4c00bc248287e4ab89b492225e0d054973725549
[sfrench/cifs-2.6.git] / drivers / md / raid5-cache.c
1 /*
2  * Copyright (C) 2015 Shaohua Li <shli@fb.com>
3  * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  */
15 #include <linux/kernel.h>
16 #include <linux/wait.h>
17 #include <linux/blkdev.h>
18 #include <linux/slab.h>
19 #include <linux/raid/md_p.h>
20 #include <linux/crc32c.h>
21 #include <linux/random.h>
22 #include <linux/kthread.h>
23 #include <linux/types.h>
24 #include "md.h"
25 #include "raid5.h"
26 #include "bitmap.h"
27 #include "raid5-log.h"
28
29 /*
30  * metadata/data stored in disk with 4k size unit (a block) regardless
31  * underneath hardware sector size. only works with PAGE_SIZE == 4096
32  */
33 #define BLOCK_SECTORS (8)
34 #define BLOCK_SECTOR_SHIFT (3)
35
36 /*
37  * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
38  *
39  * In write through mode, the reclaim runs every log->max_free_space.
40  * This can prevent the recovery scans for too long
41  */
42 #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
43 #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
44
45 /* wake up reclaim thread periodically */
46 #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
47 /* start flush with these full stripes */
48 #define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
49 /* reclaim stripes in groups */
50 #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
51
52 /*
53  * We only need 2 bios per I/O unit to make progress, but ensure we
54  * have a few more available to not get too tight.
55  */
56 #define R5L_POOL_SIZE   4
57
58 static char *r5c_journal_mode_str[] = {"write-through",
59                                        "write-back"};
60 /*
61  * raid5 cache state machine
62  *
63  * With the RAID cache, each stripe works in two phases:
64  *      - caching phase
65  *      - writing-out phase
66  *
67  * These two phases are controlled by bit STRIPE_R5C_CACHING:
68  *   if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
69  *   if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
70  *
71  * When there is no journal, or the journal is in write-through mode,
72  * the stripe is always in writing-out phase.
73  *
74  * For write-back journal, the stripe is sent to caching phase on write
75  * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
76  * the write-out phase by clearing STRIPE_R5C_CACHING.
77  *
78  * Stripes in caching phase do not write the raid disks. Instead, all
79  * writes are committed from the log device. Therefore, a stripe in
80  * caching phase handles writes as:
81  *      - write to log device
82  *      - return IO
83  *
84  * Stripes in writing-out phase handle writes as:
85  *      - calculate parity
86  *      - write pending data and parity to journal
87  *      - write data and parity to raid disks
88  *      - return IO for pending writes
89  */
90
91 struct r5l_log {
92         struct md_rdev *rdev;
93
94         u32 uuid_checksum;
95
96         sector_t device_size;           /* log device size, round to
97                                          * BLOCK_SECTORS */
98         sector_t max_free_space;        /* reclaim run if free space is at
99                                          * this size */
100
101         sector_t last_checkpoint;       /* log tail. where recovery scan
102                                          * starts from */
103         u64 last_cp_seq;                /* log tail sequence */
104
105         sector_t log_start;             /* log head. where new data appends */
106         u64 seq;                        /* log head sequence */
107
108         sector_t next_checkpoint;
109
110         struct mutex io_mutex;
111         struct r5l_io_unit *current_io; /* current io_unit accepting new data */
112
113         spinlock_t io_list_lock;
114         struct list_head running_ios;   /* io_units which are still running,
115                                          * and have not yet been completely
116                                          * written to the log */
117         struct list_head io_end_ios;    /* io_units which have been completely
118                                          * written to the log but not yet written
119                                          * to the RAID */
120         struct list_head flushing_ios;  /* io_units which are waiting for log
121                                          * cache flush */
122         struct list_head finished_ios;  /* io_units which settle down in log disk */
123         struct bio flush_bio;
124
125         struct list_head no_mem_stripes;   /* pending stripes, -ENOMEM */
126
127         struct kmem_cache *io_kc;
128         mempool_t *io_pool;
129         struct bio_set *bs;
130         mempool_t *meta_pool;
131
132         struct md_thread *reclaim_thread;
133         unsigned long reclaim_target;   /* number of space that need to be
134                                          * reclaimed.  if it's 0, reclaim spaces
135                                          * used by io_units which are in
136                                          * IO_UNIT_STRIPE_END state (eg, reclaim
137                                          * dones't wait for specific io_unit
138                                          * switching to IO_UNIT_STRIPE_END
139                                          * state) */
140         wait_queue_head_t iounit_wait;
141
142         struct list_head no_space_stripes; /* pending stripes, log has no space */
143         spinlock_t no_space_stripes_lock;
144
145         bool need_cache_flush;
146
147         /* for r5c_cache */
148         enum r5c_journal_mode r5c_journal_mode;
149
150         /* all stripes in r5cache, in the order of seq at sh->log_start */
151         struct list_head stripe_in_journal_list;
152
153         spinlock_t stripe_in_journal_lock;
154         atomic_t stripe_in_journal_count;
155
156         /* to submit async io_units, to fulfill ordering of flush */
157         struct work_struct deferred_io_work;
158         /* to disable write back during in degraded mode */
159         struct work_struct disable_writeback_work;
160
161         /* to for chunk_aligned_read in writeback mode, details below */
162         spinlock_t tree_lock;
163         struct radix_tree_root big_stripe_tree;
164 };
165
166 /*
167  * Enable chunk_aligned_read() with write back cache.
168  *
169  * Each chunk may contain more than one stripe (for example, a 256kB
170  * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
171  * chunk_aligned_read, these stripes are grouped into one "big_stripe".
172  * For each big_stripe, we count how many stripes of this big_stripe
173  * are in the write back cache. These data are tracked in a radix tree
174  * (big_stripe_tree). We use radix_tree item pointer as the counter.
175  * r5c_tree_index() is used to calculate keys for the radix tree.
176  *
177  * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
178  * big_stripe of each chunk in the tree. If this big_stripe is in the
179  * tree, chunk_aligned_read() aborts. This look up is protected by
180  * rcu_read_lock().
181  *
182  * It is necessary to remember whether a stripe is counted in
183  * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
184  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
185  * two flags are set, the stripe is counted in big_stripe_tree. This
186  * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
187  * r5c_try_caching_write(); and moving clear_bit of
188  * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
189  * r5c_finish_stripe_write_out().
190  */
191
192 /*
193  * radix tree requests lowest 2 bits of data pointer to be 2b'00.
194  * So it is necessary to left shift the counter by 2 bits before using it
195  * as data pointer of the tree.
196  */
197 #define R5C_RADIX_COUNT_SHIFT 2
198
199 /*
200  * calculate key for big_stripe_tree
201  *
202  * sect: align_bi->bi_iter.bi_sector or sh->sector
203  */
204 static inline sector_t r5c_tree_index(struct r5conf *conf,
205                                       sector_t sect)
206 {
207         sector_t offset;
208
209         offset = sector_div(sect, conf->chunk_sectors);
210         return sect;
211 }
212
213 /*
214  * an IO range starts from a meta data block and end at the next meta data
215  * block. The io unit's the meta data block tracks data/parity followed it. io
216  * unit is written to log disk with normal write, as we always flush log disk
217  * first and then start move data to raid disks, there is no requirement to
218  * write io unit with FLUSH/FUA
219  */
220 struct r5l_io_unit {
221         struct r5l_log *log;
222
223         struct page *meta_page; /* store meta block */
224         int meta_offset;        /* current offset in meta_page */
225
226         struct bio *current_bio;/* current_bio accepting new data */
227
228         atomic_t pending_stripe;/* how many stripes not flushed to raid */
229         u64 seq;                /* seq number of the metablock */
230         sector_t log_start;     /* where the io_unit starts */
231         sector_t log_end;       /* where the io_unit ends */
232         struct list_head log_sibling; /* log->running_ios */
233         struct list_head stripe_list; /* stripes added to the io_unit */
234
235         int state;
236         bool need_split_bio;
237         struct bio *split_bio;
238
239         unsigned int has_flush:1;      /* include flush request */
240         unsigned int has_fua:1;        /* include fua request */
241         unsigned int has_null_flush:1; /* include empty flush request */
242         /*
243          * io isn't sent yet, flush/fua request can only be submitted till it's
244          * the first IO in running_ios list
245          */
246         unsigned int io_deferred:1;
247
248         struct bio_list flush_barriers;   /* size == 0 flush bios */
249 };
250
251 /* r5l_io_unit state */
252 enum r5l_io_unit_state {
253         IO_UNIT_RUNNING = 0,    /* accepting new IO */
254         IO_UNIT_IO_START = 1,   /* io_unit bio start writing to log,
255                                  * don't accepting new bio */
256         IO_UNIT_IO_END = 2,     /* io_unit bio finish writing to log */
257         IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
258 };
259
260 bool r5c_is_writeback(struct r5l_log *log)
261 {
262         return (log != NULL &&
263                 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
264 }
265
266 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
267 {
268         start += inc;
269         if (start >= log->device_size)
270                 start = start - log->device_size;
271         return start;
272 }
273
274 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
275                                   sector_t end)
276 {
277         if (end >= start)
278                 return end - start;
279         else
280                 return end + log->device_size - start;
281 }
282
283 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
284 {
285         sector_t used_size;
286
287         used_size = r5l_ring_distance(log, log->last_checkpoint,
288                                         log->log_start);
289
290         return log->device_size > used_size + size;
291 }
292
293 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
294                                     enum r5l_io_unit_state state)
295 {
296         if (WARN_ON(io->state >= state))
297                 return;
298         io->state = state;
299 }
300
301 static void
302 r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
303 {
304         struct bio *wbi, *wbi2;
305
306         wbi = dev->written;
307         dev->written = NULL;
308         while (wbi && wbi->bi_iter.bi_sector <
309                dev->sector + STRIPE_SECTORS) {
310                 wbi2 = r5_next_bio(wbi, dev->sector);
311                 md_write_end(conf->mddev);
312                 bio_endio(wbi);
313                 wbi = wbi2;
314         }
315 }
316
317 void r5c_handle_cached_data_endio(struct r5conf *conf,
318                                   struct stripe_head *sh, int disks)
319 {
320         int i;
321
322         for (i = sh->disks; i--; ) {
323                 if (sh->dev[i].written) {
324                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
325                         r5c_return_dev_pending_writes(conf, &sh->dev[i]);
326                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
327                                         STRIPE_SECTORS,
328                                         !test_bit(STRIPE_DEGRADED, &sh->state),
329                                         0);
330                 }
331         }
332 }
333
334 void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
335
336 /* Check whether we should flush some stripes to free up stripe cache */
337 void r5c_check_stripe_cache_usage(struct r5conf *conf)
338 {
339         int total_cached;
340
341         if (!r5c_is_writeback(conf->log))
342                 return;
343
344         total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
345                 atomic_read(&conf->r5c_cached_full_stripes);
346
347         /*
348          * The following condition is true for either of the following:
349          *   - stripe cache pressure high:
350          *          total_cached > 3/4 min_nr_stripes ||
351          *          empty_inactive_list_nr > 0
352          *   - stripe cache pressure moderate:
353          *          total_cached > 1/2 min_nr_stripes
354          */
355         if (total_cached > conf->min_nr_stripes * 1 / 2 ||
356             atomic_read(&conf->empty_inactive_list_nr) > 0)
357                 r5l_wake_reclaim(conf->log, 0);
358 }
359
360 /*
361  * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
362  * stripes in the cache
363  */
364 void r5c_check_cached_full_stripe(struct r5conf *conf)
365 {
366         if (!r5c_is_writeback(conf->log))
367                 return;
368
369         /*
370          * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
371          * or a full stripe (chunk size / 4k stripes).
372          */
373         if (atomic_read(&conf->r5c_cached_full_stripes) >=
374             min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
375                 conf->chunk_sectors >> STRIPE_SHIFT))
376                 r5l_wake_reclaim(conf->log, 0);
377 }
378
379 /*
380  * Total log space (in sectors) needed to flush all data in cache
381  *
382  * To avoid deadlock due to log space, it is necessary to reserve log
383  * space to flush critical stripes (stripes that occupying log space near
384  * last_checkpoint). This function helps check how much log space is
385  * required to flush all cached stripes.
386  *
387  * To reduce log space requirements, two mechanisms are used to give cache
388  * flush higher priorities:
389  *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
390  *       stripes ALREADY in journal can be flushed w/o pending writes;
391  *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
392  *       can be delayed (r5l_add_no_space_stripe).
393  *
394  * In cache flush, the stripe goes through 1 and then 2. For a stripe that
395  * already passed 1, flushing it requires at most (conf->max_degraded + 1)
396  * pages of journal space. For stripes that has not passed 1, flushing it
397  * requires (conf->raid_disks + 1) pages of journal space. There are at
398  * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
399  * required to flush all cached stripes (in pages) is:
400  *
401  *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
402  *     (group_cnt + 1) * (raid_disks + 1)
403  * or
404  *     (stripe_in_journal_count) * (max_degraded + 1) +
405  *     (group_cnt + 1) * (raid_disks - max_degraded)
406  */
407 static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
408 {
409         struct r5l_log *log = conf->log;
410
411         if (!r5c_is_writeback(log))
412                 return 0;
413
414         return BLOCK_SECTORS *
415                 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
416                  (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
417 }
418
419 /*
420  * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
421  *
422  * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
423  * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
424  * device is less than 2x of reclaim_required_space.
425  */
426 static inline void r5c_update_log_state(struct r5l_log *log)
427 {
428         struct r5conf *conf = log->rdev->mddev->private;
429         sector_t free_space;
430         sector_t reclaim_space;
431         bool wake_reclaim = false;
432
433         if (!r5c_is_writeback(log))
434                 return;
435
436         free_space = r5l_ring_distance(log, log->log_start,
437                                        log->last_checkpoint);
438         reclaim_space = r5c_log_required_to_flush_cache(conf);
439         if (free_space < 2 * reclaim_space)
440                 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
441         else {
442                 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
443                         wake_reclaim = true;
444                 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
445         }
446         if (free_space < 3 * reclaim_space)
447                 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
448         else
449                 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
450
451         if (wake_reclaim)
452                 r5l_wake_reclaim(log, 0);
453 }
454
455 /*
456  * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
457  * This function should only be called in write-back mode.
458  */
459 void r5c_make_stripe_write_out(struct stripe_head *sh)
460 {
461         struct r5conf *conf = sh->raid_conf;
462         struct r5l_log *log = conf->log;
463
464         BUG_ON(!r5c_is_writeback(log));
465
466         WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
467         clear_bit(STRIPE_R5C_CACHING, &sh->state);
468
469         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
470                 atomic_inc(&conf->preread_active_stripes);
471 }
472
473 static void r5c_handle_data_cached(struct stripe_head *sh)
474 {
475         int i;
476
477         for (i = sh->disks; i--; )
478                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
479                         set_bit(R5_InJournal, &sh->dev[i].flags);
480                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
481                 }
482         clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
483 }
484
485 /*
486  * this journal write must contain full parity,
487  * it may also contain some data pages
488  */
489 static void r5c_handle_parity_cached(struct stripe_head *sh)
490 {
491         int i;
492
493         for (i = sh->disks; i--; )
494                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
495                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
496 }
497
498 /*
499  * Setting proper flags after writing (or flushing) data and/or parity to the
500  * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
501  */
502 static void r5c_finish_cache_stripe(struct stripe_head *sh)
503 {
504         struct r5l_log *log = sh->raid_conf->log;
505
506         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
507                 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
508                 /*
509                  * Set R5_InJournal for parity dev[pd_idx]. This means
510                  * all data AND parity in the journal. For RAID 6, it is
511                  * NOT necessary to set the flag for dev[qd_idx], as the
512                  * two parities are written out together.
513                  */
514                 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
515         } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
516                 r5c_handle_data_cached(sh);
517         } else {
518                 r5c_handle_parity_cached(sh);
519                 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
520         }
521 }
522
523 static void r5l_io_run_stripes(struct r5l_io_unit *io)
524 {
525         struct stripe_head *sh, *next;
526
527         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
528                 list_del_init(&sh->log_list);
529
530                 r5c_finish_cache_stripe(sh);
531
532                 set_bit(STRIPE_HANDLE, &sh->state);
533                 raid5_release_stripe(sh);
534         }
535 }
536
537 static void r5l_log_run_stripes(struct r5l_log *log)
538 {
539         struct r5l_io_unit *io, *next;
540
541         assert_spin_locked(&log->io_list_lock);
542
543         list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
544                 /* don't change list order */
545                 if (io->state < IO_UNIT_IO_END)
546                         break;
547
548                 list_move_tail(&io->log_sibling, &log->finished_ios);
549                 r5l_io_run_stripes(io);
550         }
551 }
552
553 static void r5l_move_to_end_ios(struct r5l_log *log)
554 {
555         struct r5l_io_unit *io, *next;
556
557         assert_spin_locked(&log->io_list_lock);
558
559         list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
560                 /* don't change list order */
561                 if (io->state < IO_UNIT_IO_END)
562                         break;
563                 list_move_tail(&io->log_sibling, &log->io_end_ios);
564         }
565 }
566
567 static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
568 static void r5l_log_endio(struct bio *bio)
569 {
570         struct r5l_io_unit *io = bio->bi_private;
571         struct r5l_io_unit *io_deferred;
572         struct r5l_log *log = io->log;
573         unsigned long flags;
574
575         if (bio->bi_error)
576                 md_error(log->rdev->mddev, log->rdev);
577
578         bio_put(bio);
579         mempool_free(io->meta_page, log->meta_pool);
580
581         spin_lock_irqsave(&log->io_list_lock, flags);
582         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
583         if (log->need_cache_flush && !list_empty(&io->stripe_list))
584                 r5l_move_to_end_ios(log);
585         else
586                 r5l_log_run_stripes(log);
587         if (!list_empty(&log->running_ios)) {
588                 /*
589                  * FLUSH/FUA io_unit is deferred because of ordering, now we
590                  * can dispatch it
591                  */
592                 io_deferred = list_first_entry(&log->running_ios,
593                                                struct r5l_io_unit, log_sibling);
594                 if (io_deferred->io_deferred)
595                         schedule_work(&log->deferred_io_work);
596         }
597
598         spin_unlock_irqrestore(&log->io_list_lock, flags);
599
600         if (log->need_cache_flush)
601                 md_wakeup_thread(log->rdev->mddev->thread);
602
603         if (io->has_null_flush) {
604                 struct bio *bi;
605
606                 WARN_ON(bio_list_empty(&io->flush_barriers));
607                 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
608                         bio_endio(bi);
609                         atomic_dec(&io->pending_stripe);
610                 }
611         }
612
613         /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
614         if (atomic_read(&io->pending_stripe) == 0)
615                 __r5l_stripe_write_finished(io);
616 }
617
618 static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
619 {
620         unsigned long flags;
621
622         spin_lock_irqsave(&log->io_list_lock, flags);
623         __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
624         spin_unlock_irqrestore(&log->io_list_lock, flags);
625
626         /*
627          * In case of journal device failures, submit_bio will get error
628          * and calls endio, then active stripes will continue write
629          * process. Therefore, it is not necessary to check Faulty bit
630          * of journal device here.
631          *
632          * We can't check split_bio after current_bio is submitted. If
633          * io->split_bio is null, after current_bio is submitted, current_bio
634          * might already be completed and the io_unit is freed. We submit
635          * split_bio first to avoid the issue.
636          */
637         if (io->split_bio) {
638                 if (io->has_flush)
639                         io->split_bio->bi_opf |= REQ_PREFLUSH;
640                 if (io->has_fua)
641                         io->split_bio->bi_opf |= REQ_FUA;
642                 submit_bio(io->split_bio);
643         }
644
645         if (io->has_flush)
646                 io->current_bio->bi_opf |= REQ_PREFLUSH;
647         if (io->has_fua)
648                 io->current_bio->bi_opf |= REQ_FUA;
649         submit_bio(io->current_bio);
650 }
651
652 /* deferred io_unit will be dispatched here */
653 static void r5l_submit_io_async(struct work_struct *work)
654 {
655         struct r5l_log *log = container_of(work, struct r5l_log,
656                                            deferred_io_work);
657         struct r5l_io_unit *io = NULL;
658         unsigned long flags;
659
660         spin_lock_irqsave(&log->io_list_lock, flags);
661         if (!list_empty(&log->running_ios)) {
662                 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
663                                       log_sibling);
664                 if (!io->io_deferred)
665                         io = NULL;
666                 else
667                         io->io_deferred = 0;
668         }
669         spin_unlock_irqrestore(&log->io_list_lock, flags);
670         if (io)
671                 r5l_do_submit_io(log, io);
672 }
673
674 static void r5c_disable_writeback_async(struct work_struct *work)
675 {
676         struct r5l_log *log = container_of(work, struct r5l_log,
677                                            disable_writeback_work);
678         struct mddev *mddev = log->rdev->mddev;
679
680         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
681                 return;
682         pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
683                 mdname(mddev));
684
685         /* wait superblock change before suspend */
686         wait_event(mddev->sb_wait,
687                    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
688
689         mddev_suspend(mddev);
690         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
691         mddev_resume(mddev);
692 }
693
694 static void r5l_submit_current_io(struct r5l_log *log)
695 {
696         struct r5l_io_unit *io = log->current_io;
697         struct bio *bio;
698         struct r5l_meta_block *block;
699         unsigned long flags;
700         u32 crc;
701         bool do_submit = true;
702
703         if (!io)
704                 return;
705
706         block = page_address(io->meta_page);
707         block->meta_size = cpu_to_le32(io->meta_offset);
708         crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
709         block->checksum = cpu_to_le32(crc);
710         bio = io->current_bio;
711
712         log->current_io = NULL;
713         spin_lock_irqsave(&log->io_list_lock, flags);
714         if (io->has_flush || io->has_fua) {
715                 if (io != list_first_entry(&log->running_ios,
716                                            struct r5l_io_unit, log_sibling)) {
717                         io->io_deferred = 1;
718                         do_submit = false;
719                 }
720         }
721         spin_unlock_irqrestore(&log->io_list_lock, flags);
722         if (do_submit)
723                 r5l_do_submit_io(log, io);
724 }
725
726 static struct bio *r5l_bio_alloc(struct r5l_log *log)
727 {
728         struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
729
730         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
731         bio->bi_bdev = log->rdev->bdev;
732         bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
733
734         return bio;
735 }
736
737 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
738 {
739         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
740
741         r5c_update_log_state(log);
742         /*
743          * If we filled up the log device start from the beginning again,
744          * which will require a new bio.
745          *
746          * Note: for this to work properly the log size needs to me a multiple
747          * of BLOCK_SECTORS.
748          */
749         if (log->log_start == 0)
750                 io->need_split_bio = true;
751
752         io->log_end = log->log_start;
753 }
754
755 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
756 {
757         struct r5l_io_unit *io;
758         struct r5l_meta_block *block;
759
760         io = mempool_alloc(log->io_pool, GFP_ATOMIC);
761         if (!io)
762                 return NULL;
763         memset(io, 0, sizeof(*io));
764
765         io->log = log;
766         INIT_LIST_HEAD(&io->log_sibling);
767         INIT_LIST_HEAD(&io->stripe_list);
768         bio_list_init(&io->flush_barriers);
769         io->state = IO_UNIT_RUNNING;
770
771         io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
772         block = page_address(io->meta_page);
773         clear_page(block);
774         block->magic = cpu_to_le32(R5LOG_MAGIC);
775         block->version = R5LOG_VERSION;
776         block->seq = cpu_to_le64(log->seq);
777         block->position = cpu_to_le64(log->log_start);
778
779         io->log_start = log->log_start;
780         io->meta_offset = sizeof(struct r5l_meta_block);
781         io->seq = log->seq++;
782
783         io->current_bio = r5l_bio_alloc(log);
784         io->current_bio->bi_end_io = r5l_log_endio;
785         io->current_bio->bi_private = io;
786         bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
787
788         r5_reserve_log_entry(log, io);
789
790         spin_lock_irq(&log->io_list_lock);
791         list_add_tail(&io->log_sibling, &log->running_ios);
792         spin_unlock_irq(&log->io_list_lock);
793
794         return io;
795 }
796
797 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
798 {
799         if (log->current_io &&
800             log->current_io->meta_offset + payload_size > PAGE_SIZE)
801                 r5l_submit_current_io(log);
802
803         if (!log->current_io) {
804                 log->current_io = r5l_new_meta(log);
805                 if (!log->current_io)
806                         return -ENOMEM;
807         }
808
809         return 0;
810 }
811
812 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
813                                     sector_t location,
814                                     u32 checksum1, u32 checksum2,
815                                     bool checksum2_valid)
816 {
817         struct r5l_io_unit *io = log->current_io;
818         struct r5l_payload_data_parity *payload;
819
820         payload = page_address(io->meta_page) + io->meta_offset;
821         payload->header.type = cpu_to_le16(type);
822         payload->header.flags = cpu_to_le16(0);
823         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
824                                     (PAGE_SHIFT - 9));
825         payload->location = cpu_to_le64(location);
826         payload->checksum[0] = cpu_to_le32(checksum1);
827         if (checksum2_valid)
828                 payload->checksum[1] = cpu_to_le32(checksum2);
829
830         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
831                 sizeof(__le32) * (1 + !!checksum2_valid);
832 }
833
834 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
835 {
836         struct r5l_io_unit *io = log->current_io;
837
838         if (io->need_split_bio) {
839                 BUG_ON(io->split_bio);
840                 io->split_bio = io->current_bio;
841                 io->current_bio = r5l_bio_alloc(log);
842                 bio_chain(io->current_bio, io->split_bio);
843                 io->need_split_bio = false;
844         }
845
846         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
847                 BUG();
848
849         r5_reserve_log_entry(log, io);
850 }
851
852 static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
853 {
854         struct mddev *mddev = log->rdev->mddev;
855         struct r5conf *conf = mddev->private;
856         struct r5l_io_unit *io;
857         struct r5l_payload_flush *payload;
858         int meta_size;
859
860         /*
861          * payload_flush requires extra writes to the journal.
862          * To avoid handling the extra IO in quiesce, just skip
863          * flush_payload
864          */
865         if (conf->quiesce)
866                 return;
867
868         mutex_lock(&log->io_mutex);
869         meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
870
871         if (r5l_get_meta(log, meta_size)) {
872                 mutex_unlock(&log->io_mutex);
873                 return;
874         }
875
876         /* current implementation is one stripe per flush payload */
877         io = log->current_io;
878         payload = page_address(io->meta_page) + io->meta_offset;
879         payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
880         payload->header.flags = cpu_to_le16(0);
881         payload->size = cpu_to_le32(sizeof(__le64));
882         payload->flush_stripes[0] = cpu_to_le64(sect);
883         io->meta_offset += meta_size;
884         mutex_unlock(&log->io_mutex);
885 }
886
887 static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
888                            int data_pages, int parity_pages)
889 {
890         int i;
891         int meta_size;
892         int ret;
893         struct r5l_io_unit *io;
894
895         meta_size =
896                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
897                  * data_pages) +
898                 sizeof(struct r5l_payload_data_parity) +
899                 sizeof(__le32) * parity_pages;
900
901         ret = r5l_get_meta(log, meta_size);
902         if (ret)
903                 return ret;
904
905         io = log->current_io;
906
907         if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
908                 io->has_flush = 1;
909
910         for (i = 0; i < sh->disks; i++) {
911                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
912                     test_bit(R5_InJournal, &sh->dev[i].flags))
913                         continue;
914                 if (i == sh->pd_idx || i == sh->qd_idx)
915                         continue;
916                 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
917                     log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
918                         io->has_fua = 1;
919                         /*
920                          * we need to flush journal to make sure recovery can
921                          * reach the data with fua flag
922                          */
923                         io->has_flush = 1;
924                 }
925                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
926                                         raid5_compute_blocknr(sh, i, 0),
927                                         sh->dev[i].log_checksum, 0, false);
928                 r5l_append_payload_page(log, sh->dev[i].page);
929         }
930
931         if (parity_pages == 2) {
932                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
933                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
934                                         sh->dev[sh->qd_idx].log_checksum, true);
935                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
936                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
937         } else if (parity_pages == 1) {
938                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
939                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
940                                         0, false);
941                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
942         } else  /* Just writing data, not parity, in caching phase */
943                 BUG_ON(parity_pages != 0);
944
945         list_add_tail(&sh->log_list, &io->stripe_list);
946         atomic_inc(&io->pending_stripe);
947         sh->log_io = io;
948
949         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
950                 return 0;
951
952         if (sh->log_start == MaxSector) {
953                 BUG_ON(!list_empty(&sh->r5c));
954                 sh->log_start = io->log_start;
955                 spin_lock_irq(&log->stripe_in_journal_lock);
956                 list_add_tail(&sh->r5c,
957                               &log->stripe_in_journal_list);
958                 spin_unlock_irq(&log->stripe_in_journal_lock);
959                 atomic_inc(&log->stripe_in_journal_count);
960         }
961         return 0;
962 }
963
964 /* add stripe to no_space_stripes, and then wake up reclaim */
965 static inline void r5l_add_no_space_stripe(struct r5l_log *log,
966                                            struct stripe_head *sh)
967 {
968         spin_lock(&log->no_space_stripes_lock);
969         list_add_tail(&sh->log_list, &log->no_space_stripes);
970         spin_unlock(&log->no_space_stripes_lock);
971 }
972
973 /*
974  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
975  * data from log to raid disks), so we shouldn't wait for reclaim here
976  */
977 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
978 {
979         struct r5conf *conf = sh->raid_conf;
980         int write_disks = 0;
981         int data_pages, parity_pages;
982         int reserve;
983         int i;
984         int ret = 0;
985         bool wake_reclaim = false;
986
987         if (!log)
988                 return -EAGAIN;
989         /* Don't support stripe batch */
990         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
991             test_bit(STRIPE_SYNCING, &sh->state)) {
992                 /* the stripe is written to log, we start writing it to raid */
993                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
994                 return -EAGAIN;
995         }
996
997         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
998
999         for (i = 0; i < sh->disks; i++) {
1000                 void *addr;
1001
1002                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1003                     test_bit(R5_InJournal, &sh->dev[i].flags))
1004                         continue;
1005
1006                 write_disks++;
1007                 /* checksum is already calculated in last run */
1008                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1009                         continue;
1010                 addr = kmap_atomic(sh->dev[i].page);
1011                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1012                                                     addr, PAGE_SIZE);
1013                 kunmap_atomic(addr);
1014         }
1015         parity_pages = 1 + !!(sh->qd_idx >= 0);
1016         data_pages = write_disks - parity_pages;
1017
1018         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1019         /*
1020          * The stripe must enter state machine again to finish the write, so
1021          * don't delay.
1022          */
1023         clear_bit(STRIPE_DELAYED, &sh->state);
1024         atomic_inc(&sh->count);
1025
1026         mutex_lock(&log->io_mutex);
1027         /* meta + data */
1028         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1029
1030         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1031                 if (!r5l_has_free_space(log, reserve)) {
1032                         r5l_add_no_space_stripe(log, sh);
1033                         wake_reclaim = true;
1034                 } else {
1035                         ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1036                         if (ret) {
1037                                 spin_lock_irq(&log->io_list_lock);
1038                                 list_add_tail(&sh->log_list,
1039                                               &log->no_mem_stripes);
1040                                 spin_unlock_irq(&log->io_list_lock);
1041                         }
1042                 }
1043         } else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1044                 /*
1045                  * log space critical, do not process stripes that are
1046                  * not in cache yet (sh->log_start == MaxSector).
1047                  */
1048                 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1049                     sh->log_start == MaxSector) {
1050                         r5l_add_no_space_stripe(log, sh);
1051                         wake_reclaim = true;
1052                         reserve = 0;
1053                 } else if (!r5l_has_free_space(log, reserve)) {
1054                         if (sh->log_start == log->last_checkpoint)
1055                                 BUG();
1056                         else
1057                                 r5l_add_no_space_stripe(log, sh);
1058                 } else {
1059                         ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1060                         if (ret) {
1061                                 spin_lock_irq(&log->io_list_lock);
1062                                 list_add_tail(&sh->log_list,
1063                                               &log->no_mem_stripes);
1064                                 spin_unlock_irq(&log->io_list_lock);
1065                         }
1066                 }
1067         }
1068
1069         mutex_unlock(&log->io_mutex);
1070         if (wake_reclaim)
1071                 r5l_wake_reclaim(log, reserve);
1072         return 0;
1073 }
1074
1075 void r5l_write_stripe_run(struct r5l_log *log)
1076 {
1077         if (!log)
1078                 return;
1079         mutex_lock(&log->io_mutex);
1080         r5l_submit_current_io(log);
1081         mutex_unlock(&log->io_mutex);
1082 }
1083
1084 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1085 {
1086         if (!log)
1087                 return -ENODEV;
1088
1089         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1090                 /*
1091                  * in write through (journal only)
1092                  * we flush log disk cache first, then write stripe data to
1093                  * raid disks. So if bio is finished, the log disk cache is
1094                  * flushed already. The recovery guarantees we can recovery
1095                  * the bio from log disk, so we don't need to flush again
1096                  */
1097                 if (bio->bi_iter.bi_size == 0) {
1098                         bio_endio(bio);
1099                         return 0;
1100                 }
1101                 bio->bi_opf &= ~REQ_PREFLUSH;
1102         } else {
1103                 /* write back (with cache) */
1104                 if (bio->bi_iter.bi_size == 0) {
1105                         mutex_lock(&log->io_mutex);
1106                         r5l_get_meta(log, 0);
1107                         bio_list_add(&log->current_io->flush_barriers, bio);
1108                         log->current_io->has_flush = 1;
1109                         log->current_io->has_null_flush = 1;
1110                         atomic_inc(&log->current_io->pending_stripe);
1111                         r5l_submit_current_io(log);
1112                         mutex_unlock(&log->io_mutex);
1113                         return 0;
1114                 }
1115         }
1116         return -EAGAIN;
1117 }
1118
1119 /* This will run after log space is reclaimed */
1120 static void r5l_run_no_space_stripes(struct r5l_log *log)
1121 {
1122         struct stripe_head *sh;
1123
1124         spin_lock(&log->no_space_stripes_lock);
1125         while (!list_empty(&log->no_space_stripes)) {
1126                 sh = list_first_entry(&log->no_space_stripes,
1127                                       struct stripe_head, log_list);
1128                 list_del_init(&sh->log_list);
1129                 set_bit(STRIPE_HANDLE, &sh->state);
1130                 raid5_release_stripe(sh);
1131         }
1132         spin_unlock(&log->no_space_stripes_lock);
1133 }
1134
1135 /*
1136  * calculate new last_checkpoint
1137  * for write through mode, returns log->next_checkpoint
1138  * for write back, returns log_start of first sh in stripe_in_journal_list
1139  */
1140 static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1141 {
1142         struct stripe_head *sh;
1143         struct r5l_log *log = conf->log;
1144         sector_t new_cp;
1145         unsigned long flags;
1146
1147         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1148                 return log->next_checkpoint;
1149
1150         spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1151         if (list_empty(&conf->log->stripe_in_journal_list)) {
1152                 /* all stripes flushed */
1153                 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1154                 return log->next_checkpoint;
1155         }
1156         sh = list_first_entry(&conf->log->stripe_in_journal_list,
1157                               struct stripe_head, r5c);
1158         new_cp = sh->log_start;
1159         spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1160         return new_cp;
1161 }
1162
1163 static sector_t r5l_reclaimable_space(struct r5l_log *log)
1164 {
1165         struct r5conf *conf = log->rdev->mddev->private;
1166
1167         return r5l_ring_distance(log, log->last_checkpoint,
1168                                  r5c_calculate_new_cp(conf));
1169 }
1170
1171 static void r5l_run_no_mem_stripe(struct r5l_log *log)
1172 {
1173         struct stripe_head *sh;
1174
1175         assert_spin_locked(&log->io_list_lock);
1176
1177         if (!list_empty(&log->no_mem_stripes)) {
1178                 sh = list_first_entry(&log->no_mem_stripes,
1179                                       struct stripe_head, log_list);
1180                 list_del_init(&sh->log_list);
1181                 set_bit(STRIPE_HANDLE, &sh->state);
1182                 raid5_release_stripe(sh);
1183         }
1184 }
1185
1186 static bool r5l_complete_finished_ios(struct r5l_log *log)
1187 {
1188         struct r5l_io_unit *io, *next;
1189         bool found = false;
1190
1191         assert_spin_locked(&log->io_list_lock);
1192
1193         list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1194                 /* don't change list order */
1195                 if (io->state < IO_UNIT_STRIPE_END)
1196                         break;
1197
1198                 log->next_checkpoint = io->log_start;
1199
1200                 list_del(&io->log_sibling);
1201                 mempool_free(io, log->io_pool);
1202                 r5l_run_no_mem_stripe(log);
1203
1204                 found = true;
1205         }
1206
1207         return found;
1208 }
1209
1210 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1211 {
1212         struct r5l_log *log = io->log;
1213         struct r5conf *conf = log->rdev->mddev->private;
1214         unsigned long flags;
1215
1216         spin_lock_irqsave(&log->io_list_lock, flags);
1217         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1218
1219         if (!r5l_complete_finished_ios(log)) {
1220                 spin_unlock_irqrestore(&log->io_list_lock, flags);
1221                 return;
1222         }
1223
1224         if (r5l_reclaimable_space(log) > log->max_free_space ||
1225             test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1226                 r5l_wake_reclaim(log, 0);
1227
1228         spin_unlock_irqrestore(&log->io_list_lock, flags);
1229         wake_up(&log->iounit_wait);
1230 }
1231
1232 void r5l_stripe_write_finished(struct stripe_head *sh)
1233 {
1234         struct r5l_io_unit *io;
1235
1236         io = sh->log_io;
1237         sh->log_io = NULL;
1238
1239         if (io && atomic_dec_and_test(&io->pending_stripe))
1240                 __r5l_stripe_write_finished(io);
1241 }
1242
1243 static void r5l_log_flush_endio(struct bio *bio)
1244 {
1245         struct r5l_log *log = container_of(bio, struct r5l_log,
1246                 flush_bio);
1247         unsigned long flags;
1248         struct r5l_io_unit *io;
1249
1250         if (bio->bi_error)
1251                 md_error(log->rdev->mddev, log->rdev);
1252
1253         spin_lock_irqsave(&log->io_list_lock, flags);
1254         list_for_each_entry(io, &log->flushing_ios, log_sibling)
1255                 r5l_io_run_stripes(io);
1256         list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1257         spin_unlock_irqrestore(&log->io_list_lock, flags);
1258 }
1259
1260 /*
1261  * Starting dispatch IO to raid.
1262  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1263  * broken meta in the middle of a log causes recovery can't find meta at the
1264  * head of log. If operations require meta at the head persistent in log, we
1265  * must make sure meta before it persistent in log too. A case is:
1266  *
1267  * stripe data/parity is in log, we start write stripe to raid disks. stripe
1268  * data/parity must be persistent in log before we do the write to raid disks.
1269  *
1270  * The solution is we restrictly maintain io_unit list order. In this case, we
1271  * only write stripes of an io_unit to raid disks till the io_unit is the first
1272  * one whose data/parity is in log.
1273  */
1274 void r5l_flush_stripe_to_raid(struct r5l_log *log)
1275 {
1276         bool do_flush;
1277
1278         if (!log || !log->need_cache_flush)
1279                 return;
1280
1281         spin_lock_irq(&log->io_list_lock);
1282         /* flush bio is running */
1283         if (!list_empty(&log->flushing_ios)) {
1284                 spin_unlock_irq(&log->io_list_lock);
1285                 return;
1286         }
1287         list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1288         do_flush = !list_empty(&log->flushing_ios);
1289         spin_unlock_irq(&log->io_list_lock);
1290
1291         if (!do_flush)
1292                 return;
1293         bio_reset(&log->flush_bio);
1294         log->flush_bio.bi_bdev = log->rdev->bdev;
1295         log->flush_bio.bi_end_io = r5l_log_flush_endio;
1296         log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1297         submit_bio(&log->flush_bio);
1298 }
1299
1300 static void r5l_write_super(struct r5l_log *log, sector_t cp);
1301 static void r5l_write_super_and_discard_space(struct r5l_log *log,
1302         sector_t end)
1303 {
1304         struct block_device *bdev = log->rdev->bdev;
1305         struct mddev *mddev;
1306
1307         r5l_write_super(log, end);
1308
1309         if (!blk_queue_discard(bdev_get_queue(bdev)))
1310                 return;
1311
1312         mddev = log->rdev->mddev;
1313         /*
1314          * Discard could zero data, so before discard we must make sure
1315          * superblock is updated to new log tail. Updating superblock (either
1316          * directly call md_update_sb() or depend on md thread) must hold
1317          * reconfig mutex. On the other hand, raid5_quiesce is called with
1318          * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1319          * for all IO finish, hence waitting for reclaim thread, while reclaim
1320          * thread is calling this function and waitting for reconfig mutex. So
1321          * there is a deadlock. We workaround this issue with a trylock.
1322          * FIXME: we could miss discard if we can't take reconfig mutex
1323          */
1324         set_mask_bits(&mddev->sb_flags, 0,
1325                 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1326         if (!mddev_trylock(mddev))
1327                 return;
1328         md_update_sb(mddev, 1);
1329         mddev_unlock(mddev);
1330
1331         /* discard IO error really doesn't matter, ignore it */
1332         if (log->last_checkpoint < end) {
1333                 blkdev_issue_discard(bdev,
1334                                 log->last_checkpoint + log->rdev->data_offset,
1335                                 end - log->last_checkpoint, GFP_NOIO, 0);
1336         } else {
1337                 blkdev_issue_discard(bdev,
1338                                 log->last_checkpoint + log->rdev->data_offset,
1339                                 log->device_size - log->last_checkpoint,
1340                                 GFP_NOIO, 0);
1341                 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1342                                 GFP_NOIO, 0);
1343         }
1344 }
1345
1346 /*
1347  * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1348  * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1349  *
1350  * must hold conf->device_lock
1351  */
1352 static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1353 {
1354         BUG_ON(list_empty(&sh->lru));
1355         BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1356         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1357
1358         /*
1359          * The stripe is not ON_RELEASE_LIST, so it is safe to call
1360          * raid5_release_stripe() while holding conf->device_lock
1361          */
1362         BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1363         assert_spin_locked(&conf->device_lock);
1364
1365         list_del_init(&sh->lru);
1366         atomic_inc(&sh->count);
1367
1368         set_bit(STRIPE_HANDLE, &sh->state);
1369         atomic_inc(&conf->active_stripes);
1370         r5c_make_stripe_write_out(sh);
1371
1372         if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1373                 atomic_inc(&conf->r5c_flushing_partial_stripes);
1374         else
1375                 atomic_inc(&conf->r5c_flushing_full_stripes);
1376         raid5_release_stripe(sh);
1377 }
1378
1379 /*
1380  * if num == 0, flush all full stripes
1381  * if num > 0, flush all full stripes. If less than num full stripes are
1382  *             flushed, flush some partial stripes until totally num stripes are
1383  *             flushed or there is no more cached stripes.
1384  */
1385 void r5c_flush_cache(struct r5conf *conf, int num)
1386 {
1387         int count;
1388         struct stripe_head *sh, *next;
1389
1390         assert_spin_locked(&conf->device_lock);
1391         if (!conf->log)
1392                 return;
1393
1394         count = 0;
1395         list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1396                 r5c_flush_stripe(conf, sh);
1397                 count++;
1398         }
1399
1400         if (count >= num)
1401                 return;
1402         list_for_each_entry_safe(sh, next,
1403                                  &conf->r5c_partial_stripe_list, lru) {
1404                 r5c_flush_stripe(conf, sh);
1405                 if (++count >= num)
1406                         break;
1407         }
1408 }
1409
1410 static void r5c_do_reclaim(struct r5conf *conf)
1411 {
1412         struct r5l_log *log = conf->log;
1413         struct stripe_head *sh;
1414         int count = 0;
1415         unsigned long flags;
1416         int total_cached;
1417         int stripes_to_flush;
1418         int flushing_partial, flushing_full;
1419
1420         if (!r5c_is_writeback(log))
1421                 return;
1422
1423         flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1424         flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1425         total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1426                 atomic_read(&conf->r5c_cached_full_stripes) -
1427                 flushing_full - flushing_partial;
1428
1429         if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1430             atomic_read(&conf->empty_inactive_list_nr) > 0)
1431                 /*
1432                  * if stripe cache pressure high, flush all full stripes and
1433                  * some partial stripes
1434                  */
1435                 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1436         else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1437                  atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1438                  R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1439                 /*
1440                  * if stripe cache pressure moderate, or if there is many full
1441                  * stripes,flush all full stripes
1442                  */
1443                 stripes_to_flush = 0;
1444         else
1445                 /* no need to flush */
1446                 stripes_to_flush = -1;
1447
1448         if (stripes_to_flush >= 0) {
1449                 spin_lock_irqsave(&conf->device_lock, flags);
1450                 r5c_flush_cache(conf, stripes_to_flush);
1451                 spin_unlock_irqrestore(&conf->device_lock, flags);
1452         }
1453
1454         /* if log space is tight, flush stripes on stripe_in_journal_list */
1455         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1456                 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1457                 spin_lock(&conf->device_lock);
1458                 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1459                         /*
1460                          * stripes on stripe_in_journal_list could be in any
1461                          * state of the stripe_cache state machine. In this
1462                          * case, we only want to flush stripe on
1463                          * r5c_cached_full/partial_stripes. The following
1464                          * condition makes sure the stripe is on one of the
1465                          * two lists.
1466                          */
1467                         if (!list_empty(&sh->lru) &&
1468                             !test_bit(STRIPE_HANDLE, &sh->state) &&
1469                             atomic_read(&sh->count) == 0) {
1470                                 r5c_flush_stripe(conf, sh);
1471                                 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1472                                         break;
1473                         }
1474                 }
1475                 spin_unlock(&conf->device_lock);
1476                 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1477         }
1478
1479         if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1480                 r5l_run_no_space_stripes(log);
1481
1482         md_wakeup_thread(conf->mddev->thread);
1483 }
1484
1485 static void r5l_do_reclaim(struct r5l_log *log)
1486 {
1487         struct r5conf *conf = log->rdev->mddev->private;
1488         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1489         sector_t reclaimable;
1490         sector_t next_checkpoint;
1491         bool write_super;
1492
1493         spin_lock_irq(&log->io_list_lock);
1494         write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1495                 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1496         /*
1497          * move proper io_unit to reclaim list. We should not change the order.
1498          * reclaimable/unreclaimable io_unit can be mixed in the list, we
1499          * shouldn't reuse space of an unreclaimable io_unit
1500          */
1501         while (1) {
1502                 reclaimable = r5l_reclaimable_space(log);
1503                 if (reclaimable >= reclaim_target ||
1504                     (list_empty(&log->running_ios) &&
1505                      list_empty(&log->io_end_ios) &&
1506                      list_empty(&log->flushing_ios) &&
1507                      list_empty(&log->finished_ios)))
1508                         break;
1509
1510                 md_wakeup_thread(log->rdev->mddev->thread);
1511                 wait_event_lock_irq(log->iounit_wait,
1512                                     r5l_reclaimable_space(log) > reclaimable,
1513                                     log->io_list_lock);
1514         }
1515
1516         next_checkpoint = r5c_calculate_new_cp(conf);
1517         spin_unlock_irq(&log->io_list_lock);
1518
1519         if (reclaimable == 0 || !write_super)
1520                 return;
1521
1522         /*
1523          * write_super will flush cache of each raid disk. We must write super
1524          * here, because the log area might be reused soon and we don't want to
1525          * confuse recovery
1526          */
1527         r5l_write_super_and_discard_space(log, next_checkpoint);
1528
1529         mutex_lock(&log->io_mutex);
1530         log->last_checkpoint = next_checkpoint;
1531         r5c_update_log_state(log);
1532         mutex_unlock(&log->io_mutex);
1533
1534         r5l_run_no_space_stripes(log);
1535 }
1536
1537 static void r5l_reclaim_thread(struct md_thread *thread)
1538 {
1539         struct mddev *mddev = thread->mddev;
1540         struct r5conf *conf = mddev->private;
1541         struct r5l_log *log = conf->log;
1542
1543         if (!log)
1544                 return;
1545         r5c_do_reclaim(conf);
1546         r5l_do_reclaim(log);
1547 }
1548
1549 void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1550 {
1551         unsigned long target;
1552         unsigned long new = (unsigned long)space; /* overflow in theory */
1553
1554         if (!log)
1555                 return;
1556         do {
1557                 target = log->reclaim_target;
1558                 if (new < target)
1559                         return;
1560         } while (cmpxchg(&log->reclaim_target, target, new) != target);
1561         md_wakeup_thread(log->reclaim_thread);
1562 }
1563
1564 void r5l_quiesce(struct r5l_log *log, int state)
1565 {
1566         struct mddev *mddev;
1567         if (!log || state == 2)
1568                 return;
1569         if (state == 0)
1570                 kthread_unpark(log->reclaim_thread->tsk);
1571         else if (state == 1) {
1572                 /* make sure r5l_write_super_and_discard_space exits */
1573                 mddev = log->rdev->mddev;
1574                 wake_up(&mddev->sb_wait);
1575                 kthread_park(log->reclaim_thread->tsk);
1576                 r5l_wake_reclaim(log, MaxSector);
1577                 r5l_do_reclaim(log);
1578         }
1579 }
1580
1581 bool r5l_log_disk_error(struct r5conf *conf)
1582 {
1583         struct r5l_log *log;
1584         bool ret;
1585         /* don't allow write if journal disk is missing */
1586         rcu_read_lock();
1587         log = rcu_dereference(conf->log);
1588
1589         if (!log)
1590                 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1591         else
1592                 ret = test_bit(Faulty, &log->rdev->flags);
1593         rcu_read_unlock();
1594         return ret;
1595 }
1596
1597 #define R5L_RECOVERY_PAGE_POOL_SIZE 256
1598
1599 struct r5l_recovery_ctx {
1600         struct page *meta_page;         /* current meta */
1601         sector_t meta_total_blocks;     /* total size of current meta and data */
1602         sector_t pos;                   /* recovery position */
1603         u64 seq;                        /* recovery position seq */
1604         int data_parity_stripes;        /* number of data_parity stripes */
1605         int data_only_stripes;          /* number of data_only stripes */
1606         struct list_head cached_list;
1607
1608         /*
1609          * read ahead page pool (ra_pool)
1610          * in recovery, log is read sequentially. It is not efficient to
1611          * read every page with sync_page_io(). The read ahead page pool
1612          * reads multiple pages with one IO, so further log read can
1613          * just copy data from the pool.
1614          */
1615         struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1616         sector_t pool_offset;   /* offset of first page in the pool */
1617         int total_pages;        /* total allocated pages */
1618         int valid_pages;        /* pages with valid data */
1619         struct bio *ra_bio;     /* bio to do the read ahead */
1620 };
1621
1622 static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1623                                             struct r5l_recovery_ctx *ctx)
1624 {
1625         struct page *page;
1626
1627         ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs);
1628         if (!ctx->ra_bio)
1629                 return -ENOMEM;
1630
1631         ctx->valid_pages = 0;
1632         ctx->total_pages = 0;
1633         while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1634                 page = alloc_page(GFP_KERNEL);
1635
1636                 if (!page)
1637                         break;
1638                 ctx->ra_pool[ctx->total_pages] = page;
1639                 ctx->total_pages += 1;
1640         }
1641
1642         if (ctx->total_pages == 0) {
1643                 bio_put(ctx->ra_bio);
1644                 return -ENOMEM;
1645         }
1646
1647         ctx->pool_offset = 0;
1648         return 0;
1649 }
1650
1651 static 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         bio_put(ctx->ra_bio);
1659 }
1660
1661 /*
1662  * fetch ctx->valid_pages pages from offset
1663  * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1664  * However, if the offset is close to the end of the journal device,
1665  * ctx->valid_pages could be smaller than ctx->total_pages
1666  */
1667 static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1668                                       struct r5l_recovery_ctx *ctx,
1669                                       sector_t offset)
1670 {
1671         bio_reset(ctx->ra_bio);
1672         ctx->ra_bio->bi_bdev = log->rdev->bdev;
1673         bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1674         ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1675
1676         ctx->valid_pages = 0;
1677         ctx->pool_offset = offset;
1678
1679         while (ctx->valid_pages < ctx->total_pages) {
1680                 bio_add_page(ctx->ra_bio,
1681                              ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1682                 ctx->valid_pages += 1;
1683
1684                 offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1685
1686                 if (offset == 0)  /* reached end of the device */
1687                         break;
1688         }
1689
1690         return submit_bio_wait(ctx->ra_bio);
1691 }
1692
1693 /*
1694  * try read a page from the read ahead page pool, if the page is not in the
1695  * pool, call r5l_recovery_fetch_ra_pool
1696  */
1697 static int r5l_recovery_read_page(struct r5l_log *log,
1698                                   struct r5l_recovery_ctx *ctx,
1699                                   struct page *page,
1700                                   sector_t offset)
1701 {
1702         int ret;
1703
1704         if (offset < ctx->pool_offset ||
1705             offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1706                 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1707                 if (ret)
1708                         return ret;
1709         }
1710
1711         BUG_ON(offset < ctx->pool_offset ||
1712                offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1713
1714         memcpy(page_address(page),
1715                page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1716                                          BLOCK_SECTOR_SHIFT]),
1717                PAGE_SIZE);
1718         return 0;
1719 }
1720
1721 static int r5l_recovery_read_meta_block(struct r5l_log *log,
1722                                         struct r5l_recovery_ctx *ctx)
1723 {
1724         struct page *page = ctx->meta_page;
1725         struct r5l_meta_block *mb;
1726         u32 crc, stored_crc;
1727         int ret;
1728
1729         ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1730         if (ret != 0)
1731                 return ret;
1732
1733         mb = page_address(page);
1734         stored_crc = le32_to_cpu(mb->checksum);
1735         mb->checksum = 0;
1736
1737         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1738             le64_to_cpu(mb->seq) != ctx->seq ||
1739             mb->version != R5LOG_VERSION ||
1740             le64_to_cpu(mb->position) != ctx->pos)
1741                 return -EINVAL;
1742
1743         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1744         if (stored_crc != crc)
1745                 return -EINVAL;
1746
1747         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1748                 return -EINVAL;
1749
1750         ctx->meta_total_blocks = BLOCK_SECTORS;
1751
1752         return 0;
1753 }
1754
1755 static void
1756 r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1757                                      struct page *page,
1758                                      sector_t pos, u64 seq)
1759 {
1760         struct r5l_meta_block *mb;
1761
1762         mb = page_address(page);
1763         clear_page(mb);
1764         mb->magic = cpu_to_le32(R5LOG_MAGIC);
1765         mb->version = R5LOG_VERSION;
1766         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1767         mb->seq = cpu_to_le64(seq);
1768         mb->position = cpu_to_le64(pos);
1769 }
1770
1771 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1772                                           u64 seq)
1773 {
1774         struct page *page;
1775         struct r5l_meta_block *mb;
1776
1777         page = alloc_page(GFP_KERNEL);
1778         if (!page)
1779                 return -ENOMEM;
1780         r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1781         mb = page_address(page);
1782         mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1783                                              mb, PAGE_SIZE));
1784         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1785                           REQ_FUA, false)) {
1786                 __free_page(page);
1787                 return -EIO;
1788         }
1789         __free_page(page);
1790         return 0;
1791 }
1792
1793 /*
1794  * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1795  * to mark valid (potentially not flushed) data in the journal.
1796  *
1797  * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1798  * so there should not be any mismatch here.
1799  */
1800 static void r5l_recovery_load_data(struct r5l_log *log,
1801                                    struct stripe_head *sh,
1802                                    struct r5l_recovery_ctx *ctx,
1803                                    struct r5l_payload_data_parity *payload,
1804                                    sector_t log_offset)
1805 {
1806         struct mddev *mddev = log->rdev->mddev;
1807         struct r5conf *conf = mddev->private;
1808         int dd_idx;
1809
1810         raid5_compute_sector(conf,
1811                              le64_to_cpu(payload->location), 0,
1812                              &dd_idx, sh);
1813         r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1814         sh->dev[dd_idx].log_checksum =
1815                 le32_to_cpu(payload->checksum[0]);
1816         ctx->meta_total_blocks += BLOCK_SECTORS;
1817
1818         set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1819         set_bit(STRIPE_R5C_CACHING, &sh->state);
1820 }
1821
1822 static void r5l_recovery_load_parity(struct r5l_log *log,
1823                                      struct stripe_head *sh,
1824                                      struct r5l_recovery_ctx *ctx,
1825                                      struct r5l_payload_data_parity *payload,
1826                                      sector_t log_offset)
1827 {
1828         struct mddev *mddev = log->rdev->mddev;
1829         struct r5conf *conf = mddev->private;
1830
1831         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1832         r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1833         sh->dev[sh->pd_idx].log_checksum =
1834                 le32_to_cpu(payload->checksum[0]);
1835         set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1836
1837         if (sh->qd_idx >= 0) {
1838                 r5l_recovery_read_page(
1839                         log, ctx, sh->dev[sh->qd_idx].page,
1840                         r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1841                 sh->dev[sh->qd_idx].log_checksum =
1842                         le32_to_cpu(payload->checksum[1]);
1843                 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1844         }
1845         clear_bit(STRIPE_R5C_CACHING, &sh->state);
1846 }
1847
1848 static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1849 {
1850         int i;
1851
1852         sh->state = 0;
1853         sh->log_start = MaxSector;
1854         for (i = sh->disks; i--; )
1855                 sh->dev[i].flags = 0;
1856 }
1857
1858 static void
1859 r5l_recovery_replay_one_stripe(struct r5conf *conf,
1860                                struct stripe_head *sh,
1861                                struct r5l_recovery_ctx *ctx)
1862 {
1863         struct md_rdev *rdev, *rrdev;
1864         int disk_index;
1865         int data_count = 0;
1866
1867         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1868                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1869                         continue;
1870                 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1871                         continue;
1872                 data_count++;
1873         }
1874
1875         /*
1876          * stripes that only have parity must have been flushed
1877          * before the crash that we are now recovering from, so
1878          * there is nothing more to recovery.
1879          */
1880         if (data_count == 0)
1881                 goto out;
1882
1883         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1884                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1885                         continue;
1886
1887                 /* in case device is broken */
1888                 rcu_read_lock();
1889                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1890                 if (rdev) {
1891                         atomic_inc(&rdev->nr_pending);
1892                         rcu_read_unlock();
1893                         sync_page_io(rdev, sh->sector, PAGE_SIZE,
1894                                      sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1895                                      false);
1896                         rdev_dec_pending(rdev, rdev->mddev);
1897                         rcu_read_lock();
1898                 }
1899                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1900                 if (rrdev) {
1901                         atomic_inc(&rrdev->nr_pending);
1902                         rcu_read_unlock();
1903                         sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1904                                      sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1905                                      false);
1906                         rdev_dec_pending(rrdev, rrdev->mddev);
1907                         rcu_read_lock();
1908                 }
1909                 rcu_read_unlock();
1910         }
1911         ctx->data_parity_stripes++;
1912 out:
1913         r5l_recovery_reset_stripe(sh);
1914 }
1915
1916 static struct stripe_head *
1917 r5c_recovery_alloc_stripe(struct r5conf *conf,
1918                           sector_t stripe_sect)
1919 {
1920         struct stripe_head *sh;
1921
1922         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1923         if (!sh)
1924                 return NULL;  /* no more stripe available */
1925
1926         r5l_recovery_reset_stripe(sh);
1927
1928         return sh;
1929 }
1930
1931 static struct stripe_head *
1932 r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1933 {
1934         struct stripe_head *sh;
1935
1936         list_for_each_entry(sh, list, lru)
1937                 if (sh->sector == sect)
1938                         return sh;
1939         return NULL;
1940 }
1941
1942 static void
1943 r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1944                           struct r5l_recovery_ctx *ctx)
1945 {
1946         struct stripe_head *sh, *next;
1947
1948         list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1949                 r5l_recovery_reset_stripe(sh);
1950                 list_del_init(&sh->lru);
1951                 raid5_release_stripe(sh);
1952         }
1953 }
1954
1955 static void
1956 r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1957                             struct r5l_recovery_ctx *ctx)
1958 {
1959         struct stripe_head *sh, *next;
1960
1961         list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1962                 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1963                         r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1964                         list_del_init(&sh->lru);
1965                         raid5_release_stripe(sh);
1966                 }
1967 }
1968
1969 /* if matches return 0; otherwise return -EINVAL */
1970 static int
1971 r5l_recovery_verify_data_checksum(struct r5l_log *log,
1972                                   struct r5l_recovery_ctx *ctx,
1973                                   struct page *page,
1974                                   sector_t log_offset, __le32 log_checksum)
1975 {
1976         void *addr;
1977         u32 checksum;
1978
1979         r5l_recovery_read_page(log, ctx, page, log_offset);
1980         addr = kmap_atomic(page);
1981         checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1982         kunmap_atomic(addr);
1983         return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1984 }
1985
1986 /*
1987  * before loading data to stripe cache, we need verify checksum for all data,
1988  * if there is mismatch for any data page, we drop all data in the mata block
1989  */
1990 static int
1991 r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1992                                          struct r5l_recovery_ctx *ctx)
1993 {
1994         struct mddev *mddev = log->rdev->mddev;
1995         struct r5conf *conf = mddev->private;
1996         struct r5l_meta_block *mb = page_address(ctx->meta_page);
1997         sector_t mb_offset = sizeof(struct r5l_meta_block);
1998         sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1999         struct page *page;
2000         struct r5l_payload_data_parity *payload;
2001         struct r5l_payload_flush *payload_flush;
2002
2003         page = alloc_page(GFP_KERNEL);
2004         if (!page)
2005                 return -ENOMEM;
2006
2007         while (mb_offset < le32_to_cpu(mb->meta_size)) {
2008                 payload = (void *)mb + mb_offset;
2009                 payload_flush = (void *)mb + mb_offset;
2010
2011                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2012                         if (r5l_recovery_verify_data_checksum(
2013                                     log, ctx, page, log_offset,
2014                                     payload->checksum[0]) < 0)
2015                                 goto mismatch;
2016                 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2017                         if (r5l_recovery_verify_data_checksum(
2018                                     log, ctx, page, log_offset,
2019                                     payload->checksum[0]) < 0)
2020                                 goto mismatch;
2021                         if (conf->max_degraded == 2 && /* q for RAID 6 */
2022                             r5l_recovery_verify_data_checksum(
2023                                     log, ctx, page,
2024                                     r5l_ring_add(log, log_offset,
2025                                                  BLOCK_SECTORS),
2026                                     payload->checksum[1]) < 0)
2027                                 goto mismatch;
2028                 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2029                         /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2030                 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2031                         goto mismatch;
2032
2033                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2034                         mb_offset += sizeof(struct r5l_payload_flush) +
2035                                 le32_to_cpu(payload_flush->size);
2036                 } else {
2037                         /* DATA or PARITY payload */
2038                         log_offset = r5l_ring_add(log, log_offset,
2039                                                   le32_to_cpu(payload->size));
2040                         mb_offset += sizeof(struct r5l_payload_data_parity) +
2041                                 sizeof(__le32) *
2042                                 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2043                 }
2044
2045         }
2046
2047         put_page(page);
2048         return 0;
2049
2050 mismatch:
2051         put_page(page);
2052         return -EINVAL;
2053 }
2054
2055 /*
2056  * Analyze all data/parity pages in one meta block
2057  * Returns:
2058  * 0 for success
2059  * -EINVAL for unknown playload type
2060  * -EAGAIN for checksum mismatch of data page
2061  * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2062  */
2063 static int
2064 r5c_recovery_analyze_meta_block(struct r5l_log *log,
2065                                 struct r5l_recovery_ctx *ctx,
2066                                 struct list_head *cached_stripe_list)
2067 {
2068         struct mddev *mddev = log->rdev->mddev;
2069         struct r5conf *conf = mddev->private;
2070         struct r5l_meta_block *mb;
2071         struct r5l_payload_data_parity *payload;
2072         struct r5l_payload_flush *payload_flush;
2073         int mb_offset;
2074         sector_t log_offset;
2075         sector_t stripe_sect;
2076         struct stripe_head *sh;
2077         int ret;
2078
2079         /*
2080          * for mismatch in data blocks, we will drop all data in this mb, but
2081          * we will still read next mb for other data with FLUSH flag, as
2082          * io_unit could finish out of order.
2083          */
2084         ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2085         if (ret == -EINVAL)
2086                 return -EAGAIN;
2087         else if (ret)
2088                 return ret;   /* -ENOMEM duo to alloc_page() failed */
2089
2090         mb = page_address(ctx->meta_page);
2091         mb_offset = sizeof(struct r5l_meta_block);
2092         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2093
2094         while (mb_offset < le32_to_cpu(mb->meta_size)) {
2095                 int dd;
2096
2097                 payload = (void *)mb + mb_offset;
2098                 payload_flush = (void *)mb + mb_offset;
2099
2100                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2101                         int i, count;
2102
2103                         count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2104                         for (i = 0; i < count; ++i) {
2105                                 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2106                                 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2107                                                                 stripe_sect);
2108                                 if (sh) {
2109                                         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2110                                         r5l_recovery_reset_stripe(sh);
2111                                         list_del_init(&sh->lru);
2112                                         raid5_release_stripe(sh);
2113                                 }
2114                         }
2115
2116                         mb_offset += sizeof(struct r5l_payload_flush) +
2117                                 le32_to_cpu(payload_flush->size);
2118                         continue;
2119                 }
2120
2121                 /* DATA or PARITY payload */
2122                 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2123                         raid5_compute_sector(
2124                                 conf, le64_to_cpu(payload->location), 0, &dd,
2125                                 NULL)
2126                         : le64_to_cpu(payload->location);
2127
2128                 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2129                                                 stripe_sect);
2130
2131                 if (!sh) {
2132                         sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
2133                         /*
2134                          * cannot get stripe from raid5_get_active_stripe
2135                          * try replay some stripes
2136                          */
2137                         if (!sh) {
2138                                 r5c_recovery_replay_stripes(
2139                                         cached_stripe_list, ctx);
2140                                 sh = r5c_recovery_alloc_stripe(
2141                                         conf, stripe_sect);
2142                         }
2143                         if (!sh) {
2144                                 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2145                                         mdname(mddev),
2146                                         conf->min_nr_stripes * 2);
2147                                 raid5_set_cache_size(mddev,
2148                                                      conf->min_nr_stripes * 2);
2149                                 sh = r5c_recovery_alloc_stripe(conf,
2150                                                                stripe_sect);
2151                         }
2152                         if (!sh) {
2153                                 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2154                                        mdname(mddev));
2155                                 return -ENOMEM;
2156                         }
2157                         list_add_tail(&sh->lru, cached_stripe_list);
2158                 }
2159
2160                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2161                         if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2162                             test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2163                                 r5l_recovery_replay_one_stripe(conf, sh, ctx);
2164                                 list_move_tail(&sh->lru, cached_stripe_list);
2165                         }
2166                         r5l_recovery_load_data(log, sh, ctx, payload,
2167                                                log_offset);
2168                 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2169                         r5l_recovery_load_parity(log, sh, ctx, payload,
2170                                                  log_offset);
2171                 else
2172                         return -EINVAL;
2173
2174                 log_offset = r5l_ring_add(log, log_offset,
2175                                           le32_to_cpu(payload->size));
2176
2177                 mb_offset += sizeof(struct r5l_payload_data_parity) +
2178                         sizeof(__le32) *
2179                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2180         }
2181
2182         return 0;
2183 }
2184
2185 /*
2186  * Load the stripe into cache. The stripe will be written out later by
2187  * the stripe cache state machine.
2188  */
2189 static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2190                                          struct stripe_head *sh)
2191 {
2192         struct r5dev *dev;
2193         int i;
2194
2195         for (i = sh->disks; i--; ) {
2196                 dev = sh->dev + i;
2197                 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2198                         set_bit(R5_InJournal, &dev->flags);
2199                         set_bit(R5_UPTODATE, &dev->flags);
2200                 }
2201         }
2202 }
2203
2204 /*
2205  * Scan through the log for all to-be-flushed data
2206  *
2207  * For stripes with data and parity, namely Data-Parity stripe
2208  * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2209  *
2210  * For stripes with only data, namely Data-Only stripe
2211  * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2212  *
2213  * For a stripe, if we see data after parity, we should discard all previous
2214  * data and parity for this stripe, as these data are already flushed to
2215  * the array.
2216  *
2217  * At the end of the scan, we return the new journal_tail, which points to
2218  * first data-only stripe on the journal device, or next invalid meta block.
2219  */
2220 static int r5c_recovery_flush_log(struct r5l_log *log,
2221                                   struct r5l_recovery_ctx *ctx)
2222 {
2223         struct stripe_head *sh;
2224         int ret = 0;
2225
2226         /* scan through the log */
2227         while (1) {
2228                 if (r5l_recovery_read_meta_block(log, ctx))
2229                         break;
2230
2231                 ret = r5c_recovery_analyze_meta_block(log, ctx,
2232                                                       &ctx->cached_list);
2233                 /*
2234                  * -EAGAIN means mismatch in data block, in this case, we still
2235                  * try scan the next metablock
2236                  */
2237                 if (ret && ret != -EAGAIN)
2238                         break;   /* ret == -EINVAL or -ENOMEM */
2239                 ctx->seq++;
2240                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2241         }
2242
2243         if (ret == -ENOMEM) {
2244                 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2245                 return ret;
2246         }
2247
2248         /* replay data-parity stripes */
2249         r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2250
2251         /* load data-only stripes to stripe cache */
2252         list_for_each_entry(sh, &ctx->cached_list, lru) {
2253                 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2254                 r5c_recovery_load_one_stripe(log, sh);
2255                 ctx->data_only_stripes++;
2256         }
2257
2258         return 0;
2259 }
2260
2261 /*
2262  * we did a recovery. Now ctx.pos points to an invalid meta block. New
2263  * log will start here. but we can't let superblock point to last valid
2264  * meta block. The log might looks like:
2265  * | meta 1| meta 2| meta 3|
2266  * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2267  * superblock points to meta 1, we write a new valid meta 2n.  if crash
2268  * happens again, new recovery will start from meta 1. Since meta 2n is
2269  * valid now, recovery will think meta 3 is valid, which is wrong.
2270  * The solution is we create a new meta in meta2 with its seq == meta
2271  * 1's seq + 10000 and let superblock points to meta2. The same recovery
2272  * will not think meta 3 is a valid meta, because its seq doesn't match
2273  */
2274
2275 /*
2276  * Before recovery, the log looks like the following
2277  *
2278  *   ---------------------------------------------
2279  *   |           valid log        | invalid log  |
2280  *   ---------------------------------------------
2281  *   ^
2282  *   |- log->last_checkpoint
2283  *   |- log->last_cp_seq
2284  *
2285  * Now we scan through the log until we see invalid entry
2286  *
2287  *   ---------------------------------------------
2288  *   |           valid log        | invalid log  |
2289  *   ---------------------------------------------
2290  *   ^                            ^
2291  *   |- log->last_checkpoint      |- ctx->pos
2292  *   |- log->last_cp_seq          |- ctx->seq
2293  *
2294  * From this point, we need to increase seq number by 10 to avoid
2295  * confusing next recovery.
2296  *
2297  *   ---------------------------------------------
2298  *   |           valid log        | invalid log  |
2299  *   ---------------------------------------------
2300  *   ^                              ^
2301  *   |- log->last_checkpoint        |- ctx->pos+1
2302  *   |- log->last_cp_seq            |- ctx->seq+10001
2303  *
2304  * However, it is not safe to start the state machine yet, because data only
2305  * parities are not yet secured in RAID. To save these data only parities, we
2306  * rewrite them from seq+11.
2307  *
2308  *   -----------------------------------------------------------------
2309  *   |           valid log        | data only stripes | invalid log  |
2310  *   -----------------------------------------------------------------
2311  *   ^                                                ^
2312  *   |- log->last_checkpoint                          |- ctx->pos+n
2313  *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2314  *
2315  * If failure happens again during this process, the recovery can safe start
2316  * again from log->last_checkpoint.
2317  *
2318  * Once data only stripes are rewritten to journal, we move log_tail
2319  *
2320  *   -----------------------------------------------------------------
2321  *   |     old log        |    data only stripes    | invalid log  |
2322  *   -----------------------------------------------------------------
2323  *                        ^                         ^
2324  *                        |- log->last_checkpoint   |- ctx->pos+n
2325  *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2326  *
2327  * Then we can safely start the state machine. If failure happens from this
2328  * point on, the recovery will start from new log->last_checkpoint.
2329  */
2330 static int
2331 r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2332                                        struct r5l_recovery_ctx *ctx)
2333 {
2334         struct stripe_head *sh;
2335         struct mddev *mddev = log->rdev->mddev;
2336         struct page *page;
2337         sector_t next_checkpoint = MaxSector;
2338
2339         page = alloc_page(GFP_KERNEL);
2340         if (!page) {
2341                 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2342                        mdname(mddev));
2343                 return -ENOMEM;
2344         }
2345
2346         WARN_ON(list_empty(&ctx->cached_list));
2347
2348         list_for_each_entry(sh, &ctx->cached_list, lru) {
2349                 struct r5l_meta_block *mb;
2350                 int i;
2351                 int offset;
2352                 sector_t write_pos;
2353
2354                 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2355                 r5l_recovery_create_empty_meta_block(log, page,
2356                                                      ctx->pos, ctx->seq);
2357                 mb = page_address(page);
2358                 offset = le32_to_cpu(mb->meta_size);
2359                 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2360
2361                 for (i = sh->disks; i--; ) {
2362                         struct r5dev *dev = &sh->dev[i];
2363                         struct r5l_payload_data_parity *payload;
2364                         void *addr;
2365
2366                         if (test_bit(R5_InJournal, &dev->flags)) {
2367                                 payload = (void *)mb + offset;
2368                                 payload->header.type = cpu_to_le16(
2369                                         R5LOG_PAYLOAD_DATA);
2370                                 payload->size = cpu_to_le32(BLOCK_SECTORS);
2371                                 payload->location = cpu_to_le64(
2372                                         raid5_compute_blocknr(sh, i, 0));
2373                                 addr = kmap_atomic(dev->page);
2374                                 payload->checksum[0] = cpu_to_le32(
2375                                         crc32c_le(log->uuid_checksum, addr,
2376                                                   PAGE_SIZE));
2377                                 kunmap_atomic(addr);
2378                                 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2379                                              dev->page, REQ_OP_WRITE, 0, false);
2380                                 write_pos = r5l_ring_add(log, write_pos,
2381                                                          BLOCK_SECTORS);
2382                                 offset += sizeof(__le32) +
2383                                         sizeof(struct r5l_payload_data_parity);
2384
2385                         }
2386                 }
2387                 mb->meta_size = cpu_to_le32(offset);
2388                 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2389                                                      mb, PAGE_SIZE));
2390                 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2391                              REQ_OP_WRITE, REQ_FUA, false);
2392                 sh->log_start = ctx->pos;
2393                 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2394                 atomic_inc(&log->stripe_in_journal_count);
2395                 ctx->pos = write_pos;
2396                 ctx->seq += 1;
2397                 next_checkpoint = sh->log_start;
2398         }
2399         log->next_checkpoint = next_checkpoint;
2400         __free_page(page);
2401         return 0;
2402 }
2403
2404 static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2405                                                  struct r5l_recovery_ctx *ctx)
2406 {
2407         struct mddev *mddev = log->rdev->mddev;
2408         struct r5conf *conf = mddev->private;
2409         struct stripe_head *sh, *next;
2410
2411         if (ctx->data_only_stripes == 0)
2412                 return;
2413
2414         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2415
2416         list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2417                 r5c_make_stripe_write_out(sh);
2418                 set_bit(STRIPE_HANDLE, &sh->state);
2419                 list_del_init(&sh->lru);
2420                 raid5_release_stripe(sh);
2421         }
2422
2423         md_wakeup_thread(conf->mddev->thread);
2424         /* reuse conf->wait_for_quiescent in recovery */
2425         wait_event(conf->wait_for_quiescent,
2426                    atomic_read(&conf->active_stripes) == 0);
2427
2428         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2429 }
2430
2431 static int r5l_recovery_log(struct r5l_log *log)
2432 {
2433         struct mddev *mddev = log->rdev->mddev;
2434         struct r5l_recovery_ctx *ctx;
2435         int ret;
2436         sector_t pos;
2437
2438         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2439         if (!ctx)
2440                 return -ENOMEM;
2441
2442         ctx->pos = log->last_checkpoint;
2443         ctx->seq = log->last_cp_seq;
2444         INIT_LIST_HEAD(&ctx->cached_list);
2445         ctx->meta_page = alloc_page(GFP_KERNEL);
2446
2447         if (!ctx->meta_page) {
2448                 ret =  -ENOMEM;
2449                 goto meta_page;
2450         }
2451
2452         if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2453                 ret = -ENOMEM;
2454                 goto ra_pool;
2455         }
2456
2457         ret = r5c_recovery_flush_log(log, ctx);
2458
2459         if (ret)
2460                 goto error;
2461
2462         pos = ctx->pos;
2463         ctx->seq += 10000;
2464
2465         if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2466                 pr_debug("md/raid:%s: starting from clean shutdown\n",
2467                          mdname(mddev));
2468         else
2469                 pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2470                          mdname(mddev), ctx->data_only_stripes,
2471                          ctx->data_parity_stripes);
2472
2473         if (ctx->data_only_stripes == 0) {
2474                 log->next_checkpoint = ctx->pos;
2475                 r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2476                 ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2477         } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2478                 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2479                        mdname(mddev));
2480                 ret =  -EIO;
2481                 goto error;
2482         }
2483
2484         log->log_start = ctx->pos;
2485         log->seq = ctx->seq;
2486         log->last_checkpoint = pos;
2487         r5l_write_super(log, pos);
2488
2489         r5c_recovery_flush_data_only_stripes(log, ctx);
2490         ret = 0;
2491 error:
2492         r5l_recovery_free_ra_pool(log, ctx);
2493 ra_pool:
2494         __free_page(ctx->meta_page);
2495 meta_page:
2496         kfree(ctx);
2497         return ret;
2498 }
2499
2500 static void r5l_write_super(struct r5l_log *log, sector_t cp)
2501 {
2502         struct mddev *mddev = log->rdev->mddev;
2503
2504         log->rdev->journal_tail = cp;
2505         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2506 }
2507
2508 static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2509 {
2510         struct r5conf *conf = mddev->private;
2511         int ret;
2512
2513         if (!conf->log)
2514                 return 0;
2515
2516         switch (conf->log->r5c_journal_mode) {
2517         case R5C_JOURNAL_MODE_WRITE_THROUGH:
2518                 ret = snprintf(
2519                         page, PAGE_SIZE, "[%s] %s\n",
2520                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2521                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2522                 break;
2523         case R5C_JOURNAL_MODE_WRITE_BACK:
2524                 ret = snprintf(
2525                         page, PAGE_SIZE, "%s [%s]\n",
2526                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2527                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2528                 break;
2529         default:
2530                 ret = 0;
2531         }
2532         return ret;
2533 }
2534
2535 /*
2536  * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2537  *
2538  * @mode as defined in 'enum r5c_journal_mode'.
2539  *
2540  */
2541 int r5c_journal_mode_set(struct mddev *mddev, int mode)
2542 {
2543         struct r5conf *conf = mddev->private;
2544         struct r5l_log *log = conf->log;
2545
2546         if (!log)
2547                 return -ENODEV;
2548
2549         if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2550             mode > R5C_JOURNAL_MODE_WRITE_BACK)
2551                 return -EINVAL;
2552
2553         if (raid5_calc_degraded(conf) > 0 &&
2554             mode == R5C_JOURNAL_MODE_WRITE_BACK)
2555                 return -EINVAL;
2556
2557         mddev_suspend(mddev);
2558         conf->log->r5c_journal_mode = mode;
2559         mddev_resume(mddev);
2560
2561         pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2562                  mdname(mddev), mode, r5c_journal_mode_str[mode]);
2563         return 0;
2564 }
2565 EXPORT_SYMBOL(r5c_journal_mode_set);
2566
2567 static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2568                                       const char *page, size_t length)
2569 {
2570         int mode = ARRAY_SIZE(r5c_journal_mode_str);
2571         size_t len = length;
2572
2573         if (len < 2)
2574                 return -EINVAL;
2575
2576         if (page[len - 1] == '\n')
2577                 len--;
2578
2579         while (mode--)
2580                 if (strlen(r5c_journal_mode_str[mode]) == len &&
2581                     !strncmp(page, r5c_journal_mode_str[mode], len))
2582                         break;
2583
2584         return r5c_journal_mode_set(mddev, mode) ?: length;
2585 }
2586
2587 struct md_sysfs_entry
2588 r5c_journal_mode = __ATTR(journal_mode, 0644,
2589                           r5c_journal_mode_show, r5c_journal_mode_store);
2590
2591 /*
2592  * Try handle write operation in caching phase. This function should only
2593  * be called in write-back mode.
2594  *
2595  * If all outstanding writes can be handled in caching phase, returns 0
2596  * If writes requires write-out phase, call r5c_make_stripe_write_out()
2597  * and returns -EAGAIN
2598  */
2599 int r5c_try_caching_write(struct r5conf *conf,
2600                           struct stripe_head *sh,
2601                           struct stripe_head_state *s,
2602                           int disks)
2603 {
2604         struct r5l_log *log = conf->log;
2605         int i;
2606         struct r5dev *dev;
2607         int to_cache = 0;
2608         void **pslot;
2609         sector_t tree_index;
2610         int ret;
2611         uintptr_t refcount;
2612
2613         BUG_ON(!r5c_is_writeback(log));
2614
2615         if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2616                 /*
2617                  * There are two different scenarios here:
2618                  *  1. The stripe has some data cached, and it is sent to
2619                  *     write-out phase for reclaim
2620                  *  2. The stripe is clean, and this is the first write
2621                  *
2622                  * For 1, return -EAGAIN, so we continue with
2623                  * handle_stripe_dirtying().
2624                  *
2625                  * For 2, set STRIPE_R5C_CACHING and continue with caching
2626                  * write.
2627                  */
2628
2629                 /* case 1: anything injournal or anything in written */
2630                 if (s->injournal > 0 || s->written > 0)
2631                         return -EAGAIN;
2632                 /* case 2 */
2633                 set_bit(STRIPE_R5C_CACHING, &sh->state);
2634         }
2635
2636         /*
2637          * When run in degraded mode, array is set to write-through mode.
2638          * This check helps drain pending write safely in the transition to
2639          * write-through mode.
2640          *
2641          * When a stripe is syncing, the write is also handled in write
2642          * through mode.
2643          */
2644         if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2645                 r5c_make_stripe_write_out(sh);
2646                 return -EAGAIN;
2647         }
2648
2649         for (i = disks; i--; ) {
2650                 dev = &sh->dev[i];
2651                 /* if non-overwrite, use writing-out phase */
2652                 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2653                     !test_bit(R5_InJournal, &dev->flags)) {
2654                         r5c_make_stripe_write_out(sh);
2655                         return -EAGAIN;
2656                 }
2657         }
2658
2659         /* if the stripe is not counted in big_stripe_tree, add it now */
2660         if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2661             !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2662                 tree_index = r5c_tree_index(conf, sh->sector);
2663                 spin_lock(&log->tree_lock);
2664                 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2665                                                tree_index);
2666                 if (pslot) {
2667                         refcount = (uintptr_t)radix_tree_deref_slot_protected(
2668                                 pslot, &log->tree_lock) >>
2669                                 R5C_RADIX_COUNT_SHIFT;
2670                         radix_tree_replace_slot(
2671                                 &log->big_stripe_tree, pslot,
2672                                 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2673                 } else {
2674                         /*
2675                          * this radix_tree_insert can fail safely, so no
2676                          * need to call radix_tree_preload()
2677                          */
2678                         ret = radix_tree_insert(
2679                                 &log->big_stripe_tree, tree_index,
2680                                 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2681                         if (ret) {
2682                                 spin_unlock(&log->tree_lock);
2683                                 r5c_make_stripe_write_out(sh);
2684                                 return -EAGAIN;
2685                         }
2686                 }
2687                 spin_unlock(&log->tree_lock);
2688
2689                 /*
2690                  * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2691                  * counted in the radix tree
2692                  */
2693                 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2694                 atomic_inc(&conf->r5c_cached_partial_stripes);
2695         }
2696
2697         for (i = disks; i--; ) {
2698                 dev = &sh->dev[i];
2699                 if (dev->towrite) {
2700                         set_bit(R5_Wantwrite, &dev->flags);
2701                         set_bit(R5_Wantdrain, &dev->flags);
2702                         set_bit(R5_LOCKED, &dev->flags);
2703                         to_cache++;
2704                 }
2705         }
2706
2707         if (to_cache) {
2708                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2709                 /*
2710                  * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2711                  * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2712                  * r5c_handle_data_cached()
2713                  */
2714                 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2715         }
2716
2717         return 0;
2718 }
2719
2720 /*
2721  * free extra pages (orig_page) we allocated for prexor
2722  */
2723 void r5c_release_extra_page(struct stripe_head *sh)
2724 {
2725         struct r5conf *conf = sh->raid_conf;
2726         int i;
2727         bool using_disk_info_extra_page;
2728
2729         using_disk_info_extra_page =
2730                 sh->dev[0].orig_page == conf->disks[0].extra_page;
2731
2732         for (i = sh->disks; i--; )
2733                 if (sh->dev[i].page != sh->dev[i].orig_page) {
2734                         struct page *p = sh->dev[i].orig_page;
2735
2736                         sh->dev[i].orig_page = sh->dev[i].page;
2737                         clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2738
2739                         if (!using_disk_info_extra_page)
2740                                 put_page(p);
2741                 }
2742
2743         if (using_disk_info_extra_page) {
2744                 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2745                 md_wakeup_thread(conf->mddev->thread);
2746         }
2747 }
2748
2749 void r5c_use_extra_page(struct stripe_head *sh)
2750 {
2751         struct r5conf *conf = sh->raid_conf;
2752         int i;
2753         struct r5dev *dev;
2754
2755         for (i = sh->disks; i--; ) {
2756                 dev = &sh->dev[i];
2757                 if (dev->orig_page != dev->page)
2758                         put_page(dev->orig_page);
2759                 dev->orig_page = conf->disks[i].extra_page;
2760         }
2761 }
2762
2763 /*
2764  * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2765  * stripe is committed to RAID disks.
2766  */
2767 void r5c_finish_stripe_write_out(struct r5conf *conf,
2768                                  struct stripe_head *sh,
2769                                  struct stripe_head_state *s)
2770 {
2771         struct r5l_log *log = conf->log;
2772         int i;
2773         int do_wakeup = 0;
2774         sector_t tree_index;
2775         void **pslot;
2776         uintptr_t refcount;
2777
2778         if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2779                 return;
2780
2781         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2782         clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2783
2784         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2785                 return;
2786
2787         for (i = sh->disks; i--; ) {
2788                 clear_bit(R5_InJournal, &sh->dev[i].flags);
2789                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2790                         do_wakeup = 1;
2791         }
2792
2793         /*
2794          * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2795          * We updated R5_InJournal, so we also update s->injournal.
2796          */
2797         s->injournal = 0;
2798
2799         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2800                 if (atomic_dec_and_test(&conf->pending_full_writes))
2801                         md_wakeup_thread(conf->mddev->thread);
2802
2803         if (do_wakeup)
2804                 wake_up(&conf->wait_for_overlap);
2805
2806         spin_lock_irq(&log->stripe_in_journal_lock);
2807         list_del_init(&sh->r5c);
2808         spin_unlock_irq(&log->stripe_in_journal_lock);
2809         sh->log_start = MaxSector;
2810
2811         atomic_dec(&log->stripe_in_journal_count);
2812         r5c_update_log_state(log);
2813
2814         /* stop counting this stripe in big_stripe_tree */
2815         if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2816             test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2817                 tree_index = r5c_tree_index(conf, sh->sector);
2818                 spin_lock(&log->tree_lock);
2819                 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2820                                                tree_index);
2821                 BUG_ON(pslot == NULL);
2822                 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2823                         pslot, &log->tree_lock) >>
2824                         R5C_RADIX_COUNT_SHIFT;
2825                 if (refcount == 1)
2826                         radix_tree_delete(&log->big_stripe_tree, tree_index);
2827                 else
2828                         radix_tree_replace_slot(
2829                                 &log->big_stripe_tree, pslot,
2830                                 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2831                 spin_unlock(&log->tree_lock);
2832         }
2833
2834         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2835                 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2836                 atomic_dec(&conf->r5c_flushing_partial_stripes);
2837                 atomic_dec(&conf->r5c_cached_partial_stripes);
2838         }
2839
2840         if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2841                 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2842                 atomic_dec(&conf->r5c_flushing_full_stripes);
2843                 atomic_dec(&conf->r5c_cached_full_stripes);
2844         }
2845
2846         r5l_append_flush_payload(log, sh->sector);
2847         /* stripe is flused to raid disks, we can do resync now */
2848         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2849                 set_bit(STRIPE_HANDLE, &sh->state);
2850 }
2851
2852 int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2853 {
2854         struct r5conf *conf = sh->raid_conf;
2855         int pages = 0;
2856         int reserve;
2857         int i;
2858         int ret = 0;
2859
2860         BUG_ON(!log);
2861
2862         for (i = 0; i < sh->disks; i++) {
2863                 void *addr;
2864
2865                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2866                         continue;
2867                 addr = kmap_atomic(sh->dev[i].page);
2868                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2869                                                     addr, PAGE_SIZE);
2870                 kunmap_atomic(addr);
2871                 pages++;
2872         }
2873         WARN_ON(pages == 0);
2874
2875         /*
2876          * The stripe must enter state machine again to call endio, so
2877          * don't delay.
2878          */
2879         clear_bit(STRIPE_DELAYED, &sh->state);
2880         atomic_inc(&sh->count);
2881
2882         mutex_lock(&log->io_mutex);
2883         /* meta + data */
2884         reserve = (1 + pages) << (PAGE_SHIFT - 9);
2885
2886         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2887             sh->log_start == MaxSector)
2888                 r5l_add_no_space_stripe(log, sh);
2889         else if (!r5l_has_free_space(log, reserve)) {
2890                 if (sh->log_start == log->last_checkpoint)
2891                         BUG();
2892                 else
2893                         r5l_add_no_space_stripe(log, sh);
2894         } else {
2895                 ret = r5l_log_stripe(log, sh, pages, 0);
2896                 if (ret) {
2897                         spin_lock_irq(&log->io_list_lock);
2898                         list_add_tail(&sh->log_list, &log->no_mem_stripes);
2899                         spin_unlock_irq(&log->io_list_lock);
2900                 }
2901         }
2902
2903         mutex_unlock(&log->io_mutex);
2904         return 0;
2905 }
2906
2907 /* check whether this big stripe is in write back cache. */
2908 bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2909 {
2910         struct r5l_log *log = conf->log;
2911         sector_t tree_index;
2912         void *slot;
2913
2914         if (!log)
2915                 return false;
2916
2917         WARN_ON_ONCE(!rcu_read_lock_held());
2918         tree_index = r5c_tree_index(conf, sect);
2919         slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2920         return slot != NULL;
2921 }
2922
2923 static int r5l_load_log(struct r5l_log *log)
2924 {
2925         struct md_rdev *rdev = log->rdev;
2926         struct page *page;
2927         struct r5l_meta_block *mb;
2928         sector_t cp = log->rdev->journal_tail;
2929         u32 stored_crc, expected_crc;
2930         bool create_super = false;
2931         int ret = 0;
2932
2933         /* Make sure it's valid */
2934         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2935                 cp = 0;
2936         page = alloc_page(GFP_KERNEL);
2937         if (!page)
2938                 return -ENOMEM;
2939
2940         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2941                 ret = -EIO;
2942                 goto ioerr;
2943         }
2944         mb = page_address(page);
2945
2946         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2947             mb->version != R5LOG_VERSION) {
2948                 create_super = true;
2949                 goto create;
2950         }
2951         stored_crc = le32_to_cpu(mb->checksum);
2952         mb->checksum = 0;
2953         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2954         if (stored_crc != expected_crc) {
2955                 create_super = true;
2956                 goto create;
2957         }
2958         if (le64_to_cpu(mb->position) != cp) {
2959                 create_super = true;
2960                 goto create;
2961         }
2962 create:
2963         if (create_super) {
2964                 log->last_cp_seq = prandom_u32();
2965                 cp = 0;
2966                 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2967                 /*
2968                  * Make sure super points to correct address. Log might have
2969                  * data very soon. If super hasn't correct log tail address,
2970                  * recovery can't find the log
2971                  */
2972                 r5l_write_super(log, cp);
2973         } else
2974                 log->last_cp_seq = le64_to_cpu(mb->seq);
2975
2976         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
2977         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
2978         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
2979                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
2980         log->last_checkpoint = cp;
2981
2982         __free_page(page);
2983
2984         if (create_super) {
2985                 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
2986                 log->seq = log->last_cp_seq + 1;
2987                 log->next_checkpoint = cp;
2988         } else
2989                 ret = r5l_recovery_log(log);
2990
2991         r5c_update_log_state(log);
2992         return ret;
2993 ioerr:
2994         __free_page(page);
2995         return ret;
2996 }
2997
2998 void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
2999 {
3000         struct r5conf *conf = mddev->private;
3001         struct r5l_log *log = conf->log;
3002
3003         if (!log)
3004                 return;
3005
3006         if ((raid5_calc_degraded(conf) > 0 ||
3007              test_bit(Journal, &rdev->flags)) &&
3008             conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3009                 schedule_work(&log->disable_writeback_work);
3010 }
3011
3012 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3013 {
3014         struct request_queue *q = bdev_get_queue(rdev->bdev);
3015         struct r5l_log *log;
3016         char b[BDEVNAME_SIZE];
3017
3018         pr_debug("md/raid:%s: using device %s as journal\n",
3019                  mdname(conf->mddev), bdevname(rdev->bdev, b));
3020
3021         if (PAGE_SIZE != 4096)
3022                 return -EINVAL;
3023
3024         /*
3025          * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3026          * raid_disks r5l_payload_data_parity.
3027          *
3028          * Write journal and cache does not work for very big array
3029          * (raid_disks > 203)
3030          */
3031         if (sizeof(struct r5l_meta_block) +
3032             ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3033              conf->raid_disks) > PAGE_SIZE) {
3034                 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3035                        mdname(conf->mddev), conf->raid_disks);
3036                 return -EINVAL;
3037         }
3038
3039         log = kzalloc(sizeof(*log), GFP_KERNEL);
3040         if (!log)
3041                 return -ENOMEM;
3042         log->rdev = rdev;
3043
3044         log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
3045
3046         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3047                                        sizeof(rdev->mddev->uuid));
3048
3049         mutex_init(&log->io_mutex);
3050
3051         spin_lock_init(&log->io_list_lock);
3052         INIT_LIST_HEAD(&log->running_ios);
3053         INIT_LIST_HEAD(&log->io_end_ios);
3054         INIT_LIST_HEAD(&log->flushing_ios);
3055         INIT_LIST_HEAD(&log->finished_ios);
3056         bio_init(&log->flush_bio, NULL, 0);
3057
3058         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3059         if (!log->io_kc)
3060                 goto io_kc;
3061
3062         log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
3063         if (!log->io_pool)
3064                 goto io_pool;
3065
3066         log->bs = bioset_create(R5L_POOL_SIZE, 0);
3067         if (!log->bs)
3068                 goto io_bs;
3069
3070         log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
3071         if (!log->meta_pool)
3072                 goto out_mempool;
3073
3074         spin_lock_init(&log->tree_lock);
3075         INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3076
3077         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3078                                                  log->rdev->mddev, "reclaim");
3079         if (!log->reclaim_thread)
3080                 goto reclaim_thread;
3081         log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3082
3083         init_waitqueue_head(&log->iounit_wait);
3084
3085         INIT_LIST_HEAD(&log->no_mem_stripes);
3086
3087         INIT_LIST_HEAD(&log->no_space_stripes);
3088         spin_lock_init(&log->no_space_stripes_lock);
3089
3090         INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3091         INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3092
3093         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3094         INIT_LIST_HEAD(&log->stripe_in_journal_list);
3095         spin_lock_init(&log->stripe_in_journal_lock);
3096         atomic_set(&log->stripe_in_journal_count, 0);
3097
3098         rcu_assign_pointer(conf->log, log);
3099
3100         if (r5l_load_log(log))
3101                 goto error;
3102
3103         set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3104         return 0;
3105
3106 error:
3107         rcu_assign_pointer(conf->log, NULL);
3108         md_unregister_thread(&log->reclaim_thread);
3109 reclaim_thread:
3110         mempool_destroy(log->meta_pool);
3111 out_mempool:
3112         bioset_free(log->bs);
3113 io_bs:
3114         mempool_destroy(log->io_pool);
3115 io_pool:
3116         kmem_cache_destroy(log->io_kc);
3117 io_kc:
3118         kfree(log);
3119         return -EINVAL;
3120 }
3121
3122 void r5l_exit_log(struct r5conf *conf)
3123 {
3124         struct r5l_log *log = conf->log;
3125
3126         conf->log = NULL;
3127         synchronize_rcu();
3128
3129         flush_work(&log->disable_writeback_work);
3130         md_unregister_thread(&log->reclaim_thread);
3131         mempool_destroy(log->meta_pool);
3132         bioset_free(log->bs);
3133         mempool_destroy(log->io_pool);
3134         kmem_cache_destroy(log->io_kc);
3135         kfree(log);
3136 }