
2649
2650 BUG_ON(!r5c_is_writeback(log));
2651
2652 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2653 /*
2654 * There are two different scenarios here:
2655 * 1. The stripe has some data cached, and it is sent to
2656 * write-out phase for reclaim
2657 * 2. The stripe is clean, and this is the first write
2658 *
2659 * For 1, return -EAGAIN, so we continue with
2660 * handle_stripe_dirtying().
2661 *
2662 * For 2, set STRIPE_R5C_CACHING and continue with caching
2663 * write.
2664 */
2665
2666 /* case 1: anything injournal or anything in written */
2667 if (s->injournal > 0 || s->written > 0)
2668 return -EAGAIN;
2669 /* case 2 */
2670 set_bit(STRIPE_R5C_CACHING, &sh->state);
2671 }
2672
2673 /*
2674 * When run in degraded mode, array is set to write-through mode.
2675 * This check helps drain pending write safely in the transition to
2676 * write-through mode.
2677 *
2678 * When a stripe is syncing, the write is also handled in write
2679 * through mode.
2680 */
2681 if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2682 r5c_make_stripe_write_out(sh);
2683 return -EAGAIN;
2684 }
2685
2686 for (i = disks; i--; ) {
2687 dev = &sh->dev[i];
2688 /* if non-overwrite, use writing-out phase */
2689 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2690 !test_bit(R5_InJournal, &dev->flags)) {
2691 r5c_make_stripe_write_out(sh);
2692 return -EAGAIN;
2693 }
2694 }
2695
2696 /* if the stripe is not counted in big_stripe_tree, add it now */
2697 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2698 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2699 tree_index = r5c_tree_index(conf, sh->sector);
2700 spin_lock(&log->tree_lock);
2701 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2702 tree_index);
2703 if (pslot) {
2704 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2705 pslot, &log->tree_lock) >>
2706 R5C_RADIX_COUNT_SHIFT;
2707 radix_tree_replace_slot(
2708 &log->big_stripe_tree, pslot,
2709 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2710 } else {
2711 /*
2712 * this radix_tree_insert can fail safely, so no
2713 * need to call radix_tree_preload()
2714 */
2715 ret = radix_tree_insert(
2716 &log->big_stripe_tree, tree_index,
2717 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2718 if (ret) {
2719 spin_unlock(&log->tree_lock);
2720 r5c_make_stripe_write_out(sh);
2721 return -EAGAIN;
2722 }
2723 }
2724 spin_unlock(&log->tree_lock);
2725
2726 /*
2727 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2728 * counted in the radix tree
2729 */
2730 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2731 atomic_inc(&conf->r5c_cached_partial_stripes);
2732 }
2733
2734 for (i = disks; i--; ) {
2735 dev = &sh->dev[i];
2736 if (dev->towrite) {
2737 set_bit(R5_Wantwrite, &dev->flags);
2738 set_bit(R5_Wantdrain, &dev->flags);
2739 set_bit(R5_LOCKED, &dev->flags);
2740 to_cache++;
2741 }
2742 }
2743
2744 if (to_cache) {
2745 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2746 /*
2747 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2748 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2749 * r5c_handle_data_cached()
2750 */
2751 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2752 }
2753
2754 return 0;
2755}
2756
2757/*
2758 * free extra pages (orig_page) we allocated for prexor
2759 */
2760void r5c_release_extra_page(struct stripe_head *sh)
2761{
2762 struct r5conf *conf = sh->raid_conf;
2763 int i;
2764 bool using_disk_info_extra_page;
2765
2766 using_disk_info_extra_page =
2767 sh->dev[0].orig_page == conf->disks[0].extra_page;
2768
2769 for (i = sh->disks; i--; )
2770 if (sh->dev[i].page != sh->dev[i].orig_page) {
2771 struct page *p = sh->dev[i].orig_page;
2772
2773 sh->dev[i].orig_page = sh->dev[i].page;
2774 clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2775
2776 if (!using_disk_info_extra_page)
2777 put_page(p);
2778 }
2779
2780 if (using_disk_info_extra_page) {
2781 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2782 md_wakeup_thread(conf->mddev->thread);
2783 }
2784}
2785
2786void r5c_use_extra_page(struct stripe_head *sh)
2787{
2788 struct r5conf *conf = sh->raid_conf;
2789 int i;
2790 struct r5dev *dev;
2791
2792 for (i = sh->disks; i--; ) {
2793 dev = &sh->dev[i];
2794 if (dev->orig_page != dev->page)
2795 put_page(dev->orig_page);
2796 dev->orig_page = conf->disks[i].extra_page;
2797 }
2798}
2799
2800/*
2801 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2802 * stripe is committed to RAID disks.
2803 */
2804void r5c_finish_stripe_write_out(struct r5conf *conf,
2805 struct stripe_head *sh,
2806 struct stripe_head_state *s)
2807{
2808 struct r5l_log *log = conf->log;
2809 int i;
2810 int do_wakeup = 0;
2811 sector_t tree_index;
2812 void **pslot;
2813 uintptr_t refcount;
2814
2815 if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2816 return;
2817
2818 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2819 clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2820
2821 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2822 return;
2823
2824 for (i = sh->disks; i--; ) {
2825 clear_bit(R5_InJournal, &sh->dev[i].flags);
2826 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2827 do_wakeup = 1;
2828 }
2829
2830 /*
2831 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2832 * We updated R5_InJournal, so we also update s->injournal.
2833 */
2834 s->injournal = 0;
2835
2836 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2837 if (atomic_dec_and_test(&conf->pending_full_writes))
2838 md_wakeup_thread(conf->mddev->thread);
2839
2840 if (do_wakeup)
2841 wake_up(&conf->wait_for_overlap);
2842
2843 spin_lock_irq(&log->stripe_in_journal_lock);
2844 list_del_init(&sh->r5c);
2845 spin_unlock_irq(&log->stripe_in_journal_lock);
2846 sh->log_start = MaxSector;
2847
2848 atomic_dec(&log->stripe_in_journal_count);
2849 r5c_update_log_state(log);
2850
2851 /* stop counting this stripe in big_stripe_tree */
2852 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2853 test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2854 tree_index = r5c_tree_index(conf, sh->sector);
2855 spin_lock(&log->tree_lock);
2856 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2857 tree_index);
2858 BUG_ON(pslot == NULL);
2859 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2860 pslot, &log->tree_lock) >>
2861 R5C_RADIX_COUNT_SHIFT;
2862 if (refcount == 1)
2863 radix_tree_delete(&log->big_stripe_tree, tree_index);
2864 else
2865 radix_tree_replace_slot(
2866 &log->big_stripe_tree, pslot,
2867 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2868 spin_unlock(&log->tree_lock);
2869 }
2870
2871 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2872 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2873 atomic_dec(&conf->r5c_flushing_partial_stripes);
2874 atomic_dec(&conf->r5c_cached_partial_stripes);
2875 }
2876
2877 if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2878 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2879 atomic_dec(&conf->r5c_flushing_full_stripes);
2880 atomic_dec(&conf->r5c_cached_full_stripes);
2881 }
2882
2883 r5l_append_flush_payload(log, sh->sector);
2884 /* stripe is flused to raid disks, we can do resync now */
2885 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2886 set_bit(STRIPE_HANDLE, &sh->state);
2887}
2888
2889int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2890{
2891 struct r5conf *conf = sh->raid_conf;
2892 int pages = 0;
2893 int reserve;
2894 int i;
2895 int ret = 0;
2896
2897 BUG_ON(!log);
2898
2899 for (i = 0; i < sh->disks; i++) {
2900 void *addr;
2901
2902 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2903 continue;
2904 addr = kmap_atomic(sh->dev[i].page);
2905 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2906 addr, PAGE_SIZE);
2907 kunmap_atomic(addr);
2908 pages++;
2909 }
2910 WARN_ON(pages == 0);
2911
2912 /*
2913 * The stripe must enter state machine again to call endio, so
2914 * don't delay.
2915 */
2916 clear_bit(STRIPE_DELAYED, &sh->state);
2917 atomic_inc(&sh->count);
2918
2919 mutex_lock(&log->io_mutex);
2920 /* meta + data */
2921 reserve = (1 + pages) << (PAGE_SHIFT - 9);
2922
2923 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2924 sh->log_start == MaxSector)
2925 r5l_add_no_space_stripe(log, sh);
2926 else if (!r5l_has_free_space(log, reserve)) {
2927 if (sh->log_start == log->last_checkpoint)
2928 BUG();
2929 else
2930 r5l_add_no_space_stripe(log, sh);
2931 } else {
2932 ret = r5l_log_stripe(log, sh, pages, 0);
2933 if (ret) {
2934 spin_lock_irq(&log->io_list_lock);
2935 list_add_tail(&sh->log_list, &log->no_mem_stripes);
2936 spin_unlock_irq(&log->io_list_lock);
2937 }
2938 }
2939
2940 mutex_unlock(&log->io_mutex);
2941 return 0;
2942}
2943
2944/* check whether this big stripe is in write back cache. */
2945bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2946{
2947 struct r5l_log *log = conf->log;
2948 sector_t tree_index;
2949 void *slot;
2950
2951 if (!log)
2952 return false;
2953
2954 WARN_ON_ONCE(!rcu_read_lock_held());
2955 tree_index = r5c_tree_index(conf, sect);
2956 slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2957 return slot != NULL;
2958}
2959
2960static int r5l_load_log(struct r5l_log *log)
2961{
2962 struct md_rdev *rdev = log->rdev;
2963 struct page *page;
2964 struct r5l_meta_block *mb;
2965 sector_t cp = log->rdev->journal_tail;
2966 u32 stored_crc, expected_crc;
2967 bool create_super = false;
2968 int ret = 0;
2969
2970 /* Make sure it's valid */
2971 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2972 cp = 0;
2973 page = alloc_page(GFP_KERNEL);
2974 if (!page)
2975 return -ENOMEM;
2976
2977 if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2978 ret = -EIO;
2979 goto ioerr;
2980 }
2981 mb = page_address(page);
2982
2983 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2984 mb->version != R5LOG_VERSION) {
2985 create_super = true;
2986 goto create;
2987 }
2988 stored_crc = le32_to_cpu(mb->checksum);
2989 mb->checksum = 0;
2990 expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2991 if (stored_crc != expected_crc) {
2992 create_super = true;
2993 goto create;
2994 }
2995 if (le64_to_cpu(mb->position) != cp) {
2996 create_super = true;
2997 goto create;
2998 }
2999create:
3000 if (create_super) {
3001 log->last_cp_seq = prandom_u32();
3002 cp = 0;
3003 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
3004 /*
3005 * Make sure super points to correct address. Log might have
3006 * data very soon. If super hasn't correct log tail address,
3007 * recovery can't find the log
3008 */
3009 r5l_write_super(log, cp);
3010 } else
3011 log->last_cp_seq = le64_to_cpu(mb->seq);
3012
3013 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3014 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3015 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3016 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3017 log->last_checkpoint = cp;
3018
3019 __free_page(page);
3020
3021 if (create_super) {
3022 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3023 log->seq = log->last_cp_seq + 1;
3024 log->next_checkpoint = cp;
3025 } else
3026 ret = r5l_recovery_log(log);
3027
3028 r5c_update_log_state(log);
3029 return ret;
3030ioerr:
3031 __free_page(page);
3032 return ret;
3033}
3034
3035int r5l_start(struct r5l_log *log)
3036{
3037 int ret;
3038
3039 if (!log)
3040 return 0;
3041
3042 ret = r5l_load_log(log);
3043 if (ret) {
3044 struct mddev *mddev = log->rdev->mddev;
3045 struct r5conf *conf = mddev->private;
3046
3047 r5l_exit_log(conf);
3048 }
3049 return ret;
3050}
3051
3052void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3053{
3054 struct r5conf *conf = mddev->private;
3055 struct r5l_log *log = conf->log;
3056
3057 if (!log)
3058 return;
3059
3060 if ((raid5_calc_degraded(conf) > 0 ||
3061 test_bit(Journal, &rdev->flags)) &&
3062 conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3063 schedule_work(&log->disable_writeback_work);
3064}
3065
3066int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3067{
3068 struct request_queue *q = bdev_get_queue(rdev->bdev);
3069 struct r5l_log *log;
3070 char b[BDEVNAME_SIZE];
3071 int ret;
3072
3073 pr_debug("md/raid:%s: using device %s as journal\n",
3074 mdname(conf->mddev), bdevname(rdev->bdev, b));
3075
3076 if (PAGE_SIZE != 4096)
3077 return -EINVAL;
3078
3079 /*
3080 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3081 * raid_disks r5l_payload_data_parity.
3082 *
3083 * Write journal and cache does not work for very big array
3084 * (raid_disks > 203)
3085 */
3086 if (sizeof(struct r5l_meta_block) +
3087 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3088 conf->raid_disks) > PAGE_SIZE) {
3089 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3090 mdname(conf->mddev), conf->raid_disks);
3091 return -EINVAL;
3092 }
3093
3094 log = kzalloc(sizeof(*log), GFP_KERNEL);
3095 if (!log)
3096 return -ENOMEM;
3097 log->rdev = rdev;
3098
3099 log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
3100
3101 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3102 sizeof(rdev->mddev->uuid));
3103
3104 mutex_init(&log->io_mutex);
3105
3106 spin_lock_init(&log->io_list_lock);
3107 INIT_LIST_HEAD(&log->running_ios);
3108 INIT_LIST_HEAD(&log->io_end_ios);
3109 INIT_LIST_HEAD(&log->flushing_ios);
3110 INIT_LIST_HEAD(&log->finished_ios);
3111 bio_init(&log->flush_bio, NULL, 0);
3112
3113 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3114 if (!log->io_kc)
3115 goto io_kc;
3116
3117 ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3118 if (ret)
3119 goto io_pool;
3120
3121 ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3122 if (ret)
3123 goto io_bs;
3124
3125 ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3126 if (ret)
3127 goto out_mempool;
3128
3129 spin_lock_init(&log->tree_lock);
3130 INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3131
3132 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3133 log->rdev->mddev, "reclaim");
3134 if (!log->reclaim_thread)
3135 goto reclaim_thread;
3136 log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3137
3138 init_waitqueue_head(&log->iounit_wait);
3139
3140 INIT_LIST_HEAD(&log->no_mem_stripes);
3141
3142 INIT_LIST_HEAD(&log->no_space_stripes);
3143 spin_lock_init(&log->no_space_stripes_lock);
3144
3145 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3146 INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3147
3148 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3149 INIT_LIST_HEAD(&log->stripe_in_journal_list);
3150 spin_lock_init(&log->stripe_in_journal_lock);
3151 atomic_set(&log->stripe_in_journal_count, 0);
3152
3153 rcu_assign_pointer(conf->log, log);
3154
3155 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3156 return 0;
3157
3158reclaim_thread:
3159 mempool_exit(&log->meta_pool);
3160out_mempool:
3161 bioset_exit(&log->bs);
3162io_bs:
3163 mempool_exit(&log->io_pool);
3164io_pool:
3165 kmem_cache_destroy(log->io_kc);
3166io_kc:
3167 kfree(log);
3168 return -EINVAL;
3169}
3170
3171void r5l_exit_log(struct r5conf *conf)
3172{
3173 struct r5l_log *log = conf->log;
3174
3175 conf->log = NULL;
3176 synchronize_rcu();
3177
3178 /* Ensure disable_writeback_work wakes up and exits */
3179 wake_up(&conf->mddev->sb_wait);
3180 flush_work(&log->disable_writeback_work);
3181 md_unregister_thread(&log->reclaim_thread);
3182 mempool_exit(&log->meta_pool);
3183 bioset_exit(&log->bs);
3184 mempool_exit(&log->io_pool);
3185 kmem_cache_destroy(log->io_kc);
3186 kfree(log);
3187}
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 *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 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
317 RAID5_STRIPE_SECTORS(conf),
318 !test_bit(STRIPE_DEGRADED, &sh->state),
319 0);
320 }
321 }
322}
323
324void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
325
326/* Check whether we should flush some stripes to free up stripe cache */
327void r5c_check_stripe_cache_usage(struct r5conf *conf)
328{
329 int total_cached;
330
331 if (!r5c_is_writeback(conf->log))
332 return;
333
334 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
335 atomic_read(&conf->r5c_cached_full_stripes);
336
337 /*
338 * The following condition is true for either of the following:
339 * - stripe cache pressure high:
340 * total_cached > 3/4 min_nr_stripes ||
341 * empty_inactive_list_nr > 0
342 * - stripe cache pressure moderate:
343 * total_cached > 1/2 min_nr_stripes
344 */
345 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
346 atomic_read(&conf->empty_inactive_list_nr) > 0)
347 r5l_wake_reclaim(conf->log, 0);
348}
349
350/*
351 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
352 * stripes in the cache
353 */
354void r5c_check_cached_full_stripe(struct r5conf *conf)
355{
356 if (!r5c_is_writeback(conf->log))
357 return;
358
359 /*
360 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
361 * or a full stripe (chunk size / 4k stripes).
362 */
363 if (atomic_read(&conf->r5c_cached_full_stripes) >=
364 min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
365 conf->chunk_sectors >> RAID5_STRIPE_SHIFT(conf)))
366 r5l_wake_reclaim(conf->log, 0);
367}
368
369/*
370 * Total log space (in sectors) needed to flush all data in cache
371 *
372 * To avoid deadlock due to log space, it is necessary to reserve log
373 * space to flush critical stripes (stripes that occupying log space near
374 * last_checkpoint). This function helps check how much log space is
375 * required to flush all cached stripes.
376 *
377 * To reduce log space requirements, two mechanisms are used to give cache
378 * flush higher priorities:
379 * 1. In handle_stripe_dirtying() and schedule_reconstruction(),
380 * stripes ALREADY in journal can be flushed w/o pending writes;
381 * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
382 * can be delayed (r5l_add_no_space_stripe).
383 *
384 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
385 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
386 * pages of journal space. For stripes that has not passed 1, flushing it
387 * requires (conf->raid_disks + 1) pages of journal space. There are at
388 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
389 * required to flush all cached stripes (in pages) is:
390 *
391 * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
392 * (group_cnt + 1) * (raid_disks + 1)
393 * or
394 * (stripe_in_journal_count) * (max_degraded + 1) +
395 * (group_cnt + 1) * (raid_disks - max_degraded)
396 */
397static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
398{
399 struct r5l_log *log = conf->log;
400
401 if (!r5c_is_writeback(log))
402 return 0;
403
404 return BLOCK_SECTORS *
405 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
406 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
407}
408
409/*
410 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
411 *
412 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
413 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
414 * device is less than 2x of reclaim_required_space.
415 */
416static inline void r5c_update_log_state(struct r5l_log *log)
417{
418 struct r5conf *conf = log->rdev->mddev->private;
419 sector_t free_space;
420 sector_t reclaim_space;
421 bool wake_reclaim = false;
422
423 if (!r5c_is_writeback(log))
424 return;
425
426 free_space = r5l_ring_distance(log, log->log_start,
427 log->last_checkpoint);
428 reclaim_space = r5c_log_required_to_flush_cache(conf);
429 if (free_space < 2 * reclaim_space)
430 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
431 else {
432 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
433 wake_reclaim = true;
434 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
435 }
436 if (free_space < 3 * reclaim_space)
437 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
438 else
439 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
440
441 if (wake_reclaim)
442 r5l_wake_reclaim(log, 0);
443}
444
445/*
446 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
447 * This function should only be called in write-back mode.
448 */
449void r5c_make_stripe_write_out(struct stripe_head *sh)
450{
451 struct r5conf *conf = sh->raid_conf;
452 struct r5l_log *log = conf->log;
453
454 BUG_ON(!r5c_is_writeback(log));
455
456 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
457 clear_bit(STRIPE_R5C_CACHING, &sh->state);
458
459 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
460 atomic_inc(&conf->preread_active_stripes);
461}
462
463static void r5c_handle_data_cached(struct stripe_head *sh)
464{
465 int i;
466
467 for (i = sh->disks; i--; )
468 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
469 set_bit(R5_InJournal, &sh->dev[i].flags);
470 clear_bit(R5_LOCKED, &sh->dev[i].flags);
471 }
472 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
473}
474
475/*
476 * this journal write must contain full parity,
477 * it may also contain some data pages
478 */
479static void r5c_handle_parity_cached(struct stripe_head *sh)
480{
481 int i;
482
483 for (i = sh->disks; i--; )
484 if (test_bit(R5_InJournal, &sh->dev[i].flags))
485 set_bit(R5_Wantwrite, &sh->dev[i].flags);
486}
487
488/*
489 * Setting proper flags after writing (or flushing) data and/or parity to the
490 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
491 */
492static void r5c_finish_cache_stripe(struct stripe_head *sh)
493{
494 struct r5l_log *log = sh->raid_conf->log;
495
496 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
497 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
498 /*
499 * Set R5_InJournal for parity dev[pd_idx]. This means
500 * all data AND parity in the journal. For RAID 6, it is
501 * NOT necessary to set the flag for dev[qd_idx], as the
502 * two parities are written out together.
503 */
504 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
505 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
506 r5c_handle_data_cached(sh);
507 } else {
508 r5c_handle_parity_cached(sh);
509 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
510 }
511}
512
513static void r5l_io_run_stripes(struct r5l_io_unit *io)
514{
515 struct stripe_head *sh, *next;
516
517 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
518 list_del_init(&sh->log_list);
519
520 r5c_finish_cache_stripe(sh);
521
522 set_bit(STRIPE_HANDLE, &sh->state);
523 raid5_release_stripe(sh);
524 }
525}
526
527static void r5l_log_run_stripes(struct r5l_log *log)
528{
529 struct r5l_io_unit *io, *next;
530
531 lockdep_assert_held(&log->io_list_lock);
532
533 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
534 /* don't change list order */
535 if (io->state < IO_UNIT_IO_END)
536 break;
537
538 list_move_tail(&io->log_sibling, &log->finished_ios);
539 r5l_io_run_stripes(io);
540 }
541}
542
543static void r5l_move_to_end_ios(struct r5l_log *log)
544{
545 struct r5l_io_unit *io, *next;
546
547 lockdep_assert_held(&log->io_list_lock);
548
549 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
550 /* don't change list order */
551 if (io->state < IO_UNIT_IO_END)
552 break;
553 list_move_tail(&io->log_sibling, &log->io_end_ios);
554 }
555}
556
557static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
558static void r5l_log_endio(struct bio *bio)
559{
560 struct r5l_io_unit *io = bio->bi_private;
561 struct r5l_io_unit *io_deferred;
562 struct r5l_log *log = io->log;
563 unsigned long flags;
564 bool has_null_flush;
565 bool has_flush_payload;
566
567 if (bio->bi_status)
568 md_error(log->rdev->mddev, log->rdev);
569
570 bio_put(bio);
571 mempool_free(io->meta_page, &log->meta_pool);
572
573 spin_lock_irqsave(&log->io_list_lock, flags);
574 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
575
576 /*
577 * if the io doesn't not have null_flush or flush payload,
578 * it is not safe to access it after releasing io_list_lock.
579 * Therefore, it is necessary to check the condition with
580 * the lock held.
581 */
582 has_null_flush = io->has_null_flush;
583 has_flush_payload = io->has_flush_payload;
584
585 if (log->need_cache_flush && !list_empty(&io->stripe_list))
586 r5l_move_to_end_ios(log);
587 else
588 r5l_log_run_stripes(log);
589 if (!list_empty(&log->running_ios)) {
590 /*
591 * FLUSH/FUA io_unit is deferred because of ordering, now we
592 * can dispatch it
593 */
594 io_deferred = list_first_entry(&log->running_ios,
595 struct r5l_io_unit, log_sibling);
596 if (io_deferred->io_deferred)
597 schedule_work(&log->deferred_io_work);
598 }
599
600 spin_unlock_irqrestore(&log->io_list_lock, flags);
601
602 if (log->need_cache_flush)
603 md_wakeup_thread(log->rdev->mddev->thread);
604
605 /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
606 if (has_null_flush) {
607 struct bio *bi;
608
609 WARN_ON(bio_list_empty(&io->flush_barriers));
610 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
611 bio_endio(bi);
612 if (atomic_dec_and_test(&io->pending_stripe)) {
613 __r5l_stripe_write_finished(io);
614 return;
615 }
616 }
617 }
618 /* decrease pending_stripe for flush payload */
619 if (has_flush_payload)
620 if (atomic_dec_and_test(&io->pending_stripe))
621 __r5l_stripe_write_finished(io);
622}
623
624static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
625{
626 unsigned long flags;
627
628 spin_lock_irqsave(&log->io_list_lock, flags);
629 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
630 spin_unlock_irqrestore(&log->io_list_lock, flags);
631
632 /*
633 * In case of journal device failures, submit_bio will get error
634 * and calls endio, then active stripes will continue write
635 * process. Therefore, it is not necessary to check Faulty bit
636 * of journal device here.
637 *
638 * We can't check split_bio after current_bio is submitted. If
639 * io->split_bio is null, after current_bio is submitted, current_bio
640 * might already be completed and the io_unit is freed. We submit
641 * split_bio first to avoid the issue.
642 */
643 if (io->split_bio) {
644 if (io->has_flush)
645 io->split_bio->bi_opf |= REQ_PREFLUSH;
646 if (io->has_fua)
647 io->split_bio->bi_opf |= REQ_FUA;
648 submit_bio(io->split_bio);
649 }
650
651 if (io->has_flush)
652 io->current_bio->bi_opf |= REQ_PREFLUSH;
653 if (io->has_fua)
654 io->current_bio->bi_opf |= REQ_FUA;
655 submit_bio(io->current_bio);
656}
657
658/* deferred io_unit will be dispatched here */
659static void r5l_submit_io_async(struct work_struct *work)
660{
661 struct r5l_log *log = container_of(work, struct r5l_log,
662 deferred_io_work);
663 struct r5l_io_unit *io = NULL;
664 unsigned long flags;
665
666 spin_lock_irqsave(&log->io_list_lock, flags);
667 if (!list_empty(&log->running_ios)) {
668 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
669 log_sibling);
670 if (!io->io_deferred)
671 io = NULL;
672 else
673 io->io_deferred = 0;
674 }
675 spin_unlock_irqrestore(&log->io_list_lock, flags);
676 if (io)
677 r5l_do_submit_io(log, io);
678}
679
680static void r5c_disable_writeback_async(struct work_struct *work)
681{
682 struct r5l_log *log = container_of(work, struct r5l_log,
683 disable_writeback_work);
684 struct mddev *mddev = log->rdev->mddev;
685 struct r5conf *conf = mddev->private;
686 int locked = 0;
687
688 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
689 return;
690 pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
691 mdname(mddev));
692
693 /* wait superblock change before suspend */
694 wait_event(mddev->sb_wait,
695 conf->log == NULL ||
696 (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) &&
697 (locked = mddev_trylock(mddev))));
698 if (locked) {
699 mddev_suspend(mddev);
700 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
701 mddev_resume(mddev);
702 mddev_unlock(mddev);
703 }
704}
705
706static void r5l_submit_current_io(struct r5l_log *log)
707{
708 struct r5l_io_unit *io = log->current_io;
709 struct r5l_meta_block *block;
710 unsigned long flags;
711 u32 crc;
712 bool do_submit = true;
713
714 if (!io)
715 return;
716
717 block = page_address(io->meta_page);
718 block->meta_size = cpu_to_le32(io->meta_offset);
719 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
720 block->checksum = cpu_to_le32(crc);
721
722 log->current_io = NULL;
723 spin_lock_irqsave(&log->io_list_lock, flags);
724 if (io->has_flush || io->has_fua) {
725 if (io != list_first_entry(&log->running_ios,
726 struct r5l_io_unit, log_sibling)) {
727 io->io_deferred = 1;
728 do_submit = false;
729 }
730 }
731 spin_unlock_irqrestore(&log->io_list_lock, flags);
732 if (do_submit)
733 r5l_do_submit_io(log, io);
734}
735
736static struct bio *r5l_bio_alloc(struct r5l_log *log)
737{
738 struct bio *bio = bio_alloc_bioset(log->rdev->bdev, BIO_MAX_VECS,
739 REQ_OP_WRITE, GFP_NOIO, &log->bs);
740
741 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
742
743 return bio;
744}
745
746static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
747{
748 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
749
750 r5c_update_log_state(log);
751 /*
752 * If we filled up the log device start from the beginning again,
753 * which will require a new bio.
754 *
755 * Note: for this to work properly the log size needs to me a multiple
756 * of BLOCK_SECTORS.
757 */
758 if (log->log_start == 0)
759 io->need_split_bio = true;
760
761 io->log_end = log->log_start;
762}
763
764static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
765{
766 struct r5l_io_unit *io;
767 struct r5l_meta_block *block;
768
769 io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
770 if (!io)
771 return NULL;
772 memset(io, 0, sizeof(*io));
773
774 io->log = log;
775 INIT_LIST_HEAD(&io->log_sibling);
776 INIT_LIST_HEAD(&io->stripe_list);
777 bio_list_init(&io->flush_barriers);
778 io->state = IO_UNIT_RUNNING;
779
780 io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
781 block = page_address(io->meta_page);
782 clear_page(block);
783 block->magic = cpu_to_le32(R5LOG_MAGIC);
784 block->version = R5LOG_VERSION;
785 block->seq = cpu_to_le64(log->seq);
786 block->position = cpu_to_le64(log->log_start);
787
788 io->log_start = log->log_start;
789 io->meta_offset = sizeof(struct r5l_meta_block);
790 io->seq = log->seq++;
791
792 io->current_bio = r5l_bio_alloc(log);
793 io->current_bio->bi_end_io = r5l_log_endio;
794 io->current_bio->bi_private = io;
795 bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
796
797 r5_reserve_log_entry(log, io);
798
799 spin_lock_irq(&log->io_list_lock);
800 list_add_tail(&io->log_sibling, &log->running_ios);
801 spin_unlock_irq(&log->io_list_lock);
802
803 return io;
804}
805
806static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
807{
808 if (log->current_io &&
809 log->current_io->meta_offset + payload_size > PAGE_SIZE)
810 r5l_submit_current_io(log);
811
812 if (!log->current_io) {
813 log->current_io = r5l_new_meta(log);
814 if (!log->current_io)
815 return -ENOMEM;
816 }
817
818 return 0;
819}
820
821static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
822 sector_t location,
823 u32 checksum1, u32 checksum2,
824 bool checksum2_valid)
825{
826 struct r5l_io_unit *io = log->current_io;
827 struct r5l_payload_data_parity *payload;
828
829 payload = page_address(io->meta_page) + io->meta_offset;
830 payload->header.type = cpu_to_le16(type);
831 payload->header.flags = cpu_to_le16(0);
832 payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
833 (PAGE_SHIFT - 9));
834 payload->location = cpu_to_le64(location);
835 payload->checksum[0] = cpu_to_le32(checksum1);
836 if (checksum2_valid)
837 payload->checksum[1] = cpu_to_le32(checksum2);
838
839 io->meta_offset += sizeof(struct r5l_payload_data_parity) +
840 sizeof(__le32) * (1 + !!checksum2_valid);
841}
842
843static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
844{
845 struct r5l_io_unit *io = log->current_io;
846
847 if (io->need_split_bio) {
848 BUG_ON(io->split_bio);
849 io->split_bio = io->current_bio;
850 io->current_bio = r5l_bio_alloc(log);
851 bio_chain(io->current_bio, io->split_bio);
852 io->need_split_bio = false;
853 }
854
855 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
856 BUG();
857
858 r5_reserve_log_entry(log, io);
859}
860
861static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
862{
863 struct mddev *mddev = log->rdev->mddev;
864 struct r5conf *conf = mddev->private;
865 struct r5l_io_unit *io;
866 struct r5l_payload_flush *payload;
867 int meta_size;
868
869 /*
870 * payload_flush requires extra writes to the journal.
871 * To avoid handling the extra IO in quiesce, just skip
872 * flush_payload
873 */
874 if (conf->quiesce)
875 return;
876
877 mutex_lock(&log->io_mutex);
878 meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
879
880 if (r5l_get_meta(log, meta_size)) {
881 mutex_unlock(&log->io_mutex);
882 return;
883 }
884
885 /* current implementation is one stripe per flush payload */
886 io = log->current_io;
887 payload = page_address(io->meta_page) + io->meta_offset;
888 payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
889 payload->header.flags = cpu_to_le16(0);
890 payload->size = cpu_to_le32(sizeof(__le64));
891 payload->flush_stripes[0] = cpu_to_le64(sect);
892 io->meta_offset += meta_size;
893 /* multiple flush payloads count as one pending_stripe */
894 if (!io->has_flush_payload) {
895 io->has_flush_payload = 1;
896 atomic_inc(&io->pending_stripe);
897 }
898 mutex_unlock(&log->io_mutex);
899}
900
901static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
902 int data_pages, int parity_pages)
903{
904 int i;
905 int meta_size;
906 int ret;
907 struct r5l_io_unit *io;
908
909 meta_size =
910 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
911 * data_pages) +
912 sizeof(struct r5l_payload_data_parity) +
913 sizeof(__le32) * parity_pages;
914
915 ret = r5l_get_meta(log, meta_size);
916 if (ret)
917 return ret;
918
919 io = log->current_io;
920
921 if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
922 io->has_flush = 1;
923
924 for (i = 0; i < sh->disks; i++) {
925 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
926 test_bit(R5_InJournal, &sh->dev[i].flags))
927 continue;
928 if (i == sh->pd_idx || i == sh->qd_idx)
929 continue;
930 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
931 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
932 io->has_fua = 1;
933 /*
934 * we need to flush journal to make sure recovery can
935 * reach the data with fua flag
936 */
937 io->has_flush = 1;
938 }
939 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
940 raid5_compute_blocknr(sh, i, 0),
941 sh->dev[i].log_checksum, 0, false);
942 r5l_append_payload_page(log, sh->dev[i].page);
943 }
944
945 if (parity_pages == 2) {
946 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
947 sh->sector, sh->dev[sh->pd_idx].log_checksum,
948 sh->dev[sh->qd_idx].log_checksum, true);
949 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
950 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
951 } else if (parity_pages == 1) {
952 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
953 sh->sector, sh->dev[sh->pd_idx].log_checksum,
954 0, false);
955 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
956 } else /* Just writing data, not parity, in caching phase */
957 BUG_ON(parity_pages != 0);
958
959 list_add_tail(&sh->log_list, &io->stripe_list);
960 atomic_inc(&io->pending_stripe);
961 sh->log_io = io;
962
963 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
964 return 0;
965
966 if (sh->log_start == MaxSector) {
967 BUG_ON(!list_empty(&sh->r5c));
968 sh->log_start = io->log_start;
969 spin_lock_irq(&log->stripe_in_journal_lock);
970 list_add_tail(&sh->r5c,
971 &log->stripe_in_journal_list);
972 spin_unlock_irq(&log->stripe_in_journal_lock);
973 atomic_inc(&log->stripe_in_journal_count);
974 }
975 return 0;
976}
977
978/* add stripe to no_space_stripes, and then wake up reclaim */
979static inline void r5l_add_no_space_stripe(struct r5l_log *log,
980 struct stripe_head *sh)
981{
982 spin_lock(&log->no_space_stripes_lock);
983 list_add_tail(&sh->log_list, &log->no_space_stripes);
984 spin_unlock(&log->no_space_stripes_lock);
985}
986
987/*
988 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
989 * data from log to raid disks), so we shouldn't wait for reclaim here
990 */
991int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
992{
993 struct r5conf *conf = sh->raid_conf;
994 int write_disks = 0;
995 int data_pages, parity_pages;
996 int reserve;
997 int i;
998 int ret = 0;
999 bool wake_reclaim = false;
1000
1001 if (!log)
1002 return -EAGAIN;
1003 /* Don't support stripe batch */
1004 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1005 test_bit(STRIPE_SYNCING, &sh->state)) {
1006 /* the stripe is written to log, we start writing it to raid */
1007 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1008 return -EAGAIN;
1009 }
1010
1011 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1012
1013 for (i = 0; i < sh->disks; i++) {
1014 void *addr;
1015
1016 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1017 test_bit(R5_InJournal, &sh->dev[i].flags))
1018 continue;
1019
1020 write_disks++;
1021 /* checksum is already calculated in last run */
1022 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1023 continue;
1024 addr = kmap_atomic(sh->dev[i].page);
1025 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1026 addr, PAGE_SIZE);
1027 kunmap_atomic(addr);
1028 }
1029 parity_pages = 1 + !!(sh->qd_idx >= 0);
1030 data_pages = write_disks - parity_pages;
1031
1032 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1033 /*
1034 * The stripe must enter state machine again to finish the write, so
1035 * don't delay.
1036 */
1037 clear_bit(STRIPE_DELAYED, &sh->state);
1038 atomic_inc(&sh->count);
1039
1040 mutex_lock(&log->io_mutex);
1041 /* meta + data */
1042 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1043
1044 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1045 if (!r5l_has_free_space(log, reserve)) {
1046 r5l_add_no_space_stripe(log, sh);
1047 wake_reclaim = true;
1048 } else {
1049 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1050 if (ret) {
1051 spin_lock_irq(&log->io_list_lock);
1052 list_add_tail(&sh->log_list,
1053 &log->no_mem_stripes);
1054 spin_unlock_irq(&log->io_list_lock);
1055 }
1056 }
1057 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
1058 /*
1059 * log space critical, do not process stripes that are
1060 * not in cache yet (sh->log_start == MaxSector).
1061 */
1062 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1063 sh->log_start == MaxSector) {
1064 r5l_add_no_space_stripe(log, sh);
1065 wake_reclaim = true;
1066 reserve = 0;
1067 } else if (!r5l_has_free_space(log, reserve)) {
1068 if (sh->log_start == log->last_checkpoint)
1069 BUG();
1070 else
1071 r5l_add_no_space_stripe(log, sh);
1072 } else {
1073 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1074 if (ret) {
1075 spin_lock_irq(&log->io_list_lock);
1076 list_add_tail(&sh->log_list,
1077 &log->no_mem_stripes);
1078 spin_unlock_irq(&log->io_list_lock);
1079 }
1080 }
1081 }
1082
1083 mutex_unlock(&log->io_mutex);
1084 if (wake_reclaim)
1085 r5l_wake_reclaim(log, reserve);
1086 return 0;
1087}
1088
1089void r5l_write_stripe_run(struct r5l_log *log)
1090{
1091 if (!log)
1092 return;
1093 mutex_lock(&log->io_mutex);
1094 r5l_submit_current_io(log);
1095 mutex_unlock(&log->io_mutex);
1096}
1097
1098int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1099{
1100 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1101 /*
1102 * in write through (journal only)
1103 * we flush log disk cache first, then write stripe data to
1104 * raid disks. So if bio is finished, the log disk cache is
1105 * flushed already. The recovery guarantees we can recovery
1106 * the bio from log disk, so we don't need to flush again
1107 */
1108 if (bio->bi_iter.bi_size == 0) {
1109 bio_endio(bio);
1110 return 0;
1111 }
1112 bio->bi_opf &= ~REQ_PREFLUSH;
1113 } else {
1114 /* write back (with cache) */
1115 if (bio->bi_iter.bi_size == 0) {
1116 mutex_lock(&log->io_mutex);
1117 r5l_get_meta(log, 0);
1118 bio_list_add(&log->current_io->flush_barriers, bio);
1119 log->current_io->has_flush = 1;
1120 log->current_io->has_null_flush = 1;
1121 atomic_inc(&log->current_io->pending_stripe);
1122 r5l_submit_current_io(log);
1123 mutex_unlock(&log->io_mutex);
1124 return 0;
1125 }
1126 }
1127 return -EAGAIN;
1128}
1129
1130/* This will run after log space is reclaimed */
1131static void r5l_run_no_space_stripes(struct r5l_log *log)
1132{
1133 struct stripe_head *sh;
1134
1135 spin_lock(&log->no_space_stripes_lock);
1136 while (!list_empty(&log->no_space_stripes)) {
1137 sh = list_first_entry(&log->no_space_stripes,
1138 struct stripe_head, log_list);
1139 list_del_init(&sh->log_list);
1140 set_bit(STRIPE_HANDLE, &sh->state);
1141 raid5_release_stripe(sh);
1142 }
1143 spin_unlock(&log->no_space_stripes_lock);
1144}
1145
1146/*
1147 * calculate new last_checkpoint
1148 * for write through mode, returns log->next_checkpoint
1149 * for write back, returns log_start of first sh in stripe_in_journal_list
1150 */
1151static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1152{
1153 struct stripe_head *sh;
1154 struct r5l_log *log = conf->log;
1155 sector_t new_cp;
1156 unsigned long flags;
1157
1158 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1159 return log->next_checkpoint;
1160
1161 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1162 if (list_empty(&conf->log->stripe_in_journal_list)) {
1163 /* all stripes flushed */
1164 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1165 return log->next_checkpoint;
1166 }
1167 sh = list_first_entry(&conf->log->stripe_in_journal_list,
1168 struct stripe_head, r5c);
1169 new_cp = sh->log_start;
1170 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1171 return new_cp;
1172}
1173
1174static sector_t r5l_reclaimable_space(struct r5l_log *log)
1175{
1176 struct r5conf *conf = log->rdev->mddev->private;
1177
1178 return r5l_ring_distance(log, log->last_checkpoint,
1179 r5c_calculate_new_cp(conf));
1180}
1181
1182static void r5l_run_no_mem_stripe(struct r5l_log *log)
1183{
1184 struct stripe_head *sh;
1185
1186 lockdep_assert_held(&log->io_list_lock);
1187
1188 if (!list_empty(&log->no_mem_stripes)) {
1189 sh = list_first_entry(&log->no_mem_stripes,
1190 struct stripe_head, log_list);
1191 list_del_init(&sh->log_list);
1192 set_bit(STRIPE_HANDLE, &sh->state);
1193 raid5_release_stripe(sh);
1194 }
1195}
1196
1197static bool r5l_complete_finished_ios(struct r5l_log *log)
1198{
1199 struct r5l_io_unit *io, *next;
1200 bool found = false;
1201
1202 lockdep_assert_held(&log->io_list_lock);
1203
1204 list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1205 /* don't change list order */
1206 if (io->state < IO_UNIT_STRIPE_END)
1207 break;
1208
1209 log->next_checkpoint = io->log_start;
1210
1211 list_del(&io->log_sibling);
1212 mempool_free(io, &log->io_pool);
1213 r5l_run_no_mem_stripe(log);
1214
1215 found = true;
1216 }
1217
1218 return found;
1219}
1220
1221static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1222{
1223 struct r5l_log *log = io->log;
1224 struct r5conf *conf = log->rdev->mddev->private;
1225 unsigned long flags;
1226
1227 spin_lock_irqsave(&log->io_list_lock, flags);
1228 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1229
1230 if (!r5l_complete_finished_ios(log)) {
1231 spin_unlock_irqrestore(&log->io_list_lock, flags);
1232 return;
1233 }
1234
1235 if (r5l_reclaimable_space(log) > log->max_free_space ||
1236 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1237 r5l_wake_reclaim(log, 0);
1238
1239 spin_unlock_irqrestore(&log->io_list_lock, flags);
1240 wake_up(&log->iounit_wait);
1241}
1242
1243void r5l_stripe_write_finished(struct stripe_head *sh)
1244{
1245 struct r5l_io_unit *io;
1246
1247 io = sh->log_io;
1248 sh->log_io = NULL;
1249
1250 if (io && atomic_dec_and_test(&io->pending_stripe))
1251 __r5l_stripe_write_finished(io);
1252}
1253
1254static void r5l_log_flush_endio(struct bio *bio)
1255{
1256 struct r5l_log *log = container_of(bio, struct r5l_log,
1257 flush_bio);
1258 unsigned long flags;
1259 struct r5l_io_unit *io;
1260
1261 if (bio->bi_status)
1262 md_error(log->rdev->mddev, log->rdev);
1263
1264 spin_lock_irqsave(&log->io_list_lock, flags);
1265 list_for_each_entry(io, &log->flushing_ios, log_sibling)
1266 r5l_io_run_stripes(io);
1267 list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1268 spin_unlock_irqrestore(&log->io_list_lock, flags);
1269
1270 bio_uninit(bio);
1271}
1272
1273/*
1274 * Starting dispatch IO to raid.
1275 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1276 * broken meta in the middle of a log causes recovery can't find meta at the
1277 * head of log. If operations require meta at the head persistent in log, we
1278 * must make sure meta before it persistent in log too. A case is:
1279 *
1280 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1281 * data/parity must be persistent in log before we do the write to raid disks.
1282 *
1283 * The solution is we restrictly maintain io_unit list order. In this case, we
1284 * only write stripes of an io_unit to raid disks till the io_unit is the first
1285 * one whose data/parity is in log.
1286 */
1287void r5l_flush_stripe_to_raid(struct r5l_log *log)
1288{
1289 bool do_flush;
1290
1291 if (!log || !log->need_cache_flush)
1292 return;
1293
1294 spin_lock_irq(&log->io_list_lock);
1295 /* flush bio is running */
1296 if (!list_empty(&log->flushing_ios)) {
1297 spin_unlock_irq(&log->io_list_lock);
1298 return;
1299 }
1300 list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1301 do_flush = !list_empty(&log->flushing_ios);
1302 spin_unlock_irq(&log->io_list_lock);
1303
1304 if (!do_flush)
1305 return;
1306 bio_init(&log->flush_bio, log->rdev->bdev, NULL, 0,
1307 REQ_OP_WRITE | REQ_PREFLUSH);
1308 log->flush_bio.bi_end_io = r5l_log_flush_endio;
1309 submit_bio(&log->flush_bio);
1310}
1311
1312static void r5l_write_super(struct r5l_log *log, sector_t cp);
1313static void r5l_write_super_and_discard_space(struct r5l_log *log,
1314 sector_t end)
1315{
1316 struct block_device *bdev = log->rdev->bdev;
1317 struct mddev *mddev;
1318
1319 r5l_write_super(log, end);
1320
1321 if (!bdev_max_discard_sectors(bdev))
1322 return;
1323
1324 mddev = log->rdev->mddev;
1325 /*
1326 * Discard could zero data, so before discard we must make sure
1327 * superblock is updated to new log tail. Updating superblock (either
1328 * directly call md_update_sb() or depend on md thread) must hold
1329 * reconfig mutex. On the other hand, raid5_quiesce is called with
1330 * reconfig_mutex hold. The first step of raid5_quiesce() is waiting
1331 * for all IO finish, hence waiting for reclaim thread, while reclaim
1332 * thread is calling this function and waiting for reconfig mutex. So
1333 * there is a deadlock. We workaround this issue with a trylock.
1334 * FIXME: we could miss discard if we can't take reconfig mutex
1335 */
1336 set_mask_bits(&mddev->sb_flags, 0,
1337 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1338 if (!mddev_trylock(mddev))
1339 return;
1340 md_update_sb(mddev, 1);
1341 mddev_unlock(mddev);
1342
1343 /* discard IO error really doesn't matter, ignore it */
1344 if (log->last_checkpoint < end) {
1345 blkdev_issue_discard(bdev,
1346 log->last_checkpoint + log->rdev->data_offset,
1347 end - log->last_checkpoint, GFP_NOIO);
1348 } else {
1349 blkdev_issue_discard(bdev,
1350 log->last_checkpoint + log->rdev->data_offset,
1351 log->device_size - log->last_checkpoint,
1352 GFP_NOIO);
1353 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1354 GFP_NOIO);
1355 }
1356}
1357
1358/*
1359 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1360 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1361 *
1362 * must hold conf->device_lock
1363 */
1364static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1365{
1366 BUG_ON(list_empty(&sh->lru));
1367 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1368 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1369
1370 /*
1371 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1372 * raid5_release_stripe() while holding conf->device_lock
1373 */
1374 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1375 lockdep_assert_held(&conf->device_lock);
1376
1377 list_del_init(&sh->lru);
1378 atomic_inc(&sh->count);
1379
1380 set_bit(STRIPE_HANDLE, &sh->state);
1381 atomic_inc(&conf->active_stripes);
1382 r5c_make_stripe_write_out(sh);
1383
1384 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1385 atomic_inc(&conf->r5c_flushing_partial_stripes);
1386 else
1387 atomic_inc(&conf->r5c_flushing_full_stripes);
1388 raid5_release_stripe(sh);
1389}
1390
1391/*
1392 * if num == 0, flush all full stripes
1393 * if num > 0, flush all full stripes. If less than num full stripes are
1394 * flushed, flush some partial stripes until totally num stripes are
1395 * flushed or there is no more cached stripes.
1396 */
1397void r5c_flush_cache(struct r5conf *conf, int num)
1398{
1399 int count;
1400 struct stripe_head *sh, *next;
1401
1402 lockdep_assert_held(&conf->device_lock);
1403 if (!conf->log)
1404 return;
1405
1406 count = 0;
1407 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1408 r5c_flush_stripe(conf, sh);
1409 count++;
1410 }
1411
1412 if (count >= num)
1413 return;
1414 list_for_each_entry_safe(sh, next,
1415 &conf->r5c_partial_stripe_list, lru) {
1416 r5c_flush_stripe(conf, sh);
1417 if (++count >= num)
1418 break;
1419 }
1420}
1421
1422static void r5c_do_reclaim(struct r5conf *conf)
1423{
1424 struct r5l_log *log = conf->log;
1425 struct stripe_head *sh;
1426 int count = 0;
1427 unsigned long flags;
1428 int total_cached;
1429 int stripes_to_flush;
1430 int flushing_partial, flushing_full;
1431
1432 if (!r5c_is_writeback(log))
1433 return;
1434
1435 flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1436 flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1437 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1438 atomic_read(&conf->r5c_cached_full_stripes) -
1439 flushing_full - flushing_partial;
1440
1441 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1442 atomic_read(&conf->empty_inactive_list_nr) > 0)
1443 /*
1444 * if stripe cache pressure high, flush all full stripes and
1445 * some partial stripes
1446 */
1447 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1448 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1449 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1450 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1451 /*
1452 * if stripe cache pressure moderate, or if there is many full
1453 * stripes,flush all full stripes
1454 */
1455 stripes_to_flush = 0;
1456 else
1457 /* no need to flush */
1458 stripes_to_flush = -1;
1459
1460 if (stripes_to_flush >= 0) {
1461 spin_lock_irqsave(&conf->device_lock, flags);
1462 r5c_flush_cache(conf, stripes_to_flush);
1463 spin_unlock_irqrestore(&conf->device_lock, flags);
1464 }
1465
1466 /* if log space is tight, flush stripes on stripe_in_journal_list */
1467 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1468 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1469 spin_lock(&conf->device_lock);
1470 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1471 /*
1472 * stripes on stripe_in_journal_list could be in any
1473 * state of the stripe_cache state machine. In this
1474 * case, we only want to flush stripe on
1475 * r5c_cached_full/partial_stripes. The following
1476 * condition makes sure the stripe is on one of the
1477 * two lists.
1478 */
1479 if (!list_empty(&sh->lru) &&
1480 !test_bit(STRIPE_HANDLE, &sh->state) &&
1481 atomic_read(&sh->count) == 0) {
1482 r5c_flush_stripe(conf, sh);
1483 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1484 break;
1485 }
1486 }
1487 spin_unlock(&conf->device_lock);
1488 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1489 }
1490
1491 if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1492 r5l_run_no_space_stripes(log);
1493
1494 md_wakeup_thread(conf->mddev->thread);
1495}
1496
1497static void r5l_do_reclaim(struct r5l_log *log)
1498{
1499 struct r5conf *conf = log->rdev->mddev->private;
1500 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1501 sector_t reclaimable;
1502 sector_t next_checkpoint;
1503 bool write_super;
1504
1505 spin_lock_irq(&log->io_list_lock);
1506 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1507 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1508 /*
1509 * move proper io_unit to reclaim list. We should not change the order.
1510 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1511 * shouldn't reuse space of an unreclaimable io_unit
1512 */
1513 while (1) {
1514 reclaimable = r5l_reclaimable_space(log);
1515 if (reclaimable >= reclaim_target ||
1516 (list_empty(&log->running_ios) &&
1517 list_empty(&log->io_end_ios) &&
1518 list_empty(&log->flushing_ios) &&
1519 list_empty(&log->finished_ios)))
1520 break;
1521
1522 md_wakeup_thread(log->rdev->mddev->thread);
1523 wait_event_lock_irq(log->iounit_wait,
1524 r5l_reclaimable_space(log) > reclaimable,
1525 log->io_list_lock);
1526 }
1527
1528 next_checkpoint = r5c_calculate_new_cp(conf);
1529 spin_unlock_irq(&log->io_list_lock);
1530
1531 if (reclaimable == 0 || !write_super)
1532 return;
1533
1534 /*
1535 * write_super will flush cache of each raid disk. We must write super
1536 * here, because the log area might be reused soon and we don't want to
1537 * confuse recovery
1538 */
1539 r5l_write_super_and_discard_space(log, next_checkpoint);
1540
1541 mutex_lock(&log->io_mutex);
1542 log->last_checkpoint = next_checkpoint;
1543 r5c_update_log_state(log);
1544 mutex_unlock(&log->io_mutex);
1545
1546 r5l_run_no_space_stripes(log);
1547}
1548
1549static void r5l_reclaim_thread(struct md_thread *thread)
1550{
1551 struct mddev *mddev = thread->mddev;
1552 struct r5conf *conf = mddev->private;
1553 struct r5l_log *log = conf->log;
1554
1555 if (!log)
1556 return;
1557 r5c_do_reclaim(conf);
1558 r5l_do_reclaim(log);
1559}
1560
1561void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1562{
1563 unsigned long target;
1564 unsigned long new = (unsigned long)space; /* overflow in theory */
1565
1566 if (!log)
1567 return;
1568
1569 target = READ_ONCE(log->reclaim_target);
1570 do {
1571 if (new < target)
1572 return;
1573 } while (!try_cmpxchg(&log->reclaim_target, &target, new));
1574 md_wakeup_thread(log->reclaim_thread);
1575}
1576
1577void r5l_quiesce(struct r5l_log *log, int quiesce)
1578{
1579 struct mddev *mddev;
1580
1581 if (quiesce) {
1582 /* make sure r5l_write_super_and_discard_space exits */
1583 mddev = log->rdev->mddev;
1584 wake_up(&mddev->sb_wait);
1585 kthread_park(log->reclaim_thread->tsk);
1586 r5l_wake_reclaim(log, MaxSector);
1587 r5l_do_reclaim(log);
1588 } else
1589 kthread_unpark(log->reclaim_thread->tsk);
1590}
1591
1592bool r5l_log_disk_error(struct r5conf *conf)
1593{
1594 struct r5l_log *log = conf->log;
1595
1596 /* don't allow write if journal disk is missing */
1597 if (!log)
1598 return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1599 else
1600 return test_bit(Faulty, &log->rdev->flags);
1601}
1602
1603#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1604
1605struct r5l_recovery_ctx {
1606 struct page *meta_page; /* current meta */
1607 sector_t meta_total_blocks; /* total size of current meta and data */
1608 sector_t pos; /* recovery position */
1609 u64 seq; /* recovery position seq */
1610 int data_parity_stripes; /* number of data_parity stripes */
1611 int data_only_stripes; /* number of data_only stripes */
1612 struct list_head cached_list;
1613
1614 /*
1615 * read ahead page pool (ra_pool)
1616 * in recovery, log is read sequentially. It is not efficient to
1617 * read every page with sync_page_io(). The read ahead page pool
1618 * reads multiple pages with one IO, so further log read can
1619 * just copy data from the pool.
1620 */
1621 struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1622 struct bio_vec ra_bvec[R5L_RECOVERY_PAGE_POOL_SIZE];
1623 sector_t pool_offset; /* offset of first page in the pool */
1624 int total_pages; /* total allocated pages */
1625 int valid_pages; /* pages with valid data */
1626};
1627
1628static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1629 struct r5l_recovery_ctx *ctx)
1630{
1631 struct page *page;
1632
1633 ctx->valid_pages = 0;
1634 ctx->total_pages = 0;
1635 while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1636 page = alloc_page(GFP_KERNEL);
1637
1638 if (!page)
1639 break;
1640 ctx->ra_pool[ctx->total_pages] = page;
1641 ctx->total_pages += 1;
1642 }
1643
1644 if (ctx->total_pages == 0)
1645 return -ENOMEM;
1646
1647 ctx->pool_offset = 0;
1648 return 0;
1649}
1650
1651static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1652 struct r5l_recovery_ctx *ctx)
1653{
1654 int i;
1655
1656 for (i = 0; i < ctx->total_pages; ++i)
1657 put_page(ctx->ra_pool[i]);
1658}
1659
1660/*
1661 * fetch ctx->valid_pages pages from offset
1662 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1663 * However, if the offset is close to the end of the journal device,
1664 * ctx->valid_pages could be smaller than ctx->total_pages
1665 */
1666static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1667 struct r5l_recovery_ctx *ctx,
1668 sector_t offset)
1669{
1670 struct bio bio;
1671 int ret;
1672
1673 bio_init(&bio, log->rdev->bdev, ctx->ra_bvec,
1674 R5L_RECOVERY_PAGE_POOL_SIZE, REQ_OP_READ);
1675 bio.bi_iter.bi_sector = log->rdev->data_offset + offset;
1676
1677 ctx->valid_pages = 0;
1678 ctx->pool_offset = offset;
1679
1680 while (ctx->valid_pages < ctx->total_pages) {
1681 __bio_add_page(&bio, ctx->ra_pool[ctx->valid_pages], PAGE_SIZE,
1682 0);
1683 ctx->valid_pages += 1;
1684
1685 offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1686
1687 if (offset == 0) /* reached end of the device */
1688 break;
1689 }
1690
1691 ret = submit_bio_wait(&bio);
1692 bio_uninit(&bio);
1693 return ret;
1694}
1695
1696/*
1697 * try read a page from the read ahead page pool, if the page is not in the
1698 * pool, call r5l_recovery_fetch_ra_pool
1699 */
1700static int r5l_recovery_read_page(struct r5l_log *log,
1701 struct r5l_recovery_ctx *ctx,
1702 struct page *page,
1703 sector_t offset)
1704{
1705 int ret;
1706
1707 if (offset < ctx->pool_offset ||
1708 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1709 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1710 if (ret)
1711 return ret;
1712 }
1713
1714 BUG_ON(offset < ctx->pool_offset ||
1715 offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1716
1717 memcpy(page_address(page),
1718 page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1719 BLOCK_SECTOR_SHIFT]),
1720 PAGE_SIZE);
1721 return 0;
1722}
1723
1724static int r5l_recovery_read_meta_block(struct r5l_log *log,
1725 struct r5l_recovery_ctx *ctx)
1726{
1727 struct page *page = ctx->meta_page;
1728 struct r5l_meta_block *mb;
1729 u32 crc, stored_crc;
1730 int ret;
1731
1732 ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1733 if (ret != 0)
1734 return ret;
1735
1736 mb = page_address(page);
1737 stored_crc = le32_to_cpu(mb->checksum);
1738 mb->checksum = 0;
1739
1740 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1741 le64_to_cpu(mb->seq) != ctx->seq ||
1742 mb->version != R5LOG_VERSION ||
1743 le64_to_cpu(mb->position) != ctx->pos)
1744 return -EINVAL;
1745
1746 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1747 if (stored_crc != crc)
1748 return -EINVAL;
1749
1750 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1751 return -EINVAL;
1752
1753 ctx->meta_total_blocks = BLOCK_SECTORS;
1754
1755 return 0;
1756}
1757
1758static void
1759r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1760 struct page *page,
1761 sector_t pos, u64 seq)
1762{
1763 struct r5l_meta_block *mb;
1764
1765 mb = page_address(page);
1766 clear_page(mb);
1767 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1768 mb->version = R5LOG_VERSION;
1769 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1770 mb->seq = cpu_to_le64(seq);
1771 mb->position = cpu_to_le64(pos);
1772}
1773
1774static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1775 u64 seq)
1776{
1777 struct page *page;
1778 struct r5l_meta_block *mb;
1779
1780 page = alloc_page(GFP_KERNEL);
1781 if (!page)
1782 return -ENOMEM;
1783 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1784 mb = page_address(page);
1785 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1786 mb, PAGE_SIZE));
1787 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE |
1788 REQ_SYNC | REQ_FUA, false)) {
1789 __free_page(page);
1790 return -EIO;
1791 }
1792 __free_page(page);
1793 return 0;
1794}
1795
1796/*
1797 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1798 * to mark valid (potentially not flushed) data in the journal.
1799 *
1800 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1801 * so there should not be any mismatch here.
1802 */
1803static void r5l_recovery_load_data(struct r5l_log *log,
1804 struct stripe_head *sh,
1805 struct r5l_recovery_ctx *ctx,
1806 struct r5l_payload_data_parity *payload,
1807 sector_t log_offset)
1808{
1809 struct mddev *mddev = log->rdev->mddev;
1810 struct r5conf *conf = mddev->private;
1811 int dd_idx;
1812
1813 raid5_compute_sector(conf,
1814 le64_to_cpu(payload->location), 0,
1815 &dd_idx, sh);
1816 r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1817 sh->dev[dd_idx].log_checksum =
1818 le32_to_cpu(payload->checksum[0]);
1819 ctx->meta_total_blocks += BLOCK_SECTORS;
1820
1821 set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1822 set_bit(STRIPE_R5C_CACHING, &sh->state);
1823}
1824
1825static void r5l_recovery_load_parity(struct r5l_log *log,
1826 struct stripe_head *sh,
1827 struct r5l_recovery_ctx *ctx,
1828 struct r5l_payload_data_parity *payload,
1829 sector_t log_offset)
1830{
1831 struct mddev *mddev = log->rdev->mddev;
1832 struct r5conf *conf = mddev->private;
1833
1834 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1835 r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1836 sh->dev[sh->pd_idx].log_checksum =
1837 le32_to_cpu(payload->checksum[0]);
1838 set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1839
1840 if (sh->qd_idx >= 0) {
1841 r5l_recovery_read_page(
1842 log, ctx, sh->dev[sh->qd_idx].page,
1843 r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1844 sh->dev[sh->qd_idx].log_checksum =
1845 le32_to_cpu(payload->checksum[1]);
1846 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1847 }
1848 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1849}
1850
1851static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1852{
1853 int i;
1854
1855 sh->state = 0;
1856 sh->log_start = MaxSector;
1857 for (i = sh->disks; i--; )
1858 sh->dev[i].flags = 0;
1859}
1860
1861static void
1862r5l_recovery_replay_one_stripe(struct r5conf *conf,
1863 struct stripe_head *sh,
1864 struct r5l_recovery_ctx *ctx)
1865{
1866 struct md_rdev *rdev, *rrdev;
1867 int disk_index;
1868 int data_count = 0;
1869
1870 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1871 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1872 continue;
1873 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1874 continue;
1875 data_count++;
1876 }
1877
1878 /*
1879 * stripes that only have parity must have been flushed
1880 * before the crash that we are now recovering from, so
1881 * there is nothing more to recovery.
1882 */
1883 if (data_count == 0)
1884 goto out;
1885
1886 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1887 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1888 continue;
1889
1890 /* in case device is broken */
1891 rcu_read_lock();
1892 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1893 if (rdev) {
1894 atomic_inc(&rdev->nr_pending);
1895 rcu_read_unlock();
1896 sync_page_io(rdev, sh->sector, PAGE_SIZE,
1897 sh->dev[disk_index].page, REQ_OP_WRITE,
1898 false);
1899 rdev_dec_pending(rdev, rdev->mddev);
1900 rcu_read_lock();
1901 }
1902 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1903 if (rrdev) {
1904 atomic_inc(&rrdev->nr_pending);
1905 rcu_read_unlock();
1906 sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1907 sh->dev[disk_index].page, REQ_OP_WRITE,
1908 false);
1909 rdev_dec_pending(rrdev, rrdev->mddev);
1910 rcu_read_lock();
1911 }
1912 rcu_read_unlock();
1913 }
1914 ctx->data_parity_stripes++;
1915out:
1916 r5l_recovery_reset_stripe(sh);
1917}
1918
1919static struct stripe_head *
1920r5c_recovery_alloc_stripe(
1921 struct r5conf *conf,
1922 sector_t stripe_sect,
1923 int noblock)
1924{
1925 struct stripe_head *sh;
1926
1927 sh = raid5_get_active_stripe(conf, NULL, stripe_sect,
1928 noblock ? R5_GAS_NOBLOCK : 0);
1929 if (!sh)
1930 return NULL; /* no more stripe available */
1931
1932 r5l_recovery_reset_stripe(sh);
1933
1934 return sh;
1935}
1936
1937static struct stripe_head *
1938r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1939{
1940 struct stripe_head *sh;
1941
1942 list_for_each_entry(sh, list, lru)
1943 if (sh->sector == sect)
1944 return sh;
1945 return NULL;
1946}
1947
1948static void
1949r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1950 struct r5l_recovery_ctx *ctx)
1951{
1952 struct stripe_head *sh, *next;
1953
1954 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1955 r5l_recovery_reset_stripe(sh);
1956 list_del_init(&sh->lru);
1957 raid5_release_stripe(sh);
1958 }
1959}
1960
1961static void
1962r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1963 struct r5l_recovery_ctx *ctx)
1964{
1965 struct stripe_head *sh, *next;
1966
1967 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1968 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1969 r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1970 list_del_init(&sh->lru);
1971 raid5_release_stripe(sh);
1972 }
1973}
1974
1975/* if matches return 0; otherwise return -EINVAL */
1976static int
1977r5l_recovery_verify_data_checksum(struct r5l_log *log,
1978 struct r5l_recovery_ctx *ctx,
1979 struct page *page,
1980 sector_t log_offset, __le32 log_checksum)
1981{
1982 void *addr;
1983 u32 checksum;
1984
1985 r5l_recovery_read_page(log, ctx, page, log_offset);
1986 addr = kmap_atomic(page);
1987 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1988 kunmap_atomic(addr);
1989 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1990}
1991
1992/*
1993 * before loading data to stripe cache, we need verify checksum for all data,
1994 * if there is mismatch for any data page, we drop all data in the mata block
1995 */
1996static int
1997r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1998 struct r5l_recovery_ctx *ctx)
1999{
2000 struct mddev *mddev = log->rdev->mddev;
2001 struct r5conf *conf = mddev->private;
2002 struct r5l_meta_block *mb = page_address(ctx->meta_page);
2003 sector_t mb_offset = sizeof(struct r5l_meta_block);
2004 sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2005 struct page *page;
2006 struct r5l_payload_data_parity *payload;
2007 struct r5l_payload_flush *payload_flush;
2008
2009 page = alloc_page(GFP_KERNEL);
2010 if (!page)
2011 return -ENOMEM;
2012
2013 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2014 payload = (void *)mb + mb_offset;
2015 payload_flush = (void *)mb + mb_offset;
2016
2017 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2018 if (r5l_recovery_verify_data_checksum(
2019 log, ctx, page, log_offset,
2020 payload->checksum[0]) < 0)
2021 goto mismatch;
2022 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2023 if (r5l_recovery_verify_data_checksum(
2024 log, ctx, page, log_offset,
2025 payload->checksum[0]) < 0)
2026 goto mismatch;
2027 if (conf->max_degraded == 2 && /* q for RAID 6 */
2028 r5l_recovery_verify_data_checksum(
2029 log, ctx, page,
2030 r5l_ring_add(log, log_offset,
2031 BLOCK_SECTORS),
2032 payload->checksum[1]) < 0)
2033 goto mismatch;
2034 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2035 /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2036 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2037 goto mismatch;
2038
2039 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2040 mb_offset += sizeof(struct r5l_payload_flush) +
2041 le32_to_cpu(payload_flush->size);
2042 } else {
2043 /* DATA or PARITY payload */
2044 log_offset = r5l_ring_add(log, log_offset,
2045 le32_to_cpu(payload->size));
2046 mb_offset += sizeof(struct r5l_payload_data_parity) +
2047 sizeof(__le32) *
2048 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2049 }
2050
2051 }
2052
2053 put_page(page);
2054 return 0;
2055
2056mismatch:
2057 put_page(page);
2058 return -EINVAL;
2059}
2060
2061/*
2062 * Analyze all data/parity pages in one meta block
2063 * Returns:
2064 * 0 for success
2065 * -EINVAL for unknown playload type
2066 * -EAGAIN for checksum mismatch of data page
2067 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2068 */
2069static int
2070r5c_recovery_analyze_meta_block(struct r5l_log *log,
2071 struct r5l_recovery_ctx *ctx,
2072 struct list_head *cached_stripe_list)
2073{
2074 struct mddev *mddev = log->rdev->mddev;
2075 struct r5conf *conf = mddev->private;
2076 struct r5l_meta_block *mb;
2077 struct r5l_payload_data_parity *payload;
2078 struct r5l_payload_flush *payload_flush;
2079 int mb_offset;
2080 sector_t log_offset;
2081 sector_t stripe_sect;
2082 struct stripe_head *sh;
2083 int ret;
2084
2085 /*
2086 * for mismatch in data blocks, we will drop all data in this mb, but
2087 * we will still read next mb for other data with FLUSH flag, as
2088 * io_unit could finish out of order.
2089 */
2090 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2091 if (ret == -EINVAL)
2092 return -EAGAIN;
2093 else if (ret)
2094 return ret; /* -ENOMEM duo to alloc_page() failed */
2095
2096 mb = page_address(ctx->meta_page);
2097 mb_offset = sizeof(struct r5l_meta_block);
2098 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2099
2100 while (mb_offset < le32_to_cpu(mb->meta_size)) {
2101 int dd;
2102
2103 payload = (void *)mb + mb_offset;
2104 payload_flush = (void *)mb + mb_offset;
2105
2106 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2107 int i, count;
2108
2109 count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2110 for (i = 0; i < count; ++i) {
2111 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2112 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2113 stripe_sect);
2114 if (sh) {
2115 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2116 r5l_recovery_reset_stripe(sh);
2117 list_del_init(&sh->lru);
2118 raid5_release_stripe(sh);
2119 }
2120 }
2121
2122 mb_offset += sizeof(struct r5l_payload_flush) +
2123 le32_to_cpu(payload_flush->size);
2124 continue;
2125 }
2126
2127 /* DATA or PARITY payload */
2128 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2129 raid5_compute_sector(
2130 conf, le64_to_cpu(payload->location), 0, &dd,
2131 NULL)
2132 : le64_to_cpu(payload->location);
2133
2134 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2135 stripe_sect);
2136
2137 if (!sh) {
2138 sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2139 /*
2140 * cannot get stripe from raid5_get_active_stripe
2141 * try replay some stripes
2142 */
2143 if (!sh) {
2144 r5c_recovery_replay_stripes(
2145 cached_stripe_list, ctx);
2146 sh = r5c_recovery_alloc_stripe(
2147 conf, stripe_sect, 1);
2148 }
2149 if (!sh) {
2150 int new_size = conf->min_nr_stripes * 2;
2151 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2152 mdname(mddev),
2153 new_size);
2154 ret = raid5_set_cache_size(mddev, new_size);
2155 if (conf->min_nr_stripes <= new_size / 2) {
2156 pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2157 mdname(mddev),
2158 ret,
2159 new_size,
2160 conf->min_nr_stripes,
2161 conf->max_nr_stripes);
2162 return -ENOMEM;
2163 }
2164 sh = r5c_recovery_alloc_stripe(
2165 conf, stripe_sect, 0);
2166 }
2167 if (!sh) {
2168 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2169 mdname(mddev));
2170 return -ENOMEM;
2171 }
2172 list_add_tail(&sh->lru, cached_stripe_list);
2173 }
2174
2175 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2176 if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2177 test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2178 r5l_recovery_replay_one_stripe(conf, sh, ctx);
2179 list_move_tail(&sh->lru, cached_stripe_list);
2180 }
2181 r5l_recovery_load_data(log, sh, ctx, payload,
2182 log_offset);
2183 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2184 r5l_recovery_load_parity(log, sh, ctx, payload,
2185 log_offset);
2186 else
2187 return -EINVAL;
2188
2189 log_offset = r5l_ring_add(log, log_offset,
2190 le32_to_cpu(payload->size));
2191
2192 mb_offset += sizeof(struct r5l_payload_data_parity) +
2193 sizeof(__le32) *
2194 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2195 }
2196
2197 return 0;
2198}
2199
2200/*
2201 * Load the stripe into cache. The stripe will be written out later by
2202 * the stripe cache state machine.
2203 */
2204static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2205 struct stripe_head *sh)
2206{
2207 struct r5dev *dev;
2208 int i;
2209
2210 for (i = sh->disks; i--; ) {
2211 dev = sh->dev + i;
2212 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2213 set_bit(R5_InJournal, &dev->flags);
2214 set_bit(R5_UPTODATE, &dev->flags);
2215 }
2216 }
2217}
2218
2219/*
2220 * Scan through the log for all to-be-flushed data
2221 *
2222 * For stripes with data and parity, namely Data-Parity stripe
2223 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2224 *
2225 * For stripes with only data, namely Data-Only stripe
2226 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2227 *
2228 * For a stripe, if we see data after parity, we should discard all previous
2229 * data and parity for this stripe, as these data are already flushed to
2230 * the array.
2231 *
2232 * At the end of the scan, we return the new journal_tail, which points to
2233 * first data-only stripe on the journal device, or next invalid meta block.
2234 */
2235static int r5c_recovery_flush_log(struct r5l_log *log,
2236 struct r5l_recovery_ctx *ctx)
2237{
2238 struct stripe_head *sh;
2239 int ret = 0;
2240
2241 /* scan through the log */
2242 while (1) {
2243 if (r5l_recovery_read_meta_block(log, ctx))
2244 break;
2245
2246 ret = r5c_recovery_analyze_meta_block(log, ctx,
2247 &ctx->cached_list);
2248 /*
2249 * -EAGAIN means mismatch in data block, in this case, we still
2250 * try scan the next metablock
2251 */
2252 if (ret && ret != -EAGAIN)
2253 break; /* ret == -EINVAL or -ENOMEM */
2254 ctx->seq++;
2255 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2256 }
2257
2258 if (ret == -ENOMEM) {
2259 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2260 return ret;
2261 }
2262
2263 /* replay data-parity stripes */
2264 r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2265
2266 /* load data-only stripes to stripe cache */
2267 list_for_each_entry(sh, &ctx->cached_list, lru) {
2268 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2269 r5c_recovery_load_one_stripe(log, sh);
2270 ctx->data_only_stripes++;
2271 }
2272
2273 return 0;
2274}
2275
2276/*
2277 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2278 * log will start here. but we can't let superblock point to last valid
2279 * meta block. The log might looks like:
2280 * | meta 1| meta 2| meta 3|
2281 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2282 * superblock points to meta 1, we write a new valid meta 2n. if crash
2283 * happens again, new recovery will start from meta 1. Since meta 2n is
2284 * valid now, recovery will think meta 3 is valid, which is wrong.
2285 * The solution is we create a new meta in meta2 with its seq == meta
2286 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2287 * will not think meta 3 is a valid meta, because its seq doesn't match
2288 */
2289
2290/*
2291 * Before recovery, the log looks like the following
2292 *
2293 * ---------------------------------------------
2294 * | valid log | invalid log |
2295 * ---------------------------------------------
2296 * ^
2297 * |- log->last_checkpoint
2298 * |- log->last_cp_seq
2299 *
2300 * Now we scan through the log until we see invalid entry
2301 *
2302 * ---------------------------------------------
2303 * | valid log | invalid log |
2304 * ---------------------------------------------
2305 * ^ ^
2306 * |- log->last_checkpoint |- ctx->pos
2307 * |- log->last_cp_seq |- ctx->seq
2308 *
2309 * From this point, we need to increase seq number by 10 to avoid
2310 * confusing next recovery.
2311 *
2312 * ---------------------------------------------
2313 * | valid log | invalid log |
2314 * ---------------------------------------------
2315 * ^ ^
2316 * |- log->last_checkpoint |- ctx->pos+1
2317 * |- log->last_cp_seq |- ctx->seq+10001
2318 *
2319 * However, it is not safe to start the state machine yet, because data only
2320 * parities are not yet secured in RAID. To save these data only parities, we
2321 * rewrite them from seq+11.
2322 *
2323 * -----------------------------------------------------------------
2324 * | valid log | data only stripes | invalid log |
2325 * -----------------------------------------------------------------
2326 * ^ ^
2327 * |- log->last_checkpoint |- ctx->pos+n
2328 * |- log->last_cp_seq |- ctx->seq+10000+n
2329 *
2330 * If failure happens again during this process, the recovery can safe start
2331 * again from log->last_checkpoint.
2332 *
2333 * Once data only stripes are rewritten to journal, we move log_tail
2334 *
2335 * -----------------------------------------------------------------
2336 * | old log | data only stripes | invalid log |
2337 * -----------------------------------------------------------------
2338 * ^ ^
2339 * |- log->last_checkpoint |- ctx->pos+n
2340 * |- log->last_cp_seq |- ctx->seq+10000+n
2341 *
2342 * Then we can safely start the state machine. If failure happens from this
2343 * point on, the recovery will start from new log->last_checkpoint.
2344 */
2345static int
2346r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2347 struct r5l_recovery_ctx *ctx)
2348{
2349 struct stripe_head *sh;
2350 struct mddev *mddev = log->rdev->mddev;
2351 struct page *page;
2352 sector_t next_checkpoint = MaxSector;
2353
2354 page = alloc_page(GFP_KERNEL);
2355 if (!page) {
2356 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2357 mdname(mddev));
2358 return -ENOMEM;
2359 }
2360
2361 WARN_ON(list_empty(&ctx->cached_list));
2362
2363 list_for_each_entry(sh, &ctx->cached_list, lru) {
2364 struct r5l_meta_block *mb;
2365 int i;
2366 int offset;
2367 sector_t write_pos;
2368
2369 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2370 r5l_recovery_create_empty_meta_block(log, page,
2371 ctx->pos, ctx->seq);
2372 mb = page_address(page);
2373 offset = le32_to_cpu(mb->meta_size);
2374 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2375
2376 for (i = sh->disks; i--; ) {
2377 struct r5dev *dev = &sh->dev[i];
2378 struct r5l_payload_data_parity *payload;
2379 void *addr;
2380
2381 if (test_bit(R5_InJournal, &dev->flags)) {
2382 payload = (void *)mb + offset;
2383 payload->header.type = cpu_to_le16(
2384 R5LOG_PAYLOAD_DATA);
2385 payload->size = cpu_to_le32(BLOCK_SECTORS);
2386 payload->location = cpu_to_le64(
2387 raid5_compute_blocknr(sh, i, 0));
2388 addr = kmap_atomic(dev->page);
2389 payload->checksum[0] = cpu_to_le32(
2390 crc32c_le(log->uuid_checksum, addr,
2391 PAGE_SIZE));
2392 kunmap_atomic(addr);
2393 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2394 dev->page, REQ_OP_WRITE, false);
2395 write_pos = r5l_ring_add(log, write_pos,
2396 BLOCK_SECTORS);
2397 offset += sizeof(__le32) +
2398 sizeof(struct r5l_payload_data_parity);
2399
2400 }
2401 }
2402 mb->meta_size = cpu_to_le32(offset);
2403 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2404 mb, PAGE_SIZE));
2405 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2406 REQ_OP_WRITE | REQ_SYNC | REQ_FUA, false);
2407 sh->log_start = ctx->pos;
2408 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2409 atomic_inc(&log->stripe_in_journal_count);
2410 ctx->pos = write_pos;
2411 ctx->seq += 1;
2412 next_checkpoint = sh->log_start;
2413 }
2414 log->next_checkpoint = next_checkpoint;
2415 __free_page(page);
2416 return 0;
2417}
2418
2419static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2420 struct r5l_recovery_ctx *ctx)
2421{
2422 struct mddev *mddev = log->rdev->mddev;
2423 struct r5conf *conf = mddev->private;
2424 struct stripe_head *sh, *next;
2425 bool cleared_pending = false;
2426
2427 if (ctx->data_only_stripes == 0)
2428 return;
2429
2430 if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2431 cleared_pending = true;
2432 clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2433 }
2434 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2435
2436 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2437 r5c_make_stripe_write_out(sh);
2438 set_bit(STRIPE_HANDLE, &sh->state);
2439 list_del_init(&sh->lru);
2440 raid5_release_stripe(sh);
2441 }
2442
2443 /* reuse conf->wait_for_quiescent in recovery */
2444 wait_event(conf->wait_for_quiescent,
2445 atomic_read(&conf->active_stripes) == 0);
2446
2447 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2448 if (cleared_pending)
2449 set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2450}
2451
2452static int r5l_recovery_log(struct r5l_log *log)
2453{
2454 struct mddev *mddev = log->rdev->mddev;
2455 struct r5l_recovery_ctx *ctx;
2456 int ret;
2457 sector_t pos;
2458
2459 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2460 if (!ctx)
2461 return -ENOMEM;
2462
2463 ctx->pos = log->last_checkpoint;
2464 ctx->seq = log->last_cp_seq;
2465 INIT_LIST_HEAD(&ctx->cached_list);
2466 ctx->meta_page = alloc_page(GFP_KERNEL);
2467
2468 if (!ctx->meta_page) {
2469 ret = -ENOMEM;
2470 goto meta_page;
2471 }
2472
2473 if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2474 ret = -ENOMEM;
2475 goto ra_pool;
2476 }
2477
2478 ret = r5c_recovery_flush_log(log, ctx);
2479
2480 if (ret)
2481 goto error;
2482
2483 pos = ctx->pos;
2484 ctx->seq += 10000;
2485
2486 if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2487 pr_info("md/raid:%s: starting from clean shutdown\n",
2488 mdname(mddev));
2489 else
2490 pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2491 mdname(mddev), ctx->data_only_stripes,
2492 ctx->data_parity_stripes);
2493
2494 if (ctx->data_only_stripes == 0) {
2495 log->next_checkpoint = ctx->pos;
2496 r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2497 ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2498 } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2499 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2500 mdname(mddev));
2501 ret = -EIO;
2502 goto error;
2503 }
2504
2505 log->log_start = ctx->pos;
2506 log->seq = ctx->seq;
2507 log->last_checkpoint = pos;
2508 r5l_write_super(log, pos);
2509
2510 r5c_recovery_flush_data_only_stripes(log, ctx);
2511 ret = 0;
2512error:
2513 r5l_recovery_free_ra_pool(log, ctx);
2514ra_pool:
2515 __free_page(ctx->meta_page);
2516meta_page:
2517 kfree(ctx);
2518 return ret;
2519}
2520
2521static void r5l_write_super(struct r5l_log *log, sector_t cp)
2522{
2523 struct mddev *mddev = log->rdev->mddev;
2524
2525 log->rdev->journal_tail = cp;
2526 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2527}
2528
2529static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2530{
2531 struct r5conf *conf;
2532 int ret;
2533
2534 ret = mddev_lock(mddev);
2535 if (ret)
2536 return ret;
2537
2538 conf = mddev->private;
2539 if (!conf || !conf->log)
2540 goto out_unlock;
2541
2542 switch (conf->log->r5c_journal_mode) {
2543 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2544 ret = snprintf(
2545 page, PAGE_SIZE, "[%s] %s\n",
2546 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2547 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2548 break;
2549 case R5C_JOURNAL_MODE_WRITE_BACK:
2550 ret = snprintf(
2551 page, PAGE_SIZE, "%s [%s]\n",
2552 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2553 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2554 break;
2555 default:
2556 ret = 0;
2557 }
2558
2559out_unlock:
2560 mddev_unlock(mddev);
2561 return ret;
2562}
2563
2564/*
2565 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2566 *
2567 * @mode as defined in 'enum r5c_journal_mode'.
2568 *
2569 */
2570int r5c_journal_mode_set(struct mddev *mddev, int mode)
2571{
2572 struct r5conf *conf;
2573
2574 if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2575 mode > R5C_JOURNAL_MODE_WRITE_BACK)
2576 return -EINVAL;
2577
2578 conf = mddev->private;
2579 if (!conf || !conf->log)
2580 return -ENODEV;
2581
2582 if (raid5_calc_degraded(conf) > 0 &&
2583 mode == R5C_JOURNAL_MODE_WRITE_BACK)
2584 return -EINVAL;
2585
2586 mddev_suspend(mddev);
2587 conf->log->r5c_journal_mode = mode;
2588 mddev_resume(mddev);
2589
2590 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2591 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2592 return 0;
2593}
2594EXPORT_SYMBOL(r5c_journal_mode_set);
2595
2596static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2597 const char *page, size_t length)
2598{
2599 int mode = ARRAY_SIZE(r5c_journal_mode_str);
2600 size_t len = length;
2601 int ret;
2602
2603 if (len < 2)
2604 return -EINVAL;
2605
2606 if (page[len - 1] == '\n')
2607 len--;
2608
2609 while (mode--)
2610 if (strlen(r5c_journal_mode_str[mode]) == len &&
2611 !strncmp(page, r5c_journal_mode_str[mode], len))
2612 break;
2613 ret = mddev_lock(mddev);
2614 if (ret)
2615 return ret;
2616 ret = r5c_journal_mode_set(mddev, mode);
2617 mddev_unlock(mddev);
2618 return ret ?: length;
2619}
2620
2621struct md_sysfs_entry
2622r5c_journal_mode = __ATTR(journal_mode, 0644,
2623 r5c_journal_mode_show, r5c_journal_mode_store);
2624
2625/*
2626 * Try handle write operation in caching phase. This function should only
2627 * be called in write-back mode.
2628 *
2629 * If all outstanding writes can be handled in caching phase, returns 0
2630 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2631 * and returns -EAGAIN
2632 */
2633int r5c_try_caching_write(struct r5conf *conf,
2634 struct stripe_head *sh,
2635 struct stripe_head_state *s,
2636 int disks)
2637{
2638 struct r5l_log *log = conf->log;
2639 int i;
2640 struct r5dev *dev;
2641 int to_cache = 0;
2642 void __rcu **pslot;
2643 sector_t tree_index;
2644 int ret;
2645 uintptr_t refcount;
2646
2647 BUG_ON(!r5c_is_writeback(log));
2648
2649 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2650 /*
2651 * There are two different scenarios here:
2652 * 1. The stripe has some data cached, and it is sent to
2653 * write-out phase for reclaim
2654 * 2. The stripe is clean, and this is the first write
2655 *
2656 * For 1, return -EAGAIN, so we continue with
2657 * handle_stripe_dirtying().
2658 *
2659 * For 2, set STRIPE_R5C_CACHING and continue with caching
2660 * write.
2661 */
2662
2663 /* case 1: anything injournal or anything in written */
2664 if (s->injournal > 0 || s->written > 0)
2665 return -EAGAIN;
2666 /* case 2 */
2667 set_bit(STRIPE_R5C_CACHING, &sh->state);
2668 }
2669
2670 /*
2671 * When run in degraded mode, array is set to write-through mode.
2672 * This check helps drain pending write safely in the transition to
2673 * write-through mode.
2674 *
2675 * When a stripe is syncing, the write is also handled in write
2676 * through mode.
2677 */
2678 if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2679 r5c_make_stripe_write_out(sh);
2680 return -EAGAIN;
2681 }
2682
2683 for (i = disks; i--; ) {
2684 dev = &sh->dev[i];
2685 /* if non-overwrite, use writing-out phase */
2686 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2687 !test_bit(R5_InJournal, &dev->flags)) {
2688 r5c_make_stripe_write_out(sh);
2689 return -EAGAIN;
2690 }
2691 }
2692
2693 /* if the stripe is not counted in big_stripe_tree, add it now */
2694 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2695 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2696 tree_index = r5c_tree_index(conf, sh->sector);
2697 spin_lock(&log->tree_lock);
2698 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2699 tree_index);
2700 if (pslot) {
2701 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2702 pslot, &log->tree_lock) >>
2703 R5C_RADIX_COUNT_SHIFT;
2704 radix_tree_replace_slot(
2705 &log->big_stripe_tree, pslot,
2706 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2707 } else {
2708 /*
2709 * this radix_tree_insert can fail safely, so no
2710 * need to call radix_tree_preload()
2711 */
2712 ret = radix_tree_insert(
2713 &log->big_stripe_tree, tree_index,
2714 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2715 if (ret) {
2716 spin_unlock(&log->tree_lock);
2717 r5c_make_stripe_write_out(sh);
2718 return -EAGAIN;
2719 }
2720 }
2721 spin_unlock(&log->tree_lock);
2722
2723 /*
2724 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2725 * counted in the radix tree
2726 */
2727 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2728 atomic_inc(&conf->r5c_cached_partial_stripes);
2729 }
2730
2731 for (i = disks; i--; ) {
2732 dev = &sh->dev[i];
2733 if (dev->towrite) {
2734 set_bit(R5_Wantwrite, &dev->flags);
2735 set_bit(R5_Wantdrain, &dev->flags);
2736 set_bit(R5_LOCKED, &dev->flags);
2737 to_cache++;
2738 }
2739 }
2740
2741 if (to_cache) {
2742 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2743 /*
2744 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2745 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2746 * r5c_handle_data_cached()
2747 */
2748 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2749 }
2750
2751 return 0;
2752}
2753
2754/*
2755 * free extra pages (orig_page) we allocated for prexor
2756 */
2757void r5c_release_extra_page(struct stripe_head *sh)
2758{
2759 struct r5conf *conf = sh->raid_conf;
2760 int i;
2761 bool using_disk_info_extra_page;
2762
2763 using_disk_info_extra_page =
2764 sh->dev[0].orig_page == conf->disks[0].extra_page;
2765
2766 for (i = sh->disks; i--; )
2767 if (sh->dev[i].page != sh->dev[i].orig_page) {
2768 struct page *p = sh->dev[i].orig_page;
2769
2770 sh->dev[i].orig_page = sh->dev[i].page;
2771 clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2772
2773 if (!using_disk_info_extra_page)
2774 put_page(p);
2775 }
2776
2777 if (using_disk_info_extra_page) {
2778 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2779 md_wakeup_thread(conf->mddev->thread);
2780 }
2781}
2782
2783void r5c_use_extra_page(struct stripe_head *sh)
2784{
2785 struct r5conf *conf = sh->raid_conf;
2786 int i;
2787 struct r5dev *dev;
2788
2789 for (i = sh->disks; i--; ) {
2790 dev = &sh->dev[i];
2791 if (dev->orig_page != dev->page)
2792 put_page(dev->orig_page);
2793 dev->orig_page = conf->disks[i].extra_page;
2794 }
2795}
2796
2797/*
2798 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2799 * stripe is committed to RAID disks.
2800 */
2801void r5c_finish_stripe_write_out(struct r5conf *conf,
2802 struct stripe_head *sh,
2803 struct stripe_head_state *s)
2804{
2805 struct r5l_log *log = conf->log;
2806 int i;
2807 int do_wakeup = 0;
2808 sector_t tree_index;
2809 void __rcu **pslot;
2810 uintptr_t refcount;
2811
2812 if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2813 return;
2814
2815 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2816 clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2817
2818 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2819 return;
2820
2821 for (i = sh->disks; i--; ) {
2822 clear_bit(R5_InJournal, &sh->dev[i].flags);
2823 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2824 do_wakeup = 1;
2825 }
2826
2827 /*
2828 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2829 * We updated R5_InJournal, so we also update s->injournal.
2830 */
2831 s->injournal = 0;
2832
2833 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2834 if (atomic_dec_and_test(&conf->pending_full_writes))
2835 md_wakeup_thread(conf->mddev->thread);
2836
2837 if (do_wakeup)
2838 wake_up(&conf->wait_for_overlap);
2839
2840 spin_lock_irq(&log->stripe_in_journal_lock);
2841 list_del_init(&sh->r5c);
2842 spin_unlock_irq(&log->stripe_in_journal_lock);
2843 sh->log_start = MaxSector;
2844
2845 atomic_dec(&log->stripe_in_journal_count);
2846 r5c_update_log_state(log);
2847
2848 /* stop counting this stripe in big_stripe_tree */
2849 if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2850 test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2851 tree_index = r5c_tree_index(conf, sh->sector);
2852 spin_lock(&log->tree_lock);
2853 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2854 tree_index);
2855 BUG_ON(pslot == NULL);
2856 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2857 pslot, &log->tree_lock) >>
2858 R5C_RADIX_COUNT_SHIFT;
2859 if (refcount == 1)
2860 radix_tree_delete(&log->big_stripe_tree, tree_index);
2861 else
2862 radix_tree_replace_slot(
2863 &log->big_stripe_tree, pslot,
2864 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2865 spin_unlock(&log->tree_lock);
2866 }
2867
2868 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2869 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2870 atomic_dec(&conf->r5c_flushing_partial_stripes);
2871 atomic_dec(&conf->r5c_cached_partial_stripes);
2872 }
2873
2874 if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2875 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2876 atomic_dec(&conf->r5c_flushing_full_stripes);
2877 atomic_dec(&conf->r5c_cached_full_stripes);
2878 }
2879
2880 r5l_append_flush_payload(log, sh->sector);
2881 /* stripe is flused to raid disks, we can do resync now */
2882 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2883 set_bit(STRIPE_HANDLE, &sh->state);
2884}
2885
2886int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2887{
2888 struct r5conf *conf = sh->raid_conf;
2889 int pages = 0;
2890 int reserve;
2891 int i;
2892 int ret = 0;
2893
2894 BUG_ON(!log);
2895
2896 for (i = 0; i < sh->disks; i++) {
2897 void *addr;
2898
2899 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2900 continue;
2901 addr = kmap_atomic(sh->dev[i].page);
2902 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2903 addr, PAGE_SIZE);
2904 kunmap_atomic(addr);
2905 pages++;
2906 }
2907 WARN_ON(pages == 0);
2908
2909 /*
2910 * The stripe must enter state machine again to call endio, so
2911 * don't delay.
2912 */
2913 clear_bit(STRIPE_DELAYED, &sh->state);
2914 atomic_inc(&sh->count);
2915
2916 mutex_lock(&log->io_mutex);
2917 /* meta + data */
2918 reserve = (1 + pages) << (PAGE_SHIFT - 9);
2919
2920 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2921 sh->log_start == MaxSector)
2922 r5l_add_no_space_stripe(log, sh);
2923 else if (!r5l_has_free_space(log, reserve)) {
2924 if (sh->log_start == log->last_checkpoint)
2925 BUG();
2926 else
2927 r5l_add_no_space_stripe(log, sh);
2928 } else {
2929 ret = r5l_log_stripe(log, sh, pages, 0);
2930 if (ret) {
2931 spin_lock_irq(&log->io_list_lock);
2932 list_add_tail(&sh->log_list, &log->no_mem_stripes);
2933 spin_unlock_irq(&log->io_list_lock);
2934 }
2935 }
2936
2937 mutex_unlock(&log->io_mutex);
2938 return 0;
2939}
2940
2941/* check whether this big stripe is in write back cache. */
2942bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2943{
2944 struct r5l_log *log = conf->log;
2945 sector_t tree_index;
2946 void *slot;
2947
2948 if (!log)
2949 return false;
2950
2951 WARN_ON_ONCE(!rcu_read_lock_held());
2952 tree_index = r5c_tree_index(conf, sect);
2953 slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2954 return slot != NULL;
2955}
2956
2957static int r5l_load_log(struct r5l_log *log)
2958{
2959 struct md_rdev *rdev = log->rdev;
2960 struct page *page;
2961 struct r5l_meta_block *mb;
2962 sector_t cp = log->rdev->journal_tail;
2963 u32 stored_crc, expected_crc;
2964 bool create_super = false;
2965 int ret = 0;
2966
2967 /* Make sure it's valid */
2968 if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2969 cp = 0;
2970 page = alloc_page(GFP_KERNEL);
2971 if (!page)
2972 return -ENOMEM;
2973
2974 if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, false)) {
2975 ret = -EIO;
2976 goto ioerr;
2977 }
2978 mb = page_address(page);
2979
2980 if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2981 mb->version != R5LOG_VERSION) {
2982 create_super = true;
2983 goto create;
2984 }
2985 stored_crc = le32_to_cpu(mb->checksum);
2986 mb->checksum = 0;
2987 expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2988 if (stored_crc != expected_crc) {
2989 create_super = true;
2990 goto create;
2991 }
2992 if (le64_to_cpu(mb->position) != cp) {
2993 create_super = true;
2994 goto create;
2995 }
2996create:
2997 if (create_super) {
2998 log->last_cp_seq = get_random_u32();
2999 cp = 0;
3000 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
3001 /*
3002 * Make sure super points to correct address. Log might have
3003 * data very soon. If super hasn't correct log tail address,
3004 * recovery can't find the log
3005 */
3006 r5l_write_super(log, cp);
3007 } else
3008 log->last_cp_seq = le64_to_cpu(mb->seq);
3009
3010 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3011 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3012 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3013 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3014 log->last_checkpoint = cp;
3015
3016 __free_page(page);
3017
3018 if (create_super) {
3019 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3020 log->seq = log->last_cp_seq + 1;
3021 log->next_checkpoint = cp;
3022 } else
3023 ret = r5l_recovery_log(log);
3024
3025 r5c_update_log_state(log);
3026 return ret;
3027ioerr:
3028 __free_page(page);
3029 return ret;
3030}
3031
3032int r5l_start(struct r5l_log *log)
3033{
3034 int ret;
3035
3036 if (!log)
3037 return 0;
3038
3039 ret = r5l_load_log(log);
3040 if (ret) {
3041 struct mddev *mddev = log->rdev->mddev;
3042 struct r5conf *conf = mddev->private;
3043
3044 r5l_exit_log(conf);
3045 }
3046 return ret;
3047}
3048
3049void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3050{
3051 struct r5conf *conf = mddev->private;
3052 struct r5l_log *log = conf->log;
3053
3054 if (!log)
3055 return;
3056
3057 if ((raid5_calc_degraded(conf) > 0 ||
3058 test_bit(Journal, &rdev->flags)) &&
3059 conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3060 schedule_work(&log->disable_writeback_work);
3061}
3062
3063int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3064{
3065 struct r5l_log *log;
3066 int ret;
3067
3068 pr_debug("md/raid:%s: using device %pg as journal\n",
3069 mdname(conf->mddev), rdev->bdev);
3070
3071 if (PAGE_SIZE != 4096)
3072 return -EINVAL;
3073
3074 /*
3075 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3076 * raid_disks r5l_payload_data_parity.
3077 *
3078 * Write journal and cache does not work for very big array
3079 * (raid_disks > 203)
3080 */
3081 if (sizeof(struct r5l_meta_block) +
3082 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3083 conf->raid_disks) > PAGE_SIZE) {
3084 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3085 mdname(conf->mddev), conf->raid_disks);
3086 return -EINVAL;
3087 }
3088
3089 log = kzalloc(sizeof(*log), GFP_KERNEL);
3090 if (!log)
3091 return -ENOMEM;
3092 log->rdev = rdev;
3093 log->need_cache_flush = bdev_write_cache(rdev->bdev);
3094 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3095 sizeof(rdev->mddev->uuid));
3096
3097 mutex_init(&log->io_mutex);
3098
3099 spin_lock_init(&log->io_list_lock);
3100 INIT_LIST_HEAD(&log->running_ios);
3101 INIT_LIST_HEAD(&log->io_end_ios);
3102 INIT_LIST_HEAD(&log->flushing_ios);
3103 INIT_LIST_HEAD(&log->finished_ios);
3104
3105 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3106 if (!log->io_kc)
3107 goto io_kc;
3108
3109 ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3110 if (ret)
3111 goto io_pool;
3112
3113 ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3114 if (ret)
3115 goto io_bs;
3116
3117 ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3118 if (ret)
3119 goto out_mempool;
3120
3121 spin_lock_init(&log->tree_lock);
3122 INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3123
3124 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3125 log->rdev->mddev, "reclaim");
3126 if (!log->reclaim_thread)
3127 goto reclaim_thread;
3128 log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3129
3130 init_waitqueue_head(&log->iounit_wait);
3131
3132 INIT_LIST_HEAD(&log->no_mem_stripes);
3133
3134 INIT_LIST_HEAD(&log->no_space_stripes);
3135 spin_lock_init(&log->no_space_stripes_lock);
3136
3137 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3138 INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3139
3140 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3141 INIT_LIST_HEAD(&log->stripe_in_journal_list);
3142 spin_lock_init(&log->stripe_in_journal_lock);
3143 atomic_set(&log->stripe_in_journal_count, 0);
3144
3145 conf->log = log;
3146
3147 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3148 return 0;
3149
3150reclaim_thread:
3151 mempool_exit(&log->meta_pool);
3152out_mempool:
3153 bioset_exit(&log->bs);
3154io_bs:
3155 mempool_exit(&log->io_pool);
3156io_pool:
3157 kmem_cache_destroy(log->io_kc);
3158io_kc:
3159 kfree(log);
3160 return -EINVAL;
3161}
3162
3163void r5l_exit_log(struct r5conf *conf)
3164{
3165 struct r5l_log *log = conf->log;
3166
3167 /* Ensure disable_writeback_work wakes up and exits */
3168 wake_up(&conf->mddev->sb_wait);
3169 flush_work(&log->disable_writeback_work);
3170 md_unregister_thread(&log->reclaim_thread);
3171
3172 conf->log = NULL;
3173
3174 mempool_exit(&log->meta_pool);
3175 bioset_exit(&log->bs);
3176 mempool_exit(&log->io_pool);
3177 kmem_cache_destroy(log->io_kc);
3178 kfree(log);
3179}