8b8f9db05e916a8c7512b62dcf1bb529277293fb
[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 "md-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 null flush request */
242         unsigned int has_flush_payload:1;       /* include flush payload  */
243         /*
244          * io isn't sent yet, flush/fua request can only be submitted till it's
245          * the first IO in running_ios list
246          */
247         unsigned int io_deferred:1;
248
249         struct bio_list flush_barriers;   /* size == 0 flush bios */
250 };
251
252 /* r5l_io_unit state */
253 enum r5l_io_unit_state {
254         IO_UNIT_RUNNING = 0,    /* accepting new IO */
255         IO_UNIT_IO_START = 1,   /* io_unit bio start writing to log,
256                                  * don't accepting new bio */
257         IO_UNIT_IO_END = 2,     /* io_unit bio finish writing to log */
258         IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
259 };
260
261 bool r5c_is_writeback(struct r5l_log *log)
262 {
263         return (log != NULL &&
264                 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
265 }
266
267 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
268 {
269         start += inc;
270         if (start >= log->device_size)
271                 start = start - log->device_size;
272         return start;
273 }
274
275 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
276                                   sector_t end)
277 {
278         if (end >= start)
279                 return end - start;
280         else
281                 return end + log->device_size - start;
282 }
283
284 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
285 {
286         sector_t used_size;
287
288         used_size = r5l_ring_distance(log, log->last_checkpoint,
289                                         log->log_start);
290
291         return log->device_size > used_size + size;
292 }
293
294 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
295                                     enum r5l_io_unit_state state)
296 {
297         if (WARN_ON(io->state >= state))
298                 return;
299         io->state = state;
300 }
301
302 static void
303 r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
304 {
305         struct bio *wbi, *wbi2;
306
307         wbi = dev->written;
308         dev->written = NULL;
309         while (wbi && wbi->bi_iter.bi_sector <
310                dev->sector + STRIPE_SECTORS) {
311                 wbi2 = r5_next_bio(wbi, dev->sector);
312                 md_write_end(conf->mddev);
313                 bio_endio(wbi);
314                 wbi = wbi2;
315         }
316 }
317
318 void r5c_handle_cached_data_endio(struct r5conf *conf,
319                                   struct stripe_head *sh, int disks)
320 {
321         int i;
322
323         for (i = sh->disks; i--; ) {
324                 if (sh->dev[i].written) {
325                         set_bit(R5_UPTODATE, &sh->dev[i].flags);
326                         r5c_return_dev_pending_writes(conf, &sh->dev[i]);
327                         bitmap_endwrite(conf->mddev->bitmap, sh->sector,
328                                         STRIPE_SECTORS,
329                                         !test_bit(STRIPE_DEGRADED, &sh->state),
330                                         0);
331                 }
332         }
333 }
334
335 void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
336
337 /* Check whether we should flush some stripes to free up stripe cache */
338 void r5c_check_stripe_cache_usage(struct r5conf *conf)
339 {
340         int total_cached;
341
342         if (!r5c_is_writeback(conf->log))
343                 return;
344
345         total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
346                 atomic_read(&conf->r5c_cached_full_stripes);
347
348         /*
349          * The following condition is true for either of the following:
350          *   - stripe cache pressure high:
351          *          total_cached > 3/4 min_nr_stripes ||
352          *          empty_inactive_list_nr > 0
353          *   - stripe cache pressure moderate:
354          *          total_cached > 1/2 min_nr_stripes
355          */
356         if (total_cached > conf->min_nr_stripes * 1 / 2 ||
357             atomic_read(&conf->empty_inactive_list_nr) > 0)
358                 r5l_wake_reclaim(conf->log, 0);
359 }
360
361 /*
362  * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
363  * stripes in the cache
364  */
365 void r5c_check_cached_full_stripe(struct r5conf *conf)
366 {
367         if (!r5c_is_writeback(conf->log))
368                 return;
369
370         /*
371          * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
372          * or a full stripe (chunk size / 4k stripes).
373          */
374         if (atomic_read(&conf->r5c_cached_full_stripes) >=
375             min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
376                 conf->chunk_sectors >> STRIPE_SHIFT))
377                 r5l_wake_reclaim(conf->log, 0);
378 }
379
380 /*
381  * Total log space (in sectors) needed to flush all data in cache
382  *
383  * To avoid deadlock due to log space, it is necessary to reserve log
384  * space to flush critical stripes (stripes that occupying log space near
385  * last_checkpoint). This function helps check how much log space is
386  * required to flush all cached stripes.
387  *
388  * To reduce log space requirements, two mechanisms are used to give cache
389  * flush higher priorities:
390  *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
391  *       stripes ALREADY in journal can be flushed w/o pending writes;
392  *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
393  *       can be delayed (r5l_add_no_space_stripe).
394  *
395  * In cache flush, the stripe goes through 1 and then 2. For a stripe that
396  * already passed 1, flushing it requires at most (conf->max_degraded + 1)
397  * pages of journal space. For stripes that has not passed 1, flushing it
398  * requires (conf->raid_disks + 1) pages of journal space. There are at
399  * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
400  * required to flush all cached stripes (in pages) is:
401  *
402  *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
403  *     (group_cnt + 1) * (raid_disks + 1)
404  * or
405  *     (stripe_in_journal_count) * (max_degraded + 1) +
406  *     (group_cnt + 1) * (raid_disks - max_degraded)
407  */
408 static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
409 {
410         struct r5l_log *log = conf->log;
411
412         if (!r5c_is_writeback(log))
413                 return 0;
414
415         return BLOCK_SECTORS *
416                 ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
417                  (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
418 }
419
420 /*
421  * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
422  *
423  * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
424  * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
425  * device is less than 2x of reclaim_required_space.
426  */
427 static inline void r5c_update_log_state(struct r5l_log *log)
428 {
429         struct r5conf *conf = log->rdev->mddev->private;
430         sector_t free_space;
431         sector_t reclaim_space;
432         bool wake_reclaim = false;
433
434         if (!r5c_is_writeback(log))
435                 return;
436
437         free_space = r5l_ring_distance(log, log->log_start,
438                                        log->last_checkpoint);
439         reclaim_space = r5c_log_required_to_flush_cache(conf);
440         if (free_space < 2 * reclaim_space)
441                 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
442         else {
443                 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
444                         wake_reclaim = true;
445                 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
446         }
447         if (free_space < 3 * reclaim_space)
448                 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
449         else
450                 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
451
452         if (wake_reclaim)
453                 r5l_wake_reclaim(log, 0);
454 }
455
456 /*
457  * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
458  * This function should only be called in write-back mode.
459  */
460 void r5c_make_stripe_write_out(struct stripe_head *sh)
461 {
462         struct r5conf *conf = sh->raid_conf;
463         struct r5l_log *log = conf->log;
464
465         BUG_ON(!r5c_is_writeback(log));
466
467         WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
468         clear_bit(STRIPE_R5C_CACHING, &sh->state);
469
470         if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
471                 atomic_inc(&conf->preread_active_stripes);
472 }
473
474 static void r5c_handle_data_cached(struct stripe_head *sh)
475 {
476         int i;
477
478         for (i = sh->disks; i--; )
479                 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
480                         set_bit(R5_InJournal, &sh->dev[i].flags);
481                         clear_bit(R5_LOCKED, &sh->dev[i].flags);
482                 }
483         clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
484 }
485
486 /*
487  * this journal write must contain full parity,
488  * it may also contain some data pages
489  */
490 static void r5c_handle_parity_cached(struct stripe_head *sh)
491 {
492         int i;
493
494         for (i = sh->disks; i--; )
495                 if (test_bit(R5_InJournal, &sh->dev[i].flags))
496                         set_bit(R5_Wantwrite, &sh->dev[i].flags);
497 }
498
499 /*
500  * Setting proper flags after writing (or flushing) data and/or parity to the
501  * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
502  */
503 static void r5c_finish_cache_stripe(struct stripe_head *sh)
504 {
505         struct r5l_log *log = sh->raid_conf->log;
506
507         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
508                 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
509                 /*
510                  * Set R5_InJournal for parity dev[pd_idx]. This means
511                  * all data AND parity in the journal. For RAID 6, it is
512                  * NOT necessary to set the flag for dev[qd_idx], as the
513                  * two parities are written out together.
514                  */
515                 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
516         } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
517                 r5c_handle_data_cached(sh);
518         } else {
519                 r5c_handle_parity_cached(sh);
520                 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
521         }
522 }
523
524 static void r5l_io_run_stripes(struct r5l_io_unit *io)
525 {
526         struct stripe_head *sh, *next;
527
528         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
529                 list_del_init(&sh->log_list);
530
531                 r5c_finish_cache_stripe(sh);
532
533                 set_bit(STRIPE_HANDLE, &sh->state);
534                 raid5_release_stripe(sh);
535         }
536 }
537
538 static void r5l_log_run_stripes(struct r5l_log *log)
539 {
540         struct r5l_io_unit *io, *next;
541
542         lockdep_assert_held(&log->io_list_lock);
543
544         list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
545                 /* don't change list order */
546                 if (io->state < IO_UNIT_IO_END)
547                         break;
548
549                 list_move_tail(&io->log_sibling, &log->finished_ios);
550                 r5l_io_run_stripes(io);
551         }
552 }
553
554 static void r5l_move_to_end_ios(struct r5l_log *log)
555 {
556         struct r5l_io_unit *io, *next;
557
558         lockdep_assert_held(&log->io_list_lock);
559
560         list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
561                 /* don't change list order */
562                 if (io->state < IO_UNIT_IO_END)
563                         break;
564                 list_move_tail(&io->log_sibling, &log->io_end_ios);
565         }
566 }
567
568 static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
569 static void r5l_log_endio(struct bio *bio)
570 {
571         struct r5l_io_unit *io = bio->bi_private;
572         struct r5l_io_unit *io_deferred;
573         struct r5l_log *log = io->log;
574         unsigned long flags;
575         bool has_null_flush;
576         bool has_flush_payload;
577
578         if (bio->bi_status)
579                 md_error(log->rdev->mddev, log->rdev);
580
581         bio_put(bio);
582         mempool_free(io->meta_page, log->meta_pool);
583
584         spin_lock_irqsave(&log->io_list_lock, flags);
585         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
586
587         /*
588          * if the io doesn't not have null_flush or flush payload,
589          * it is not safe to access it after releasing io_list_lock.
590          * Therefore, it is necessary to check the condition with
591          * the lock held.
592          */
593         has_null_flush = io->has_null_flush;
594         has_flush_payload = io->has_flush_payload;
595
596         if (log->need_cache_flush && !list_empty(&io->stripe_list))
597                 r5l_move_to_end_ios(log);
598         else
599                 r5l_log_run_stripes(log);
600         if (!list_empty(&log->running_ios)) {
601                 /*
602                  * FLUSH/FUA io_unit is deferred because of ordering, now we
603                  * can dispatch it
604                  */
605                 io_deferred = list_first_entry(&log->running_ios,
606                                                struct r5l_io_unit, log_sibling);
607                 if (io_deferred->io_deferred)
608                         schedule_work(&log->deferred_io_work);
609         }
610
611         spin_unlock_irqrestore(&log->io_list_lock, flags);
612
613         if (log->need_cache_flush)
614                 md_wakeup_thread(log->rdev->mddev->thread);
615
616         /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
617         if (has_null_flush) {
618                 struct bio *bi;
619
620                 WARN_ON(bio_list_empty(&io->flush_barriers));
621                 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
622                         bio_endio(bi);
623                         if (atomic_dec_and_test(&io->pending_stripe)) {
624                                 __r5l_stripe_write_finished(io);
625                                 return;
626                         }
627                 }
628         }
629         /* decrease pending_stripe for flush payload */
630         if (has_flush_payload)
631                 if (atomic_dec_and_test(&io->pending_stripe))
632                         __r5l_stripe_write_finished(io);
633 }
634
635 static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
636 {
637         unsigned long flags;
638
639         spin_lock_irqsave(&log->io_list_lock, flags);
640         __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
641         spin_unlock_irqrestore(&log->io_list_lock, flags);
642
643         /*
644          * In case of journal device failures, submit_bio will get error
645          * and calls endio, then active stripes will continue write
646          * process. Therefore, it is not necessary to check Faulty bit
647          * of journal device here.
648          *
649          * We can't check split_bio after current_bio is submitted. If
650          * io->split_bio is null, after current_bio is submitted, current_bio
651          * might already be completed and the io_unit is freed. We submit
652          * split_bio first to avoid the issue.
653          */
654         if (io->split_bio) {
655                 if (io->has_flush)
656                         io->split_bio->bi_opf |= REQ_PREFLUSH;
657                 if (io->has_fua)
658                         io->split_bio->bi_opf |= REQ_FUA;
659                 submit_bio(io->split_bio);
660         }
661
662         if (io->has_flush)
663                 io->current_bio->bi_opf |= REQ_PREFLUSH;
664         if (io->has_fua)
665                 io->current_bio->bi_opf |= REQ_FUA;
666         submit_bio(io->current_bio);
667 }
668
669 /* deferred io_unit will be dispatched here */
670 static void r5l_submit_io_async(struct work_struct *work)
671 {
672         struct r5l_log *log = container_of(work, struct r5l_log,
673                                            deferred_io_work);
674         struct r5l_io_unit *io = NULL;
675         unsigned long flags;
676
677         spin_lock_irqsave(&log->io_list_lock, flags);
678         if (!list_empty(&log->running_ios)) {
679                 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
680                                       log_sibling);
681                 if (!io->io_deferred)
682                         io = NULL;
683                 else
684                         io->io_deferred = 0;
685         }
686         spin_unlock_irqrestore(&log->io_list_lock, flags);
687         if (io)
688                 r5l_do_submit_io(log, io);
689 }
690
691 static void r5c_disable_writeback_async(struct work_struct *work)
692 {
693         struct r5l_log *log = container_of(work, struct r5l_log,
694                                            disable_writeback_work);
695         struct mddev *mddev = log->rdev->mddev;
696         struct r5conf *conf = mddev->private;
697         int locked = 0;
698
699         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
700                 return;
701         pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
702                 mdname(mddev));
703
704         /* wait superblock change before suspend */
705         wait_event(mddev->sb_wait,
706                    conf->log == NULL ||
707                    (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) &&
708                     (locked = mddev_trylock(mddev))));
709         if (locked) {
710                 mddev_suspend(mddev);
711                 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
712                 mddev_resume(mddev);
713                 mddev_unlock(mddev);
714         }
715 }
716
717 static void r5l_submit_current_io(struct r5l_log *log)
718 {
719         struct r5l_io_unit *io = log->current_io;
720         struct bio *bio;
721         struct r5l_meta_block *block;
722         unsigned long flags;
723         u32 crc;
724         bool do_submit = true;
725
726         if (!io)
727                 return;
728
729         block = page_address(io->meta_page);
730         block->meta_size = cpu_to_le32(io->meta_offset);
731         crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
732         block->checksum = cpu_to_le32(crc);
733         bio = io->current_bio;
734
735         log->current_io = NULL;
736         spin_lock_irqsave(&log->io_list_lock, flags);
737         if (io->has_flush || io->has_fua) {
738                 if (io != list_first_entry(&log->running_ios,
739                                            struct r5l_io_unit, log_sibling)) {
740                         io->io_deferred = 1;
741                         do_submit = false;
742                 }
743         }
744         spin_unlock_irqrestore(&log->io_list_lock, flags);
745         if (do_submit)
746                 r5l_do_submit_io(log, io);
747 }
748
749 static struct bio *r5l_bio_alloc(struct r5l_log *log)
750 {
751         struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
752
753         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
754         bio_set_dev(bio, log->rdev->bdev);
755         bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
756
757         return bio;
758 }
759
760 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
761 {
762         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
763
764         r5c_update_log_state(log);
765         /*
766          * If we filled up the log device start from the beginning again,
767          * which will require a new bio.
768          *
769          * Note: for this to work properly the log size needs to me a multiple
770          * of BLOCK_SECTORS.
771          */
772         if (log->log_start == 0)
773                 io->need_split_bio = true;
774
775         io->log_end = log->log_start;
776 }
777
778 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
779 {
780         struct r5l_io_unit *io;
781         struct r5l_meta_block *block;
782
783         io = mempool_alloc(log->io_pool, GFP_ATOMIC);
784         if (!io)
785                 return NULL;
786         memset(io, 0, sizeof(*io));
787
788         io->log = log;
789         INIT_LIST_HEAD(&io->log_sibling);
790         INIT_LIST_HEAD(&io->stripe_list);
791         bio_list_init(&io->flush_barriers);
792         io->state = IO_UNIT_RUNNING;
793
794         io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
795         block = page_address(io->meta_page);
796         clear_page(block);
797         block->magic = cpu_to_le32(R5LOG_MAGIC);
798         block->version = R5LOG_VERSION;
799         block->seq = cpu_to_le64(log->seq);
800         block->position = cpu_to_le64(log->log_start);
801
802         io->log_start = log->log_start;
803         io->meta_offset = sizeof(struct r5l_meta_block);
804         io->seq = log->seq++;
805
806         io->current_bio = r5l_bio_alloc(log);
807         io->current_bio->bi_end_io = r5l_log_endio;
808         io->current_bio->bi_private = io;
809         bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
810
811         r5_reserve_log_entry(log, io);
812
813         spin_lock_irq(&log->io_list_lock);
814         list_add_tail(&io->log_sibling, &log->running_ios);
815         spin_unlock_irq(&log->io_list_lock);
816
817         return io;
818 }
819
820 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
821 {
822         if (log->current_io &&
823             log->current_io->meta_offset + payload_size > PAGE_SIZE)
824                 r5l_submit_current_io(log);
825
826         if (!log->current_io) {
827                 log->current_io = r5l_new_meta(log);
828                 if (!log->current_io)
829                         return -ENOMEM;
830         }
831
832         return 0;
833 }
834
835 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
836                                     sector_t location,
837                                     u32 checksum1, u32 checksum2,
838                                     bool checksum2_valid)
839 {
840         struct r5l_io_unit *io = log->current_io;
841         struct r5l_payload_data_parity *payload;
842
843         payload = page_address(io->meta_page) + io->meta_offset;
844         payload->header.type = cpu_to_le16(type);
845         payload->header.flags = cpu_to_le16(0);
846         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
847                                     (PAGE_SHIFT - 9));
848         payload->location = cpu_to_le64(location);
849         payload->checksum[0] = cpu_to_le32(checksum1);
850         if (checksum2_valid)
851                 payload->checksum[1] = cpu_to_le32(checksum2);
852
853         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
854                 sizeof(__le32) * (1 + !!checksum2_valid);
855 }
856
857 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
858 {
859         struct r5l_io_unit *io = log->current_io;
860
861         if (io->need_split_bio) {
862                 BUG_ON(io->split_bio);
863                 io->split_bio = io->current_bio;
864                 io->current_bio = r5l_bio_alloc(log);
865                 bio_chain(io->current_bio, io->split_bio);
866                 io->need_split_bio = false;
867         }
868
869         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
870                 BUG();
871
872         r5_reserve_log_entry(log, io);
873 }
874
875 static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
876 {
877         struct mddev *mddev = log->rdev->mddev;
878         struct r5conf *conf = mddev->private;
879         struct r5l_io_unit *io;
880         struct r5l_payload_flush *payload;
881         int meta_size;
882
883         /*
884          * payload_flush requires extra writes to the journal.
885          * To avoid handling the extra IO in quiesce, just skip
886          * flush_payload
887          */
888         if (conf->quiesce)
889                 return;
890
891         mutex_lock(&log->io_mutex);
892         meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
893
894         if (r5l_get_meta(log, meta_size)) {
895                 mutex_unlock(&log->io_mutex);
896                 return;
897         }
898
899         /* current implementation is one stripe per flush payload */
900         io = log->current_io;
901         payload = page_address(io->meta_page) + io->meta_offset;
902         payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
903         payload->header.flags = cpu_to_le16(0);
904         payload->size = cpu_to_le32(sizeof(__le64));
905         payload->flush_stripes[0] = cpu_to_le64(sect);
906         io->meta_offset += meta_size;
907         /* multiple flush payloads count as one pending_stripe */
908         if (!io->has_flush_payload) {
909                 io->has_flush_payload = 1;
910                 atomic_inc(&io->pending_stripe);
911         }
912         mutex_unlock(&log->io_mutex);
913 }
914
915 static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
916                            int data_pages, int parity_pages)
917 {
918         int i;
919         int meta_size;
920         int ret;
921         struct r5l_io_unit *io;
922
923         meta_size =
924                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
925                  * data_pages) +
926                 sizeof(struct r5l_payload_data_parity) +
927                 sizeof(__le32) * parity_pages;
928
929         ret = r5l_get_meta(log, meta_size);
930         if (ret)
931                 return ret;
932
933         io = log->current_io;
934
935         if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
936                 io->has_flush = 1;
937
938         for (i = 0; i < sh->disks; i++) {
939                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
940                     test_bit(R5_InJournal, &sh->dev[i].flags))
941                         continue;
942                 if (i == sh->pd_idx || i == sh->qd_idx)
943                         continue;
944                 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
945                     log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
946                         io->has_fua = 1;
947                         /*
948                          * we need to flush journal to make sure recovery can
949                          * reach the data with fua flag
950                          */
951                         io->has_flush = 1;
952                 }
953                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
954                                         raid5_compute_blocknr(sh, i, 0),
955                                         sh->dev[i].log_checksum, 0, false);
956                 r5l_append_payload_page(log, sh->dev[i].page);
957         }
958
959         if (parity_pages == 2) {
960                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
961                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
962                                         sh->dev[sh->qd_idx].log_checksum, true);
963                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
964                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
965         } else if (parity_pages == 1) {
966                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
967                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
968                                         0, false);
969                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
970         } else  /* Just writing data, not parity, in caching phase */
971                 BUG_ON(parity_pages != 0);
972
973         list_add_tail(&sh->log_list, &io->stripe_list);
974         atomic_inc(&io->pending_stripe);
975         sh->log_io = io;
976
977         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
978                 return 0;
979
980         if (sh->log_start == MaxSector) {
981                 BUG_ON(!list_empty(&sh->r5c));
982                 sh->log_start = io->log_start;
983                 spin_lock_irq(&log->stripe_in_journal_lock);
984                 list_add_tail(&sh->r5c,
985                               &log->stripe_in_journal_list);
986                 spin_unlock_irq(&log->stripe_in_journal_lock);
987                 atomic_inc(&log->stripe_in_journal_count);
988         }
989         return 0;
990 }
991
992 /* add stripe to no_space_stripes, and then wake up reclaim */
993 static inline void r5l_add_no_space_stripe(struct r5l_log *log,
994                                            struct stripe_head *sh)
995 {
996         spin_lock(&log->no_space_stripes_lock);
997         list_add_tail(&sh->log_list, &log->no_space_stripes);
998         spin_unlock(&log->no_space_stripes_lock);
999 }
1000
1001 /*
1002  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
1003  * data from log to raid disks), so we shouldn't wait for reclaim here
1004  */
1005 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
1006 {
1007         struct r5conf *conf = sh->raid_conf;
1008         int write_disks = 0;
1009         int data_pages, parity_pages;
1010         int reserve;
1011         int i;
1012         int ret = 0;
1013         bool wake_reclaim = false;
1014
1015         if (!log)
1016                 return -EAGAIN;
1017         /* Don't support stripe batch */
1018         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1019             test_bit(STRIPE_SYNCING, &sh->state)) {
1020                 /* the stripe is written to log, we start writing it to raid */
1021                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1022                 return -EAGAIN;
1023         }
1024
1025         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1026
1027         for (i = 0; i < sh->disks; i++) {
1028                 void *addr;
1029
1030                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1031                     test_bit(R5_InJournal, &sh->dev[i].flags))
1032                         continue;
1033
1034                 write_disks++;
1035                 /* checksum is already calculated in last run */
1036                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1037                         continue;
1038                 addr = kmap_atomic(sh->dev[i].page);
1039                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1040                                                     addr, PAGE_SIZE);
1041                 kunmap_atomic(addr);
1042         }
1043         parity_pages = 1 + !!(sh->qd_idx >= 0);
1044         data_pages = write_disks - parity_pages;
1045
1046         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1047         /*
1048          * The stripe must enter state machine again to finish the write, so
1049          * don't delay.
1050          */
1051         clear_bit(STRIPE_DELAYED, &sh->state);
1052         atomic_inc(&sh->count);
1053
1054         mutex_lock(&log->io_mutex);
1055         /* meta + data */
1056         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1057
1058         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1059                 if (!r5l_has_free_space(log, reserve)) {
1060                         r5l_add_no_space_stripe(log, sh);
1061                         wake_reclaim = true;
1062                 } else {
1063                         ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1064                         if (ret) {
1065                                 spin_lock_irq(&log->io_list_lock);
1066                                 list_add_tail(&sh->log_list,
1067                                               &log->no_mem_stripes);
1068                                 spin_unlock_irq(&log->io_list_lock);
1069                         }
1070                 }
1071         } else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1072                 /*
1073                  * log space critical, do not process stripes that are
1074                  * not in cache yet (sh->log_start == MaxSector).
1075                  */
1076                 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1077                     sh->log_start == MaxSector) {
1078                         r5l_add_no_space_stripe(log, sh);
1079                         wake_reclaim = true;
1080                         reserve = 0;
1081                 } else if (!r5l_has_free_space(log, reserve)) {
1082                         if (sh->log_start == log->last_checkpoint)
1083                                 BUG();
1084                         else
1085                                 r5l_add_no_space_stripe(log, sh);
1086                 } else {
1087                         ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1088                         if (ret) {
1089                                 spin_lock_irq(&log->io_list_lock);
1090                                 list_add_tail(&sh->log_list,
1091                                               &log->no_mem_stripes);
1092                                 spin_unlock_irq(&log->io_list_lock);
1093                         }
1094                 }
1095         }
1096
1097         mutex_unlock(&log->io_mutex);
1098         if (wake_reclaim)
1099                 r5l_wake_reclaim(log, reserve);
1100         return 0;
1101 }
1102
1103 void r5l_write_stripe_run(struct r5l_log *log)
1104 {
1105         if (!log)
1106                 return;
1107         mutex_lock(&log->io_mutex);
1108         r5l_submit_current_io(log);
1109         mutex_unlock(&log->io_mutex);
1110 }
1111
1112 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1113 {
1114         if (!log)
1115                 return -ENODEV;
1116
1117         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1118                 /*
1119                  * in write through (journal only)
1120                  * we flush log disk cache first, then write stripe data to
1121                  * raid disks. So if bio is finished, the log disk cache is
1122                  * flushed already. The recovery guarantees we can recovery
1123                  * the bio from log disk, so we don't need to flush again
1124                  */
1125                 if (bio->bi_iter.bi_size == 0) {
1126                         bio_endio(bio);
1127                         return 0;
1128                 }
1129                 bio->bi_opf &= ~REQ_PREFLUSH;
1130         } else {
1131                 /* write back (with cache) */
1132                 if (bio->bi_iter.bi_size == 0) {
1133                         mutex_lock(&log->io_mutex);
1134                         r5l_get_meta(log, 0);
1135                         bio_list_add(&log->current_io->flush_barriers, bio);
1136                         log->current_io->has_flush = 1;
1137                         log->current_io->has_null_flush = 1;
1138                         atomic_inc(&log->current_io->pending_stripe);
1139                         r5l_submit_current_io(log);
1140                         mutex_unlock(&log->io_mutex);
1141                         return 0;
1142                 }
1143         }
1144         return -EAGAIN;
1145 }
1146
1147 /* This will run after log space is reclaimed */
1148 static void r5l_run_no_space_stripes(struct r5l_log *log)
1149 {
1150         struct stripe_head *sh;
1151
1152         spin_lock(&log->no_space_stripes_lock);
1153         while (!list_empty(&log->no_space_stripes)) {
1154                 sh = list_first_entry(&log->no_space_stripes,
1155                                       struct stripe_head, log_list);
1156                 list_del_init(&sh->log_list);
1157                 set_bit(STRIPE_HANDLE, &sh->state);
1158                 raid5_release_stripe(sh);
1159         }
1160         spin_unlock(&log->no_space_stripes_lock);
1161 }
1162
1163 /*
1164  * calculate new last_checkpoint
1165  * for write through mode, returns log->next_checkpoint
1166  * for write back, returns log_start of first sh in stripe_in_journal_list
1167  */
1168 static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1169 {
1170         struct stripe_head *sh;
1171         struct r5l_log *log = conf->log;
1172         sector_t new_cp;
1173         unsigned long flags;
1174
1175         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1176                 return log->next_checkpoint;
1177
1178         spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1179         if (list_empty(&conf->log->stripe_in_journal_list)) {
1180                 /* all stripes flushed */
1181                 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1182                 return log->next_checkpoint;
1183         }
1184         sh = list_first_entry(&conf->log->stripe_in_journal_list,
1185                               struct stripe_head, r5c);
1186         new_cp = sh->log_start;
1187         spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1188         return new_cp;
1189 }
1190
1191 static sector_t r5l_reclaimable_space(struct r5l_log *log)
1192 {
1193         struct r5conf *conf = log->rdev->mddev->private;
1194
1195         return r5l_ring_distance(log, log->last_checkpoint,
1196                                  r5c_calculate_new_cp(conf));
1197 }
1198
1199 static void r5l_run_no_mem_stripe(struct r5l_log *log)
1200 {
1201         struct stripe_head *sh;
1202
1203         lockdep_assert_held(&log->io_list_lock);
1204
1205         if (!list_empty(&log->no_mem_stripes)) {
1206                 sh = list_first_entry(&log->no_mem_stripes,
1207                                       struct stripe_head, log_list);
1208                 list_del_init(&sh->log_list);
1209                 set_bit(STRIPE_HANDLE, &sh->state);
1210                 raid5_release_stripe(sh);
1211         }
1212 }
1213
1214 static bool r5l_complete_finished_ios(struct r5l_log *log)
1215 {
1216         struct r5l_io_unit *io, *next;
1217         bool found = false;
1218
1219         lockdep_assert_held(&log->io_list_lock);
1220
1221         list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1222                 /* don't change list order */
1223                 if (io->state < IO_UNIT_STRIPE_END)
1224                         break;
1225
1226                 log->next_checkpoint = io->log_start;
1227
1228                 list_del(&io->log_sibling);
1229                 mempool_free(io, log->io_pool);
1230                 r5l_run_no_mem_stripe(log);
1231
1232                 found = true;
1233         }
1234
1235         return found;
1236 }
1237
1238 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1239 {
1240         struct r5l_log *log = io->log;
1241         struct r5conf *conf = log->rdev->mddev->private;
1242         unsigned long flags;
1243
1244         spin_lock_irqsave(&log->io_list_lock, flags);
1245         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1246
1247         if (!r5l_complete_finished_ios(log)) {
1248                 spin_unlock_irqrestore(&log->io_list_lock, flags);
1249                 return;
1250         }
1251
1252         if (r5l_reclaimable_space(log) > log->max_free_space ||
1253             test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1254                 r5l_wake_reclaim(log, 0);
1255
1256         spin_unlock_irqrestore(&log->io_list_lock, flags);
1257         wake_up(&log->iounit_wait);
1258 }
1259
1260 void r5l_stripe_write_finished(struct stripe_head *sh)
1261 {
1262         struct r5l_io_unit *io;
1263
1264         io = sh->log_io;
1265         sh->log_io = NULL;
1266
1267         if (io && atomic_dec_and_test(&io->pending_stripe))
1268                 __r5l_stripe_write_finished(io);
1269 }
1270
1271 static void r5l_log_flush_endio(struct bio *bio)
1272 {
1273         struct r5l_log *log = container_of(bio, struct r5l_log,
1274                 flush_bio);
1275         unsigned long flags;
1276         struct r5l_io_unit *io;
1277
1278         if (bio->bi_status)
1279                 md_error(log->rdev->mddev, log->rdev);
1280
1281         spin_lock_irqsave(&log->io_list_lock, flags);
1282         list_for_each_entry(io, &log->flushing_ios, log_sibling)
1283                 r5l_io_run_stripes(io);
1284         list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1285         spin_unlock_irqrestore(&log->io_list_lock, flags);
1286 }
1287
1288 /*
1289  * Starting dispatch IO to raid.
1290  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1291  * broken meta in the middle of a log causes recovery can't find meta at the
1292  * head of log. If operations require meta at the head persistent in log, we
1293  * must make sure meta before it persistent in log too. A case is:
1294  *
1295  * stripe data/parity is in log, we start write stripe to raid disks. stripe
1296  * data/parity must be persistent in log before we do the write to raid disks.
1297  *
1298  * The solution is we restrictly maintain io_unit list order. In this case, we
1299  * only write stripes of an io_unit to raid disks till the io_unit is the first
1300  * one whose data/parity is in log.
1301  */
1302 void r5l_flush_stripe_to_raid(struct r5l_log *log)
1303 {
1304         bool do_flush;
1305
1306         if (!log || !log->need_cache_flush)
1307                 return;
1308
1309         spin_lock_irq(&log->io_list_lock);
1310         /* flush bio is running */
1311         if (!list_empty(&log->flushing_ios)) {
1312                 spin_unlock_irq(&log->io_list_lock);
1313                 return;
1314         }
1315         list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1316         do_flush = !list_empty(&log->flushing_ios);
1317         spin_unlock_irq(&log->io_list_lock);
1318
1319         if (!do_flush)
1320                 return;
1321         bio_reset(&log->flush_bio);
1322         bio_set_dev(&log->flush_bio, log->rdev->bdev);
1323         log->flush_bio.bi_end_io = r5l_log_flush_endio;
1324         log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1325         submit_bio(&log->flush_bio);
1326 }
1327
1328 static void r5l_write_super(struct r5l_log *log, sector_t cp);
1329 static void r5l_write_super_and_discard_space(struct r5l_log *log,
1330         sector_t end)
1331 {
1332         struct block_device *bdev = log->rdev->bdev;
1333         struct mddev *mddev;
1334
1335         r5l_write_super(log, end);
1336
1337         if (!blk_queue_discard(bdev_get_queue(bdev)))
1338                 return;
1339
1340         mddev = log->rdev->mddev;
1341         /*
1342          * Discard could zero data, so before discard we must make sure
1343          * superblock is updated to new log tail. Updating superblock (either
1344          * directly call md_update_sb() or depend on md thread) must hold
1345          * reconfig mutex. On the other hand, raid5_quiesce is called with
1346          * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1347          * for all IO finish, hence waitting for reclaim thread, while reclaim
1348          * thread is calling this function and waitting for reconfig mutex. So
1349          * there is a deadlock. We workaround this issue with a trylock.
1350          * FIXME: we could miss discard if we can't take reconfig mutex
1351          */
1352         set_mask_bits(&mddev->sb_flags, 0,
1353                 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1354         if (!mddev_trylock(mddev))
1355                 return;
1356         md_update_sb(mddev, 1);
1357         mddev_unlock(mddev);
1358
1359         /* discard IO error really doesn't matter, ignore it */
1360         if (log->last_checkpoint < end) {
1361                 blkdev_issue_discard(bdev,
1362                                 log->last_checkpoint + log->rdev->data_offset,
1363                                 end - log->last_checkpoint, GFP_NOIO, 0);
1364         } else {
1365                 blkdev_issue_discard(bdev,
1366                                 log->last_checkpoint + log->rdev->data_offset,
1367                                 log->device_size - log->last_checkpoint,
1368                                 GFP_NOIO, 0);
1369                 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1370                                 GFP_NOIO, 0);
1371         }
1372 }
1373
1374 /*
1375  * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1376  * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1377  *
1378  * must hold conf->device_lock
1379  */
1380 static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1381 {
1382         BUG_ON(list_empty(&sh->lru));
1383         BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1384         BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1385
1386         /*
1387          * The stripe is not ON_RELEASE_LIST, so it is safe to call
1388          * raid5_release_stripe() while holding conf->device_lock
1389          */
1390         BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1391         lockdep_assert_held(&conf->device_lock);
1392
1393         list_del_init(&sh->lru);
1394         atomic_inc(&sh->count);
1395
1396         set_bit(STRIPE_HANDLE, &sh->state);
1397         atomic_inc(&conf->active_stripes);
1398         r5c_make_stripe_write_out(sh);
1399
1400         if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1401                 atomic_inc(&conf->r5c_flushing_partial_stripes);
1402         else
1403                 atomic_inc(&conf->r5c_flushing_full_stripes);
1404         raid5_release_stripe(sh);
1405 }
1406
1407 /*
1408  * if num == 0, flush all full stripes
1409  * if num > 0, flush all full stripes. If less than num full stripes are
1410  *             flushed, flush some partial stripes until totally num stripes are
1411  *             flushed or there is no more cached stripes.
1412  */
1413 void r5c_flush_cache(struct r5conf *conf, int num)
1414 {
1415         int count;
1416         struct stripe_head *sh, *next;
1417
1418         lockdep_assert_held(&conf->device_lock);
1419         if (!conf->log)
1420                 return;
1421
1422         count = 0;
1423         list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1424                 r5c_flush_stripe(conf, sh);
1425                 count++;
1426         }
1427
1428         if (count >= num)
1429                 return;
1430         list_for_each_entry_safe(sh, next,
1431                                  &conf->r5c_partial_stripe_list, lru) {
1432                 r5c_flush_stripe(conf, sh);
1433                 if (++count >= num)
1434                         break;
1435         }
1436 }
1437
1438 static void r5c_do_reclaim(struct r5conf *conf)
1439 {
1440         struct r5l_log *log = conf->log;
1441         struct stripe_head *sh;
1442         int count = 0;
1443         unsigned long flags;
1444         int total_cached;
1445         int stripes_to_flush;
1446         int flushing_partial, flushing_full;
1447
1448         if (!r5c_is_writeback(log))
1449                 return;
1450
1451         flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1452         flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1453         total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1454                 atomic_read(&conf->r5c_cached_full_stripes) -
1455                 flushing_full - flushing_partial;
1456
1457         if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1458             atomic_read(&conf->empty_inactive_list_nr) > 0)
1459                 /*
1460                  * if stripe cache pressure high, flush all full stripes and
1461                  * some partial stripes
1462                  */
1463                 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1464         else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1465                  atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1466                  R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1467                 /*
1468                  * if stripe cache pressure moderate, or if there is many full
1469                  * stripes,flush all full stripes
1470                  */
1471                 stripes_to_flush = 0;
1472         else
1473                 /* no need to flush */
1474                 stripes_to_flush = -1;
1475
1476         if (stripes_to_flush >= 0) {
1477                 spin_lock_irqsave(&conf->device_lock, flags);
1478                 r5c_flush_cache(conf, stripes_to_flush);
1479                 spin_unlock_irqrestore(&conf->device_lock, flags);
1480         }
1481
1482         /* if log space is tight, flush stripes on stripe_in_journal_list */
1483         if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1484                 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1485                 spin_lock(&conf->device_lock);
1486                 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1487                         /*
1488                          * stripes on stripe_in_journal_list could be in any
1489                          * state of the stripe_cache state machine. In this
1490                          * case, we only want to flush stripe on
1491                          * r5c_cached_full/partial_stripes. The following
1492                          * condition makes sure the stripe is on one of the
1493                          * two lists.
1494                          */
1495                         if (!list_empty(&sh->lru) &&
1496                             !test_bit(STRIPE_HANDLE, &sh->state) &&
1497                             atomic_read(&sh->count) == 0) {
1498                                 r5c_flush_stripe(conf, sh);
1499                                 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1500                                         break;
1501                         }
1502                 }
1503                 spin_unlock(&conf->device_lock);
1504                 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1505         }
1506
1507         if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1508                 r5l_run_no_space_stripes(log);
1509
1510         md_wakeup_thread(conf->mddev->thread);
1511 }
1512
1513 static void r5l_do_reclaim(struct r5l_log *log)
1514 {
1515         struct r5conf *conf = log->rdev->mddev->private;
1516         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1517         sector_t reclaimable;
1518         sector_t next_checkpoint;
1519         bool write_super;
1520
1521         spin_lock_irq(&log->io_list_lock);
1522         write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1523                 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1524         /*
1525          * move proper io_unit to reclaim list. We should not change the order.
1526          * reclaimable/unreclaimable io_unit can be mixed in the list, we
1527          * shouldn't reuse space of an unreclaimable io_unit
1528          */
1529         while (1) {
1530                 reclaimable = r5l_reclaimable_space(log);
1531                 if (reclaimable >= reclaim_target ||
1532                     (list_empty(&log->running_ios) &&
1533                      list_empty(&log->io_end_ios) &&
1534                      list_empty(&log->flushing_ios) &&
1535                      list_empty(&log->finished_ios)))
1536                         break;
1537
1538                 md_wakeup_thread(log->rdev->mddev->thread);
1539                 wait_event_lock_irq(log->iounit_wait,
1540                                     r5l_reclaimable_space(log) > reclaimable,
1541                                     log->io_list_lock);
1542         }
1543
1544         next_checkpoint = r5c_calculate_new_cp(conf);
1545         spin_unlock_irq(&log->io_list_lock);
1546
1547         if (reclaimable == 0 || !write_super)
1548                 return;
1549
1550         /*
1551          * write_super will flush cache of each raid disk. We must write super
1552          * here, because the log area might be reused soon and we don't want to
1553          * confuse recovery
1554          */
1555         r5l_write_super_and_discard_space(log, next_checkpoint);
1556
1557         mutex_lock(&log->io_mutex);
1558         log->last_checkpoint = next_checkpoint;
1559         r5c_update_log_state(log);
1560         mutex_unlock(&log->io_mutex);
1561
1562         r5l_run_no_space_stripes(log);
1563 }
1564
1565 static void r5l_reclaim_thread(struct md_thread *thread)
1566 {
1567         struct mddev *mddev = thread->mddev;
1568         struct r5conf *conf = mddev->private;
1569         struct r5l_log *log = conf->log;
1570
1571         if (!log)
1572                 return;
1573         r5c_do_reclaim(conf);
1574         r5l_do_reclaim(log);
1575 }
1576
1577 void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1578 {
1579         unsigned long target;
1580         unsigned long new = (unsigned long)space; /* overflow in theory */
1581
1582         if (!log)
1583                 return;
1584         do {
1585                 target = log->reclaim_target;
1586                 if (new < target)
1587                         return;
1588         } while (cmpxchg(&log->reclaim_target, target, new) != target);
1589         md_wakeup_thread(log->reclaim_thread);
1590 }
1591
1592 void r5l_quiesce(struct r5l_log *log, int quiesce)
1593 {
1594         struct mddev *mddev;
1595         if (!log)
1596                 return;
1597
1598         if (quiesce) {
1599                 /* make sure r5l_write_super_and_discard_space exits */
1600                 mddev = log->rdev->mddev;
1601                 wake_up(&mddev->sb_wait);
1602                 kthread_park(log->reclaim_thread->tsk);
1603                 r5l_wake_reclaim(log, MaxSector);
1604                 r5l_do_reclaim(log);
1605         } else
1606                 kthread_unpark(log->reclaim_thread->tsk);
1607 }
1608
1609 bool r5l_log_disk_error(struct r5conf *conf)
1610 {
1611         struct r5l_log *log;
1612         bool ret;
1613         /* don't allow write if journal disk is missing */
1614         rcu_read_lock();
1615         log = rcu_dereference(conf->log);
1616
1617         if (!log)
1618                 ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1619         else
1620                 ret = test_bit(Faulty, &log->rdev->flags);
1621         rcu_read_unlock();
1622         return ret;
1623 }
1624
1625 #define R5L_RECOVERY_PAGE_POOL_SIZE 256
1626
1627 struct r5l_recovery_ctx {
1628         struct page *meta_page;         /* current meta */
1629         sector_t meta_total_blocks;     /* total size of current meta and data */
1630         sector_t pos;                   /* recovery position */
1631         u64 seq;                        /* recovery position seq */
1632         int data_parity_stripes;        /* number of data_parity stripes */
1633         int data_only_stripes;          /* number of data_only stripes */
1634         struct list_head cached_list;
1635
1636         /*
1637          * read ahead page pool (ra_pool)
1638          * in recovery, log is read sequentially. It is not efficient to
1639          * read every page with sync_page_io(). The read ahead page pool
1640          * reads multiple pages with one IO, so further log read can
1641          * just copy data from the pool.
1642          */
1643         struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1644         sector_t pool_offset;   /* offset of first page in the pool */
1645         int total_pages;        /* total allocated pages */
1646         int valid_pages;        /* pages with valid data */
1647         struct bio *ra_bio;     /* bio to do the read ahead */
1648 };
1649
1650 static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1651                                             struct r5l_recovery_ctx *ctx)
1652 {
1653         struct page *page;
1654
1655         ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs);
1656         if (!ctx->ra_bio)
1657                 return -ENOMEM;
1658
1659         ctx->valid_pages = 0;
1660         ctx->total_pages = 0;
1661         while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1662                 page = alloc_page(GFP_KERNEL);
1663
1664                 if (!page)
1665                         break;
1666                 ctx->ra_pool[ctx->total_pages] = page;
1667                 ctx->total_pages += 1;
1668         }
1669
1670         if (ctx->total_pages == 0) {
1671                 bio_put(ctx->ra_bio);
1672                 return -ENOMEM;
1673         }
1674
1675         ctx->pool_offset = 0;
1676         return 0;
1677 }
1678
1679 static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1680                                         struct r5l_recovery_ctx *ctx)
1681 {
1682         int i;
1683
1684         for (i = 0; i < ctx->total_pages; ++i)
1685                 put_page(ctx->ra_pool[i]);
1686         bio_put(ctx->ra_bio);
1687 }
1688
1689 /*
1690  * fetch ctx->valid_pages pages from offset
1691  * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1692  * However, if the offset is close to the end of the journal device,
1693  * ctx->valid_pages could be smaller than ctx->total_pages
1694  */
1695 static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1696                                       struct r5l_recovery_ctx *ctx,
1697                                       sector_t offset)
1698 {
1699         bio_reset(ctx->ra_bio);
1700         bio_set_dev(ctx->ra_bio, log->rdev->bdev);
1701         bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1702         ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
1703
1704         ctx->valid_pages = 0;
1705         ctx->pool_offset = offset;
1706
1707         while (ctx->valid_pages < ctx->total_pages) {
1708                 bio_add_page(ctx->ra_bio,
1709                              ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1710                 ctx->valid_pages += 1;
1711
1712                 offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1713
1714                 if (offset == 0)  /* reached end of the device */
1715                         break;
1716         }
1717
1718         return submit_bio_wait(ctx->ra_bio);
1719 }
1720
1721 /*
1722  * try read a page from the read ahead page pool, if the page is not in the
1723  * pool, call r5l_recovery_fetch_ra_pool
1724  */
1725 static int r5l_recovery_read_page(struct r5l_log *log,
1726                                   struct r5l_recovery_ctx *ctx,
1727                                   struct page *page,
1728                                   sector_t offset)
1729 {
1730         int ret;
1731
1732         if (offset < ctx->pool_offset ||
1733             offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1734                 ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1735                 if (ret)
1736                         return ret;
1737         }
1738
1739         BUG_ON(offset < ctx->pool_offset ||
1740                offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1741
1742         memcpy(page_address(page),
1743                page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1744                                          BLOCK_SECTOR_SHIFT]),
1745                PAGE_SIZE);
1746         return 0;
1747 }
1748
1749 static int r5l_recovery_read_meta_block(struct r5l_log *log,
1750                                         struct r5l_recovery_ctx *ctx)
1751 {
1752         struct page *page = ctx->meta_page;
1753         struct r5l_meta_block *mb;
1754         u32 crc, stored_crc;
1755         int ret;
1756
1757         ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1758         if (ret != 0)
1759                 return ret;
1760
1761         mb = page_address(page);
1762         stored_crc = le32_to_cpu(mb->checksum);
1763         mb->checksum = 0;
1764
1765         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1766             le64_to_cpu(mb->seq) != ctx->seq ||
1767             mb->version != R5LOG_VERSION ||
1768             le64_to_cpu(mb->position) != ctx->pos)
1769                 return -EINVAL;
1770
1771         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1772         if (stored_crc != crc)
1773                 return -EINVAL;
1774
1775         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1776                 return -EINVAL;
1777
1778         ctx->meta_total_blocks = BLOCK_SECTORS;
1779
1780         return 0;
1781 }
1782
1783 static void
1784 r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1785                                      struct page *page,
1786                                      sector_t pos, u64 seq)
1787 {
1788         struct r5l_meta_block *mb;
1789
1790         mb = page_address(page);
1791         clear_page(mb);
1792         mb->magic = cpu_to_le32(R5LOG_MAGIC);
1793         mb->version = R5LOG_VERSION;
1794         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1795         mb->seq = cpu_to_le64(seq);
1796         mb->position = cpu_to_le64(pos);
1797 }
1798
1799 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1800                                           u64 seq)
1801 {
1802         struct page *page;
1803         struct r5l_meta_block *mb;
1804
1805         page = alloc_page(GFP_KERNEL);
1806         if (!page)
1807                 return -ENOMEM;
1808         r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1809         mb = page_address(page);
1810         mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1811                                              mb, PAGE_SIZE));
1812         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1813                           REQ_SYNC | REQ_FUA, false)) {
1814                 __free_page(page);
1815                 return -EIO;
1816         }
1817         __free_page(page);
1818         return 0;
1819 }
1820
1821 /*
1822  * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1823  * to mark valid (potentially not flushed) data in the journal.
1824  *
1825  * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1826  * so there should not be any mismatch here.
1827  */
1828 static void r5l_recovery_load_data(struct r5l_log *log,
1829                                    struct stripe_head *sh,
1830                                    struct r5l_recovery_ctx *ctx,
1831                                    struct r5l_payload_data_parity *payload,
1832                                    sector_t log_offset)
1833 {
1834         struct mddev *mddev = log->rdev->mddev;
1835         struct r5conf *conf = mddev->private;
1836         int dd_idx;
1837
1838         raid5_compute_sector(conf,
1839                              le64_to_cpu(payload->location), 0,
1840                              &dd_idx, sh);
1841         r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1842         sh->dev[dd_idx].log_checksum =
1843                 le32_to_cpu(payload->checksum[0]);
1844         ctx->meta_total_blocks += BLOCK_SECTORS;
1845
1846         set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1847         set_bit(STRIPE_R5C_CACHING, &sh->state);
1848 }
1849
1850 static void r5l_recovery_load_parity(struct r5l_log *log,
1851                                      struct stripe_head *sh,
1852                                      struct r5l_recovery_ctx *ctx,
1853                                      struct r5l_payload_data_parity *payload,
1854                                      sector_t log_offset)
1855 {
1856         struct mddev *mddev = log->rdev->mddev;
1857         struct r5conf *conf = mddev->private;
1858
1859         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1860         r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1861         sh->dev[sh->pd_idx].log_checksum =
1862                 le32_to_cpu(payload->checksum[0]);
1863         set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1864
1865         if (sh->qd_idx >= 0) {
1866                 r5l_recovery_read_page(
1867                         log, ctx, sh->dev[sh->qd_idx].page,
1868                         r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1869                 sh->dev[sh->qd_idx].log_checksum =
1870                         le32_to_cpu(payload->checksum[1]);
1871                 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1872         }
1873         clear_bit(STRIPE_R5C_CACHING, &sh->state);
1874 }
1875
1876 static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1877 {
1878         int i;
1879
1880         sh->state = 0;
1881         sh->log_start = MaxSector;
1882         for (i = sh->disks; i--; )
1883                 sh->dev[i].flags = 0;
1884 }
1885
1886 static void
1887 r5l_recovery_replay_one_stripe(struct r5conf *conf,
1888                                struct stripe_head *sh,
1889                                struct r5l_recovery_ctx *ctx)
1890 {
1891         struct md_rdev *rdev, *rrdev;
1892         int disk_index;
1893         int data_count = 0;
1894
1895         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1896                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1897                         continue;
1898                 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1899                         continue;
1900                 data_count++;
1901         }
1902
1903         /*
1904          * stripes that only have parity must have been flushed
1905          * before the crash that we are now recovering from, so
1906          * there is nothing more to recovery.
1907          */
1908         if (data_count == 0)
1909                 goto out;
1910
1911         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1912                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1913                         continue;
1914
1915                 /* in case device is broken */
1916                 rcu_read_lock();
1917                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1918                 if (rdev) {
1919                         atomic_inc(&rdev->nr_pending);
1920                         rcu_read_unlock();
1921                         sync_page_io(rdev, sh->sector, PAGE_SIZE,
1922                                      sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1923                                      false);
1924                         rdev_dec_pending(rdev, rdev->mddev);
1925                         rcu_read_lock();
1926                 }
1927                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1928                 if (rrdev) {
1929                         atomic_inc(&rrdev->nr_pending);
1930                         rcu_read_unlock();
1931                         sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1932                                      sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1933                                      false);
1934                         rdev_dec_pending(rrdev, rrdev->mddev);
1935                         rcu_read_lock();
1936                 }
1937                 rcu_read_unlock();
1938         }
1939         ctx->data_parity_stripes++;
1940 out:
1941         r5l_recovery_reset_stripe(sh);
1942 }
1943
1944 static struct stripe_head *
1945 r5c_recovery_alloc_stripe(struct r5conf *conf,
1946                           sector_t stripe_sect)
1947 {
1948         struct stripe_head *sh;
1949
1950         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1951         if (!sh)
1952                 return NULL;  /* no more stripe available */
1953
1954         r5l_recovery_reset_stripe(sh);
1955
1956         return sh;
1957 }
1958
1959 static struct stripe_head *
1960 r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1961 {
1962         struct stripe_head *sh;
1963
1964         list_for_each_entry(sh, list, lru)
1965                 if (sh->sector == sect)
1966                         return sh;
1967         return NULL;
1968 }
1969
1970 static void
1971 r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1972                           struct r5l_recovery_ctx *ctx)
1973 {
1974         struct stripe_head *sh, *next;
1975
1976         list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1977                 r5l_recovery_reset_stripe(sh);
1978                 list_del_init(&sh->lru);
1979                 raid5_release_stripe(sh);
1980         }
1981 }
1982
1983 static void
1984 r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1985                             struct r5l_recovery_ctx *ctx)
1986 {
1987         struct stripe_head *sh, *next;
1988
1989         list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1990                 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1991                         r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1992                         list_del_init(&sh->lru);
1993                         raid5_release_stripe(sh);
1994                 }
1995 }
1996
1997 /* if matches return 0; otherwise return -EINVAL */
1998 static int
1999 r5l_recovery_verify_data_checksum(struct r5l_log *log,
2000                                   struct r5l_recovery_ctx *ctx,
2001                                   struct page *page,
2002                                   sector_t log_offset, __le32 log_checksum)
2003 {
2004         void *addr;
2005         u32 checksum;
2006
2007         r5l_recovery_read_page(log, ctx, page, log_offset);
2008         addr = kmap_atomic(page);
2009         checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
2010         kunmap_atomic(addr);
2011         return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
2012 }
2013
2014 /*
2015  * before loading data to stripe cache, we need verify checksum for all data,
2016  * if there is mismatch for any data page, we drop all data in the mata block
2017  */
2018 static int
2019 r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
2020                                          struct r5l_recovery_ctx *ctx)
2021 {
2022         struct mddev *mddev = log->rdev->mddev;
2023         struct r5conf *conf = mddev->private;
2024         struct r5l_meta_block *mb = page_address(ctx->meta_page);
2025         sector_t mb_offset = sizeof(struct r5l_meta_block);
2026         sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2027         struct page *page;
2028         struct r5l_payload_data_parity *payload;
2029         struct r5l_payload_flush *payload_flush;
2030
2031         page = alloc_page(GFP_KERNEL);
2032         if (!page)
2033                 return -ENOMEM;
2034
2035         while (mb_offset < le32_to_cpu(mb->meta_size)) {
2036                 payload = (void *)mb + mb_offset;
2037                 payload_flush = (void *)mb + mb_offset;
2038
2039                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2040                         if (r5l_recovery_verify_data_checksum(
2041                                     log, ctx, page, log_offset,
2042                                     payload->checksum[0]) < 0)
2043                                 goto mismatch;
2044                 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2045                         if (r5l_recovery_verify_data_checksum(
2046                                     log, ctx, page, log_offset,
2047                                     payload->checksum[0]) < 0)
2048                                 goto mismatch;
2049                         if (conf->max_degraded == 2 && /* q for RAID 6 */
2050                             r5l_recovery_verify_data_checksum(
2051                                     log, ctx, page,
2052                                     r5l_ring_add(log, log_offset,
2053                                                  BLOCK_SECTORS),
2054                                     payload->checksum[1]) < 0)
2055                                 goto mismatch;
2056                 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2057                         /* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2058                 } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2059                         goto mismatch;
2060
2061                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2062                         mb_offset += sizeof(struct r5l_payload_flush) +
2063                                 le32_to_cpu(payload_flush->size);
2064                 } else {
2065                         /* DATA or PARITY payload */
2066                         log_offset = r5l_ring_add(log, log_offset,
2067                                                   le32_to_cpu(payload->size));
2068                         mb_offset += sizeof(struct r5l_payload_data_parity) +
2069                                 sizeof(__le32) *
2070                                 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2071                 }
2072
2073         }
2074
2075         put_page(page);
2076         return 0;
2077
2078 mismatch:
2079         put_page(page);
2080         return -EINVAL;
2081 }
2082
2083 /*
2084  * Analyze all data/parity pages in one meta block
2085  * Returns:
2086  * 0 for success
2087  * -EINVAL for unknown playload type
2088  * -EAGAIN for checksum mismatch of data page
2089  * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2090  */
2091 static int
2092 r5c_recovery_analyze_meta_block(struct r5l_log *log,
2093                                 struct r5l_recovery_ctx *ctx,
2094                                 struct list_head *cached_stripe_list)
2095 {
2096         struct mddev *mddev = log->rdev->mddev;
2097         struct r5conf *conf = mddev->private;
2098         struct r5l_meta_block *mb;
2099         struct r5l_payload_data_parity *payload;
2100         struct r5l_payload_flush *payload_flush;
2101         int mb_offset;
2102         sector_t log_offset;
2103         sector_t stripe_sect;
2104         struct stripe_head *sh;
2105         int ret;
2106
2107         /*
2108          * for mismatch in data blocks, we will drop all data in this mb, but
2109          * we will still read next mb for other data with FLUSH flag, as
2110          * io_unit could finish out of order.
2111          */
2112         ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2113         if (ret == -EINVAL)
2114                 return -EAGAIN;
2115         else if (ret)
2116                 return ret;   /* -ENOMEM duo to alloc_page() failed */
2117
2118         mb = page_address(ctx->meta_page);
2119         mb_offset = sizeof(struct r5l_meta_block);
2120         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2121
2122         while (mb_offset < le32_to_cpu(mb->meta_size)) {
2123                 int dd;
2124
2125                 payload = (void *)mb + mb_offset;
2126                 payload_flush = (void *)mb + mb_offset;
2127
2128                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2129                         int i, count;
2130
2131                         count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2132                         for (i = 0; i < count; ++i) {
2133                                 stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2134                                 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2135                                                                 stripe_sect);
2136                                 if (sh) {
2137                                         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2138                                         r5l_recovery_reset_stripe(sh);
2139                                         list_del_init(&sh->lru);
2140                                         raid5_release_stripe(sh);
2141                                 }
2142                         }
2143
2144                         mb_offset += sizeof(struct r5l_payload_flush) +
2145                                 le32_to_cpu(payload_flush->size);
2146                         continue;
2147                 }
2148
2149                 /* DATA or PARITY payload */
2150                 stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2151                         raid5_compute_sector(
2152                                 conf, le64_to_cpu(payload->location), 0, &dd,
2153                                 NULL)
2154                         : le64_to_cpu(payload->location);
2155
2156                 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2157                                                 stripe_sect);
2158
2159                 if (!sh) {
2160                         sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
2161                         /*
2162                          * cannot get stripe from raid5_get_active_stripe
2163                          * try replay some stripes
2164                          */
2165                         if (!sh) {
2166                                 r5c_recovery_replay_stripes(
2167                                         cached_stripe_list, ctx);
2168                                 sh = r5c_recovery_alloc_stripe(
2169                                         conf, stripe_sect);
2170                         }
2171                         if (!sh) {
2172                                 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2173                                         mdname(mddev),
2174                                         conf->min_nr_stripes * 2);
2175                                 raid5_set_cache_size(mddev,
2176                                                      conf->min_nr_stripes * 2);
2177                                 sh = r5c_recovery_alloc_stripe(conf,
2178                                                                stripe_sect);
2179                         }
2180                         if (!sh) {
2181                                 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2182                                        mdname(mddev));
2183                                 return -ENOMEM;
2184                         }
2185                         list_add_tail(&sh->lru, cached_stripe_list);
2186                 }
2187
2188                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2189                         if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2190                             test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2191                                 r5l_recovery_replay_one_stripe(conf, sh, ctx);
2192                                 list_move_tail(&sh->lru, cached_stripe_list);
2193                         }
2194                         r5l_recovery_load_data(log, sh, ctx, payload,
2195                                                log_offset);
2196                 } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2197                         r5l_recovery_load_parity(log, sh, ctx, payload,
2198                                                  log_offset);
2199                 else
2200                         return -EINVAL;
2201
2202                 log_offset = r5l_ring_add(log, log_offset,
2203                                           le32_to_cpu(payload->size));
2204
2205                 mb_offset += sizeof(struct r5l_payload_data_parity) +
2206                         sizeof(__le32) *
2207                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2208         }
2209
2210         return 0;
2211 }
2212
2213 /*
2214  * Load the stripe into cache. The stripe will be written out later by
2215  * the stripe cache state machine.
2216  */
2217 static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2218                                          struct stripe_head *sh)
2219 {
2220         struct r5dev *dev;
2221         int i;
2222
2223         for (i = sh->disks; i--; ) {
2224                 dev = sh->dev + i;
2225                 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2226                         set_bit(R5_InJournal, &dev->flags);
2227                         set_bit(R5_UPTODATE, &dev->flags);
2228                 }
2229         }
2230 }
2231
2232 /*
2233  * Scan through the log for all to-be-flushed data
2234  *
2235  * For stripes with data and parity, namely Data-Parity stripe
2236  * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2237  *
2238  * For stripes with only data, namely Data-Only stripe
2239  * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2240  *
2241  * For a stripe, if we see data after parity, we should discard all previous
2242  * data and parity for this stripe, as these data are already flushed to
2243  * the array.
2244  *
2245  * At the end of the scan, we return the new journal_tail, which points to
2246  * first data-only stripe on the journal device, or next invalid meta block.
2247  */
2248 static int r5c_recovery_flush_log(struct r5l_log *log,
2249                                   struct r5l_recovery_ctx *ctx)
2250 {
2251         struct stripe_head *sh;
2252         int ret = 0;
2253
2254         /* scan through the log */
2255         while (1) {
2256                 if (r5l_recovery_read_meta_block(log, ctx))
2257                         break;
2258
2259                 ret = r5c_recovery_analyze_meta_block(log, ctx,
2260                                                       &ctx->cached_list);
2261                 /*
2262                  * -EAGAIN means mismatch in data block, in this case, we still
2263                  * try scan the next metablock
2264                  */
2265                 if (ret && ret != -EAGAIN)
2266                         break;   /* ret == -EINVAL or -ENOMEM */
2267                 ctx->seq++;
2268                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2269         }
2270
2271         if (ret == -ENOMEM) {
2272                 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2273                 return ret;
2274         }
2275
2276         /* replay data-parity stripes */
2277         r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2278
2279         /* load data-only stripes to stripe cache */
2280         list_for_each_entry(sh, &ctx->cached_list, lru) {
2281                 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2282                 r5c_recovery_load_one_stripe(log, sh);
2283                 ctx->data_only_stripes++;
2284         }
2285
2286         return 0;
2287 }
2288
2289 /*
2290  * we did a recovery. Now ctx.pos points to an invalid meta block. New
2291  * log will start here. but we can't let superblock point to last valid
2292  * meta block. The log might looks like:
2293  * | meta 1| meta 2| meta 3|
2294  * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2295  * superblock points to meta 1, we write a new valid meta 2n.  if crash
2296  * happens again, new recovery will start from meta 1. Since meta 2n is
2297  * valid now, recovery will think meta 3 is valid, which is wrong.
2298  * The solution is we create a new meta in meta2 with its seq == meta
2299  * 1's seq + 10000 and let superblock points to meta2. The same recovery
2300  * will not think meta 3 is a valid meta, because its seq doesn't match
2301  */
2302
2303 /*
2304  * Before recovery, the log looks like the following
2305  *
2306  *   ---------------------------------------------
2307  *   |           valid log        | invalid log  |
2308  *   ---------------------------------------------
2309  *   ^
2310  *   |- log->last_checkpoint
2311  *   |- log->last_cp_seq
2312  *
2313  * Now we scan through the log until we see invalid entry
2314  *
2315  *   ---------------------------------------------
2316  *   |           valid log        | invalid log  |
2317  *   ---------------------------------------------
2318  *   ^                            ^
2319  *   |- log->last_checkpoint      |- ctx->pos
2320  *   |- log->last_cp_seq          |- ctx->seq
2321  *
2322  * From this point, we need to increase seq number by 10 to avoid
2323  * confusing next recovery.
2324  *
2325  *   ---------------------------------------------
2326  *   |           valid log        | invalid log  |
2327  *   ---------------------------------------------
2328  *   ^                              ^
2329  *   |- log->last_checkpoint        |- ctx->pos+1
2330  *   |- log->last_cp_seq            |- ctx->seq+10001
2331  *
2332  * However, it is not safe to start the state machine yet, because data only
2333  * parities are not yet secured in RAID. To save these data only parities, we
2334  * rewrite them from seq+11.
2335  *
2336  *   -----------------------------------------------------------------
2337  *   |           valid log        | data only stripes | invalid log  |
2338  *   -----------------------------------------------------------------
2339  *   ^                                                ^
2340  *   |- log->last_checkpoint                          |- ctx->pos+n
2341  *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2342  *
2343  * If failure happens again during this process, the recovery can safe start
2344  * again from log->last_checkpoint.
2345  *
2346  * Once data only stripes are rewritten to journal, we move log_tail
2347  *
2348  *   -----------------------------------------------------------------
2349  *   |     old log        |    data only stripes    | invalid log  |
2350  *   -----------------------------------------------------------------
2351  *                        ^                         ^
2352  *                        |- log->last_checkpoint   |- ctx->pos+n
2353  *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2354  *
2355  * Then we can safely start the state machine. If failure happens from this
2356  * point on, the recovery will start from new log->last_checkpoint.
2357  */
2358 static int
2359 r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2360                                        struct r5l_recovery_ctx *ctx)
2361 {
2362         struct stripe_head *sh;
2363         struct mddev *mddev = log->rdev->mddev;
2364         struct page *page;
2365         sector_t next_checkpoint = MaxSector;
2366
2367         page = alloc_page(GFP_KERNEL);
2368         if (!page) {
2369                 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2370                        mdname(mddev));
2371                 return -ENOMEM;
2372         }
2373
2374         WARN_ON(list_empty(&ctx->cached_list));
2375
2376         list_for_each_entry(sh, &ctx->cached_list, lru) {
2377                 struct r5l_meta_block *mb;
2378                 int i;
2379                 int offset;
2380                 sector_t write_pos;
2381
2382                 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2383                 r5l_recovery_create_empty_meta_block(log, page,
2384                                                      ctx->pos, ctx->seq);
2385                 mb = page_address(page);
2386                 offset = le32_to_cpu(mb->meta_size);
2387                 write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2388
2389                 for (i = sh->disks; i--; ) {
2390                         struct r5dev *dev = &sh->dev[i];
2391                         struct r5l_payload_data_parity *payload;
2392                         void *addr;
2393
2394                         if (test_bit(R5_InJournal, &dev->flags)) {
2395                                 payload = (void *)mb + offset;
2396                                 payload->header.type = cpu_to_le16(
2397                                         R5LOG_PAYLOAD_DATA);
2398                                 payload->size = cpu_to_le32(BLOCK_SECTORS);
2399                                 payload->location = cpu_to_le64(
2400                                         raid5_compute_blocknr(sh, i, 0));
2401                                 addr = kmap_atomic(dev->page);
2402                                 payload->checksum[0] = cpu_to_le32(
2403                                         crc32c_le(log->uuid_checksum, addr,
2404                                                   PAGE_SIZE));
2405                                 kunmap_atomic(addr);
2406                                 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2407                                              dev->page, REQ_OP_WRITE, 0, false);
2408                                 write_pos = r5l_ring_add(log, write_pos,
2409                                                          BLOCK_SECTORS);
2410                                 offset += sizeof(__le32) +
2411                                         sizeof(struct r5l_payload_data_parity);
2412
2413                         }
2414                 }
2415                 mb->meta_size = cpu_to_le32(offset);
2416                 mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2417                                                      mb, PAGE_SIZE));
2418                 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2419                              REQ_OP_WRITE, REQ_SYNC | REQ_FUA, false);
2420                 sh->log_start = ctx->pos;
2421                 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2422                 atomic_inc(&log->stripe_in_journal_count);
2423                 ctx->pos = write_pos;
2424                 ctx->seq += 1;
2425                 next_checkpoint = sh->log_start;
2426         }
2427         log->next_checkpoint = next_checkpoint;
2428         __free_page(page);
2429         return 0;
2430 }
2431
2432 static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2433                                                  struct r5l_recovery_ctx *ctx)
2434 {
2435         struct mddev *mddev = log->rdev->mddev;
2436         struct r5conf *conf = mddev->private;
2437         struct stripe_head *sh, *next;
2438
2439         if (ctx->data_only_stripes == 0)
2440                 return;
2441
2442         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2443
2444         list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2445                 r5c_make_stripe_write_out(sh);
2446                 set_bit(STRIPE_HANDLE, &sh->state);
2447                 list_del_init(&sh->lru);
2448                 raid5_release_stripe(sh);
2449         }
2450
2451         /* reuse conf->wait_for_quiescent in recovery */
2452         wait_event(conf->wait_for_quiescent,
2453                    atomic_read(&conf->active_stripes) == 0);
2454
2455         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2456 }
2457
2458 static int r5l_recovery_log(struct r5l_log *log)
2459 {
2460         struct mddev *mddev = log->rdev->mddev;
2461         struct r5l_recovery_ctx *ctx;
2462         int ret;
2463         sector_t pos;
2464
2465         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2466         if (!ctx)
2467                 return -ENOMEM;
2468
2469         ctx->pos = log->last_checkpoint;
2470         ctx->seq = log->last_cp_seq;
2471         INIT_LIST_HEAD(&ctx->cached_list);
2472         ctx->meta_page = alloc_page(GFP_KERNEL);
2473
2474         if (!ctx->meta_page) {
2475                 ret =  -ENOMEM;
2476                 goto meta_page;
2477         }
2478
2479         if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2480                 ret = -ENOMEM;
2481                 goto ra_pool;
2482         }
2483
2484         ret = r5c_recovery_flush_log(log, ctx);
2485
2486         if (ret)
2487                 goto error;
2488
2489         pos = ctx->pos;
2490         ctx->seq += 10000;
2491
2492         if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2493                 pr_info("md/raid:%s: starting from clean shutdown\n",
2494                          mdname(mddev));
2495         else
2496                 pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2497                          mdname(mddev), ctx->data_only_stripes,
2498                          ctx->data_parity_stripes);
2499
2500         if (ctx->data_only_stripes == 0) {
2501                 log->next_checkpoint = ctx->pos;
2502                 r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2503                 ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2504         } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2505                 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2506                        mdname(mddev));
2507                 ret =  -EIO;
2508                 goto error;
2509         }
2510
2511         log->log_start = ctx->pos;
2512         log->seq = ctx->seq;
2513         log->last_checkpoint = pos;
2514         r5l_write_super(log, pos);
2515
2516         r5c_recovery_flush_data_only_stripes(log, ctx);
2517         ret = 0;
2518 error:
2519         r5l_recovery_free_ra_pool(log, ctx);
2520 ra_pool:
2521         __free_page(ctx->meta_page);
2522 meta_page:
2523         kfree(ctx);
2524         return ret;
2525 }
2526
2527 static void r5l_write_super(struct r5l_log *log, sector_t cp)
2528 {
2529         struct mddev *mddev = log->rdev->mddev;
2530
2531         log->rdev->journal_tail = cp;
2532         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2533 }
2534
2535 static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2536 {
2537         struct r5conf *conf;
2538         int ret;
2539
2540         ret = mddev_lock(mddev);
2541         if (ret)
2542                 return ret;
2543
2544         conf = mddev->private;
2545         if (!conf || !conf->log) {
2546                 mddev_unlock(mddev);
2547                 return 0;
2548         }
2549
2550         switch (conf->log->r5c_journal_mode) {
2551         case R5C_JOURNAL_MODE_WRITE_THROUGH:
2552                 ret = snprintf(
2553                         page, PAGE_SIZE, "[%s] %s\n",
2554                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2555                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2556                 break;
2557         case R5C_JOURNAL_MODE_WRITE_BACK:
2558                 ret = snprintf(
2559                         page, PAGE_SIZE, "%s [%s]\n",
2560                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2561                         r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2562                 break;
2563         default:
2564                 ret = 0;
2565         }
2566         mddev_unlock(mddev);
2567         return ret;
2568 }
2569
2570 /*
2571  * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2572  *
2573  * @mode as defined in 'enum r5c_journal_mode'.
2574  *
2575  */
2576 int r5c_journal_mode_set(struct mddev *mddev, int mode)
2577 {
2578         struct r5conf *conf;
2579
2580         if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2581             mode > R5C_JOURNAL_MODE_WRITE_BACK)
2582                 return -EINVAL;
2583
2584         conf = mddev->private;
2585         if (!conf || !conf->log)
2586                 return -ENODEV;
2587
2588         if (raid5_calc_degraded(conf) > 0 &&
2589             mode == R5C_JOURNAL_MODE_WRITE_BACK)
2590                 return -EINVAL;
2591
2592         mddev_suspend(mddev);
2593         conf->log->r5c_journal_mode = mode;
2594         mddev_resume(mddev);
2595
2596         pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2597                  mdname(mddev), mode, r5c_journal_mode_str[mode]);
2598         return 0;
2599 }
2600 EXPORT_SYMBOL(r5c_journal_mode_set);
2601
2602 static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2603                                       const char *page, size_t length)
2604 {
2605         int mode = ARRAY_SIZE(r5c_journal_mode_str);
2606         size_t len = length;
2607         int ret;
2608
2609         if (len < 2)
2610                 return -EINVAL;
2611
2612         if (page[len - 1] == '\n')
2613                 len--;
2614
2615         while (mode--)
2616                 if (strlen(r5c_journal_mode_str[mode]) == len &&
2617                     !strncmp(page, r5c_journal_mode_str[mode], len))
2618                         break;
2619         ret = mddev_lock(mddev);
2620         if (ret)
2621                 return ret;
2622         ret = r5c_journal_mode_set(mddev, mode);
2623         mddev_unlock(mddev);
2624         return ret ?: length;
2625 }
2626
2627 struct md_sysfs_entry
2628 r5c_journal_mode = __ATTR(journal_mode, 0644,
2629                           r5c_journal_mode_show, r5c_journal_mode_store);
2630
2631 /*
2632  * Try handle write operation in caching phase. This function should only
2633  * be called in write-back mode.
2634  *
2635  * If all outstanding writes can be handled in caching phase, returns 0
2636  * If writes requires write-out phase, call r5c_make_stripe_write_out()
2637  * and returns -EAGAIN
2638  */
2639 int r5c_try_caching_write(struct r5conf *conf,
2640                           struct stripe_head *sh,
2641                           struct stripe_head_state *s,
2642                           int disks)
2643 {
2644         struct r5l_log *log = conf->log;
2645         int i;
2646         struct r5dev *dev;
2647         int to_cache = 0;
2648         void **pslot;
2649         sector_t tree_index;
2650         int ret;
2651         uintptr_t refcount;
2652
2653         BUG_ON(!r5c_is_writeback(log));
2654
2655         if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2656                 /*
2657                  * There are two different scenarios here:
2658                  *  1. The stripe has some data cached, and it is sent to
2659                  *     write-out phase for reclaim
2660                  *  2. The stripe is clean, and this is the first write
2661                  *
2662                  * For 1, return -EAGAIN, so we continue with
2663                  * handle_stripe_dirtying().
2664                  *
2665                  * For 2, set STRIPE_R5C_CACHING and continue with caching
2666                  * write.
2667                  */
2668
2669                 /* case 1: anything injournal or anything in written */
2670                 if (s->injournal > 0 || s->written > 0)
2671                         return -EAGAIN;
2672                 /* case 2 */
2673                 set_bit(STRIPE_R5C_CACHING, &sh->state);
2674         }
2675
2676         /*
2677          * When run in degraded mode, array is set to write-through mode.
2678          * This check helps drain pending write safely in the transition to
2679          * write-through mode.
2680          *
2681          * When a stripe is syncing, the write is also handled in write
2682          * through mode.
2683          */
2684         if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2685                 r5c_make_stripe_write_out(sh);
2686                 return -EAGAIN;
2687         }
2688
2689         for (i = disks; i--; ) {
2690                 dev = &sh->dev[i];
2691                 /* if non-overwrite, use writing-out phase */
2692                 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2693                     !test_bit(R5_InJournal, &dev->flags)) {
2694                         r5c_make_stripe_write_out(sh);
2695                         return -EAGAIN;
2696                 }
2697         }
2698
2699         /* if the stripe is not counted in big_stripe_tree, add it now */
2700         if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2701             !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2702                 tree_index = r5c_tree_index(conf, sh->sector);
2703                 spin_lock(&log->tree_lock);
2704                 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2705                                                tree_index);
2706                 if (pslot) {
2707                         refcount = (uintptr_t)radix_tree_deref_slot_protected(
2708                                 pslot, &log->tree_lock) >>
2709                                 R5C_RADIX_COUNT_SHIFT;
2710                         radix_tree_replace_slot(
2711                                 &log->big_stripe_tree, pslot,
2712                                 (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2713                 } else {
2714                         /*
2715                          * this radix_tree_insert can fail safely, so no
2716                          * need to call radix_tree_preload()
2717                          */
2718                         ret = radix_tree_insert(
2719                                 &log->big_stripe_tree, tree_index,
2720                                 (void *)(1 << R5C_RADIX_COUNT_SHIFT));
2721                         if (ret) {
2722                                 spin_unlock(&log->tree_lock);
2723                                 r5c_make_stripe_write_out(sh);
2724                                 return -EAGAIN;
2725                         }
2726                 }
2727                 spin_unlock(&log->tree_lock);
2728
2729                 /*
2730                  * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2731                  * counted in the radix tree
2732                  */
2733                 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2734                 atomic_inc(&conf->r5c_cached_partial_stripes);
2735         }
2736
2737         for (i = disks; i--; ) {
2738                 dev = &sh->dev[i];
2739                 if (dev->towrite) {
2740                         set_bit(R5_Wantwrite, &dev->flags);
2741                         set_bit(R5_Wantdrain, &dev->flags);
2742                         set_bit(R5_LOCKED, &dev->flags);
2743                         to_cache++;
2744                 }
2745         }
2746
2747         if (to_cache) {
2748                 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2749                 /*
2750                  * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2751                  * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2752                  * r5c_handle_data_cached()
2753                  */
2754                 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2755         }
2756
2757         return 0;
2758 }
2759
2760 /*
2761  * free extra pages (orig_page) we allocated for prexor
2762  */
2763 void r5c_release_extra_page(struct stripe_head *sh)
2764 {
2765         struct r5conf *conf = sh->raid_conf;
2766         int i;
2767         bool using_disk_info_extra_page;
2768
2769         using_disk_info_extra_page =
2770                 sh->dev[0].orig_page == conf->disks[0].extra_page;
2771
2772         for (i = sh->disks; i--; )
2773                 if (sh->dev[i].page != sh->dev[i].orig_page) {
2774                         struct page *p = sh->dev[i].orig_page;
2775
2776                         sh->dev[i].orig_page = sh->dev[i].page;
2777                         clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2778
2779                         if (!using_disk_info_extra_page)
2780                                 put_page(p);
2781                 }
2782
2783         if (using_disk_info_extra_page) {
2784                 clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2785                 md_wakeup_thread(conf->mddev->thread);
2786         }
2787 }
2788
2789 void r5c_use_extra_page(struct stripe_head *sh)
2790 {
2791         struct r5conf *conf = sh->raid_conf;
2792         int i;
2793         struct r5dev *dev;
2794
2795         for (i = sh->disks; i--; ) {
2796                 dev = &sh->dev[i];
2797                 if (dev->orig_page != dev->page)
2798                         put_page(dev->orig_page);
2799                 dev->orig_page = conf->disks[i].extra_page;
2800         }
2801 }
2802
2803 /*
2804  * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2805  * stripe is committed to RAID disks.
2806  */
2807 void r5c_finish_stripe_write_out(struct r5conf *conf,
2808                                  struct stripe_head *sh,
2809                                  struct stripe_head_state *s)
2810 {
2811         struct r5l_log *log = conf->log;
2812         int i;
2813         int do_wakeup = 0;
2814         sector_t tree_index;
2815         void **pslot;
2816         uintptr_t refcount;
2817
2818         if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2819                 return;
2820
2821         WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2822         clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2823
2824         if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2825                 return;
2826
2827         for (i = sh->disks; i--; ) {
2828                 clear_bit(R5_InJournal, &sh->dev[i].flags);
2829                 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2830                         do_wakeup = 1;
2831         }
2832
2833         /*
2834          * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2835          * We updated R5_InJournal, so we also update s->injournal.
2836          */
2837         s->injournal = 0;
2838
2839         if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2840                 if (atomic_dec_and_test(&conf->pending_full_writes))
2841                         md_wakeup_thread(conf->mddev->thread);
2842
2843         if (do_wakeup)
2844                 wake_up(&conf->wait_for_overlap);
2845
2846         spin_lock_irq(&log->stripe_in_journal_lock);
2847         list_del_init(&sh->r5c);
2848         spin_unlock_irq(&log->stripe_in_journal_lock);
2849         sh->log_start = MaxSector;
2850
2851         atomic_dec(&log->stripe_in_journal_count);
2852         r5c_update_log_state(log);
2853
2854         /* stop counting this stripe in big_stripe_tree */
2855         if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2856             test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2857                 tree_index = r5c_tree_index(conf, sh->sector);
2858                 spin_lock(&log->tree_lock);
2859                 pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2860                                                tree_index);
2861                 BUG_ON(pslot == NULL);
2862                 refcount = (uintptr_t)radix_tree_deref_slot_protected(
2863                         pslot, &log->tree_lock) >>
2864                         R5C_RADIX_COUNT_SHIFT;
2865                 if (refcount == 1)
2866                         radix_tree_delete(&log->big_stripe_tree, tree_index);
2867                 else
2868                         radix_tree_replace_slot(
2869                                 &log->big_stripe_tree, pslot,
2870                                 (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2871                 spin_unlock(&log->tree_lock);
2872         }
2873
2874         if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2875                 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2876                 atomic_dec(&conf->r5c_flushing_partial_stripes);
2877                 atomic_dec(&conf->r5c_cached_partial_stripes);
2878         }
2879
2880         if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2881                 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2882                 atomic_dec(&conf->r5c_flushing_full_stripes);
2883                 atomic_dec(&conf->r5c_cached_full_stripes);
2884         }
2885
2886         r5l_append_flush_payload(log, sh->sector);
2887         /* stripe is flused to raid disks, we can do resync now */
2888         if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2889                 set_bit(STRIPE_HANDLE, &sh->state);
2890 }
2891
2892 int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2893 {
2894         struct r5conf *conf = sh->raid_conf;
2895         int pages = 0;
2896         int reserve;
2897         int i;
2898         int ret = 0;
2899
2900         BUG_ON(!log);
2901
2902         for (i = 0; i < sh->disks; i++) {
2903                 void *addr;
2904
2905                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2906                         continue;
2907                 addr = kmap_atomic(sh->dev[i].page);
2908                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2909                                                     addr, PAGE_SIZE);
2910                 kunmap_atomic(addr);
2911                 pages++;
2912         }
2913         WARN_ON(pages == 0);
2914
2915         /*
2916          * The stripe must enter state machine again to call endio, so
2917          * don't delay.
2918          */
2919         clear_bit(STRIPE_DELAYED, &sh->state);
2920         atomic_inc(&sh->count);
2921
2922         mutex_lock(&log->io_mutex);
2923         /* meta + data */
2924         reserve = (1 + pages) << (PAGE_SHIFT - 9);
2925
2926         if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2927             sh->log_start == MaxSector)
2928                 r5l_add_no_space_stripe(log, sh);
2929         else if (!r5l_has_free_space(log, reserve)) {
2930                 if (sh->log_start == log->last_checkpoint)
2931                         BUG();
2932                 else
2933                         r5l_add_no_space_stripe(log, sh);
2934         } else {
2935                 ret = r5l_log_stripe(log, sh, pages, 0);
2936                 if (ret) {
2937                         spin_lock_irq(&log->io_list_lock);
2938                         list_add_tail(&sh->log_list, &log->no_mem_stripes);
2939                         spin_unlock_irq(&log->io_list_lock);
2940                 }
2941         }
2942
2943         mutex_unlock(&log->io_mutex);
2944         return 0;
2945 }
2946
2947 /* check whether this big stripe is in write back cache. */
2948 bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2949 {
2950         struct r5l_log *log = conf->log;
2951         sector_t tree_index;
2952         void *slot;
2953
2954         if (!log)
2955                 return false;
2956
2957         WARN_ON_ONCE(!rcu_read_lock_held());
2958         tree_index = r5c_tree_index(conf, sect);
2959         slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2960         return slot != NULL;
2961 }
2962
2963 static int r5l_load_log(struct r5l_log *log)
2964 {
2965         struct md_rdev *rdev = log->rdev;
2966         struct page *page;
2967         struct r5l_meta_block *mb;
2968         sector_t cp = log->rdev->journal_tail;
2969         u32 stored_crc, expected_crc;
2970         bool create_super = false;
2971         int ret = 0;
2972
2973         /* Make sure it's valid */
2974         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2975                 cp = 0;
2976         page = alloc_page(GFP_KERNEL);
2977         if (!page)
2978                 return -ENOMEM;
2979
2980         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2981                 ret = -EIO;
2982                 goto ioerr;
2983         }
2984         mb = page_address(page);
2985
2986         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2987             mb->version != R5LOG_VERSION) {
2988                 create_super = true;
2989                 goto create;
2990         }
2991         stored_crc = le32_to_cpu(mb->checksum);
2992         mb->checksum = 0;
2993         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2994         if (stored_crc != expected_crc) {
2995                 create_super = true;
2996                 goto create;
2997         }
2998         if (le64_to_cpu(mb->position) != cp) {
2999                 create_super = true;
3000                 goto create;
3001         }
3002 create:
3003         if (create_super) {
3004                 log->last_cp_seq = prandom_u32();
3005                 cp = 0;
3006                 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
3007                 /*
3008                  * Make sure super points to correct address. Log might have
3009                  * data very soon. If super hasn't correct log tail address,
3010                  * recovery can't find the log
3011                  */
3012                 r5l_write_super(log, cp);
3013         } else
3014                 log->last_cp_seq = le64_to_cpu(mb->seq);
3015
3016         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3017         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3018         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3019                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3020         log->last_checkpoint = cp;
3021
3022         __free_page(page);
3023
3024         if (create_super) {
3025                 log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3026                 log->seq = log->last_cp_seq + 1;
3027                 log->next_checkpoint = cp;
3028         } else
3029                 ret = r5l_recovery_log(log);
3030
3031         r5c_update_log_state(log);
3032         return ret;
3033 ioerr:
3034         __free_page(page);
3035         return ret;
3036 }
3037
3038 int r5l_start(struct r5l_log *log)
3039 {
3040         int ret;
3041
3042         if (!log)
3043                 return 0;
3044
3045         ret = r5l_load_log(log);
3046         if (ret) {
3047                 struct mddev *mddev = log->rdev->mddev;
3048                 struct r5conf *conf = mddev->private;
3049
3050                 r5l_exit_log(conf);
3051         }
3052         return ret;
3053 }
3054
3055 void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3056 {
3057         struct r5conf *conf = mddev->private;
3058         struct r5l_log *log = conf->log;
3059
3060         if (!log)
3061                 return;
3062
3063         if ((raid5_calc_degraded(conf) > 0 ||
3064              test_bit(Journal, &rdev->flags)) &&
3065             conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3066                 schedule_work(&log->disable_writeback_work);
3067 }
3068
3069 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3070 {
3071         struct request_queue *q = bdev_get_queue(rdev->bdev);
3072         struct r5l_log *log;
3073         char b[BDEVNAME_SIZE];
3074
3075         pr_debug("md/raid:%s: using device %s as journal\n",
3076                  mdname(conf->mddev), bdevname(rdev->bdev, b));
3077
3078         if (PAGE_SIZE != 4096)
3079                 return -EINVAL;
3080
3081         /*
3082          * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3083          * raid_disks r5l_payload_data_parity.
3084          *
3085          * Write journal and cache does not work for very big array
3086          * (raid_disks > 203)
3087          */
3088         if (sizeof(struct r5l_meta_block) +
3089             ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3090              conf->raid_disks) > PAGE_SIZE) {
3091                 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3092                        mdname(conf->mddev), conf->raid_disks);
3093                 return -EINVAL;
3094         }
3095
3096         log = kzalloc(sizeof(*log), GFP_KERNEL);
3097         if (!log)
3098                 return -ENOMEM;
3099         log->rdev = rdev;
3100
3101         log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
3102
3103         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3104                                        sizeof(rdev->mddev->uuid));
3105
3106         mutex_init(&log->io_mutex);
3107
3108         spin_lock_init(&log->io_list_lock);
3109         INIT_LIST_HEAD(&log->running_ios);
3110         INIT_LIST_HEAD(&log->io_end_ios);
3111         INIT_LIST_HEAD(&log->flushing_ios);
3112         INIT_LIST_HEAD(&log->finished_ios);
3113         bio_init(&log->flush_bio, NULL, 0);
3114
3115         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3116         if (!log->io_kc)
3117                 goto io_kc;
3118
3119         log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
3120         if (!log->io_pool)
3121                 goto io_pool;
3122
3123         log->bs = bioset_create(R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3124         if (!log->bs)
3125                 goto io_bs;
3126
3127         log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
3128         if (!log->meta_pool)
3129                 goto out_mempool;
3130
3131         spin_lock_init(&log->tree_lock);
3132         INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3133
3134         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3135                                                  log->rdev->mddev, "reclaim");
3136         if (!log->reclaim_thread)
3137                 goto reclaim_thread;
3138         log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3139
3140         init_waitqueue_head(&log->iounit_wait);
3141
3142         INIT_LIST_HEAD(&log->no_mem_stripes);
3143
3144         INIT_LIST_HEAD(&log->no_space_stripes);
3145         spin_lock_init(&log->no_space_stripes_lock);
3146
3147         INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3148         INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3149
3150         log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3151         INIT_LIST_HEAD(&log->stripe_in_journal_list);
3152         spin_lock_init(&log->stripe_in_journal_lock);
3153         atomic_set(&log->stripe_in_journal_count, 0);
3154
3155         rcu_assign_pointer(conf->log, log);
3156
3157         set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3158         return 0;
3159
3160         rcu_assign_pointer(conf->log, NULL);
3161         md_unregister_thread(&log->reclaim_thread);
3162 reclaim_thread:
3163         mempool_destroy(log->meta_pool);
3164 out_mempool:
3165         bioset_free(log->bs);
3166 io_bs:
3167         mempool_destroy(log->io_pool);
3168 io_pool:
3169         kmem_cache_destroy(log->io_kc);
3170 io_kc:
3171         kfree(log);
3172         return -EINVAL;
3173 }
3174
3175 void r5l_exit_log(struct r5conf *conf)
3176 {
3177         struct r5l_log *log = conf->log;
3178
3179         conf->log = NULL;
3180         synchronize_rcu();
3181
3182         /* Ensure disable_writeback_work wakes up and exits */
3183         wake_up(&conf->mddev->sb_wait);
3184         flush_work(&log->disable_writeback_work);
3185         md_unregister_thread(&log->reclaim_thread);
3186         mempool_destroy(log->meta_pool);
3187         bioset_free(log->bs);
3188         mempool_destroy(log->io_pool);
3189         kmem_cache_destroy(log->io_kc);
3190         kfree(log);
3191 }