raid5-cache.c (3bddb7f8f264ec58dc86e11ca97341c24f9d38f6) raid5-cache.c (3a83f4677539bce8eaa2bca9ee9c20e172d7ab04)
1/*
2 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
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

--- 4 unchanged lines hidden (view full) ---

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 "md.h"
23#include "raid5.h"
3 *
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for

--- 4 unchanged lines hidden (view full) ---

15#include <linux/wait.h>
16#include <linux/blkdev.h>
17#include <linux/slab.h>
18#include <linux/raid/md_p.h>
19#include <linux/crc32c.h>
20#include <linux/random.h>
21#include "md.h"
22#include "raid5.h"
24#include "bitmap.h"
25
26/*
27 * metadata/data stored in disk with 4k size unit (a block) regardless
28 * underneath hardware sector size. only works with PAGE_SIZE == 4096
29 */
30#define BLOCK_SECTORS (8)
31
32/*
23
24/*
25 * metadata/data stored in disk with 4k size unit (a block) regardless
26 * underneath hardware sector size. only works with PAGE_SIZE == 4096
27 */
28#define BLOCK_SECTORS (8)
29
30/*
33 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
34 *
35 * In write through mode, the reclaim runs every log->max_free_space.
36 * This can prevent the recovery scans for too long
31 * reclaim runs every 1/4 disk size or 10G reclaimable space. This can prevent
32 * recovery scans a very long log
37 */
38#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
39#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
40
33 */
34#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
35#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
36
41/* wake up reclaim thread periodically */
42#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
43/* start flush with these full stripes */
44#define R5C_FULL_STRIPE_FLUSH_BATCH 256
45/* reclaim stripes in groups */
46#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
47
48/*
49 * We only need 2 bios per I/O unit to make progress, but ensure we
50 * have a few more available to not get too tight.
51 */
52#define R5L_POOL_SIZE 4
53
37/*
38 * We only need 2 bios per I/O unit to make progress, but ensure we
39 * have a few more available to not get too tight.
40 */
41#define R5L_POOL_SIZE 4
42
54/*
55 * r5c journal modes of the array: write-back or write-through.
56 * write-through mode has identical behavior as existing log only
57 * implementation.
58 */
59enum r5c_journal_mode {
60 R5C_JOURNAL_MODE_WRITE_THROUGH = 0,
61 R5C_JOURNAL_MODE_WRITE_BACK = 1,
62};
63
64static char *r5c_journal_mode_str[] = {"write-through",
65 "write-back"};
66/*
67 * raid5 cache state machine
68 *
69 * With rhe RAID cache, each stripe works in two phases:
70 * - caching phase
71 * - writing-out phase
72 *
73 * These two phases are controlled by bit STRIPE_R5C_CACHING:
74 * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
75 * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
76 *
77 * When there is no journal, or the journal is in write-through mode,
78 * the stripe is always in writing-out phase.
79 *
80 * For write-back journal, the stripe is sent to caching phase on write
81 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
82 * the write-out phase by clearing STRIPE_R5C_CACHING.
83 *
84 * Stripes in caching phase do not write the raid disks. Instead, all
85 * writes are committed from the log device. Therefore, a stripe in
86 * caching phase handles writes as:
87 * - write to log device
88 * - return IO
89 *
90 * Stripes in writing-out phase handle writes as:
91 * - calculate parity
92 * - write pending data and parity to journal
93 * - write data and parity to raid disks
94 * - return IO for pending writes
95 */
96
97struct r5l_log {
98 struct md_rdev *rdev;
99
100 u32 uuid_checksum;
101
102 sector_t device_size; /* log device size, round to
103 * BLOCK_SECTORS */
104 sector_t max_free_space; /* reclaim run if free space is at

--- 40 unchanged lines hidden (view full) ---

145 * switching to IO_UNIT_STRIPE_END
146 * state) */
147 wait_queue_head_t iounit_wait;
148
149 struct list_head no_space_stripes; /* pending stripes, log has no space */
150 spinlock_t no_space_stripes_lock;
151
152 bool need_cache_flush;
43struct r5l_log {
44 struct md_rdev *rdev;
45
46 u32 uuid_checksum;
47
48 sector_t device_size; /* log device size, round to
49 * BLOCK_SECTORS */
50 sector_t max_free_space; /* reclaim run if free space is at

--- 40 unchanged lines hidden (view full) ---

91 * switching to IO_UNIT_STRIPE_END
92 * state) */
93 wait_queue_head_t iounit_wait;
94
95 struct list_head no_space_stripes; /* pending stripes, log has no space */
96 spinlock_t no_space_stripes_lock;
97
98 bool need_cache_flush;
153
154 /* for r5c_cache */
155 enum r5c_journal_mode r5c_journal_mode;
156
157 /* all stripes in r5cache, in the order of seq at sh->log_start */
158 struct list_head stripe_in_journal_list;
159
160 spinlock_t stripe_in_journal_lock;
161 atomic_t stripe_in_journal_count;
162
163 /* to submit async io_units, to fulfill ordering of flush */
164 struct work_struct deferred_io_work;
165};
166
167/*
168 * an IO range starts from a meta data block and end at the next meta data
169 * block. The io unit's the meta data block tracks data/parity followed it. io
170 * unit is written to log disk with normal write, as we always flush log disk
171 * first and then start move data to raid disks, there is no requirement to
172 * write io unit with FLUSH/FUA

--- 10 unchanged lines hidden (view full) ---

183 u64 seq; /* seq number of the metablock */
184 sector_t log_start; /* where the io_unit starts */
185 sector_t log_end; /* where the io_unit ends */
186 struct list_head log_sibling; /* log->running_ios */
187 struct list_head stripe_list; /* stripes added to the io_unit */
188
189 int state;
190 bool need_split_bio;
99};
100
101/*
102 * an IO range starts from a meta data block and end at the next meta data
103 * block. The io unit's the meta data block tracks data/parity followed it. io
104 * unit is written to log disk with normal write, as we always flush log disk
105 * first and then start move data to raid disks, there is no requirement to
106 * write io unit with FLUSH/FUA

--- 10 unchanged lines hidden (view full) ---

117 u64 seq; /* seq number of the metablock */
118 sector_t log_start; /* where the io_unit starts */
119 sector_t log_end; /* where the io_unit ends */
120 struct list_head log_sibling; /* log->running_ios */
121 struct list_head stripe_list; /* stripes added to the io_unit */
122
123 int state;
124 bool need_split_bio;
191 struct bio *split_bio;
192
193 unsigned int has_flush:1; /* include flush request */
194 unsigned int has_fua:1; /* include fua request */
195 unsigned int has_null_flush:1; /* include empty flush request */
196 /*
197 * io isn't sent yet, flush/fua request can only be submitted till it's
198 * the first IO in running_ios list
199 */
200 unsigned int io_deferred:1;
201
202 struct bio_list flush_barriers; /* size == 0 flush bios */
203};
204
205/* r5l_io_unit state */
206enum r5l_io_unit_state {
207 IO_UNIT_RUNNING = 0, /* accepting new IO */
208 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
209 * don't accepting new bio */
210 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
211 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
212};
213
125};
126
127/* r5l_io_unit state */
128enum r5l_io_unit_state {
129 IO_UNIT_RUNNING = 0, /* accepting new IO */
130 IO_UNIT_IO_START = 1, /* io_unit bio start writing to log,
131 * don't accepting new bio */
132 IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */
133 IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
134};
135
214bool r5c_is_writeback(struct r5l_log *log)
215{
216 return (log != NULL &&
217 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
218}
219
220static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
221{
222 start += inc;
223 if (start >= log->device_size)
224 start = start - log->device_size;
225 return start;
226}
227

--- 19 unchanged lines hidden (view full) ---

247static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
248 enum r5l_io_unit_state state)
249{
250 if (WARN_ON(io->state >= state))
251 return;
252 io->state = state;
253}
254
136static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
137{
138 start += inc;
139 if (start >= log->device_size)
140 start = start - log->device_size;
141 return start;
142}
143

--- 19 unchanged lines hidden (view full) ---

163static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
164 enum r5l_io_unit_state state)
165{
166 if (WARN_ON(io->state >= state))
167 return;
168 io->state = state;
169}
170
255static void
256r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev,
257 struct bio_list *return_bi)
258{
259 struct bio *wbi, *wbi2;
260
261 wbi = dev->written;
262 dev->written = NULL;
263 while (wbi && wbi->bi_iter.bi_sector <
264 dev->sector + STRIPE_SECTORS) {
265 wbi2 = r5_next_bio(wbi, dev->sector);
266 if (!raid5_dec_bi_active_stripes(wbi)) {
267 md_write_end(conf->mddev);
268 bio_list_add(return_bi, wbi);
269 }
270 wbi = wbi2;
271 }
272}
273
274void r5c_handle_cached_data_endio(struct r5conf *conf,
275 struct stripe_head *sh, int disks, struct bio_list *return_bi)
276{
277 int i;
278
279 for (i = sh->disks; i--; ) {
280 if (sh->dev[i].written) {
281 set_bit(R5_UPTODATE, &sh->dev[i].flags);
282 r5c_return_dev_pending_writes(conf, &sh->dev[i],
283 return_bi);
284 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
285 STRIPE_SECTORS,
286 !test_bit(STRIPE_DEGRADED, &sh->state),
287 0);
288 }
289 }
290}
291
292/* Check whether we should flush some stripes to free up stripe cache */
293void r5c_check_stripe_cache_usage(struct r5conf *conf)
294{
295 int total_cached;
296
297 if (!r5c_is_writeback(conf->log))
298 return;
299
300 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
301 atomic_read(&conf->r5c_cached_full_stripes);
302
303 /*
304 * The following condition is true for either of the following:
305 * - stripe cache pressure high:
306 * total_cached > 3/4 min_nr_stripes ||
307 * empty_inactive_list_nr > 0
308 * - stripe cache pressure moderate:
309 * total_cached > 1/2 min_nr_stripes
310 */
311 if (total_cached > conf->min_nr_stripes * 1 / 2 ||
312 atomic_read(&conf->empty_inactive_list_nr) > 0)
313 r5l_wake_reclaim(conf->log, 0);
314}
315
316/*
317 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
318 * stripes in the cache
319 */
320void r5c_check_cached_full_stripe(struct r5conf *conf)
321{
322 if (!r5c_is_writeback(conf->log))
323 return;
324
325 /*
326 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
327 * or a full stripe (chunk size / 4k stripes).
328 */
329 if (atomic_read(&conf->r5c_cached_full_stripes) >=
330 min(R5C_FULL_STRIPE_FLUSH_BATCH,
331 conf->chunk_sectors >> STRIPE_SHIFT))
332 r5l_wake_reclaim(conf->log, 0);
333}
334
335/*
336 * Total log space (in sectors) needed to flush all data in cache
337 *
338 * Currently, writing-out phase automatically includes all pending writes
339 * to the same sector. So the reclaim of each stripe takes up to
340 * (conf->raid_disks + 1) pages of log space.
341 *
342 * To totally avoid deadlock due to log space, the code reserves
343 * (conf->raid_disks + 1) pages for each stripe in cache, which is not
344 * necessary in most cases.
345 *
346 * To improve this, we will need writing-out phase to be able to NOT include
347 * pending writes, which will reduce the requirement to
348 * (conf->max_degraded + 1) pages per stripe in cache.
349 */
350static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
351{
352 struct r5l_log *log = conf->log;
353
354 if (!r5c_is_writeback(log))
355 return 0;
356
357 return BLOCK_SECTORS * (conf->raid_disks + 1) *
358 atomic_read(&log->stripe_in_journal_count);
359}
360
361/*
362 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
363 *
364 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
365 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
366 * device is less than 2x of reclaim_required_space.
367 */
368static inline void r5c_update_log_state(struct r5l_log *log)
369{
370 struct r5conf *conf = log->rdev->mddev->private;
371 sector_t free_space;
372 sector_t reclaim_space;
373
374 if (!r5c_is_writeback(log))
375 return;
376
377 free_space = r5l_ring_distance(log, log->log_start,
378 log->last_checkpoint);
379 reclaim_space = r5c_log_required_to_flush_cache(conf);
380 if (free_space < 2 * reclaim_space)
381 set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
382 else
383 clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
384 if (free_space < 3 * reclaim_space)
385 set_bit(R5C_LOG_TIGHT, &conf->cache_state);
386 else
387 clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
388}
389
390/*
391 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
392 * This function should only be called in write-back mode.
393 */
394void r5c_make_stripe_write_out(struct stripe_head *sh)
395{
396 struct r5conf *conf = sh->raid_conf;
397 struct r5l_log *log = conf->log;
398
399 BUG_ON(!r5c_is_writeback(log));
400
401 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
402 clear_bit(STRIPE_R5C_CACHING, &sh->state);
403
404 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
405 atomic_inc(&conf->preread_active_stripes);
406
407 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
408 BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
409 atomic_dec(&conf->r5c_cached_partial_stripes);
410 }
411
412 if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
413 BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
414 atomic_dec(&conf->r5c_cached_full_stripes);
415 }
416}
417
418static void r5c_handle_data_cached(struct stripe_head *sh)
419{
420 int i;
421
422 for (i = sh->disks; i--; )
423 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
424 set_bit(R5_InJournal, &sh->dev[i].flags);
425 clear_bit(R5_LOCKED, &sh->dev[i].flags);
426 }
427 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
428}
429
430/*
431 * this journal write must contain full parity,
432 * it may also contain some data pages
433 */
434static void r5c_handle_parity_cached(struct stripe_head *sh)
435{
436 int i;
437
438 for (i = sh->disks; i--; )
439 if (test_bit(R5_InJournal, &sh->dev[i].flags))
440 set_bit(R5_Wantwrite, &sh->dev[i].flags);
441}
442
443/*
444 * Setting proper flags after writing (or flushing) data and/or parity to the
445 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
446 */
447static void r5c_finish_cache_stripe(struct stripe_head *sh)
448{
449 struct r5l_log *log = sh->raid_conf->log;
450
451 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
452 BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
453 /*
454 * Set R5_InJournal for parity dev[pd_idx]. This means
455 * all data AND parity in the journal. For RAID 6, it is
456 * NOT necessary to set the flag for dev[qd_idx], as the
457 * two parities are written out together.
458 */
459 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
460 } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
461 r5c_handle_data_cached(sh);
462 } else {
463 r5c_handle_parity_cached(sh);
464 set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
465 }
466}
467
468static void r5l_io_run_stripes(struct r5l_io_unit *io)
469{
470 struct stripe_head *sh, *next;
471
472 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
473 list_del_init(&sh->log_list);
171static void r5l_io_run_stripes(struct r5l_io_unit *io)
172{
173 struct stripe_head *sh, *next;
174
175 list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
176 list_del_init(&sh->log_list);
474
475 r5c_finish_cache_stripe(sh);
476
477 set_bit(STRIPE_HANDLE, &sh->state);
478 raid5_release_stripe(sh);
479 }
480}
481
482static void r5l_log_run_stripes(struct r5l_log *log)
483{
484 struct r5l_io_unit *io, *next;

--- 19 unchanged lines hidden (view full) ---

504 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
505 /* don't change list order */
506 if (io->state < IO_UNIT_IO_END)
507 break;
508 list_move_tail(&io->log_sibling, &log->io_end_ios);
509 }
510}
511
177 set_bit(STRIPE_HANDLE, &sh->state);
178 raid5_release_stripe(sh);
179 }
180}
181
182static void r5l_log_run_stripes(struct r5l_log *log)
183{
184 struct r5l_io_unit *io, *next;

--- 19 unchanged lines hidden (view full) ---

204 list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
205 /* don't change list order */
206 if (io->state < IO_UNIT_IO_END)
207 break;
208 list_move_tail(&io->log_sibling, &log->io_end_ios);
209 }
210}
211
512static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
513static void r5l_log_endio(struct bio *bio)
514{
515 struct r5l_io_unit *io = bio->bi_private;
212static void r5l_log_endio(struct bio *bio)
213{
214 struct r5l_io_unit *io = bio->bi_private;
516 struct r5l_io_unit *io_deferred;
517 struct r5l_log *log = io->log;
518 unsigned long flags;
519
520 if (bio->bi_error)
521 md_error(log->rdev->mddev, log->rdev);
522
523 bio_put(bio);
524 mempool_free(io->meta_page, log->meta_pool);
525
526 spin_lock_irqsave(&log->io_list_lock, flags);
527 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
528 if (log->need_cache_flush)
529 r5l_move_to_end_ios(log);
530 else
531 r5l_log_run_stripes(log);
215 struct r5l_log *log = io->log;
216 unsigned long flags;
217
218 if (bio->bi_error)
219 md_error(log->rdev->mddev, log->rdev);
220
221 bio_put(bio);
222 mempool_free(io->meta_page, log->meta_pool);
223
224 spin_lock_irqsave(&log->io_list_lock, flags);
225 __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
226 if (log->need_cache_flush)
227 r5l_move_to_end_ios(log);
228 else
229 r5l_log_run_stripes(log);
532 if (!list_empty(&log->running_ios)) {
533 /*
534 * FLUSH/FUA io_unit is deferred because of ordering, now we
535 * can dispatch it
536 */
537 io_deferred = list_first_entry(&log->running_ios,
538 struct r5l_io_unit, log_sibling);
539 if (io_deferred->io_deferred)
540 schedule_work(&log->deferred_io_work);
541 }
542
543 spin_unlock_irqrestore(&log->io_list_lock, flags);
544
545 if (log->need_cache_flush)
546 md_wakeup_thread(log->rdev->mddev->thread);
230 spin_unlock_irqrestore(&log->io_list_lock, flags);
231
232 if (log->need_cache_flush)
233 md_wakeup_thread(log->rdev->mddev->thread);
547
548 if (io->has_null_flush) {
549 struct bio *bi;
550
551 WARN_ON(bio_list_empty(&io->flush_barriers));
552 while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
553 bio_endio(bi);
554 atomic_dec(&io->pending_stripe);
555 }
556 if (atomic_read(&io->pending_stripe) == 0)
557 __r5l_stripe_write_finished(io);
558 }
559}
560
234}
235
561static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
562{
563 unsigned long flags;
564
565 spin_lock_irqsave(&log->io_list_lock, flags);
566 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
567 spin_unlock_irqrestore(&log->io_list_lock, flags);
568
569 if (io->has_flush)
570 bio_set_op_attrs(io->current_bio, REQ_OP_WRITE, WRITE_FLUSH);
571 if (io->has_fua)
572 bio_set_op_attrs(io->current_bio, REQ_OP_WRITE, WRITE_FUA);
573 submit_bio(io->current_bio);
574
575 if (!io->split_bio)
576 return;
577
578 if (io->has_flush)
579 bio_set_op_attrs(io->split_bio, REQ_OP_WRITE, WRITE_FLUSH);
580 if (io->has_fua)
581 bio_set_op_attrs(io->split_bio, REQ_OP_WRITE, WRITE_FUA);
582 submit_bio(io->split_bio);
583}
584
585/* deferred io_unit will be dispatched here */
586static void r5l_submit_io_async(struct work_struct *work)
587{
588 struct r5l_log *log = container_of(work, struct r5l_log,
589 deferred_io_work);
590 struct r5l_io_unit *io = NULL;
591 unsigned long flags;
592
593 spin_lock_irqsave(&log->io_list_lock, flags);
594 if (!list_empty(&log->running_ios)) {
595 io = list_first_entry(&log->running_ios, struct r5l_io_unit,
596 log_sibling);
597 if (!io->io_deferred)
598 io = NULL;
599 else
600 io->io_deferred = 0;
601 }
602 spin_unlock_irqrestore(&log->io_list_lock, flags);
603 if (io)
604 r5l_do_submit_io(log, io);
605}
606
607static void r5l_submit_current_io(struct r5l_log *log)
608{
609 struct r5l_io_unit *io = log->current_io;
236static void r5l_submit_current_io(struct r5l_log *log)
237{
238 struct r5l_io_unit *io = log->current_io;
610 struct bio *bio;
611 struct r5l_meta_block *block;
612 unsigned long flags;
613 u32 crc;
239 struct r5l_meta_block *block;
240 unsigned long flags;
241 u32 crc;
614 bool do_submit = true;
615
616 if (!io)
617 return;
618
619 block = page_address(io->meta_page);
620 block->meta_size = cpu_to_le32(io->meta_offset);
621 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
622 block->checksum = cpu_to_le32(crc);
242
243 if (!io)
244 return;
245
246 block = page_address(io->meta_page);
247 block->meta_size = cpu_to_le32(io->meta_offset);
248 crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
249 block->checksum = cpu_to_le32(crc);
623 bio = io->current_bio;
624
625 log->current_io = NULL;
626 spin_lock_irqsave(&log->io_list_lock, flags);
250
251 log->current_io = NULL;
252 spin_lock_irqsave(&log->io_list_lock, flags);
627 if (io->has_flush || io->has_fua) {
628 if (io != list_first_entry(&log->running_ios,
629 struct r5l_io_unit, log_sibling)) {
630 io->io_deferred = 1;
631 do_submit = false;
632 }
633 }
253 __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
634 spin_unlock_irqrestore(&log->io_list_lock, flags);
254 spin_unlock_irqrestore(&log->io_list_lock, flags);
635 if (do_submit)
636 r5l_do_submit_io(log, io);
255
256 submit_bio(io->current_bio);
637}
638
639static struct bio *r5l_bio_alloc(struct r5l_log *log)
640{
641 struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
642
643 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
644 bio->bi_bdev = log->rdev->bdev;
645 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
646
647 return bio;
648}
649
650static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
651{
652 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
653
257}
258
259static struct bio *r5l_bio_alloc(struct r5l_log *log)
260{
261 struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
262
263 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
264 bio->bi_bdev = log->rdev->bdev;
265 bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
266
267 return bio;
268}
269
270static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
271{
272 log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
273
654 r5c_update_log_state(log);
655 /*
656 * If we filled up the log device start from the beginning again,
657 * which will require a new bio.
658 *
659 * Note: for this to work properly the log size needs to me a multiple
660 * of BLOCK_SECTORS.
661 */
662 if (log->log_start == 0)

--- 10 unchanged lines hidden (view full) ---

673 io = mempool_alloc(log->io_pool, GFP_ATOMIC);
674 if (!io)
675 return NULL;
676 memset(io, 0, sizeof(*io));
677
678 io->log = log;
679 INIT_LIST_HEAD(&io->log_sibling);
680 INIT_LIST_HEAD(&io->stripe_list);
274 /*
275 * If we filled up the log device start from the beginning again,
276 * which will require a new bio.
277 *
278 * Note: for this to work properly the log size needs to me a multiple
279 * of BLOCK_SECTORS.
280 */
281 if (log->log_start == 0)

--- 10 unchanged lines hidden (view full) ---

292 io = mempool_alloc(log->io_pool, GFP_ATOMIC);
293 if (!io)
294 return NULL;
295 memset(io, 0, sizeof(*io));
296
297 io->log = log;
298 INIT_LIST_HEAD(&io->log_sibling);
299 INIT_LIST_HEAD(&io->stripe_list);
681 bio_list_init(&io->flush_barriers);
682 io->state = IO_UNIT_RUNNING;
683
684 io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
685 block = page_address(io->meta_page);
686 clear_page(block);
687 block->magic = cpu_to_le32(R5LOG_MAGIC);
688 block->version = R5LOG_VERSION;
689 block->seq = cpu_to_le64(log->seq);

--- 54 unchanged lines hidden (view full) ---

744 sizeof(__le32) * (1 + !!checksum2_valid);
745}
746
747static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
748{
749 struct r5l_io_unit *io = log->current_io;
750
751 if (io->need_split_bio) {
300 io->state = IO_UNIT_RUNNING;
301
302 io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
303 block = page_address(io->meta_page);
304 clear_page(block);
305 block->magic = cpu_to_le32(R5LOG_MAGIC);
306 block->version = R5LOG_VERSION;
307 block->seq = cpu_to_le64(log->seq);

--- 54 unchanged lines hidden (view full) ---

362 sizeof(__le32) * (1 + !!checksum2_valid);
363}
364
365static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
366{
367 struct r5l_io_unit *io = log->current_io;
368
369 if (io->need_split_bio) {
752 BUG_ON(io->split_bio);
753 io->split_bio = io->current_bio;
370 struct bio *prev = io->current_bio;
371
754 io->current_bio = r5l_bio_alloc(log);
372 io->current_bio = r5l_bio_alloc(log);
755 bio_chain(io->current_bio, io->split_bio);
756 io->need_split_bio = false;
373 bio_chain(io->current_bio, prev);
374
375 submit_bio(prev);
757 }
758
759 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
760 BUG();
761
762 r5_reserve_log_entry(log, io);
763}
764

--- 12 unchanged lines hidden (view full) ---

777 sizeof(__le32) * parity_pages;
778
779 ret = r5l_get_meta(log, meta_size);
780 if (ret)
781 return ret;
782
783 io = log->current_io;
784
376 }
377
378 if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
379 BUG();
380
381 r5_reserve_log_entry(log, io);
382}
383

--- 12 unchanged lines hidden (view full) ---

396 sizeof(__le32) * parity_pages;
397
398 ret = r5l_get_meta(log, meta_size);
399 if (ret)
400 return ret;
401
402 io = log->current_io;
403
785 if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
786 io->has_flush = 1;
787
788 for (i = 0; i < sh->disks; i++) {
404 for (i = 0; i < sh->disks; i++) {
789 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
790 test_bit(R5_InJournal, &sh->dev[i].flags))
405 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
791 continue;
792 if (i == sh->pd_idx || i == sh->qd_idx)
793 continue;
406 continue;
407 if (i == sh->pd_idx || i == sh->qd_idx)
408 continue;
794 if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
795 log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
796 io->has_fua = 1;
797 /*
798 * we need to flush journal to make sure recovery can
799 * reach the data with fua flag
800 */
801 io->has_flush = 1;
802 }
803 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
804 raid5_compute_blocknr(sh, i, 0),
805 sh->dev[i].log_checksum, 0, false);
806 r5l_append_payload_page(log, sh->dev[i].page);
807 }
808
409 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
410 raid5_compute_blocknr(sh, i, 0),
411 sh->dev[i].log_checksum, 0, false);
412 r5l_append_payload_page(log, sh->dev[i].page);
413 }
414
809 if (parity_pages == 2) {
415 if (sh->qd_idx >= 0) {
810 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
811 sh->sector, sh->dev[sh->pd_idx].log_checksum,
812 sh->dev[sh->qd_idx].log_checksum, true);
813 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
814 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
416 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
417 sh->sector, sh->dev[sh->pd_idx].log_checksum,
418 sh->dev[sh->qd_idx].log_checksum, true);
419 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
420 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
815 } else if (parity_pages == 1) {
421 } else {
816 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
817 sh->sector, sh->dev[sh->pd_idx].log_checksum,
818 0, false);
819 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
422 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
423 sh->sector, sh->dev[sh->pd_idx].log_checksum,
424 0, false);
425 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
820 } else /* Just writing data, not parity, in caching phase */
821 BUG_ON(parity_pages != 0);
426 }
822
823 list_add_tail(&sh->log_list, &io->stripe_list);
824 atomic_inc(&io->pending_stripe);
825 sh->log_io = io;
826
427
428 list_add_tail(&sh->log_list, &io->stripe_list);
429 atomic_inc(&io->pending_stripe);
430 sh->log_io = io;
431
827 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
828 return 0;
829
830 if (sh->log_start == MaxSector) {
831 BUG_ON(!list_empty(&sh->r5c));
832 sh->log_start = io->log_start;
833 spin_lock_irq(&log->stripe_in_journal_lock);
834 list_add_tail(&sh->r5c,
835 &log->stripe_in_journal_list);
836 spin_unlock_irq(&log->stripe_in_journal_lock);
837 atomic_inc(&log->stripe_in_journal_count);
838 }
839 return 0;
840}
841
432 return 0;
433}
434
842/* add stripe to no_space_stripes, and then wake up reclaim */
843static inline void r5l_add_no_space_stripe(struct r5l_log *log,
844 struct stripe_head *sh)
845{
846 spin_lock(&log->no_space_stripes_lock);
847 list_add_tail(&sh->log_list, &log->no_space_stripes);
848 spin_unlock(&log->no_space_stripes_lock);
849}
850
435static void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
851/*
852 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
853 * data from log to raid disks), so we shouldn't wait for reclaim here
854 */
855int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
856{
436/*
437 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
438 * data from log to raid disks), so we shouldn't wait for reclaim here
439 */
440int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
441{
857 struct r5conf *conf = sh->raid_conf;
858 int write_disks = 0;
859 int data_pages, parity_pages;
442 int write_disks = 0;
443 int data_pages, parity_pages;
444 int meta_size;
860 int reserve;
861 int i;
862 int ret = 0;
445 int reserve;
446 int i;
447 int ret = 0;
863 bool wake_reclaim = false;
864
865 if (!log)
866 return -EAGAIN;
867 /* Don't support stripe batch */
868 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
869 test_bit(STRIPE_SYNCING, &sh->state)) {
870 /* the stripe is written to log, we start writing it to raid */
871 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
872 return -EAGAIN;
873 }
874
448
449 if (!log)
450 return -EAGAIN;
451 /* Don't support stripe batch */
452 if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
453 test_bit(STRIPE_SYNCING, &sh->state)) {
454 /* the stripe is written to log, we start writing it to raid */
455 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
456 return -EAGAIN;
457 }
458
875 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
876
877 for (i = 0; i < sh->disks; i++) {
878 void *addr;
879
459 for (i = 0; i < sh->disks; i++) {
460 void *addr;
461
880 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
881 test_bit(R5_InJournal, &sh->dev[i].flags))
462 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
882 continue;
463 continue;
883
884 write_disks++;
885 /* checksum is already calculated in last run */
886 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
887 continue;
888 addr = kmap_atomic(sh->dev[i].page);
889 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
890 addr, PAGE_SIZE);
891 kunmap_atomic(addr);
892 }
893 parity_pages = 1 + !!(sh->qd_idx >= 0);
894 data_pages = write_disks - parity_pages;
895
464 write_disks++;
465 /* checksum is already calculated in last run */
466 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
467 continue;
468 addr = kmap_atomic(sh->dev[i].page);
469 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
470 addr, PAGE_SIZE);
471 kunmap_atomic(addr);
472 }
473 parity_pages = 1 + !!(sh->qd_idx >= 0);
474 data_pages = write_disks - parity_pages;
475
476 meta_size =
477 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
478 * data_pages) +
479 sizeof(struct r5l_payload_data_parity) +
480 sizeof(__le32) * parity_pages;
481 /* Doesn't work with very big raid array */
482 if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE)
483 return -EINVAL;
484
896 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
897 /*
898 * The stripe must enter state machine again to finish the write, so
899 * don't delay.
900 */
901 clear_bit(STRIPE_DELAYED, &sh->state);
902 atomic_inc(&sh->count);
903
904 mutex_lock(&log->io_mutex);
905 /* meta + data */
906 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
485 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
486 /*
487 * The stripe must enter state machine again to finish the write, so
488 * don't delay.
489 */
490 clear_bit(STRIPE_DELAYED, &sh->state);
491 atomic_inc(&sh->count);
492
493 mutex_lock(&log->io_mutex);
494 /* meta + data */
495 reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
496 if (!r5l_has_free_space(log, reserve)) {
497 spin_lock(&log->no_space_stripes_lock);
498 list_add_tail(&sh->log_list, &log->no_space_stripes);
499 spin_unlock(&log->no_space_stripes_lock);
907
500
908 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
909 if (!r5l_has_free_space(log, reserve)) {
910 r5l_add_no_space_stripe(log, sh);
911 wake_reclaim = true;
912 } else {
913 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
914 if (ret) {
915 spin_lock_irq(&log->io_list_lock);
916 list_add_tail(&sh->log_list,
917 &log->no_mem_stripes);
918 spin_unlock_irq(&log->io_list_lock);
919 }
501 r5l_wake_reclaim(log, reserve);
502 } else {
503 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
504 if (ret) {
505 spin_lock_irq(&log->io_list_lock);
506 list_add_tail(&sh->log_list, &log->no_mem_stripes);
507 spin_unlock_irq(&log->io_list_lock);
920 }
508 }
921 } else { /* R5C_JOURNAL_MODE_WRITE_BACK */
922 /*
923 * log space critical, do not process stripes that are
924 * not in cache yet (sh->log_start == MaxSector).
925 */
926 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
927 sh->log_start == MaxSector) {
928 r5l_add_no_space_stripe(log, sh);
929 wake_reclaim = true;
930 reserve = 0;
931 } else if (!r5l_has_free_space(log, reserve)) {
932 if (sh->log_start == log->last_checkpoint)
933 BUG();
934 else
935 r5l_add_no_space_stripe(log, sh);
936 } else {
937 ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
938 if (ret) {
939 spin_lock_irq(&log->io_list_lock);
940 list_add_tail(&sh->log_list,
941 &log->no_mem_stripes);
942 spin_unlock_irq(&log->io_list_lock);
943 }
944 }
945 }
946
947 mutex_unlock(&log->io_mutex);
509 }
510
511 mutex_unlock(&log->io_mutex);
948 if (wake_reclaim)
949 r5l_wake_reclaim(log, reserve);
950 return 0;
951}
952
953void r5l_write_stripe_run(struct r5l_log *log)
954{
955 if (!log)
956 return;
957 mutex_lock(&log->io_mutex);
958 r5l_submit_current_io(log);
959 mutex_unlock(&log->io_mutex);
960}
961
962int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
963{
964 if (!log)
965 return -ENODEV;
512 return 0;
513}
514
515void r5l_write_stripe_run(struct r5l_log *log)
516{
517 if (!log)
518 return;
519 mutex_lock(&log->io_mutex);
520 r5l_submit_current_io(log);
521 mutex_unlock(&log->io_mutex);
522}
523
524int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
525{
526 if (!log)
527 return -ENODEV;
966
967 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
968 /*
969 * in write through (journal only)
970 * we flush log disk cache first, then write stripe data to
971 * raid disks. So if bio is finished, the log disk cache is
972 * flushed already. The recovery guarantees we can recovery
973 * the bio from log disk, so we don't need to flush again
974 */
975 if (bio->bi_iter.bi_size == 0) {
976 bio_endio(bio);
977 return 0;
978 }
979 bio->bi_opf &= ~REQ_PREFLUSH;
980 } else {
981 /* write back (with cache) */
982 if (bio->bi_iter.bi_size == 0) {
983 mutex_lock(&log->io_mutex);
984 r5l_get_meta(log, 0);
985 bio_list_add(&log->current_io->flush_barriers, bio);
986 log->current_io->has_flush = 1;
987 log->current_io->has_null_flush = 1;
988 atomic_inc(&log->current_io->pending_stripe);
989 r5l_submit_current_io(log);
990 mutex_unlock(&log->io_mutex);
991 return 0;
992 }
528 /*
529 * we flush log disk cache first, then write stripe data to raid disks.
530 * So if bio is finished, the log disk cache is flushed already. The
531 * recovery guarantees we can recovery the bio from log disk, so we
532 * don't need to flush again
533 */
534 if (bio->bi_iter.bi_size == 0) {
535 bio_endio(bio);
536 return 0;
993 }
537 }
538 bio->bi_opf &= ~REQ_PREFLUSH;
994 return -EAGAIN;
995}
996
997/* This will run after log space is reclaimed */
998static void r5l_run_no_space_stripes(struct r5l_log *log)
999{
1000 struct stripe_head *sh;
1001
1002 spin_lock(&log->no_space_stripes_lock);
1003 while (!list_empty(&log->no_space_stripes)) {
1004 sh = list_first_entry(&log->no_space_stripes,
1005 struct stripe_head, log_list);
1006 list_del_init(&sh->log_list);
1007 set_bit(STRIPE_HANDLE, &sh->state);
1008 raid5_release_stripe(sh);
1009 }
1010 spin_unlock(&log->no_space_stripes_lock);
1011}
1012
539 return -EAGAIN;
540}
541
542/* This will run after log space is reclaimed */
543static void r5l_run_no_space_stripes(struct r5l_log *log)
544{
545 struct stripe_head *sh;
546
547 spin_lock(&log->no_space_stripes_lock);
548 while (!list_empty(&log->no_space_stripes)) {
549 sh = list_first_entry(&log->no_space_stripes,
550 struct stripe_head, log_list);
551 list_del_init(&sh->log_list);
552 set_bit(STRIPE_HANDLE, &sh->state);
553 raid5_release_stripe(sh);
554 }
555 spin_unlock(&log->no_space_stripes_lock);
556}
557
1013/*
1014 * calculate new last_checkpoint
1015 * for write through mode, returns log->next_checkpoint
1016 * for write back, returns log_start of first sh in stripe_in_journal_list
1017 */
1018static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1019{
1020 struct stripe_head *sh;
1021 struct r5l_log *log = conf->log;
1022 sector_t new_cp;
1023 unsigned long flags;
1024
1025 if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1026 return log->next_checkpoint;
1027
1028 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1029 if (list_empty(&conf->log->stripe_in_journal_list)) {
1030 /* all stripes flushed */
1031 spin_unlock(&log->stripe_in_journal_lock);
1032 return log->next_checkpoint;
1033 }
1034 sh = list_first_entry(&conf->log->stripe_in_journal_list,
1035 struct stripe_head, r5c);
1036 new_cp = sh->log_start;
1037 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1038 return new_cp;
1039}
1040
1041static sector_t r5l_reclaimable_space(struct r5l_log *log)
1042{
558static sector_t r5l_reclaimable_space(struct r5l_log *log)
559{
1043 struct r5conf *conf = log->rdev->mddev->private;
1044
1045 return r5l_ring_distance(log, log->last_checkpoint,
560 return r5l_ring_distance(log, log->last_checkpoint,
1046 r5c_calculate_new_cp(conf));
561 log->next_checkpoint);
1047}
1048
1049static void r5l_run_no_mem_stripe(struct r5l_log *log)
1050{
1051 struct stripe_head *sh;
1052
1053 assert_spin_locked(&log->io_list_lock);
1054

--- 29 unchanged lines hidden (view full) ---

1084 }
1085
1086 return found;
1087}
1088
1089static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1090{
1091 struct r5l_log *log = io->log;
562}
563
564static void r5l_run_no_mem_stripe(struct r5l_log *log)
565{
566 struct stripe_head *sh;
567
568 assert_spin_locked(&log->io_list_lock);
569

--- 29 unchanged lines hidden (view full) ---

599 }
600
601 return found;
602}
603
604static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
605{
606 struct r5l_log *log = io->log;
1092 struct r5conf *conf = log->rdev->mddev->private;
1093 unsigned long flags;
1094
1095 spin_lock_irqsave(&log->io_list_lock, flags);
1096 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1097
1098 if (!r5l_complete_finished_ios(log)) {
1099 spin_unlock_irqrestore(&log->io_list_lock, flags);
1100 return;
1101 }
1102
607 unsigned long flags;
608
609 spin_lock_irqsave(&log->io_list_lock, flags);
610 __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
611
612 if (!r5l_complete_finished_ios(log)) {
613 spin_unlock_irqrestore(&log->io_list_lock, flags);
614 return;
615 }
616
1103 if (r5l_reclaimable_space(log) > log->max_free_space ||
1104 test_bit(R5C_LOG_TIGHT, &conf->cache_state))
617 if (r5l_reclaimable_space(log) > log->max_free_space)
1105 r5l_wake_reclaim(log, 0);
1106
1107 spin_unlock_irqrestore(&log->io_list_lock, flags);
1108 wake_up(&log->iounit_wait);
1109}
1110
1111void r5l_stripe_write_finished(struct stripe_head *sh)
1112{

--- 54 unchanged lines hidden (view full) ---

1167 do_flush = !list_empty(&log->flushing_ios);
1168 spin_unlock_irq(&log->io_list_lock);
1169
1170 if (!do_flush)
1171 return;
1172 bio_reset(&log->flush_bio);
1173 log->flush_bio.bi_bdev = log->rdev->bdev;
1174 log->flush_bio.bi_end_io = r5l_log_flush_endio;
618 r5l_wake_reclaim(log, 0);
619
620 spin_unlock_irqrestore(&log->io_list_lock, flags);
621 wake_up(&log->iounit_wait);
622}
623
624void r5l_stripe_write_finished(struct stripe_head *sh)
625{

--- 54 unchanged lines hidden (view full) ---

680 do_flush = !list_empty(&log->flushing_ios);
681 spin_unlock_irq(&log->io_list_lock);
682
683 if (!do_flush)
684 return;
685 bio_reset(&log->flush_bio);
686 log->flush_bio.bi_bdev = log->rdev->bdev;
687 log->flush_bio.bi_end_io = r5l_log_flush_endio;
1175 bio_set_op_attrs(&log->flush_bio, REQ_OP_WRITE, WRITE_FLUSH);
688 log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1176 submit_bio(&log->flush_bio);
1177}
1178
1179static void r5l_write_super(struct r5l_log *log, sector_t cp);
1180static void r5l_write_super_and_discard_space(struct r5l_log *log,
1181 sector_t end)
1182{
1183 struct block_device *bdev = log->rdev->bdev;

--- 33 unchanged lines hidden (view full) ---

1217 log->last_checkpoint + log->rdev->data_offset,
1218 log->device_size - log->last_checkpoint,
1219 GFP_NOIO, 0);
1220 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1221 GFP_NOIO, 0);
1222 }
1223}
1224
689 submit_bio(&log->flush_bio);
690}
691
692static void r5l_write_super(struct r5l_log *log, sector_t cp);
693static void r5l_write_super_and_discard_space(struct r5l_log *log,
694 sector_t end)
695{
696 struct block_device *bdev = log->rdev->bdev;

--- 33 unchanged lines hidden (view full) ---

730 log->last_checkpoint + log->rdev->data_offset,
731 log->device_size - log->last_checkpoint,
732 GFP_NOIO, 0);
733 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
734 GFP_NOIO, 0);
735 }
736}
737
1225/*
1226 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1227 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1228 *
1229 * must hold conf->device_lock
1230 */
1231static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1232{
1233 BUG_ON(list_empty(&sh->lru));
1234 BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1235 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1236
738
1237 /*
1238 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1239 * raid5_release_stripe() while holding conf->device_lock
1240 */
1241 BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1242 assert_spin_locked(&conf->device_lock);
1243
1244 list_del_init(&sh->lru);
1245 atomic_inc(&sh->count);
1246
1247 set_bit(STRIPE_HANDLE, &sh->state);
1248 atomic_inc(&conf->active_stripes);
1249 r5c_make_stripe_write_out(sh);
1250
1251 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
1252 atomic_inc(&conf->preread_active_stripes);
1253 raid5_release_stripe(sh);
1254}
1255
1256/*
1257 * if num == 0, flush all full stripes
1258 * if num > 0, flush all full stripes. If less than num full stripes are
1259 * flushed, flush some partial stripes until totally num stripes are
1260 * flushed or there is no more cached stripes.
1261 */
1262void r5c_flush_cache(struct r5conf *conf, int num)
1263{
1264 int count;
1265 struct stripe_head *sh, *next;
1266
1267 assert_spin_locked(&conf->device_lock);
1268 if (!conf->log)
1269 return;
1270
1271 count = 0;
1272 list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1273 r5c_flush_stripe(conf, sh);
1274 count++;
1275 }
1276
1277 if (count >= num)
1278 return;
1279 list_for_each_entry_safe(sh, next,
1280 &conf->r5c_partial_stripe_list, lru) {
1281 r5c_flush_stripe(conf, sh);
1282 if (++count >= num)
1283 break;
1284 }
1285}
1286
1287static void r5c_do_reclaim(struct r5conf *conf)
1288{
1289 struct r5l_log *log = conf->log;
1290 struct stripe_head *sh;
1291 int count = 0;
1292 unsigned long flags;
1293 int total_cached;
1294 int stripes_to_flush;
1295
1296 if (!r5c_is_writeback(log))
1297 return;
1298
1299 total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1300 atomic_read(&conf->r5c_cached_full_stripes);
1301
1302 if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1303 atomic_read(&conf->empty_inactive_list_nr) > 0)
1304 /*
1305 * if stripe cache pressure high, flush all full stripes and
1306 * some partial stripes
1307 */
1308 stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1309 else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1310 atomic_read(&conf->r5c_cached_full_stripes) >
1311 R5C_FULL_STRIPE_FLUSH_BATCH)
1312 /*
1313 * if stripe cache pressure moderate, or if there is many full
1314 * stripes,flush all full stripes
1315 */
1316 stripes_to_flush = 0;
1317 else
1318 /* no need to flush */
1319 stripes_to_flush = -1;
1320
1321 if (stripes_to_flush >= 0) {
1322 spin_lock_irqsave(&conf->device_lock, flags);
1323 r5c_flush_cache(conf, stripes_to_flush);
1324 spin_unlock_irqrestore(&conf->device_lock, flags);
1325 }
1326
1327 /* if log space is tight, flush stripes on stripe_in_journal_list */
1328 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1329 spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1330 spin_lock(&conf->device_lock);
1331 list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1332 /*
1333 * stripes on stripe_in_journal_list could be in any
1334 * state of the stripe_cache state machine. In this
1335 * case, we only want to flush stripe on
1336 * r5c_cached_full/partial_stripes. The following
1337 * condition makes sure the stripe is on one of the
1338 * two lists.
1339 */
1340 if (!list_empty(&sh->lru) &&
1341 !test_bit(STRIPE_HANDLE, &sh->state) &&
1342 atomic_read(&sh->count) == 0) {
1343 r5c_flush_stripe(conf, sh);
1344 }
1345 if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1346 break;
1347 }
1348 spin_unlock(&conf->device_lock);
1349 spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1350 }
1351 md_wakeup_thread(conf->mddev->thread);
1352}
1353
1354static void r5l_do_reclaim(struct r5l_log *log)
1355{
739static void r5l_do_reclaim(struct r5l_log *log)
740{
1356 struct r5conf *conf = log->rdev->mddev->private;
1357 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1358 sector_t reclaimable;
1359 sector_t next_checkpoint;
741 sector_t reclaim_target = xchg(&log->reclaim_target, 0);
742 sector_t reclaimable;
743 sector_t next_checkpoint;
1360 bool write_super;
744 u64 next_cp_seq;
1361
1362 spin_lock_irq(&log->io_list_lock);
745
746 spin_lock_irq(&log->io_list_lock);
1363 write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1364 reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1365 /*
1366 * move proper io_unit to reclaim list. We should not change the order.
1367 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1368 * shouldn't reuse space of an unreclaimable io_unit
1369 */
1370 while (1) {
1371 reclaimable = r5l_reclaimable_space(log);
1372 if (reclaimable >= reclaim_target ||

--- 4 unchanged lines hidden (view full) ---

1377 break;
1378
1379 md_wakeup_thread(log->rdev->mddev->thread);
1380 wait_event_lock_irq(log->iounit_wait,
1381 r5l_reclaimable_space(log) > reclaimable,
1382 log->io_list_lock);
1383 }
1384
747 /*
748 * move proper io_unit to reclaim list. We should not change the order.
749 * reclaimable/unreclaimable io_unit can be mixed in the list, we
750 * shouldn't reuse space of an unreclaimable io_unit
751 */
752 while (1) {
753 reclaimable = r5l_reclaimable_space(log);
754 if (reclaimable >= reclaim_target ||

--- 4 unchanged lines hidden (view full) ---

759 break;
760
761 md_wakeup_thread(log->rdev->mddev->thread);
762 wait_event_lock_irq(log->iounit_wait,
763 r5l_reclaimable_space(log) > reclaimable,
764 log->io_list_lock);
765 }
766
1385 next_checkpoint = r5c_calculate_new_cp(conf);
767 next_checkpoint = log->next_checkpoint;
768 next_cp_seq = log->next_cp_seq;
1386 spin_unlock_irq(&log->io_list_lock);
1387
1388 BUG_ON(reclaimable < 0);
769 spin_unlock_irq(&log->io_list_lock);
770
771 BUG_ON(reclaimable < 0);
1389
1390 if (reclaimable == 0 || !write_super)
772 if (reclaimable == 0)
1391 return;
1392
1393 /*
1394 * write_super will flush cache of each raid disk. We must write super
1395 * here, because the log area might be reused soon and we don't want to
1396 * confuse recovery
1397 */
1398 r5l_write_super_and_discard_space(log, next_checkpoint);
1399
1400 mutex_lock(&log->io_mutex);
1401 log->last_checkpoint = next_checkpoint;
773 return;
774
775 /*
776 * write_super will flush cache of each raid disk. We must write super
777 * here, because the log area might be reused soon and we don't want to
778 * confuse recovery
779 */
780 r5l_write_super_and_discard_space(log, next_checkpoint);
781
782 mutex_lock(&log->io_mutex);
783 log->last_checkpoint = next_checkpoint;
1402 r5c_update_log_state(log);
784 log->last_cp_seq = next_cp_seq;
1403 mutex_unlock(&log->io_mutex);
1404
1405 r5l_run_no_space_stripes(log);
1406}
1407
1408static void r5l_reclaim_thread(struct md_thread *thread)
1409{
1410 struct mddev *mddev = thread->mddev;
1411 struct r5conf *conf = mddev->private;
1412 struct r5l_log *log = conf->log;
1413
1414 if (!log)
1415 return;
785 mutex_unlock(&log->io_mutex);
786
787 r5l_run_no_space_stripes(log);
788}
789
790static void r5l_reclaim_thread(struct md_thread *thread)
791{
792 struct mddev *mddev = thread->mddev;
793 struct r5conf *conf = mddev->private;
794 struct r5l_log *log = conf->log;
795
796 if (!log)
797 return;
1416 r5c_do_reclaim(conf);
1417 r5l_do_reclaim(log);
1418}
1419
798 r5l_do_reclaim(log);
799}
800
1420void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
801static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1421{
1422 unsigned long target;
1423 unsigned long new = (unsigned long)space; /* overflow in theory */
1424
802{
803 unsigned long target;
804 unsigned long new = (unsigned long)space; /* overflow in theory */
805
1425 if (!log)
1426 return;
1427 do {
1428 target = log->reclaim_target;
1429 if (new < target)
1430 return;
1431 } while (cmpxchg(&log->reclaim_target, target, new) != target);
1432 md_wakeup_thread(log->reclaim_thread);
1433}
1434

--- 7 unchanged lines hidden (view full) ---

1442 * This is a special case for hotadd. In suspend, the array has
1443 * no journal. In resume, journal is initialized as well as the
1444 * reclaim thread.
1445 */
1446 if (log->reclaim_thread)
1447 return;
1448 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1449 log->rdev->mddev, "reclaim");
806 do {
807 target = log->reclaim_target;
808 if (new < target)
809 return;
810 } while (cmpxchg(&log->reclaim_target, target, new) != target);
811 md_wakeup_thread(log->reclaim_thread);
812}
813

--- 7 unchanged lines hidden (view full) ---

821 * This is a special case for hotadd. In suspend, the array has
822 * no journal. In resume, journal is initialized as well as the
823 * reclaim thread.
824 */
825 if (log->reclaim_thread)
826 return;
827 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
828 log->rdev->mddev, "reclaim");
1450 log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
1451 } else if (state == 1) {
1452 /* make sure r5l_write_super_and_discard_space exits */
1453 mddev = log->rdev->mddev;
1454 wake_up(&mddev->sb_wait);
829 } else if (state == 1) {
830 /* make sure r5l_write_super_and_discard_space exits */
831 mddev = log->rdev->mddev;
832 wake_up(&mddev->sb_wait);
1455 r5l_wake_reclaim(log, MaxSector);
833 r5l_wake_reclaim(log, -1L);
1456 md_unregister_thread(&log->reclaim_thread);
1457 r5l_do_reclaim(log);
1458 }
1459}
1460
1461bool r5l_log_disk_error(struct r5conf *conf)
1462{
1463 struct r5l_log *log;

--- 10 unchanged lines hidden (view full) ---

1474 return ret;
1475}
1476
1477struct r5l_recovery_ctx {
1478 struct page *meta_page; /* current meta */
1479 sector_t meta_total_blocks; /* total size of current meta and data */
1480 sector_t pos; /* recovery position */
1481 u64 seq; /* recovery position seq */
834 md_unregister_thread(&log->reclaim_thread);
835 r5l_do_reclaim(log);
836 }
837}
838
839bool r5l_log_disk_error(struct r5conf *conf)
840{
841 struct r5l_log *log;

--- 10 unchanged lines hidden (view full) ---

852 return ret;
853}
854
855struct r5l_recovery_ctx {
856 struct page *meta_page; /* current meta */
857 sector_t meta_total_blocks; /* total size of current meta and data */
858 sector_t pos; /* recovery position */
859 u64 seq; /* recovery position seq */
1482 int data_parity_stripes; /* number of data_parity stripes */
1483 int data_only_stripes; /* number of data_only stripes */
1484 struct list_head cached_list;
1485};
1486
860};
861
1487static int r5l_recovery_read_meta_block(struct r5l_log *log,
1488 struct r5l_recovery_ctx *ctx)
862static int r5l_read_meta_block(struct r5l_log *log,
863 struct r5l_recovery_ctx *ctx)
1489{
1490 struct page *page = ctx->meta_page;
1491 struct r5l_meta_block *mb;
1492 u32 crc, stored_crc;
1493
1494 if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, REQ_OP_READ, 0,
1495 false))
1496 return -EIO;

--- 15 unchanged lines hidden (view full) ---

1512 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1513 return -EINVAL;
1514
1515 ctx->meta_total_blocks = BLOCK_SECTORS;
1516
1517 return 0;
1518}
1519
864{
865 struct page *page = ctx->meta_page;
866 struct r5l_meta_block *mb;
867 u32 crc, stored_crc;
868
869 if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, REQ_OP_READ, 0,
870 false))
871 return -EIO;

--- 15 unchanged lines hidden (view full) ---

887 if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
888 return -EINVAL;
889
890 ctx->meta_total_blocks = BLOCK_SECTORS;
891
892 return 0;
893}
894
1520static void
1521r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1522 struct page *page,
1523 sector_t pos, u64 seq)
895static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
896 struct r5l_recovery_ctx *ctx,
897 sector_t stripe_sect,
898 int *offset, sector_t *log_offset)
1524{
899{
1525 struct r5l_meta_block *mb;
1526 u32 crc;
900 struct r5conf *conf = log->rdev->mddev->private;
901 struct stripe_head *sh;
902 struct r5l_payload_data_parity *payload;
903 int disk_index;
1527
904
1528 mb = page_address(page);
1529 clear_page(mb);
1530 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1531 mb->version = R5LOG_VERSION;
1532 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1533 mb->seq = cpu_to_le64(seq);
1534 mb->position = cpu_to_le64(pos);
1535 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1536 mb->checksum = cpu_to_le32(crc);
1537}
905 sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
906 while (1) {
907 payload = page_address(ctx->meta_page) + *offset;
1538
908
1539static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1540 u64 seq)
1541{
1542 struct page *page;
909 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
910 raid5_compute_sector(conf,
911 le64_to_cpu(payload->location), 0,
912 &disk_index, sh);
1543
913
1544 page = alloc_page(GFP_KERNEL);
1545 if (!page)
1546 return -ENOMEM;
1547 r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1548 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1549 WRITE_FUA, false)) {
1550 __free_page(page);
1551 return -EIO;
1552 }
1553 __free_page(page);
1554 return 0;
1555}
914 sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
915 sh->dev[disk_index].page, REQ_OP_READ, 0,
916 false);
917 sh->dev[disk_index].log_checksum =
918 le32_to_cpu(payload->checksum[0]);
919 set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
920 ctx->meta_total_blocks += BLOCK_SECTORS;
921 } else {
922 disk_index = sh->pd_idx;
923 sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
924 sh->dev[disk_index].page, REQ_OP_READ, 0,
925 false);
926 sh->dev[disk_index].log_checksum =
927 le32_to_cpu(payload->checksum[0]);
928 set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
1556
929
1557/*
1558 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1559 * to mark valid (potentially not flushed) data in the journal.
1560 *
1561 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1562 * so there should not be any mismatch here.
1563 */
1564static void r5l_recovery_load_data(struct r5l_log *log,
1565 struct stripe_head *sh,
1566 struct r5l_recovery_ctx *ctx,
1567 struct r5l_payload_data_parity *payload,
1568 sector_t log_offset)
1569{
1570 struct mddev *mddev = log->rdev->mddev;
1571 struct r5conf *conf = mddev->private;
1572 int dd_idx;
930 if (sh->qd_idx >= 0) {
931 disk_index = sh->qd_idx;
932 sync_page_io(log->rdev,
933 r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
934 PAGE_SIZE, sh->dev[disk_index].page,
935 REQ_OP_READ, 0, false);
936 sh->dev[disk_index].log_checksum =
937 le32_to_cpu(payload->checksum[1]);
938 set_bit(R5_Wantwrite,
939 &sh->dev[disk_index].flags);
940 }
941 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
942 }
1573
943
1574 raid5_compute_sector(conf,
1575 le64_to_cpu(payload->location), 0,
1576 &dd_idx, sh);
1577 sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1578 sh->dev[dd_idx].page, REQ_OP_READ, 0, false);
1579 sh->dev[dd_idx].log_checksum =
1580 le32_to_cpu(payload->checksum[0]);
1581 ctx->meta_total_blocks += BLOCK_SECTORS;
1582
1583 set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1584 set_bit(STRIPE_R5C_CACHING, &sh->state);
1585}
1586
1587static void r5l_recovery_load_parity(struct r5l_log *log,
1588 struct stripe_head *sh,
1589 struct r5l_recovery_ctx *ctx,
1590 struct r5l_payload_data_parity *payload,
1591 sector_t log_offset)
1592{
1593 struct mddev *mddev = log->rdev->mddev;
1594 struct r5conf *conf = mddev->private;
1595
1596 ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1597 sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1598 sh->dev[sh->pd_idx].page, REQ_OP_READ, 0, false);
1599 sh->dev[sh->pd_idx].log_checksum =
1600 le32_to_cpu(payload->checksum[0]);
1601 set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1602
1603 if (sh->qd_idx >= 0) {
1604 sync_page_io(log->rdev,
1605 r5l_ring_add(log, log_offset, BLOCK_SECTORS),
1606 PAGE_SIZE, sh->dev[sh->qd_idx].page,
1607 REQ_OP_READ, 0, false);
1608 sh->dev[sh->qd_idx].log_checksum =
1609 le32_to_cpu(payload->checksum[1]);
1610 set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
944 *log_offset = r5l_ring_add(log, *log_offset,
945 le32_to_cpu(payload->size));
946 *offset += sizeof(struct r5l_payload_data_parity) +
947 sizeof(__le32) *
948 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
949 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
950 break;
1611 }
951 }
1612 clear_bit(STRIPE_R5C_CACHING, &sh->state);
1613}
1614
952
1615static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1616{
1617 int i;
1618
1619 sh->state = 0;
1620 sh->log_start = MaxSector;
1621 for (i = sh->disks; i--; )
1622 sh->dev[i].flags = 0;
1623}
1624
1625static void
1626r5l_recovery_replay_one_stripe(struct r5conf *conf,
1627 struct stripe_head *sh,
1628 struct r5l_recovery_ctx *ctx)
1629{
1630 struct md_rdev *rdev, *rrdev;
1631 int disk_index;
1632 int data_count = 0;
1633
1634 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
953 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
954 void *addr;
955 u32 checksum;
956
1635 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1636 continue;
957 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
958 continue;
1637 if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1638 continue;
1639 data_count++;
959 addr = kmap_atomic(sh->dev[disk_index].page);
960 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
961 kunmap_atomic(addr);
962 if (checksum != sh->dev[disk_index].log_checksum)
963 goto error;
1640 }
1641
964 }
965
1642 /*
1643 * stripes that only have parity must have been flushed
1644 * before the crash that we are now recovering from, so
1645 * there is nothing more to recovery.
1646 */
1647 if (data_count == 0)
1648 goto out;
1649
1650 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
966 for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1651 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
967 struct md_rdev *rdev, *rrdev;
968
969 if (!test_and_clear_bit(R5_Wantwrite,
970 &sh->dev[disk_index].flags))
1652 continue;
1653
1654 /* in case device is broken */
971 continue;
972
973 /* in case device is broken */
1655 rcu_read_lock();
1656 rdev = rcu_dereference(conf->disks[disk_index].rdev);
974 rdev = rcu_dereference(conf->disks[disk_index].rdev);
1657 if (rdev) {
1658 atomic_inc(&rdev->nr_pending);
1659 rcu_read_unlock();
1660 sync_page_io(rdev, sh->sector, PAGE_SIZE,
975 if (rdev)
976 sync_page_io(rdev, stripe_sect, PAGE_SIZE,
1661 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1662 false);
977 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
978 false);
1663 rdev_dec_pending(rdev, rdev->mddev);
1664 rcu_read_lock();
1665 }
1666 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
979 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1667 if (rrdev) {
1668 atomic_inc(&rrdev->nr_pending);
1669 rcu_read_unlock();
1670 sync_page_io(rrdev, sh->sector, PAGE_SIZE,
980 if (rrdev)
981 sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
1671 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1672 false);
982 sh->dev[disk_index].page, REQ_OP_WRITE, 0,
983 false);
1673 rdev_dec_pending(rrdev, rrdev->mddev);
1674 rcu_read_lock();
1675 }
1676 rcu_read_unlock();
1677 }
984 }
1678 ctx->data_parity_stripes++;
1679out:
1680 r5l_recovery_reset_stripe(sh);
1681}
1682
1683static struct stripe_head *
1684r5c_recovery_alloc_stripe(struct r5conf *conf,
1685 struct list_head *recovery_list,
1686 sector_t stripe_sect,
1687 sector_t log_start)
1688{
1689 struct stripe_head *sh;
1690
1691 sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
1692 if (!sh)
1693 return NULL; /* no more stripe available */
1694
1695 r5l_recovery_reset_stripe(sh);
1696 sh->log_start = log_start;
1697
1698 return sh;
1699}
1700
1701static struct stripe_head *
1702r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1703{
1704 struct stripe_head *sh;
1705
1706 list_for_each_entry(sh, list, lru)
1707 if (sh->sector == sect)
1708 return sh;
1709 return NULL;
1710}
1711
1712static void
1713r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1714 struct r5l_recovery_ctx *ctx)
1715{
1716 struct stripe_head *sh, *next;
1717
1718 list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1719 r5l_recovery_reset_stripe(sh);
1720 list_del_init(&sh->lru);
1721 raid5_release_stripe(sh);
1722 }
1723}
1724
1725static void
1726r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1727 struct r5l_recovery_ctx *ctx)
1728{
1729 struct stripe_head *sh, *next;
1730
1731 list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1732 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1733 r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1734 list_del_init(&sh->lru);
1735 raid5_release_stripe(sh);
1736 }
1737}
1738
1739/* if matches return 0; otherwise return -EINVAL */
1740static int
1741r5l_recovery_verify_data_checksum(struct r5l_log *log, struct page *page,
1742 sector_t log_offset, __le32 log_checksum)
1743{
1744 void *addr;
1745 u32 checksum;
1746
1747 sync_page_io(log->rdev, log_offset, PAGE_SIZE,
1748 page, REQ_OP_READ, 0, false);
1749 addr = kmap_atomic(page);
1750 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1751 kunmap_atomic(addr);
1752 return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1753}
1754
1755/*
1756 * before loading data to stripe cache, we need verify checksum for all data,
1757 * if there is mismatch for any data page, we drop all data in the mata block
1758 */
1759static int
1760r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1761 struct r5l_recovery_ctx *ctx)
1762{
1763 struct mddev *mddev = log->rdev->mddev;
1764 struct r5conf *conf = mddev->private;
1765 struct r5l_meta_block *mb = page_address(ctx->meta_page);
1766 sector_t mb_offset = sizeof(struct r5l_meta_block);
1767 sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1768 struct page *page;
1769 struct r5l_payload_data_parity *payload;
1770
1771 page = alloc_page(GFP_KERNEL);
1772 if (!page)
1773 return -ENOMEM;
1774
1775 while (mb_offset < le32_to_cpu(mb->meta_size)) {
1776 payload = (void *)mb + mb_offset;
1777
1778 if (payload->header.type == R5LOG_PAYLOAD_DATA) {
1779 if (r5l_recovery_verify_data_checksum(
1780 log, page, log_offset,
1781 payload->checksum[0]) < 0)
1782 goto mismatch;
1783 } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) {
1784 if (r5l_recovery_verify_data_checksum(
1785 log, page, log_offset,
1786 payload->checksum[0]) < 0)
1787 goto mismatch;
1788 if (conf->max_degraded == 2 && /* q for RAID 6 */
1789 r5l_recovery_verify_data_checksum(
1790 log, page,
1791 r5l_ring_add(log, log_offset,
1792 BLOCK_SECTORS),
1793 payload->checksum[1]) < 0)
1794 goto mismatch;
1795 } else /* not R5LOG_PAYLOAD_DATA or R5LOG_PAYLOAD_PARITY */
1796 goto mismatch;
1797
1798 log_offset = r5l_ring_add(log, log_offset,
1799 le32_to_cpu(payload->size));
1800
1801 mb_offset += sizeof(struct r5l_payload_data_parity) +
1802 sizeof(__le32) *
1803 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
1804 }
1805
1806 put_page(page);
985 raid5_release_stripe(sh);
1807 return 0;
1808
986 return 0;
987
1809mismatch:
1810 put_page(page);
988error:
989 for (disk_index = 0; disk_index < sh->disks; disk_index++)
990 sh->dev[disk_index].flags = 0;
991 raid5_release_stripe(sh);
1811 return -EINVAL;
1812}
1813
992 return -EINVAL;
993}
994
1814/*
1815 * Analyze all data/parity pages in one meta block
1816 * Returns:
1817 * 0 for success
1818 * -EINVAL for unknown playload type
1819 * -EAGAIN for checksum mismatch of data page
1820 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
1821 */
1822static int
1823r5c_recovery_analyze_meta_block(struct r5l_log *log,
1824 struct r5l_recovery_ctx *ctx,
1825 struct list_head *cached_stripe_list)
995static int r5l_recovery_flush_one_meta(struct r5l_log *log,
996 struct r5l_recovery_ctx *ctx)
1826{
997{
1827 struct mddev *mddev = log->rdev->mddev;
1828 struct r5conf *conf = mddev->private;
1829 struct r5l_meta_block *mb;
998 struct r5conf *conf = log->rdev->mddev->private;
1830 struct r5l_payload_data_parity *payload;
999 struct r5l_payload_data_parity *payload;
1831 int mb_offset;
1000 struct r5l_meta_block *mb;
1001 int offset;
1832 sector_t log_offset;
1002 sector_t log_offset;
1833 sector_t stripe_sect;
1834 struct stripe_head *sh;
1835 int ret;
1003 sector_t stripe_sector;
1836
1004
1837 /*
1838 * for mismatch in data blocks, we will drop all data in this mb, but
1839 * we will still read next mb for other data with FLUSH flag, as
1840 * io_unit could finish out of order.
1841 */
1842 ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
1843 if (ret == -EINVAL)
1844 return -EAGAIN;
1845 else if (ret)
1846 return ret; /* -ENOMEM duo to alloc_page() failed */
1847
1848 mb = page_address(ctx->meta_page);
1005 mb = page_address(ctx->meta_page);
1849 mb_offset = sizeof(struct r5l_meta_block);
1006 offset = sizeof(struct r5l_meta_block);
1850 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1851
1007 log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
1008
1852 while (mb_offset < le32_to_cpu(mb->meta_size)) {
1009 while (offset < le32_to_cpu(mb->meta_size)) {
1853 int dd;
1854
1010 int dd;
1011
1855 payload = (void *)mb + mb_offset;
1856 stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ?
1857 raid5_compute_sector(
1858 conf, le64_to_cpu(payload->location), 0, &dd,
1859 NULL)
1860 : le64_to_cpu(payload->location);
1861
1862 sh = r5c_recovery_lookup_stripe(cached_stripe_list,
1863 stripe_sect);
1864
1865 if (!sh) {
1866 sh = r5c_recovery_alloc_stripe(conf, cached_stripe_list,
1867 stripe_sect, ctx->pos);
1868 /*
1869 * cannot get stripe from raid5_get_active_stripe
1870 * try replay some stripes
1871 */
1872 if (!sh) {
1873 r5c_recovery_replay_stripes(
1874 cached_stripe_list, ctx);
1875 sh = r5c_recovery_alloc_stripe(
1876 conf, cached_stripe_list,
1877 stripe_sect, ctx->pos);
1878 }
1879 if (!sh) {
1880 pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
1881 mdname(mddev),
1882 conf->min_nr_stripes * 2);
1883 raid5_set_cache_size(mddev,
1884 conf->min_nr_stripes * 2);
1885 sh = r5c_recovery_alloc_stripe(
1886 conf, cached_stripe_list, stripe_sect,
1887 ctx->pos);
1888 }
1889 if (!sh) {
1890 pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
1891 mdname(mddev));
1892 return -ENOMEM;
1893 }
1894 list_add_tail(&sh->lru, cached_stripe_list);
1895 }
1896
1897 if (payload->header.type == R5LOG_PAYLOAD_DATA) {
1898 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1899 r5l_recovery_replay_one_stripe(conf, sh, ctx);
1900 r5l_recovery_reset_stripe(sh);
1901 sh->log_start = ctx->pos;
1902 list_move_tail(&sh->lru, cached_stripe_list);
1903 }
1904 r5l_recovery_load_data(log, sh, ctx, payload,
1905 log_offset);
1906 } else if (payload->header.type == R5LOG_PAYLOAD_PARITY)
1907 r5l_recovery_load_parity(log, sh, ctx, payload,
1908 log_offset);
1909 else
1012 payload = (void *)mb + offset;
1013 stripe_sector = raid5_compute_sector(conf,
1014 le64_to_cpu(payload->location), 0, &dd, NULL);
1015 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
1016 &offset, &log_offset))
1910 return -EINVAL;
1017 return -EINVAL;
1911
1912 log_offset = r5l_ring_add(log, log_offset,
1913 le32_to_cpu(payload->size));
1914
1915 mb_offset += sizeof(struct r5l_payload_data_parity) +
1916 sizeof(__le32) *
1917 (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
1918 }
1018 }
1919
1920 return 0;
1921}
1922
1019 return 0;
1020}
1021
1923/*
1924 * Load the stripe into cache. The stripe will be written out later by
1925 * the stripe cache state machine.
1926 */
1927static void r5c_recovery_load_one_stripe(struct r5l_log *log,
1928 struct stripe_head *sh)
1022/* copy data/parity from log to raid disks */
1023static void r5l_recovery_flush_log(struct r5l_log *log,
1024 struct r5l_recovery_ctx *ctx)
1929{
1025{
1930 struct r5conf *conf = sh->raid_conf;
1931 struct r5dev *dev;
1932 int i;
1933
1934 for (i = sh->disks; i--; ) {
1935 dev = sh->dev + i;
1936 if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
1937 set_bit(R5_InJournal, &dev->flags);
1938 set_bit(R5_UPTODATE, &dev->flags);
1939 }
1940 }
1941 set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
1942 atomic_inc(&conf->r5c_cached_partial_stripes);
1943 list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
1944}
1945
1946/*
1947 * Scan through the log for all to-be-flushed data
1948 *
1949 * For stripes with data and parity, namely Data-Parity stripe
1950 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
1951 *
1952 * For stripes with only data, namely Data-Only stripe
1953 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
1954 *
1955 * For a stripe, if we see data after parity, we should discard all previous
1956 * data and parity for this stripe, as these data are already flushed to
1957 * the array.
1958 *
1959 * At the end of the scan, we return the new journal_tail, which points to
1960 * first data-only stripe on the journal device, or next invalid meta block.
1961 */
1962static int r5c_recovery_flush_log(struct r5l_log *log,
1963 struct r5l_recovery_ctx *ctx)
1964{
1965 struct stripe_head *sh, *next;
1966 int ret = 0;
1967
1968 /* scan through the log */
1969 while (1) {
1026 while (1) {
1970 if (r5l_recovery_read_meta_block(log, ctx))
1971 break;
1972
1973 ret = r5c_recovery_analyze_meta_block(log, ctx,
1974 &ctx->cached_list);
1975 /*
1976 * -EAGAIN means mismatch in data block, in this case, we still
1977 * try scan the next metablock
1978 */
1979 if (ret && ret != -EAGAIN)
1980 break; /* ret == -EINVAL or -ENOMEM */
1027 if (r5l_read_meta_block(log, ctx))
1028 return;
1029 if (r5l_recovery_flush_one_meta(log, ctx))
1030 return;
1981 ctx->seq++;
1982 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
1983 }
1031 ctx->seq++;
1032 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
1033 }
1984
1985 if (ret == -ENOMEM) {
1986 r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
1987 return ret;
1988 }
1989
1990 /* replay data-parity stripes */
1991 r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
1992
1993 /* load data-only stripes to stripe cache */
1994 list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
1995 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1996 r5c_recovery_load_one_stripe(log, sh);
1997 list_del_init(&sh->lru);
1998 raid5_release_stripe(sh);
1999 ctx->data_only_stripes++;
2000 }
2001
2002 return 0;
2003}
2004
1034}
1035
2005/*
2006 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2007 * log will start here. but we can't let superblock point to last valid
2008 * meta block. The log might looks like:
2009 * | meta 1| meta 2| meta 3|
2010 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2011 * superblock points to meta 1, we write a new valid meta 2n. if crash
2012 * happens again, new recovery will start from meta 1. Since meta 2n is
2013 * valid now, recovery will think meta 3 is valid, which is wrong.
2014 * The solution is we create a new meta in meta2 with its seq == meta
2015 * 1's seq + 10 and let superblock points to meta2. The same recovery will
2016 * not think meta 3 is a valid meta, because its seq doesn't match
2017 */
2018
2019/*
2020 * Before recovery, the log looks like the following
2021 *
2022 * ---------------------------------------------
2023 * | valid log | invalid log |
2024 * ---------------------------------------------
2025 * ^
2026 * |- log->last_checkpoint
2027 * |- log->last_cp_seq
2028 *
2029 * Now we scan through the log until we see invalid entry
2030 *
2031 * ---------------------------------------------
2032 * | valid log | invalid log |
2033 * ---------------------------------------------
2034 * ^ ^
2035 * |- log->last_checkpoint |- ctx->pos
2036 * |- log->last_cp_seq |- ctx->seq
2037 *
2038 * From this point, we need to increase seq number by 10 to avoid
2039 * confusing next recovery.
2040 *
2041 * ---------------------------------------------
2042 * | valid log | invalid log |
2043 * ---------------------------------------------
2044 * ^ ^
2045 * |- log->last_checkpoint |- ctx->pos+1
2046 * |- log->last_cp_seq |- ctx->seq+11
2047 *
2048 * However, it is not safe to start the state machine yet, because data only
2049 * parities are not yet secured in RAID. To save these data only parities, we
2050 * rewrite them from seq+11.
2051 *
2052 * -----------------------------------------------------------------
2053 * | valid log | data only stripes | invalid log |
2054 * -----------------------------------------------------------------
2055 * ^ ^
2056 * |- log->last_checkpoint |- ctx->pos+n
2057 * |- log->last_cp_seq |- ctx->seq+10+n
2058 *
2059 * If failure happens again during this process, the recovery can safe start
2060 * again from log->last_checkpoint.
2061 *
2062 * Once data only stripes are rewritten to journal, we move log_tail
2063 *
2064 * -----------------------------------------------------------------
2065 * | old log | data only stripes | invalid log |
2066 * -----------------------------------------------------------------
2067 * ^ ^
2068 * |- log->last_checkpoint |- ctx->pos+n
2069 * |- log->last_cp_seq |- ctx->seq+10+n
2070 *
2071 * Then we can safely start the state machine. If failure happens from this
2072 * point on, the recovery will start from new log->last_checkpoint.
2073 */
2074static int
2075r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2076 struct r5l_recovery_ctx *ctx)
1036static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1037 u64 seq)
2077{
1038{
2078 struct stripe_head *sh;
2079 struct mddev *mddev = log->rdev->mddev;
2080 struct page *page;
1039 struct page *page;
1040 struct r5l_meta_block *mb;
1041 u32 crc;
2081
1042
2082 page = alloc_page(GFP_KERNEL);
2083 if (!page) {
2084 pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2085 mdname(mddev));
1043 page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1044 if (!page)
2086 return -ENOMEM;
1045 return -ENOMEM;
2087 }
1046 mb = page_address(page);
1047 mb->magic = cpu_to_le32(R5LOG_MAGIC);
1048 mb->version = R5LOG_VERSION;
1049 mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1050 mb->seq = cpu_to_le64(seq);
1051 mb->position = cpu_to_le64(pos);
1052 crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1053 mb->checksum = cpu_to_le32(crc);
2088
1054
2089 ctx->seq += 10;
2090 list_for_each_entry(sh, &ctx->cached_list, lru) {
2091 struct r5l_meta_block *mb;
2092 int i;
2093 int offset;
2094 sector_t write_pos;
2095
2096 WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2097 r5l_recovery_create_empty_meta_block(log, page,
2098 ctx->pos, ctx->seq);
2099 mb = page_address(page);
2100 offset = le32_to_cpu(mb->meta_size);
2101 write_pos = ctx->pos + BLOCK_SECTORS;
2102
2103 for (i = sh->disks; i--; ) {
2104 struct r5dev *dev = &sh->dev[i];
2105 struct r5l_payload_data_parity *payload;
2106 void *addr;
2107
2108 if (test_bit(R5_InJournal, &dev->flags)) {
2109 payload = (void *)mb + offset;
2110 payload->header.type = cpu_to_le16(
2111 R5LOG_PAYLOAD_DATA);
2112 payload->size = BLOCK_SECTORS;
2113 payload->location = cpu_to_le64(
2114 raid5_compute_blocknr(sh, i, 0));
2115 addr = kmap_atomic(dev->page);
2116 payload->checksum[0] = cpu_to_le32(
2117 crc32c_le(log->uuid_checksum, addr,
2118 PAGE_SIZE));
2119 kunmap_atomic(addr);
2120 sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2121 dev->page, REQ_OP_WRITE, 0, false);
2122 write_pos = r5l_ring_add(log, write_pos,
2123 BLOCK_SECTORS);
2124 offset += sizeof(__le32) +
2125 sizeof(struct r5l_payload_data_parity);
2126
2127 }
2128 }
2129 mb->meta_size = cpu_to_le32(offset);
2130 mb->checksum = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2131 sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2132 REQ_OP_WRITE, WRITE_FUA, false);
2133 sh->log_start = ctx->pos;
2134 ctx->pos = write_pos;
2135 ctx->seq += 1;
1055 if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1056 REQ_FUA, false)) {
1057 __free_page(page);
1058 return -EIO;
2136 }
2137 __free_page(page);
2138 return 0;
2139}
2140
2141static int r5l_recovery_log(struct r5l_log *log)
2142{
1059 }
1060 __free_page(page);
1061 return 0;
1062}
1063
1064static int r5l_recovery_log(struct r5l_log *log)
1065{
2143 struct mddev *mddev = log->rdev->mddev;
2144 struct r5l_recovery_ctx ctx;
1066 struct r5l_recovery_ctx ctx;
2145 int ret;
2146
2147 ctx.pos = log->last_checkpoint;
2148 ctx.seq = log->last_cp_seq;
2149 ctx.meta_page = alloc_page(GFP_KERNEL);
1067
1068 ctx.pos = log->last_checkpoint;
1069 ctx.seq = log->last_cp_seq;
1070 ctx.meta_page = alloc_page(GFP_KERNEL);
2150 ctx.data_only_stripes = 0;
2151 ctx.data_parity_stripes = 0;
2152 INIT_LIST_HEAD(&ctx.cached_list);
2153
2154 if (!ctx.meta_page)
2155 return -ENOMEM;
2156
1071 if (!ctx.meta_page)
1072 return -ENOMEM;
1073
2157 ret = r5c_recovery_flush_log(log, &ctx);
1074 r5l_recovery_flush_log(log, &ctx);
2158 __free_page(ctx.meta_page);
2159
1075 __free_page(ctx.meta_page);
1076
2160 if (ret)
2161 return ret;
1077 /*
1078 * we did a recovery. Now ctx.pos points to an invalid meta block. New
1079 * log will start here. but we can't let superblock point to last valid
1080 * meta block. The log might looks like:
1081 * | meta 1| meta 2| meta 3|
1082 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
1083 * superblock points to meta 1, we write a new valid meta 2n. if crash
1084 * happens again, new recovery will start from meta 1. Since meta 2n is
1085 * valid now, recovery will think meta 3 is valid, which is wrong.
1086 * The solution is we create a new meta in meta2 with its seq == meta
1087 * 1's seq + 10 and let superblock points to meta2. The same recovery will
1088 * not think meta 3 is a valid meta, because its seq doesn't match
1089 */
1090 if (ctx.seq > log->last_cp_seq + 1) {
1091 int ret;
2162
1092
2163 if ((ctx.data_only_stripes == 0) && (ctx.data_parity_stripes == 0))
2164 pr_debug("md/raid:%s: starting from clean shutdown\n",
2165 mdname(mddev));
2166 else {
2167 pr_debug("md/raid:%s: recoverying %d data-only stripes and %d data-parity stripes\n",
2168 mdname(mddev), ctx.data_only_stripes,
2169 ctx.data_parity_stripes);
2170
2171 if (ctx.data_only_stripes > 0)
2172 if (r5c_recovery_rewrite_data_only_stripes(log, &ctx)) {
2173 pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2174 mdname(mddev));
2175 return -EIO;
2176 }
1093 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
1094 if (ret)
1095 return ret;
1096 log->seq = ctx.seq + 11;
1097 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
1098 r5l_write_super(log, ctx.pos);
1099 } else {
1100 log->log_start = ctx.pos;
1101 log->seq = ctx.seq;
2177 }
1102 }
2178
2179 log->log_start = ctx.pos;
2180 log->next_checkpoint = ctx.pos;
2181 log->seq = ctx.seq;
2182 r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq);
2183 r5l_write_super(log, ctx.pos);
2184 return 0;
2185}
2186
2187static void r5l_write_super(struct r5l_log *log, sector_t cp)
2188{
2189 struct mddev *mddev = log->rdev->mddev;
2190
2191 log->rdev->journal_tail = cp;
2192 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2193}
2194
1103 return 0;
1104}
1105
1106static void r5l_write_super(struct r5l_log *log, sector_t cp)
1107{
1108 struct mddev *mddev = log->rdev->mddev;
1109
1110 log->rdev->journal_tail = cp;
1111 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1112}
1113
2195static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2196{
2197 struct r5conf *conf = mddev->private;
2198 int ret;
2199
2200 if (!conf->log)
2201 return 0;
2202
2203 switch (conf->log->r5c_journal_mode) {
2204 case R5C_JOURNAL_MODE_WRITE_THROUGH:
2205 ret = snprintf(
2206 page, PAGE_SIZE, "[%s] %s\n",
2207 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2208 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2209 break;
2210 case R5C_JOURNAL_MODE_WRITE_BACK:
2211 ret = snprintf(
2212 page, PAGE_SIZE, "%s [%s]\n",
2213 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2214 r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2215 break;
2216 default:
2217 ret = 0;
2218 }
2219 return ret;
2220}
2221
2222static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2223 const char *page, size_t length)
2224{
2225 struct r5conf *conf = mddev->private;
2226 struct r5l_log *log = conf->log;
2227 int val = -1, i;
2228 int len = length;
2229
2230 if (!log)
2231 return -ENODEV;
2232
2233 if (len && page[len - 1] == '\n')
2234 len -= 1;
2235 for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++)
2236 if (strlen(r5c_journal_mode_str[i]) == len &&
2237 strncmp(page, r5c_journal_mode_str[i], len) == 0) {
2238 val = i;
2239 break;
2240 }
2241 if (val < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2242 val > R5C_JOURNAL_MODE_WRITE_BACK)
2243 return -EINVAL;
2244
2245 mddev_suspend(mddev);
2246 conf->log->r5c_journal_mode = val;
2247 mddev_resume(mddev);
2248
2249 pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2250 mdname(mddev), val, r5c_journal_mode_str[val]);
2251 return length;
2252}
2253
2254struct md_sysfs_entry
2255r5c_journal_mode = __ATTR(journal_mode, 0644,
2256 r5c_journal_mode_show, r5c_journal_mode_store);
2257
2258/*
2259 * Try handle write operation in caching phase. This function should only
2260 * be called in write-back mode.
2261 *
2262 * If all outstanding writes can be handled in caching phase, returns 0
2263 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2264 * and returns -EAGAIN
2265 */
2266int r5c_try_caching_write(struct r5conf *conf,
2267 struct stripe_head *sh,
2268 struct stripe_head_state *s,
2269 int disks)
2270{
2271 struct r5l_log *log = conf->log;
2272 int i;
2273 struct r5dev *dev;
2274 int to_cache = 0;
2275
2276 BUG_ON(!r5c_is_writeback(log));
2277
2278 if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2279 /*
2280 * There are two different scenarios here:
2281 * 1. The stripe has some data cached, and it is sent to
2282 * write-out phase for reclaim
2283 * 2. The stripe is clean, and this is the first write
2284 *
2285 * For 1, return -EAGAIN, so we continue with
2286 * handle_stripe_dirtying().
2287 *
2288 * For 2, set STRIPE_R5C_CACHING and continue with caching
2289 * write.
2290 */
2291
2292 /* case 1: anything injournal or anything in written */
2293 if (s->injournal > 0 || s->written > 0)
2294 return -EAGAIN;
2295 /* case 2 */
2296 set_bit(STRIPE_R5C_CACHING, &sh->state);
2297 }
2298
2299 for (i = disks; i--; ) {
2300 dev = &sh->dev[i];
2301 /* if non-overwrite, use writing-out phase */
2302 if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2303 !test_bit(R5_InJournal, &dev->flags)) {
2304 r5c_make_stripe_write_out(sh);
2305 return -EAGAIN;
2306 }
2307 }
2308
2309 for (i = disks; i--; ) {
2310 dev = &sh->dev[i];
2311 if (dev->towrite) {
2312 set_bit(R5_Wantwrite, &dev->flags);
2313 set_bit(R5_Wantdrain, &dev->flags);
2314 set_bit(R5_LOCKED, &dev->flags);
2315 to_cache++;
2316 }
2317 }
2318
2319 if (to_cache) {
2320 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2321 /*
2322 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2323 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2324 * r5c_handle_data_cached()
2325 */
2326 set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2327 }
2328
2329 return 0;
2330}
2331
2332/*
2333 * free extra pages (orig_page) we allocated for prexor
2334 */
2335void r5c_release_extra_page(struct stripe_head *sh)
2336{
2337 int i;
2338
2339 for (i = sh->disks; i--; )
2340 if (sh->dev[i].page != sh->dev[i].orig_page) {
2341 struct page *p = sh->dev[i].orig_page;
2342
2343 sh->dev[i].orig_page = sh->dev[i].page;
2344 put_page(p);
2345 }
2346}
2347
2348/*
2349 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2350 * stripe is committed to RAID disks.
2351 */
2352void r5c_finish_stripe_write_out(struct r5conf *conf,
2353 struct stripe_head *sh,
2354 struct stripe_head_state *s)
2355{
2356 int i;
2357 int do_wakeup = 0;
2358
2359 if (!conf->log ||
2360 !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2361 return;
2362
2363 WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2364 clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2365
2366 if (conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2367 return;
2368
2369 for (i = sh->disks; i--; ) {
2370 clear_bit(R5_InJournal, &sh->dev[i].flags);
2371 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2372 do_wakeup = 1;
2373 }
2374
2375 /*
2376 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2377 * We updated R5_InJournal, so we also update s->injournal.
2378 */
2379 s->injournal = 0;
2380
2381 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2382 if (atomic_dec_and_test(&conf->pending_full_writes))
2383 md_wakeup_thread(conf->mddev->thread);
2384
2385 if (do_wakeup)
2386 wake_up(&conf->wait_for_overlap);
2387
2388 if (conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2389 return;
2390
2391 spin_lock_irq(&conf->log->stripe_in_journal_lock);
2392 list_del_init(&sh->r5c);
2393 spin_unlock_irq(&conf->log->stripe_in_journal_lock);
2394 sh->log_start = MaxSector;
2395 atomic_dec(&conf->log->stripe_in_journal_count);
2396}
2397
2398int
2399r5c_cache_data(struct r5l_log *log, struct stripe_head *sh,
2400 struct stripe_head_state *s)
2401{
2402 struct r5conf *conf = sh->raid_conf;
2403 int pages = 0;
2404 int reserve;
2405 int i;
2406 int ret = 0;
2407
2408 BUG_ON(!log);
2409
2410 for (i = 0; i < sh->disks; i++) {
2411 void *addr;
2412
2413 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2414 continue;
2415 addr = kmap_atomic(sh->dev[i].page);
2416 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2417 addr, PAGE_SIZE);
2418 kunmap_atomic(addr);
2419 pages++;
2420 }
2421 WARN_ON(pages == 0);
2422
2423 /*
2424 * The stripe must enter state machine again to call endio, so
2425 * don't delay.
2426 */
2427 clear_bit(STRIPE_DELAYED, &sh->state);
2428 atomic_inc(&sh->count);
2429
2430 mutex_lock(&log->io_mutex);
2431 /* meta + data */
2432 reserve = (1 + pages) << (PAGE_SHIFT - 9);
2433
2434 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2435 sh->log_start == MaxSector)
2436 r5l_add_no_space_stripe(log, sh);
2437 else if (!r5l_has_free_space(log, reserve)) {
2438 if (sh->log_start == log->last_checkpoint)
2439 BUG();
2440 else
2441 r5l_add_no_space_stripe(log, sh);
2442 } else {
2443 ret = r5l_log_stripe(log, sh, pages, 0);
2444 if (ret) {
2445 spin_lock_irq(&log->io_list_lock);
2446 list_add_tail(&sh->log_list, &log->no_mem_stripes);
2447 spin_unlock_irq(&log->io_list_lock);
2448 }
2449 }
2450
2451 mutex_unlock(&log->io_mutex);
2452 return 0;
2453}
2454
2455static int r5l_load_log(struct r5l_log *log)
2456{
2457 struct md_rdev *rdev = log->rdev;
2458 struct page *page;
2459 struct r5l_meta_block *mb;
2460 sector_t cp = log->rdev->journal_tail;
2461 u32 stored_crc, expected_crc;
2462 bool create_super = false;

--- 27 unchanged lines hidden (view full) ---

2490 if (le64_to_cpu(mb->position) != cp) {
2491 create_super = true;
2492 goto create;
2493 }
2494create:
2495 if (create_super) {
2496 log->last_cp_seq = prandom_u32();
2497 cp = 0;
1114static int r5l_load_log(struct r5l_log *log)
1115{
1116 struct md_rdev *rdev = log->rdev;
1117 struct page *page;
1118 struct r5l_meta_block *mb;
1119 sector_t cp = log->rdev->journal_tail;
1120 u32 stored_crc, expected_crc;
1121 bool create_super = false;

--- 27 unchanged lines hidden (view full) ---

1149 if (le64_to_cpu(mb->position) != cp) {
1150 create_super = true;
1151 goto create;
1152 }
1153create:
1154 if (create_super) {
1155 log->last_cp_seq = prandom_u32();
1156 cp = 0;
2498 r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2499 /*
2500 * Make sure super points to correct address. Log might have
2501 * data very soon. If super hasn't correct log tail address,
2502 * recovery can't find the log
2503 */
2504 r5l_write_super(log, cp);
2505 } else
2506 log->last_cp_seq = le64_to_cpu(mb->seq);
2507
2508 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
2509 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
2510 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
2511 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
2512 log->last_checkpoint = cp;
1157 /*
1158 * Make sure super points to correct address. Log might have
1159 * data very soon. If super hasn't correct log tail address,
1160 * recovery can't find the log
1161 */
1162 r5l_write_super(log, cp);
1163 } else
1164 log->last_cp_seq = le64_to_cpu(mb->seq);
1165
1166 log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1167 log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1168 if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1169 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1170 log->last_checkpoint = cp;
2513 log->next_checkpoint = cp;
2514 mutex_lock(&log->io_mutex);
2515 r5c_update_log_state(log);
2516 mutex_unlock(&log->io_mutex);
2517
2518 __free_page(page);
2519
2520 return r5l_recovery_log(log);
2521ioerr:
2522 __free_page(page);
2523 return ret;
2524}
2525
2526int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
2527{
2528 struct request_queue *q = bdev_get_queue(rdev->bdev);
2529 struct r5l_log *log;
2530
2531 if (PAGE_SIZE != 4096)
2532 return -EINVAL;
1171
1172 __free_page(page);
1173
1174 return r5l_recovery_log(log);
1175ioerr:
1176 __free_page(page);
1177 return ret;
1178}
1179
1180int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1181{
1182 struct request_queue *q = bdev_get_queue(rdev->bdev);
1183 struct r5l_log *log;
1184
1185 if (PAGE_SIZE != 4096)
1186 return -EINVAL;
2533
2534 /*
2535 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
2536 * raid_disks r5l_payload_data_parity.
2537 *
2538 * Write journal and cache does not work for very big array
2539 * (raid_disks > 203)
2540 */
2541 if (sizeof(struct r5l_meta_block) +
2542 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
2543 conf->raid_disks) > PAGE_SIZE) {
2544 pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
2545 mdname(conf->mddev), conf->raid_disks);
2546 return -EINVAL;
2547 }
2548
2549 log = kzalloc(sizeof(*log), GFP_KERNEL);
2550 if (!log)
2551 return -ENOMEM;
2552 log->rdev = rdev;
2553
2554 log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
2555
2556 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
2557 sizeof(rdev->mddev->uuid));
2558
2559 mutex_init(&log->io_mutex);
2560
2561 spin_lock_init(&log->io_list_lock);
2562 INIT_LIST_HEAD(&log->running_ios);
2563 INIT_LIST_HEAD(&log->io_end_ios);
2564 INIT_LIST_HEAD(&log->flushing_ios);
2565 INIT_LIST_HEAD(&log->finished_ios);
1187 log = kzalloc(sizeof(*log), GFP_KERNEL);
1188 if (!log)
1189 return -ENOMEM;
1190 log->rdev = rdev;
1191
1192 log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
1193
1194 log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1195 sizeof(rdev->mddev->uuid));
1196
1197 mutex_init(&log->io_mutex);
1198
1199 spin_lock_init(&log->io_list_lock);
1200 INIT_LIST_HEAD(&log->running_ios);
1201 INIT_LIST_HEAD(&log->io_end_ios);
1202 INIT_LIST_HEAD(&log->flushing_ios);
1203 INIT_LIST_HEAD(&log->finished_ios);
2566 bio_init(&log->flush_bio);
1204 bio_init(&log->flush_bio, NULL, 0);
2567
2568 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
2569 if (!log->io_kc)
2570 goto io_kc;
2571
2572 log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
2573 if (!log->io_pool)
2574 goto io_pool;

--- 5 unchanged lines hidden (view full) ---

2580 log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
2581 if (!log->meta_pool)
2582 goto out_mempool;
2583
2584 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
2585 log->rdev->mddev, "reclaim");
2586 if (!log->reclaim_thread)
2587 goto reclaim_thread;
1205
1206 log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1207 if (!log->io_kc)
1208 goto io_kc;
1209
1210 log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
1211 if (!log->io_pool)
1212 goto io_pool;

--- 5 unchanged lines hidden (view full) ---

1218 log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
1219 if (!log->meta_pool)
1220 goto out_mempool;
1221
1222 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1223 log->rdev->mddev, "reclaim");
1224 if (!log->reclaim_thread)
1225 goto reclaim_thread;
2588 log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
2589
2590 init_waitqueue_head(&log->iounit_wait);
2591
2592 INIT_LIST_HEAD(&log->no_mem_stripes);
2593
2594 INIT_LIST_HEAD(&log->no_space_stripes);
2595 spin_lock_init(&log->no_space_stripes_lock);
2596
1226 init_waitqueue_head(&log->iounit_wait);
1227
1228 INIT_LIST_HEAD(&log->no_mem_stripes);
1229
1230 INIT_LIST_HEAD(&log->no_space_stripes);
1231 spin_lock_init(&log->no_space_stripes_lock);
1232
2597 INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
2598
2599 log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2600 INIT_LIST_HEAD(&log->stripe_in_journal_list);
2601 spin_lock_init(&log->stripe_in_journal_lock);
2602 atomic_set(&log->stripe_in_journal_count, 0);
2603
2604 if (r5l_load_log(log))
2605 goto error;
2606
2607 rcu_assign_pointer(conf->log, log);
2608 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
2609 return 0;
2610
2611error:

--- 23 unchanged lines hidden --- 		1233 if (r5l_load_log(log))
1234 goto error;
1235
1236 rcu_assign_pointer(conf->log, log);
1237 set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1238 return 0;
1239
1240error:

--- 23 unchanged lines hidden ---