Commit | Line | Data |
---|---|---|
f6bed0ef SL |
1 | /* |
2 | * Copyright (C) 2015 Shaohua Li <shli@fb.com> | |
b4c625c6 | 3 | * Copyright (C) 2016 Song Liu <songliubraving@fb.com> |
f6bed0ef SL |
4 | * |
5 | * This program is free software; you can redistribute it and/or modify it | |
6 | * under the terms and conditions of the GNU General Public License, | |
7 | * version 2, as published by the Free Software Foundation. | |
8 | * | |
9 | * This program is distributed in the hope it will be useful, but WITHOUT | |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for | |
12 | * more details. | |
13 | * | |
14 | */ | |
15 | #include <linux/kernel.h> | |
16 | #include <linux/wait.h> | |
17 | #include <linux/blkdev.h> | |
18 | #include <linux/slab.h> | |
19 | #include <linux/raid/md_p.h> | |
5cb2fbd6 | 20 | #include <linux/crc32c.h> |
f6bed0ef SL |
21 | #include <linux/random.h> |
22 | #include "md.h" | |
23 | #include "raid5.h" | |
1e6d690b | 24 | #include "bitmap.h" |
f6bed0ef SL |
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 | ||
0576b1c6 | 32 | /* |
a39f7afd SL |
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 | |
0576b1c6 SL |
37 | */ |
38 | #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */ | |
39 | #define RECLAIM_MAX_FREE_SPACE_SHIFT (2) | |
40 | ||
a39f7afd SL |
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 | ||
c38d29b3 CH |
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 | ||
2ded3703 SL |
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 | */ | |
59 | enum r5c_journal_mode { | |
60 | R5C_JOURNAL_MODE_WRITE_THROUGH = 0, | |
61 | R5C_JOURNAL_MODE_WRITE_BACK = 1, | |
62 | }; | |
63 | ||
2c7da14b SL |
64 | static char *r5c_journal_mode_str[] = {"write-through", |
65 | "write-back"}; | |
2ded3703 SL |
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 | ||
f6bed0ef SL |
97 | struct 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 */ | |
0576b1c6 SL |
104 | sector_t max_free_space; /* reclaim run if free space is at |
105 | * this size */ | |
f6bed0ef SL |
106 | |
107 | sector_t last_checkpoint; /* log tail. where recovery scan | |
108 | * starts from */ | |
109 | u64 last_cp_seq; /* log tail sequence */ | |
110 | ||
111 | sector_t log_start; /* log head. where new data appends */ | |
112 | u64 seq; /* log head sequence */ | |
113 | ||
17036461 CH |
114 | sector_t next_checkpoint; |
115 | u64 next_cp_seq; | |
116 | ||
f6bed0ef SL |
117 | struct mutex io_mutex; |
118 | struct r5l_io_unit *current_io; /* current io_unit accepting new data */ | |
119 | ||
120 | spinlock_t io_list_lock; | |
121 | struct list_head running_ios; /* io_units which are still running, | |
122 | * and have not yet been completely | |
123 | * written to the log */ | |
124 | struct list_head io_end_ios; /* io_units which have been completely | |
125 | * written to the log but not yet written | |
126 | * to the RAID */ | |
a8c34f91 SL |
127 | struct list_head flushing_ios; /* io_units which are waiting for log |
128 | * cache flush */ | |
04732f74 | 129 | struct list_head finished_ios; /* io_units which settle down in log disk */ |
a8c34f91 | 130 | struct bio flush_bio; |
f6bed0ef | 131 | |
5036c390 CH |
132 | struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */ |
133 | ||
f6bed0ef | 134 | struct kmem_cache *io_kc; |
5036c390 | 135 | mempool_t *io_pool; |
c38d29b3 | 136 | struct bio_set *bs; |
e8deb638 | 137 | mempool_t *meta_pool; |
f6bed0ef | 138 | |
0576b1c6 SL |
139 | struct md_thread *reclaim_thread; |
140 | unsigned long reclaim_target; /* number of space that need to be | |
141 | * reclaimed. if it's 0, reclaim spaces | |
142 | * used by io_units which are in | |
143 | * IO_UNIT_STRIPE_END state (eg, reclaim | |
144 | * dones't wait for specific io_unit | |
145 | * switching to IO_UNIT_STRIPE_END | |
146 | * state) */ | |
0fd22b45 | 147 | wait_queue_head_t iounit_wait; |
0576b1c6 | 148 | |
f6bed0ef SL |
149 | struct list_head no_space_stripes; /* pending stripes, log has no space */ |
150 | spinlock_t no_space_stripes_lock; | |
56fef7c6 CH |
151 | |
152 | bool need_cache_flush; | |
2ded3703 SL |
153 | |
154 | /* for r5c_cache */ | |
155 | enum r5c_journal_mode r5c_journal_mode; | |
a39f7afd SL |
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; | |
f6bed0ef SL |
162 | }; |
163 | ||
164 | /* | |
165 | * an IO range starts from a meta data block and end at the next meta data | |
166 | * block. The io unit's the meta data block tracks data/parity followed it. io | |
167 | * unit is written to log disk with normal write, as we always flush log disk | |
168 | * first and then start move data to raid disks, there is no requirement to | |
169 | * write io unit with FLUSH/FUA | |
170 | */ | |
171 | struct r5l_io_unit { | |
172 | struct r5l_log *log; | |
173 | ||
174 | struct page *meta_page; /* store meta block */ | |
175 | int meta_offset; /* current offset in meta_page */ | |
176 | ||
f6bed0ef SL |
177 | struct bio *current_bio;/* current_bio accepting new data */ |
178 | ||
179 | atomic_t pending_stripe;/* how many stripes not flushed to raid */ | |
180 | u64 seq; /* seq number of the metablock */ | |
181 | sector_t log_start; /* where the io_unit starts */ | |
182 | sector_t log_end; /* where the io_unit ends */ | |
183 | struct list_head log_sibling; /* log->running_ios */ | |
184 | struct list_head stripe_list; /* stripes added to the io_unit */ | |
185 | ||
186 | int state; | |
6143e2ce | 187 | bool need_split_bio; |
f6bed0ef SL |
188 | }; |
189 | ||
190 | /* r5l_io_unit state */ | |
191 | enum r5l_io_unit_state { | |
192 | IO_UNIT_RUNNING = 0, /* accepting new IO */ | |
193 | IO_UNIT_IO_START = 1, /* io_unit bio start writing to log, | |
194 | * don't accepting new bio */ | |
195 | IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */ | |
a8c34f91 | 196 | IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */ |
f6bed0ef SL |
197 | }; |
198 | ||
2ded3703 SL |
199 | bool r5c_is_writeback(struct r5l_log *log) |
200 | { | |
201 | return (log != NULL && | |
202 | log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK); | |
203 | } | |
204 | ||
f6bed0ef SL |
205 | static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc) |
206 | { | |
207 | start += inc; | |
208 | if (start >= log->device_size) | |
209 | start = start - log->device_size; | |
210 | return start; | |
211 | } | |
212 | ||
213 | static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start, | |
214 | sector_t end) | |
215 | { | |
216 | if (end >= start) | |
217 | return end - start; | |
218 | else | |
219 | return end + log->device_size - start; | |
220 | } | |
221 | ||
222 | static bool r5l_has_free_space(struct r5l_log *log, sector_t size) | |
223 | { | |
224 | sector_t used_size; | |
225 | ||
226 | used_size = r5l_ring_distance(log, log->last_checkpoint, | |
227 | log->log_start); | |
228 | ||
229 | return log->device_size > used_size + size; | |
230 | } | |
231 | ||
f6bed0ef SL |
232 | static void __r5l_set_io_unit_state(struct r5l_io_unit *io, |
233 | enum r5l_io_unit_state state) | |
234 | { | |
f6bed0ef SL |
235 | if (WARN_ON(io->state >= state)) |
236 | return; | |
237 | io->state = state; | |
f6bed0ef SL |
238 | } |
239 | ||
1e6d690b SL |
240 | static void |
241 | r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev, | |
242 | struct bio_list *return_bi) | |
243 | { | |
244 | struct bio *wbi, *wbi2; | |
245 | ||
246 | wbi = dev->written; | |
247 | dev->written = NULL; | |
248 | while (wbi && wbi->bi_iter.bi_sector < | |
249 | dev->sector + STRIPE_SECTORS) { | |
250 | wbi2 = r5_next_bio(wbi, dev->sector); | |
251 | if (!raid5_dec_bi_active_stripes(wbi)) { | |
252 | md_write_end(conf->mddev); | |
253 | bio_list_add(return_bi, wbi); | |
254 | } | |
255 | wbi = wbi2; | |
256 | } | |
257 | } | |
258 | ||
259 | void r5c_handle_cached_data_endio(struct r5conf *conf, | |
260 | struct stripe_head *sh, int disks, struct bio_list *return_bi) | |
261 | { | |
262 | int i; | |
263 | ||
264 | for (i = sh->disks; i--; ) { | |
265 | if (sh->dev[i].written) { | |
266 | set_bit(R5_UPTODATE, &sh->dev[i].flags); | |
267 | r5c_return_dev_pending_writes(conf, &sh->dev[i], | |
268 | return_bi); | |
269 | bitmap_endwrite(conf->mddev->bitmap, sh->sector, | |
270 | STRIPE_SECTORS, | |
271 | !test_bit(STRIPE_DEGRADED, &sh->state), | |
272 | 0); | |
273 | } | |
274 | } | |
275 | } | |
276 | ||
a39f7afd SL |
277 | /* Check whether we should flush some stripes to free up stripe cache */ |
278 | void r5c_check_stripe_cache_usage(struct r5conf *conf) | |
279 | { | |
280 | int total_cached; | |
281 | ||
282 | if (!r5c_is_writeback(conf->log)) | |
283 | return; | |
284 | ||
285 | total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + | |
286 | atomic_read(&conf->r5c_cached_full_stripes); | |
287 | ||
288 | /* | |
289 | * The following condition is true for either of the following: | |
290 | * - stripe cache pressure high: | |
291 | * total_cached > 3/4 min_nr_stripes || | |
292 | * empty_inactive_list_nr > 0 | |
293 | * - stripe cache pressure moderate: | |
294 | * total_cached > 1/2 min_nr_stripes | |
295 | */ | |
296 | if (total_cached > conf->min_nr_stripes * 1 / 2 || | |
297 | atomic_read(&conf->empty_inactive_list_nr) > 0) | |
298 | r5l_wake_reclaim(conf->log, 0); | |
299 | } | |
300 | ||
301 | /* | |
302 | * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full | |
303 | * stripes in the cache | |
304 | */ | |
305 | void r5c_check_cached_full_stripe(struct r5conf *conf) | |
306 | { | |
307 | if (!r5c_is_writeback(conf->log)) | |
308 | return; | |
309 | ||
310 | /* | |
311 | * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes | |
312 | * or a full stripe (chunk size / 4k stripes). | |
313 | */ | |
314 | if (atomic_read(&conf->r5c_cached_full_stripes) >= | |
315 | min(R5C_FULL_STRIPE_FLUSH_BATCH, | |
316 | conf->chunk_sectors >> STRIPE_SHIFT)) | |
317 | r5l_wake_reclaim(conf->log, 0); | |
318 | } | |
319 | ||
320 | /* | |
321 | * Total log space (in sectors) needed to flush all data in cache | |
322 | * | |
323 | * Currently, writing-out phase automatically includes all pending writes | |
324 | * to the same sector. So the reclaim of each stripe takes up to | |
325 | * (conf->raid_disks + 1) pages of log space. | |
326 | * | |
327 | * To totally avoid deadlock due to log space, the code reserves | |
328 | * (conf->raid_disks + 1) pages for each stripe in cache, which is not | |
329 | * necessary in most cases. | |
330 | * | |
331 | * To improve this, we will need writing-out phase to be able to NOT include | |
332 | * pending writes, which will reduce the requirement to | |
333 | * (conf->max_degraded + 1) pages per stripe in cache. | |
334 | */ | |
335 | static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf) | |
336 | { | |
337 | struct r5l_log *log = conf->log; | |
338 | ||
339 | if (!r5c_is_writeback(log)) | |
340 | return 0; | |
341 | ||
342 | return BLOCK_SECTORS * (conf->raid_disks + 1) * | |
343 | atomic_read(&log->stripe_in_journal_count); | |
344 | } | |
345 | ||
346 | /* | |
347 | * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL | |
348 | * | |
349 | * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of | |
350 | * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log | |
351 | * device is less than 2x of reclaim_required_space. | |
352 | */ | |
353 | static inline void r5c_update_log_state(struct r5l_log *log) | |
354 | { | |
355 | struct r5conf *conf = log->rdev->mddev->private; | |
356 | sector_t free_space; | |
357 | sector_t reclaim_space; | |
358 | ||
359 | if (!r5c_is_writeback(log)) | |
360 | return; | |
361 | ||
362 | free_space = r5l_ring_distance(log, log->log_start, | |
363 | log->last_checkpoint); | |
364 | reclaim_space = r5c_log_required_to_flush_cache(conf); | |
365 | if (free_space < 2 * reclaim_space) | |
366 | set_bit(R5C_LOG_CRITICAL, &conf->cache_state); | |
367 | else | |
368 | clear_bit(R5C_LOG_CRITICAL, &conf->cache_state); | |
369 | if (free_space < 3 * reclaim_space) | |
370 | set_bit(R5C_LOG_TIGHT, &conf->cache_state); | |
371 | else | |
372 | clear_bit(R5C_LOG_TIGHT, &conf->cache_state); | |
373 | } | |
374 | ||
2ded3703 SL |
375 | /* |
376 | * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING. | |
377 | * This function should only be called in write-back mode. | |
378 | */ | |
a39f7afd | 379 | void r5c_make_stripe_write_out(struct stripe_head *sh) |
2ded3703 SL |
380 | { |
381 | struct r5conf *conf = sh->raid_conf; | |
382 | struct r5l_log *log = conf->log; | |
383 | ||
384 | BUG_ON(!r5c_is_writeback(log)); | |
385 | ||
386 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
387 | clear_bit(STRIPE_R5C_CACHING, &sh->state); | |
1e6d690b SL |
388 | |
389 | if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) | |
390 | atomic_inc(&conf->preread_active_stripes); | |
391 | ||
392 | if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) { | |
393 | BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0); | |
394 | atomic_dec(&conf->r5c_cached_partial_stripes); | |
395 | } | |
396 | ||
397 | if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { | |
398 | BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0); | |
399 | atomic_dec(&conf->r5c_cached_full_stripes); | |
400 | } | |
401 | } | |
402 | ||
403 | static void r5c_handle_data_cached(struct stripe_head *sh) | |
404 | { | |
405 | int i; | |
406 | ||
407 | for (i = sh->disks; i--; ) | |
408 | if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) { | |
409 | set_bit(R5_InJournal, &sh->dev[i].flags); | |
410 | clear_bit(R5_LOCKED, &sh->dev[i].flags); | |
411 | } | |
412 | clear_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
413 | } | |
414 | ||
415 | /* | |
416 | * this journal write must contain full parity, | |
417 | * it may also contain some data pages | |
418 | */ | |
419 | static void r5c_handle_parity_cached(struct stripe_head *sh) | |
420 | { | |
421 | int i; | |
422 | ||
423 | for (i = sh->disks; i--; ) | |
424 | if (test_bit(R5_InJournal, &sh->dev[i].flags)) | |
425 | set_bit(R5_Wantwrite, &sh->dev[i].flags); | |
2ded3703 SL |
426 | } |
427 | ||
428 | /* | |
429 | * Setting proper flags after writing (or flushing) data and/or parity to the | |
430 | * log device. This is called from r5l_log_endio() or r5l_log_flush_endio(). | |
431 | */ | |
432 | static void r5c_finish_cache_stripe(struct stripe_head *sh) | |
433 | { | |
434 | struct r5l_log *log = sh->raid_conf->log; | |
435 | ||
436 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { | |
437 | BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
438 | /* | |
439 | * Set R5_InJournal for parity dev[pd_idx]. This means | |
440 | * all data AND parity in the journal. For RAID 6, it is | |
441 | * NOT necessary to set the flag for dev[qd_idx], as the | |
442 | * two parities are written out together. | |
443 | */ | |
444 | set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
1e6d690b SL |
445 | } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) { |
446 | r5c_handle_data_cached(sh); | |
447 | } else { | |
448 | r5c_handle_parity_cached(sh); | |
449 | set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
450 | } | |
2ded3703 SL |
451 | } |
452 | ||
d8858f43 CH |
453 | static void r5l_io_run_stripes(struct r5l_io_unit *io) |
454 | { | |
455 | struct stripe_head *sh, *next; | |
456 | ||
457 | list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) { | |
458 | list_del_init(&sh->log_list); | |
2ded3703 SL |
459 | |
460 | r5c_finish_cache_stripe(sh); | |
461 | ||
d8858f43 CH |
462 | set_bit(STRIPE_HANDLE, &sh->state); |
463 | raid5_release_stripe(sh); | |
464 | } | |
465 | } | |
466 | ||
56fef7c6 CH |
467 | static void r5l_log_run_stripes(struct r5l_log *log) |
468 | { | |
469 | struct r5l_io_unit *io, *next; | |
470 | ||
471 | assert_spin_locked(&log->io_list_lock); | |
472 | ||
473 | list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { | |
474 | /* don't change list order */ | |
475 | if (io->state < IO_UNIT_IO_END) | |
476 | break; | |
477 | ||
478 | list_move_tail(&io->log_sibling, &log->finished_ios); | |
479 | r5l_io_run_stripes(io); | |
480 | } | |
481 | } | |
482 | ||
3848c0bc CH |
483 | static void r5l_move_to_end_ios(struct r5l_log *log) |
484 | { | |
485 | struct r5l_io_unit *io, *next; | |
486 | ||
487 | assert_spin_locked(&log->io_list_lock); | |
488 | ||
489 | list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { | |
490 | /* don't change list order */ | |
491 | if (io->state < IO_UNIT_IO_END) | |
492 | break; | |
493 | list_move_tail(&io->log_sibling, &log->io_end_ios); | |
494 | } | |
495 | } | |
496 | ||
f6bed0ef SL |
497 | static void r5l_log_endio(struct bio *bio) |
498 | { | |
499 | struct r5l_io_unit *io = bio->bi_private; | |
500 | struct r5l_log *log = io->log; | |
509ffec7 | 501 | unsigned long flags; |
f6bed0ef | 502 | |
6e74a9cf SL |
503 | if (bio->bi_error) |
504 | md_error(log->rdev->mddev, log->rdev); | |
505 | ||
f6bed0ef | 506 | bio_put(bio); |
e8deb638 | 507 | mempool_free(io->meta_page, log->meta_pool); |
f6bed0ef | 508 | |
509ffec7 CH |
509 | spin_lock_irqsave(&log->io_list_lock, flags); |
510 | __r5l_set_io_unit_state(io, IO_UNIT_IO_END); | |
56fef7c6 | 511 | if (log->need_cache_flush) |
3848c0bc | 512 | r5l_move_to_end_ios(log); |
56fef7c6 CH |
513 | else |
514 | r5l_log_run_stripes(log); | |
509ffec7 CH |
515 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
516 | ||
56fef7c6 CH |
517 | if (log->need_cache_flush) |
518 | md_wakeup_thread(log->rdev->mddev->thread); | |
f6bed0ef SL |
519 | } |
520 | ||
521 | static void r5l_submit_current_io(struct r5l_log *log) | |
522 | { | |
523 | struct r5l_io_unit *io = log->current_io; | |
524 | struct r5l_meta_block *block; | |
509ffec7 | 525 | unsigned long flags; |
f6bed0ef SL |
526 | u32 crc; |
527 | ||
528 | if (!io) | |
529 | return; | |
530 | ||
531 | block = page_address(io->meta_page); | |
532 | block->meta_size = cpu_to_le32(io->meta_offset); | |
5cb2fbd6 | 533 | crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE); |
f6bed0ef SL |
534 | block->checksum = cpu_to_le32(crc); |
535 | ||
536 | log->current_io = NULL; | |
509ffec7 CH |
537 | spin_lock_irqsave(&log->io_list_lock, flags); |
538 | __r5l_set_io_unit_state(io, IO_UNIT_IO_START); | |
539 | spin_unlock_irqrestore(&log->io_list_lock, flags); | |
f6bed0ef | 540 | |
4e49ea4a | 541 | submit_bio(io->current_bio); |
f6bed0ef SL |
542 | } |
543 | ||
6143e2ce | 544 | static struct bio *r5l_bio_alloc(struct r5l_log *log) |
b349feb3 | 545 | { |
c38d29b3 | 546 | struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs); |
b349feb3 | 547 | |
796a5cf0 | 548 | bio_set_op_attrs(bio, REQ_OP_WRITE, 0); |
b349feb3 | 549 | bio->bi_bdev = log->rdev->bdev; |
1e932a37 | 550 | bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start; |
b349feb3 | 551 | |
b349feb3 CH |
552 | return bio; |
553 | } | |
554 | ||
c1b99198 CH |
555 | static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io) |
556 | { | |
557 | log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS); | |
558 | ||
a39f7afd | 559 | r5c_update_log_state(log); |
c1b99198 CH |
560 | /* |
561 | * If we filled up the log device start from the beginning again, | |
562 | * which will require a new bio. | |
563 | * | |
564 | * Note: for this to work properly the log size needs to me a multiple | |
565 | * of BLOCK_SECTORS. | |
566 | */ | |
567 | if (log->log_start == 0) | |
6143e2ce | 568 | io->need_split_bio = true; |
c1b99198 CH |
569 | |
570 | io->log_end = log->log_start; | |
571 | } | |
572 | ||
f6bed0ef SL |
573 | static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log) |
574 | { | |
575 | struct r5l_io_unit *io; | |
576 | struct r5l_meta_block *block; | |
f6bed0ef | 577 | |
5036c390 CH |
578 | io = mempool_alloc(log->io_pool, GFP_ATOMIC); |
579 | if (!io) | |
580 | return NULL; | |
581 | memset(io, 0, sizeof(*io)); | |
582 | ||
51039cd0 | 583 | io->log = log; |
51039cd0 CH |
584 | INIT_LIST_HEAD(&io->log_sibling); |
585 | INIT_LIST_HEAD(&io->stripe_list); | |
586 | io->state = IO_UNIT_RUNNING; | |
f6bed0ef | 587 | |
e8deb638 | 588 | io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO); |
f6bed0ef | 589 | block = page_address(io->meta_page); |
e8deb638 | 590 | clear_page(block); |
f6bed0ef SL |
591 | block->magic = cpu_to_le32(R5LOG_MAGIC); |
592 | block->version = R5LOG_VERSION; | |
593 | block->seq = cpu_to_le64(log->seq); | |
594 | block->position = cpu_to_le64(log->log_start); | |
595 | ||
596 | io->log_start = log->log_start; | |
597 | io->meta_offset = sizeof(struct r5l_meta_block); | |
2b8ef16e | 598 | io->seq = log->seq++; |
f6bed0ef | 599 | |
6143e2ce CH |
600 | io->current_bio = r5l_bio_alloc(log); |
601 | io->current_bio->bi_end_io = r5l_log_endio; | |
602 | io->current_bio->bi_private = io; | |
b349feb3 | 603 | bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0); |
f6bed0ef | 604 | |
c1b99198 | 605 | r5_reserve_log_entry(log, io); |
f6bed0ef SL |
606 | |
607 | spin_lock_irq(&log->io_list_lock); | |
608 | list_add_tail(&io->log_sibling, &log->running_ios); | |
609 | spin_unlock_irq(&log->io_list_lock); | |
610 | ||
611 | return io; | |
612 | } | |
613 | ||
614 | static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size) | |
615 | { | |
22581f58 CH |
616 | if (log->current_io && |
617 | log->current_io->meta_offset + payload_size > PAGE_SIZE) | |
f6bed0ef | 618 | r5l_submit_current_io(log); |
f6bed0ef | 619 | |
5036c390 | 620 | if (!log->current_io) { |
22581f58 | 621 | log->current_io = r5l_new_meta(log); |
5036c390 CH |
622 | if (!log->current_io) |
623 | return -ENOMEM; | |
624 | } | |
625 | ||
f6bed0ef SL |
626 | return 0; |
627 | } | |
628 | ||
629 | static void r5l_append_payload_meta(struct r5l_log *log, u16 type, | |
630 | sector_t location, | |
631 | u32 checksum1, u32 checksum2, | |
632 | bool checksum2_valid) | |
633 | { | |
634 | struct r5l_io_unit *io = log->current_io; | |
635 | struct r5l_payload_data_parity *payload; | |
636 | ||
637 | payload = page_address(io->meta_page) + io->meta_offset; | |
638 | payload->header.type = cpu_to_le16(type); | |
639 | payload->header.flags = cpu_to_le16(0); | |
640 | payload->size = cpu_to_le32((1 + !!checksum2_valid) << | |
641 | (PAGE_SHIFT - 9)); | |
642 | payload->location = cpu_to_le64(location); | |
643 | payload->checksum[0] = cpu_to_le32(checksum1); | |
644 | if (checksum2_valid) | |
645 | payload->checksum[1] = cpu_to_le32(checksum2); | |
646 | ||
647 | io->meta_offset += sizeof(struct r5l_payload_data_parity) + | |
648 | sizeof(__le32) * (1 + !!checksum2_valid); | |
649 | } | |
650 | ||
651 | static void r5l_append_payload_page(struct r5l_log *log, struct page *page) | |
652 | { | |
653 | struct r5l_io_unit *io = log->current_io; | |
654 | ||
6143e2ce CH |
655 | if (io->need_split_bio) { |
656 | struct bio *prev = io->current_bio; | |
b349feb3 | 657 | |
6143e2ce CH |
658 | io->current_bio = r5l_bio_alloc(log); |
659 | bio_chain(io->current_bio, prev); | |
660 | ||
4e49ea4a | 661 | submit_bio(prev); |
f6bed0ef | 662 | } |
f6bed0ef | 663 | |
6143e2ce CH |
664 | if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) |
665 | BUG(); | |
666 | ||
c1b99198 | 667 | r5_reserve_log_entry(log, io); |
f6bed0ef SL |
668 | } |
669 | ||
5036c390 | 670 | static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh, |
f6bed0ef SL |
671 | int data_pages, int parity_pages) |
672 | { | |
673 | int i; | |
674 | int meta_size; | |
5036c390 | 675 | int ret; |
f6bed0ef SL |
676 | struct r5l_io_unit *io; |
677 | ||
678 | meta_size = | |
679 | ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) | |
680 | * data_pages) + | |
681 | sizeof(struct r5l_payload_data_parity) + | |
682 | sizeof(__le32) * parity_pages; | |
683 | ||
5036c390 CH |
684 | ret = r5l_get_meta(log, meta_size); |
685 | if (ret) | |
686 | return ret; | |
687 | ||
f6bed0ef SL |
688 | io = log->current_io; |
689 | ||
690 | for (i = 0; i < sh->disks; i++) { | |
1e6d690b SL |
691 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || |
692 | test_bit(R5_InJournal, &sh->dev[i].flags)) | |
f6bed0ef SL |
693 | continue; |
694 | if (i == sh->pd_idx || i == sh->qd_idx) | |
695 | continue; | |
696 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA, | |
697 | raid5_compute_blocknr(sh, i, 0), | |
698 | sh->dev[i].log_checksum, 0, false); | |
699 | r5l_append_payload_page(log, sh->dev[i].page); | |
700 | } | |
701 | ||
2ded3703 | 702 | if (parity_pages == 2) { |
f6bed0ef SL |
703 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, |
704 | sh->sector, sh->dev[sh->pd_idx].log_checksum, | |
705 | sh->dev[sh->qd_idx].log_checksum, true); | |
706 | r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); | |
707 | r5l_append_payload_page(log, sh->dev[sh->qd_idx].page); | |
2ded3703 | 708 | } else if (parity_pages == 1) { |
f6bed0ef SL |
709 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, |
710 | sh->sector, sh->dev[sh->pd_idx].log_checksum, | |
711 | 0, false); | |
712 | r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); | |
2ded3703 SL |
713 | } else /* Just writing data, not parity, in caching phase */ |
714 | BUG_ON(parity_pages != 0); | |
f6bed0ef SL |
715 | |
716 | list_add_tail(&sh->log_list, &io->stripe_list); | |
717 | atomic_inc(&io->pending_stripe); | |
718 | sh->log_io = io; | |
5036c390 | 719 | |
a39f7afd SL |
720 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) |
721 | return 0; | |
722 | ||
723 | if (sh->log_start == MaxSector) { | |
724 | BUG_ON(!list_empty(&sh->r5c)); | |
725 | sh->log_start = io->log_start; | |
726 | spin_lock_irq(&log->stripe_in_journal_lock); | |
727 | list_add_tail(&sh->r5c, | |
728 | &log->stripe_in_journal_list); | |
729 | spin_unlock_irq(&log->stripe_in_journal_lock); | |
730 | atomic_inc(&log->stripe_in_journal_count); | |
731 | } | |
5036c390 | 732 | return 0; |
f6bed0ef SL |
733 | } |
734 | ||
a39f7afd SL |
735 | /* add stripe to no_space_stripes, and then wake up reclaim */ |
736 | static inline void r5l_add_no_space_stripe(struct r5l_log *log, | |
737 | struct stripe_head *sh) | |
738 | { | |
739 | spin_lock(&log->no_space_stripes_lock); | |
740 | list_add_tail(&sh->log_list, &log->no_space_stripes); | |
741 | spin_unlock(&log->no_space_stripes_lock); | |
742 | } | |
743 | ||
f6bed0ef SL |
744 | /* |
745 | * running in raid5d, where reclaim could wait for raid5d too (when it flushes | |
746 | * data from log to raid disks), so we shouldn't wait for reclaim here | |
747 | */ | |
748 | int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh) | |
749 | { | |
a39f7afd | 750 | struct r5conf *conf = sh->raid_conf; |
f6bed0ef SL |
751 | int write_disks = 0; |
752 | int data_pages, parity_pages; | |
f6bed0ef SL |
753 | int reserve; |
754 | int i; | |
5036c390 | 755 | int ret = 0; |
a39f7afd | 756 | bool wake_reclaim = false; |
f6bed0ef SL |
757 | |
758 | if (!log) | |
759 | return -EAGAIN; | |
760 | /* Don't support stripe batch */ | |
761 | if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) || | |
762 | test_bit(STRIPE_SYNCING, &sh->state)) { | |
763 | /* the stripe is written to log, we start writing it to raid */ | |
764 | clear_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
765 | return -EAGAIN; | |
766 | } | |
767 | ||
2ded3703 SL |
768 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); |
769 | ||
f6bed0ef SL |
770 | for (i = 0; i < sh->disks; i++) { |
771 | void *addr; | |
772 | ||
1e6d690b SL |
773 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || |
774 | test_bit(R5_InJournal, &sh->dev[i].flags)) | |
f6bed0ef | 775 | continue; |
1e6d690b | 776 | |
f6bed0ef SL |
777 | write_disks++; |
778 | /* checksum is already calculated in last run */ | |
779 | if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) | |
780 | continue; | |
781 | addr = kmap_atomic(sh->dev[i].page); | |
5cb2fbd6 SL |
782 | sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, |
783 | addr, PAGE_SIZE); | |
f6bed0ef SL |
784 | kunmap_atomic(addr); |
785 | } | |
786 | parity_pages = 1 + !!(sh->qd_idx >= 0); | |
787 | data_pages = write_disks - parity_pages; | |
788 | ||
f6bed0ef | 789 | set_bit(STRIPE_LOG_TRAPPED, &sh->state); |
253f9fd4 SL |
790 | /* |
791 | * The stripe must enter state machine again to finish the write, so | |
792 | * don't delay. | |
793 | */ | |
794 | clear_bit(STRIPE_DELAYED, &sh->state); | |
f6bed0ef SL |
795 | atomic_inc(&sh->count); |
796 | ||
797 | mutex_lock(&log->io_mutex); | |
798 | /* meta + data */ | |
799 | reserve = (1 + write_disks) << (PAGE_SHIFT - 9); | |
f6bed0ef | 800 | |
a39f7afd SL |
801 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { |
802 | if (!r5l_has_free_space(log, reserve)) { | |
803 | r5l_add_no_space_stripe(log, sh); | |
804 | wake_reclaim = true; | |
805 | } else { | |
806 | ret = r5l_log_stripe(log, sh, data_pages, parity_pages); | |
807 | if (ret) { | |
808 | spin_lock_irq(&log->io_list_lock); | |
809 | list_add_tail(&sh->log_list, | |
810 | &log->no_mem_stripes); | |
811 | spin_unlock_irq(&log->io_list_lock); | |
812 | } | |
813 | } | |
814 | } else { /* R5C_JOURNAL_MODE_WRITE_BACK */ | |
815 | /* | |
816 | * log space critical, do not process stripes that are | |
817 | * not in cache yet (sh->log_start == MaxSector). | |
818 | */ | |
819 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && | |
820 | sh->log_start == MaxSector) { | |
821 | r5l_add_no_space_stripe(log, sh); | |
822 | wake_reclaim = true; | |
823 | reserve = 0; | |
824 | } else if (!r5l_has_free_space(log, reserve)) { | |
825 | if (sh->log_start == log->last_checkpoint) | |
826 | BUG(); | |
827 | else | |
828 | r5l_add_no_space_stripe(log, sh); | |
829 | } else { | |
830 | ret = r5l_log_stripe(log, sh, data_pages, parity_pages); | |
831 | if (ret) { | |
832 | spin_lock_irq(&log->io_list_lock); | |
833 | list_add_tail(&sh->log_list, | |
834 | &log->no_mem_stripes); | |
835 | spin_unlock_irq(&log->io_list_lock); | |
836 | } | |
5036c390 | 837 | } |
f6bed0ef | 838 | } |
f6bed0ef | 839 | |
5036c390 | 840 | mutex_unlock(&log->io_mutex); |
a39f7afd SL |
841 | if (wake_reclaim) |
842 | r5l_wake_reclaim(log, reserve); | |
f6bed0ef SL |
843 | return 0; |
844 | } | |
845 | ||
846 | void r5l_write_stripe_run(struct r5l_log *log) | |
847 | { | |
848 | if (!log) | |
849 | return; | |
850 | mutex_lock(&log->io_mutex); | |
851 | r5l_submit_current_io(log); | |
852 | mutex_unlock(&log->io_mutex); | |
853 | } | |
854 | ||
828cbe98 SL |
855 | int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio) |
856 | { | |
857 | if (!log) | |
858 | return -ENODEV; | |
859 | /* | |
860 | * we flush log disk cache first, then write stripe data to raid disks. | |
861 | * So if bio is finished, the log disk cache is flushed already. The | |
862 | * recovery guarantees we can recovery the bio from log disk, so we | |
863 | * don't need to flush again | |
864 | */ | |
865 | if (bio->bi_iter.bi_size == 0) { | |
866 | bio_endio(bio); | |
867 | return 0; | |
868 | } | |
1eff9d32 | 869 | bio->bi_opf &= ~REQ_PREFLUSH; |
828cbe98 SL |
870 | return -EAGAIN; |
871 | } | |
872 | ||
f6bed0ef SL |
873 | /* This will run after log space is reclaimed */ |
874 | static void r5l_run_no_space_stripes(struct r5l_log *log) | |
875 | { | |
876 | struct stripe_head *sh; | |
877 | ||
878 | spin_lock(&log->no_space_stripes_lock); | |
879 | while (!list_empty(&log->no_space_stripes)) { | |
880 | sh = list_first_entry(&log->no_space_stripes, | |
881 | struct stripe_head, log_list); | |
882 | list_del_init(&sh->log_list); | |
883 | set_bit(STRIPE_HANDLE, &sh->state); | |
884 | raid5_release_stripe(sh); | |
885 | } | |
886 | spin_unlock(&log->no_space_stripes_lock); | |
887 | } | |
888 | ||
a39f7afd SL |
889 | /* |
890 | * calculate new last_checkpoint | |
891 | * for write through mode, returns log->next_checkpoint | |
892 | * for write back, returns log_start of first sh in stripe_in_journal_list | |
893 | */ | |
894 | static sector_t r5c_calculate_new_cp(struct r5conf *conf) | |
895 | { | |
896 | struct stripe_head *sh; | |
897 | struct r5l_log *log = conf->log; | |
898 | sector_t new_cp; | |
899 | unsigned long flags; | |
900 | ||
901 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) | |
902 | return log->next_checkpoint; | |
903 | ||
904 | spin_lock_irqsave(&log->stripe_in_journal_lock, flags); | |
905 | if (list_empty(&conf->log->stripe_in_journal_list)) { | |
906 | /* all stripes flushed */ | |
907 | spin_unlock(&log->stripe_in_journal_lock); | |
908 | return log->next_checkpoint; | |
909 | } | |
910 | sh = list_first_entry(&conf->log->stripe_in_journal_list, | |
911 | struct stripe_head, r5c); | |
912 | new_cp = sh->log_start; | |
913 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); | |
914 | return new_cp; | |
915 | } | |
916 | ||
17036461 CH |
917 | static sector_t r5l_reclaimable_space(struct r5l_log *log) |
918 | { | |
a39f7afd SL |
919 | struct r5conf *conf = log->rdev->mddev->private; |
920 | ||
17036461 | 921 | return r5l_ring_distance(log, log->last_checkpoint, |
a39f7afd | 922 | r5c_calculate_new_cp(conf)); |
17036461 CH |
923 | } |
924 | ||
5036c390 CH |
925 | static void r5l_run_no_mem_stripe(struct r5l_log *log) |
926 | { | |
927 | struct stripe_head *sh; | |
928 | ||
929 | assert_spin_locked(&log->io_list_lock); | |
930 | ||
931 | if (!list_empty(&log->no_mem_stripes)) { | |
932 | sh = list_first_entry(&log->no_mem_stripes, | |
933 | struct stripe_head, log_list); | |
934 | list_del_init(&sh->log_list); | |
935 | set_bit(STRIPE_HANDLE, &sh->state); | |
936 | raid5_release_stripe(sh); | |
937 | } | |
938 | } | |
939 | ||
04732f74 | 940 | static bool r5l_complete_finished_ios(struct r5l_log *log) |
17036461 CH |
941 | { |
942 | struct r5l_io_unit *io, *next; | |
943 | bool found = false; | |
944 | ||
945 | assert_spin_locked(&log->io_list_lock); | |
946 | ||
04732f74 | 947 | list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) { |
17036461 CH |
948 | /* don't change list order */ |
949 | if (io->state < IO_UNIT_STRIPE_END) | |
950 | break; | |
951 | ||
952 | log->next_checkpoint = io->log_start; | |
953 | log->next_cp_seq = io->seq; | |
954 | ||
955 | list_del(&io->log_sibling); | |
5036c390 CH |
956 | mempool_free(io, log->io_pool); |
957 | r5l_run_no_mem_stripe(log); | |
17036461 CH |
958 | |
959 | found = true; | |
960 | } | |
961 | ||
962 | return found; | |
963 | } | |
964 | ||
509ffec7 CH |
965 | static void __r5l_stripe_write_finished(struct r5l_io_unit *io) |
966 | { | |
967 | struct r5l_log *log = io->log; | |
a39f7afd | 968 | struct r5conf *conf = log->rdev->mddev->private; |
509ffec7 CH |
969 | unsigned long flags; |
970 | ||
971 | spin_lock_irqsave(&log->io_list_lock, flags); | |
972 | __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END); | |
17036461 | 973 | |
04732f74 | 974 | if (!r5l_complete_finished_ios(log)) { |
85f2f9a4 SL |
975 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
976 | return; | |
977 | } | |
509ffec7 | 978 | |
a39f7afd SL |
979 | if (r5l_reclaimable_space(log) > log->max_free_space || |
980 | test_bit(R5C_LOG_TIGHT, &conf->cache_state)) | |
509ffec7 CH |
981 | r5l_wake_reclaim(log, 0); |
982 | ||
509ffec7 CH |
983 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
984 | wake_up(&log->iounit_wait); | |
985 | } | |
986 | ||
0576b1c6 SL |
987 | void r5l_stripe_write_finished(struct stripe_head *sh) |
988 | { | |
989 | struct r5l_io_unit *io; | |
990 | ||
0576b1c6 | 991 | io = sh->log_io; |
0576b1c6 SL |
992 | sh->log_io = NULL; |
993 | ||
509ffec7 CH |
994 | if (io && atomic_dec_and_test(&io->pending_stripe)) |
995 | __r5l_stripe_write_finished(io); | |
0576b1c6 SL |
996 | } |
997 | ||
a8c34f91 SL |
998 | static void r5l_log_flush_endio(struct bio *bio) |
999 | { | |
1000 | struct r5l_log *log = container_of(bio, struct r5l_log, | |
1001 | flush_bio); | |
1002 | unsigned long flags; | |
1003 | struct r5l_io_unit *io; | |
a8c34f91 | 1004 | |
6e74a9cf SL |
1005 | if (bio->bi_error) |
1006 | md_error(log->rdev->mddev, log->rdev); | |
1007 | ||
a8c34f91 | 1008 | spin_lock_irqsave(&log->io_list_lock, flags); |
d8858f43 CH |
1009 | list_for_each_entry(io, &log->flushing_ios, log_sibling) |
1010 | r5l_io_run_stripes(io); | |
04732f74 | 1011 | list_splice_tail_init(&log->flushing_ios, &log->finished_ios); |
a8c34f91 SL |
1012 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
1013 | } | |
1014 | ||
0576b1c6 SL |
1015 | /* |
1016 | * Starting dispatch IO to raid. | |
1017 | * io_unit(meta) consists of a log. There is one situation we want to avoid. A | |
1018 | * broken meta in the middle of a log causes recovery can't find meta at the | |
1019 | * head of log. If operations require meta at the head persistent in log, we | |
1020 | * must make sure meta before it persistent in log too. A case is: | |
1021 | * | |
1022 | * stripe data/parity is in log, we start write stripe to raid disks. stripe | |
1023 | * data/parity must be persistent in log before we do the write to raid disks. | |
1024 | * | |
1025 | * The solution is we restrictly maintain io_unit list order. In this case, we | |
1026 | * only write stripes of an io_unit to raid disks till the io_unit is the first | |
1027 | * one whose data/parity is in log. | |
1028 | */ | |
1029 | void r5l_flush_stripe_to_raid(struct r5l_log *log) | |
1030 | { | |
a8c34f91 | 1031 | bool do_flush; |
56fef7c6 CH |
1032 | |
1033 | if (!log || !log->need_cache_flush) | |
0576b1c6 | 1034 | return; |
0576b1c6 SL |
1035 | |
1036 | spin_lock_irq(&log->io_list_lock); | |
a8c34f91 SL |
1037 | /* flush bio is running */ |
1038 | if (!list_empty(&log->flushing_ios)) { | |
1039 | spin_unlock_irq(&log->io_list_lock); | |
1040 | return; | |
0576b1c6 | 1041 | } |
a8c34f91 SL |
1042 | list_splice_tail_init(&log->io_end_ios, &log->flushing_ios); |
1043 | do_flush = !list_empty(&log->flushing_ios); | |
0576b1c6 | 1044 | spin_unlock_irq(&log->io_list_lock); |
a8c34f91 SL |
1045 | |
1046 | if (!do_flush) | |
1047 | return; | |
1048 | bio_reset(&log->flush_bio); | |
1049 | log->flush_bio.bi_bdev = log->rdev->bdev; | |
1050 | log->flush_bio.bi_end_io = r5l_log_flush_endio; | |
796a5cf0 | 1051 | bio_set_op_attrs(&log->flush_bio, REQ_OP_WRITE, WRITE_FLUSH); |
4e49ea4a | 1052 | submit_bio(&log->flush_bio); |
0576b1c6 SL |
1053 | } |
1054 | ||
0576b1c6 | 1055 | static void r5l_write_super(struct r5l_log *log, sector_t cp); |
4b482044 SL |
1056 | static void r5l_write_super_and_discard_space(struct r5l_log *log, |
1057 | sector_t end) | |
1058 | { | |
1059 | struct block_device *bdev = log->rdev->bdev; | |
1060 | struct mddev *mddev; | |
1061 | ||
1062 | r5l_write_super(log, end); | |
1063 | ||
1064 | if (!blk_queue_discard(bdev_get_queue(bdev))) | |
1065 | return; | |
1066 | ||
1067 | mddev = log->rdev->mddev; | |
1068 | /* | |
8e018c21 SL |
1069 | * Discard could zero data, so before discard we must make sure |
1070 | * superblock is updated to new log tail. Updating superblock (either | |
1071 | * directly call md_update_sb() or depend on md thread) must hold | |
1072 | * reconfig mutex. On the other hand, raid5_quiesce is called with | |
1073 | * reconfig_mutex hold. The first step of raid5_quiesce() is waitting | |
1074 | * for all IO finish, hence waitting for reclaim thread, while reclaim | |
1075 | * thread is calling this function and waitting for reconfig mutex. So | |
1076 | * there is a deadlock. We workaround this issue with a trylock. | |
1077 | * FIXME: we could miss discard if we can't take reconfig mutex | |
4b482044 | 1078 | */ |
8e018c21 SL |
1079 | set_mask_bits(&mddev->flags, 0, |
1080 | BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING)); | |
1081 | if (!mddev_trylock(mddev)) | |
1082 | return; | |
1083 | md_update_sb(mddev, 1); | |
1084 | mddev_unlock(mddev); | |
4b482044 | 1085 | |
6e74a9cf | 1086 | /* discard IO error really doesn't matter, ignore it */ |
4b482044 SL |
1087 | if (log->last_checkpoint < end) { |
1088 | blkdev_issue_discard(bdev, | |
1089 | log->last_checkpoint + log->rdev->data_offset, | |
1090 | end - log->last_checkpoint, GFP_NOIO, 0); | |
1091 | } else { | |
1092 | blkdev_issue_discard(bdev, | |
1093 | log->last_checkpoint + log->rdev->data_offset, | |
1094 | log->device_size - log->last_checkpoint, | |
1095 | GFP_NOIO, 0); | |
1096 | blkdev_issue_discard(bdev, log->rdev->data_offset, end, | |
1097 | GFP_NOIO, 0); | |
1098 | } | |
1099 | } | |
1100 | ||
a39f7afd SL |
1101 | /* |
1102 | * r5c_flush_stripe moves stripe from cached list to handle_list. When called, | |
1103 | * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes. | |
1104 | * | |
1105 | * must hold conf->device_lock | |
1106 | */ | |
1107 | static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh) | |
1108 | { | |
1109 | BUG_ON(list_empty(&sh->lru)); | |
1110 | BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
1111 | BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); | |
1112 | ||
1113 | /* | |
1114 | * The stripe is not ON_RELEASE_LIST, so it is safe to call | |
1115 | * raid5_release_stripe() while holding conf->device_lock | |
1116 | */ | |
1117 | BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state)); | |
1118 | assert_spin_locked(&conf->device_lock); | |
1119 | ||
1120 | list_del_init(&sh->lru); | |
1121 | atomic_inc(&sh->count); | |
1122 | ||
1123 | set_bit(STRIPE_HANDLE, &sh->state); | |
1124 | atomic_inc(&conf->active_stripes); | |
1125 | r5c_make_stripe_write_out(sh); | |
1126 | ||
1127 | if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) | |
1128 | atomic_inc(&conf->preread_active_stripes); | |
1129 | raid5_release_stripe(sh); | |
1130 | } | |
1131 | ||
1132 | /* | |
1133 | * if num == 0, flush all full stripes | |
1134 | * if num > 0, flush all full stripes. If less than num full stripes are | |
1135 | * flushed, flush some partial stripes until totally num stripes are | |
1136 | * flushed or there is no more cached stripes. | |
1137 | */ | |
1138 | void r5c_flush_cache(struct r5conf *conf, int num) | |
1139 | { | |
1140 | int count; | |
1141 | struct stripe_head *sh, *next; | |
1142 | ||
1143 | assert_spin_locked(&conf->device_lock); | |
1144 | if (!conf->log) | |
1145 | return; | |
1146 | ||
1147 | count = 0; | |
1148 | list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) { | |
1149 | r5c_flush_stripe(conf, sh); | |
1150 | count++; | |
1151 | } | |
1152 | ||
1153 | if (count >= num) | |
1154 | return; | |
1155 | list_for_each_entry_safe(sh, next, | |
1156 | &conf->r5c_partial_stripe_list, lru) { | |
1157 | r5c_flush_stripe(conf, sh); | |
1158 | if (++count >= num) | |
1159 | break; | |
1160 | } | |
1161 | } | |
1162 | ||
1163 | static void r5c_do_reclaim(struct r5conf *conf) | |
1164 | { | |
1165 | struct r5l_log *log = conf->log; | |
1166 | struct stripe_head *sh; | |
1167 | int count = 0; | |
1168 | unsigned long flags; | |
1169 | int total_cached; | |
1170 | int stripes_to_flush; | |
1171 | ||
1172 | if (!r5c_is_writeback(log)) | |
1173 | return; | |
1174 | ||
1175 | total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + | |
1176 | atomic_read(&conf->r5c_cached_full_stripes); | |
1177 | ||
1178 | if (total_cached > conf->min_nr_stripes * 3 / 4 || | |
1179 | atomic_read(&conf->empty_inactive_list_nr) > 0) | |
1180 | /* | |
1181 | * if stripe cache pressure high, flush all full stripes and | |
1182 | * some partial stripes | |
1183 | */ | |
1184 | stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP; | |
1185 | else if (total_cached > conf->min_nr_stripes * 1 / 2 || | |
1186 | atomic_read(&conf->r5c_cached_full_stripes) > | |
1187 | R5C_FULL_STRIPE_FLUSH_BATCH) | |
1188 | /* | |
1189 | * if stripe cache pressure moderate, or if there is many full | |
1190 | * stripes,flush all full stripes | |
1191 | */ | |
1192 | stripes_to_flush = 0; | |
1193 | else | |
1194 | /* no need to flush */ | |
1195 | stripes_to_flush = -1; | |
1196 | ||
1197 | if (stripes_to_flush >= 0) { | |
1198 | spin_lock_irqsave(&conf->device_lock, flags); | |
1199 | r5c_flush_cache(conf, stripes_to_flush); | |
1200 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1201 | } | |
1202 | ||
1203 | /* if log space is tight, flush stripes on stripe_in_journal_list */ | |
1204 | if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) { | |
1205 | spin_lock_irqsave(&log->stripe_in_journal_lock, flags); | |
1206 | spin_lock(&conf->device_lock); | |
1207 | list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) { | |
1208 | /* | |
1209 | * stripes on stripe_in_journal_list could be in any | |
1210 | * state of the stripe_cache state machine. In this | |
1211 | * case, we only want to flush stripe on | |
1212 | * r5c_cached_full/partial_stripes. The following | |
1213 | * condition makes sure the stripe is on one of the | |
1214 | * two lists. | |
1215 | */ | |
1216 | if (!list_empty(&sh->lru) && | |
1217 | !test_bit(STRIPE_HANDLE, &sh->state) && | |
1218 | atomic_read(&sh->count) == 0) { | |
1219 | r5c_flush_stripe(conf, sh); | |
1220 | } | |
1221 | if (count++ >= R5C_RECLAIM_STRIPE_GROUP) | |
1222 | break; | |
1223 | } | |
1224 | spin_unlock(&conf->device_lock); | |
1225 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); | |
1226 | } | |
1227 | md_wakeup_thread(conf->mddev->thread); | |
1228 | } | |
1229 | ||
0576b1c6 SL |
1230 | static void r5l_do_reclaim(struct r5l_log *log) |
1231 | { | |
a39f7afd | 1232 | struct r5conf *conf = log->rdev->mddev->private; |
0576b1c6 | 1233 | sector_t reclaim_target = xchg(&log->reclaim_target, 0); |
17036461 CH |
1234 | sector_t reclaimable; |
1235 | sector_t next_checkpoint; | |
a39f7afd | 1236 | bool write_super; |
0576b1c6 SL |
1237 | |
1238 | spin_lock_irq(&log->io_list_lock); | |
a39f7afd SL |
1239 | write_super = r5l_reclaimable_space(log) > log->max_free_space || |
1240 | reclaim_target != 0 || !list_empty(&log->no_space_stripes); | |
0576b1c6 SL |
1241 | /* |
1242 | * move proper io_unit to reclaim list. We should not change the order. | |
1243 | * reclaimable/unreclaimable io_unit can be mixed in the list, we | |
1244 | * shouldn't reuse space of an unreclaimable io_unit | |
1245 | */ | |
1246 | while (1) { | |
17036461 CH |
1247 | reclaimable = r5l_reclaimable_space(log); |
1248 | if (reclaimable >= reclaim_target || | |
0576b1c6 SL |
1249 | (list_empty(&log->running_ios) && |
1250 | list_empty(&log->io_end_ios) && | |
a8c34f91 | 1251 | list_empty(&log->flushing_ios) && |
04732f74 | 1252 | list_empty(&log->finished_ios))) |
0576b1c6 SL |
1253 | break; |
1254 | ||
17036461 CH |
1255 | md_wakeup_thread(log->rdev->mddev->thread); |
1256 | wait_event_lock_irq(log->iounit_wait, | |
1257 | r5l_reclaimable_space(log) > reclaimable, | |
1258 | log->io_list_lock); | |
0576b1c6 | 1259 | } |
17036461 | 1260 | |
a39f7afd | 1261 | next_checkpoint = r5c_calculate_new_cp(conf); |
0576b1c6 SL |
1262 | spin_unlock_irq(&log->io_list_lock); |
1263 | ||
17036461 | 1264 | BUG_ON(reclaimable < 0); |
a39f7afd SL |
1265 | |
1266 | if (reclaimable == 0 || !write_super) | |
0576b1c6 SL |
1267 | return; |
1268 | ||
0576b1c6 SL |
1269 | /* |
1270 | * write_super will flush cache of each raid disk. We must write super | |
1271 | * here, because the log area might be reused soon and we don't want to | |
1272 | * confuse recovery | |
1273 | */ | |
4b482044 | 1274 | r5l_write_super_and_discard_space(log, next_checkpoint); |
0576b1c6 SL |
1275 | |
1276 | mutex_lock(&log->io_mutex); | |
17036461 | 1277 | log->last_checkpoint = next_checkpoint; |
a39f7afd | 1278 | r5c_update_log_state(log); |
0576b1c6 | 1279 | mutex_unlock(&log->io_mutex); |
0576b1c6 | 1280 | |
17036461 | 1281 | r5l_run_no_space_stripes(log); |
0576b1c6 SL |
1282 | } |
1283 | ||
1284 | static void r5l_reclaim_thread(struct md_thread *thread) | |
1285 | { | |
1286 | struct mddev *mddev = thread->mddev; | |
1287 | struct r5conf *conf = mddev->private; | |
1288 | struct r5l_log *log = conf->log; | |
1289 | ||
1290 | if (!log) | |
1291 | return; | |
a39f7afd | 1292 | r5c_do_reclaim(conf); |
0576b1c6 SL |
1293 | r5l_do_reclaim(log); |
1294 | } | |
1295 | ||
a39f7afd | 1296 | void r5l_wake_reclaim(struct r5l_log *log, sector_t space) |
f6bed0ef | 1297 | { |
0576b1c6 SL |
1298 | unsigned long target; |
1299 | unsigned long new = (unsigned long)space; /* overflow in theory */ | |
1300 | ||
a39f7afd SL |
1301 | if (!log) |
1302 | return; | |
0576b1c6 SL |
1303 | do { |
1304 | target = log->reclaim_target; | |
1305 | if (new < target) | |
1306 | return; | |
1307 | } while (cmpxchg(&log->reclaim_target, target, new) != target); | |
1308 | md_wakeup_thread(log->reclaim_thread); | |
f6bed0ef SL |
1309 | } |
1310 | ||
e6c033f7 SL |
1311 | void r5l_quiesce(struct r5l_log *log, int state) |
1312 | { | |
4b482044 | 1313 | struct mddev *mddev; |
e6c033f7 SL |
1314 | if (!log || state == 2) |
1315 | return; | |
1316 | if (state == 0) { | |
16a43f6a SL |
1317 | /* |
1318 | * This is a special case for hotadd. In suspend, the array has | |
1319 | * no journal. In resume, journal is initialized as well as the | |
1320 | * reclaim thread. | |
1321 | */ | |
1322 | if (log->reclaim_thread) | |
1323 | return; | |
e6c033f7 SL |
1324 | log->reclaim_thread = md_register_thread(r5l_reclaim_thread, |
1325 | log->rdev->mddev, "reclaim"); | |
a39f7afd | 1326 | log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL; |
e6c033f7 | 1327 | } else if (state == 1) { |
4b482044 SL |
1328 | /* make sure r5l_write_super_and_discard_space exits */ |
1329 | mddev = log->rdev->mddev; | |
1330 | wake_up(&mddev->sb_wait); | |
a39f7afd | 1331 | r5l_wake_reclaim(log, MaxSector); |
e6c033f7 SL |
1332 | md_unregister_thread(&log->reclaim_thread); |
1333 | r5l_do_reclaim(log); | |
1334 | } | |
1335 | } | |
1336 | ||
6e74a9cf SL |
1337 | bool r5l_log_disk_error(struct r5conf *conf) |
1338 | { | |
f6b6ec5c SL |
1339 | struct r5l_log *log; |
1340 | bool ret; | |
7dde2ad3 | 1341 | /* don't allow write if journal disk is missing */ |
f6b6ec5c SL |
1342 | rcu_read_lock(); |
1343 | log = rcu_dereference(conf->log); | |
1344 | ||
1345 | if (!log) | |
1346 | ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags); | |
1347 | else | |
1348 | ret = test_bit(Faulty, &log->rdev->flags); | |
1349 | rcu_read_unlock(); | |
1350 | return ret; | |
6e74a9cf SL |
1351 | } |
1352 | ||
355810d1 SL |
1353 | struct r5l_recovery_ctx { |
1354 | struct page *meta_page; /* current meta */ | |
1355 | sector_t meta_total_blocks; /* total size of current meta and data */ | |
1356 | sector_t pos; /* recovery position */ | |
1357 | u64 seq; /* recovery position seq */ | |
b4c625c6 SL |
1358 | int data_parity_stripes; /* number of data_parity stripes */ |
1359 | int data_only_stripes; /* number of data_only stripes */ | |
1360 | struct list_head cached_list; | |
355810d1 SL |
1361 | }; |
1362 | ||
9ed988f5 SL |
1363 | static int r5l_recovery_read_meta_block(struct r5l_log *log, |
1364 | struct r5l_recovery_ctx *ctx) | |
355810d1 SL |
1365 | { |
1366 | struct page *page = ctx->meta_page; | |
1367 | struct r5l_meta_block *mb; | |
1368 | u32 crc, stored_crc; | |
1369 | ||
796a5cf0 MC |
1370 | if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, REQ_OP_READ, 0, |
1371 | false)) | |
355810d1 SL |
1372 | return -EIO; |
1373 | ||
1374 | mb = page_address(page); | |
1375 | stored_crc = le32_to_cpu(mb->checksum); | |
1376 | mb->checksum = 0; | |
1377 | ||
1378 | if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || | |
1379 | le64_to_cpu(mb->seq) != ctx->seq || | |
1380 | mb->version != R5LOG_VERSION || | |
1381 | le64_to_cpu(mb->position) != ctx->pos) | |
1382 | return -EINVAL; | |
1383 | ||
5cb2fbd6 | 1384 | crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); |
355810d1 SL |
1385 | if (stored_crc != crc) |
1386 | return -EINVAL; | |
1387 | ||
1388 | if (le32_to_cpu(mb->meta_size) > PAGE_SIZE) | |
1389 | return -EINVAL; | |
1390 | ||
1391 | ctx->meta_total_blocks = BLOCK_SECTORS; | |
1392 | ||
1393 | return 0; | |
1394 | } | |
1395 | ||
1396 | static int r5l_recovery_flush_one_stripe(struct r5l_log *log, | |
1397 | struct r5l_recovery_ctx *ctx, | |
1398 | sector_t stripe_sect, | |
3fd880af | 1399 | int *offset) |
355810d1 SL |
1400 | { |
1401 | struct r5conf *conf = log->rdev->mddev->private; | |
1402 | struct stripe_head *sh; | |
1403 | struct r5l_payload_data_parity *payload; | |
1404 | int disk_index; | |
1405 | ||
1406 | sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0); | |
1407 | while (1) { | |
3fd880af J |
1408 | sector_t log_offset = r5l_ring_add(log, ctx->pos, |
1409 | ctx->meta_total_blocks); | |
355810d1 SL |
1410 | payload = page_address(ctx->meta_page) + *offset; |
1411 | ||
1412 | if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) { | |
1413 | raid5_compute_sector(conf, | |
1414 | le64_to_cpu(payload->location), 0, | |
1415 | &disk_index, sh); | |
1416 | ||
3fd880af | 1417 | sync_page_io(log->rdev, log_offset, PAGE_SIZE, |
796a5cf0 MC |
1418 | sh->dev[disk_index].page, REQ_OP_READ, 0, |
1419 | false); | |
355810d1 SL |
1420 | sh->dev[disk_index].log_checksum = |
1421 | le32_to_cpu(payload->checksum[0]); | |
1422 | set_bit(R5_Wantwrite, &sh->dev[disk_index].flags); | |
355810d1 SL |
1423 | } else { |
1424 | disk_index = sh->pd_idx; | |
3fd880af | 1425 | sync_page_io(log->rdev, log_offset, PAGE_SIZE, |
796a5cf0 MC |
1426 | sh->dev[disk_index].page, REQ_OP_READ, 0, |
1427 | false); | |
355810d1 SL |
1428 | sh->dev[disk_index].log_checksum = |
1429 | le32_to_cpu(payload->checksum[0]); | |
1430 | set_bit(R5_Wantwrite, &sh->dev[disk_index].flags); | |
1431 | ||
1432 | if (sh->qd_idx >= 0) { | |
1433 | disk_index = sh->qd_idx; | |
1434 | sync_page_io(log->rdev, | |
3fd880af | 1435 | r5l_ring_add(log, log_offset, BLOCK_SECTORS), |
355810d1 | 1436 | PAGE_SIZE, sh->dev[disk_index].page, |
796a5cf0 | 1437 | REQ_OP_READ, 0, false); |
355810d1 SL |
1438 | sh->dev[disk_index].log_checksum = |
1439 | le32_to_cpu(payload->checksum[1]); | |
1440 | set_bit(R5_Wantwrite, | |
1441 | &sh->dev[disk_index].flags); | |
1442 | } | |
355810d1 SL |
1443 | } |
1444 | ||
3fd880af | 1445 | ctx->meta_total_blocks += le32_to_cpu(payload->size); |
355810d1 SL |
1446 | *offset += sizeof(struct r5l_payload_data_parity) + |
1447 | sizeof(__le32) * | |
1448 | (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); | |
1449 | if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) | |
1450 | break; | |
1451 | } | |
1452 | ||
1453 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { | |
1454 | void *addr; | |
1455 | u32 checksum; | |
1456 | ||
1457 | if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) | |
1458 | continue; | |
1459 | addr = kmap_atomic(sh->dev[disk_index].page); | |
5cb2fbd6 | 1460 | checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE); |
355810d1 SL |
1461 | kunmap_atomic(addr); |
1462 | if (checksum != sh->dev[disk_index].log_checksum) | |
1463 | goto error; | |
1464 | } | |
1465 | ||
1466 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { | |
1467 | struct md_rdev *rdev, *rrdev; | |
1468 | ||
1469 | if (!test_and_clear_bit(R5_Wantwrite, | |
1470 | &sh->dev[disk_index].flags)) | |
1471 | continue; | |
1472 | ||
1473 | /* in case device is broken */ | |
354b445b | 1474 | rcu_read_lock(); |
355810d1 | 1475 | rdev = rcu_dereference(conf->disks[disk_index].rdev); |
354b445b SL |
1476 | if (rdev) { |
1477 | atomic_inc(&rdev->nr_pending); | |
1478 | rcu_read_unlock(); | |
355810d1 | 1479 | sync_page_io(rdev, stripe_sect, PAGE_SIZE, |
796a5cf0 MC |
1480 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, |
1481 | false); | |
354b445b SL |
1482 | rdev_dec_pending(rdev, rdev->mddev); |
1483 | rcu_read_lock(); | |
1484 | } | |
355810d1 | 1485 | rrdev = rcu_dereference(conf->disks[disk_index].replacement); |
354b445b SL |
1486 | if (rrdev) { |
1487 | atomic_inc(&rrdev->nr_pending); | |
1488 | rcu_read_unlock(); | |
355810d1 | 1489 | sync_page_io(rrdev, stripe_sect, PAGE_SIZE, |
796a5cf0 MC |
1490 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, |
1491 | false); | |
354b445b SL |
1492 | rdev_dec_pending(rrdev, rrdev->mddev); |
1493 | rcu_read_lock(); | |
1494 | } | |
1495 | rcu_read_unlock(); | |
355810d1 SL |
1496 | } |
1497 | raid5_release_stripe(sh); | |
1498 | return 0; | |
1499 | ||
1500 | error: | |
1501 | for (disk_index = 0; disk_index < sh->disks; disk_index++) | |
1502 | sh->dev[disk_index].flags = 0; | |
1503 | raid5_release_stripe(sh); | |
1504 | return -EINVAL; | |
1505 | } | |
1506 | ||
1507 | static int r5l_recovery_flush_one_meta(struct r5l_log *log, | |
1508 | struct r5l_recovery_ctx *ctx) | |
1509 | { | |
1510 | struct r5conf *conf = log->rdev->mddev->private; | |
1511 | struct r5l_payload_data_parity *payload; | |
1512 | struct r5l_meta_block *mb; | |
1513 | int offset; | |
355810d1 SL |
1514 | sector_t stripe_sector; |
1515 | ||
1516 | mb = page_address(ctx->meta_page); | |
1517 | offset = sizeof(struct r5l_meta_block); | |
355810d1 SL |
1518 | |
1519 | while (offset < le32_to_cpu(mb->meta_size)) { | |
1520 | int dd; | |
1521 | ||
1522 | payload = (void *)mb + offset; | |
1523 | stripe_sector = raid5_compute_sector(conf, | |
1524 | le64_to_cpu(payload->location), 0, &dd, NULL); | |
1525 | if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector, | |
3fd880af | 1526 | &offset)) |
355810d1 SL |
1527 | return -EINVAL; |
1528 | } | |
1529 | return 0; | |
1530 | } | |
1531 | ||
1532 | /* copy data/parity from log to raid disks */ | |
1533 | static void r5l_recovery_flush_log(struct r5l_log *log, | |
1534 | struct r5l_recovery_ctx *ctx) | |
1535 | { | |
1536 | while (1) { | |
9ed988f5 | 1537 | if (r5l_recovery_read_meta_block(log, ctx)) |
355810d1 SL |
1538 | return; |
1539 | if (r5l_recovery_flush_one_meta(log, ctx)) | |
1540 | return; | |
1541 | ctx->seq++; | |
1542 | ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks); | |
1543 | } | |
1544 | } | |
1545 | ||
9ed988f5 SL |
1546 | static void |
1547 | r5l_recovery_create_empty_meta_block(struct r5l_log *log, | |
1548 | struct page *page, | |
1549 | sector_t pos, u64 seq) | |
355810d1 | 1550 | { |
355810d1 SL |
1551 | struct r5l_meta_block *mb; |
1552 | u32 crc; | |
1553 | ||
355810d1 | 1554 | mb = page_address(page); |
9ed988f5 | 1555 | clear_page(mb); |
355810d1 SL |
1556 | mb->magic = cpu_to_le32(R5LOG_MAGIC); |
1557 | mb->version = R5LOG_VERSION; | |
1558 | mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block)); | |
1559 | mb->seq = cpu_to_le64(seq); | |
1560 | mb->position = cpu_to_le64(pos); | |
5cb2fbd6 | 1561 | crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); |
355810d1 | 1562 | mb->checksum = cpu_to_le32(crc); |
9ed988f5 | 1563 | } |
355810d1 | 1564 | |
9ed988f5 SL |
1565 | static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos, |
1566 | u64 seq) | |
1567 | { | |
1568 | struct page *page; | |
1569 | ||
1570 | page = alloc_page(GFP_KERNEL); | |
1571 | if (!page) | |
1572 | return -ENOMEM; | |
1573 | r5l_recovery_create_empty_meta_block(log, page, pos, seq); | |
796a5cf0 MC |
1574 | if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE, |
1575 | WRITE_FUA, false)) { | |
355810d1 SL |
1576 | __free_page(page); |
1577 | return -EIO; | |
1578 | } | |
1579 | __free_page(page); | |
1580 | return 0; | |
1581 | } | |
1582 | ||
b4c625c6 SL |
1583 | /* |
1584 | * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite | |
1585 | * to mark valid (potentially not flushed) data in the journal. | |
1586 | * | |
1587 | * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb, | |
1588 | * so there should not be any mismatch here. | |
1589 | */ | |
1590 | static void r5l_recovery_load_data(struct r5l_log *log, | |
1591 | struct stripe_head *sh, | |
1592 | struct r5l_recovery_ctx *ctx, | |
1593 | struct r5l_payload_data_parity *payload, | |
1594 | sector_t log_offset) | |
1595 | { | |
1596 | struct mddev *mddev = log->rdev->mddev; | |
1597 | struct r5conf *conf = mddev->private; | |
1598 | int dd_idx; | |
1599 | ||
1600 | raid5_compute_sector(conf, | |
1601 | le64_to_cpu(payload->location), 0, | |
1602 | &dd_idx, sh); | |
1603 | sync_page_io(log->rdev, log_offset, PAGE_SIZE, | |
1604 | sh->dev[dd_idx].page, REQ_OP_READ, 0, false); | |
1605 | sh->dev[dd_idx].log_checksum = | |
1606 | le32_to_cpu(payload->checksum[0]); | |
1607 | ctx->meta_total_blocks += BLOCK_SECTORS; | |
1608 | ||
1609 | set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags); | |
1610 | set_bit(STRIPE_R5C_CACHING, &sh->state); | |
1611 | } | |
1612 | ||
1613 | static void r5l_recovery_load_parity(struct r5l_log *log, | |
1614 | struct stripe_head *sh, | |
1615 | struct r5l_recovery_ctx *ctx, | |
1616 | struct r5l_payload_data_parity *payload, | |
1617 | sector_t log_offset) | |
1618 | { | |
1619 | struct mddev *mddev = log->rdev->mddev; | |
1620 | struct r5conf *conf = mddev->private; | |
1621 | ||
1622 | ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded; | |
1623 | sync_page_io(log->rdev, log_offset, PAGE_SIZE, | |
1624 | sh->dev[sh->pd_idx].page, REQ_OP_READ, 0, false); | |
1625 | sh->dev[sh->pd_idx].log_checksum = | |
1626 | le32_to_cpu(payload->checksum[0]); | |
1627 | set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags); | |
1628 | ||
1629 | if (sh->qd_idx >= 0) { | |
1630 | sync_page_io(log->rdev, | |
1631 | r5l_ring_add(log, log_offset, BLOCK_SECTORS), | |
1632 | PAGE_SIZE, sh->dev[sh->qd_idx].page, | |
1633 | REQ_OP_READ, 0, false); | |
1634 | sh->dev[sh->qd_idx].log_checksum = | |
1635 | le32_to_cpu(payload->checksum[1]); | |
1636 | set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags); | |
1637 | } | |
1638 | clear_bit(STRIPE_R5C_CACHING, &sh->state); | |
1639 | } | |
1640 | ||
1641 | static void r5l_recovery_reset_stripe(struct stripe_head *sh) | |
1642 | { | |
1643 | int i; | |
1644 | ||
1645 | sh->state = 0; | |
1646 | sh->log_start = MaxSector; | |
1647 | for (i = sh->disks; i--; ) | |
1648 | sh->dev[i].flags = 0; | |
1649 | } | |
1650 | ||
1651 | static void | |
1652 | r5l_recovery_replay_one_stripe(struct r5conf *conf, | |
1653 | struct stripe_head *sh, | |
1654 | struct r5l_recovery_ctx *ctx) | |
1655 | { | |
1656 | struct md_rdev *rdev, *rrdev; | |
1657 | int disk_index; | |
1658 | int data_count = 0; | |
1659 | ||
1660 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { | |
1661 | if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) | |
1662 | continue; | |
1663 | if (disk_index == sh->qd_idx || disk_index == sh->pd_idx) | |
1664 | continue; | |
1665 | data_count++; | |
1666 | } | |
1667 | ||
1668 | /* | |
1669 | * stripes that only have parity must have been flushed | |
1670 | * before the crash that we are now recovering from, so | |
1671 | * there is nothing more to recovery. | |
1672 | */ | |
1673 | if (data_count == 0) | |
1674 | goto out; | |
1675 | ||
1676 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { | |
1677 | if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) | |
1678 | continue; | |
1679 | ||
1680 | /* in case device is broken */ | |
1681 | rcu_read_lock(); | |
1682 | rdev = rcu_dereference(conf->disks[disk_index].rdev); | |
1683 | if (rdev) { | |
1684 | atomic_inc(&rdev->nr_pending); | |
1685 | rcu_read_unlock(); | |
1686 | sync_page_io(rdev, sh->sector, PAGE_SIZE, | |
1687 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, | |
1688 | false); | |
1689 | rdev_dec_pending(rdev, rdev->mddev); | |
1690 | rcu_read_lock(); | |
1691 | } | |
1692 | rrdev = rcu_dereference(conf->disks[disk_index].replacement); | |
1693 | if (rrdev) { | |
1694 | atomic_inc(&rrdev->nr_pending); | |
1695 | rcu_read_unlock(); | |
1696 | sync_page_io(rrdev, sh->sector, PAGE_SIZE, | |
1697 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, | |
1698 | false); | |
1699 | rdev_dec_pending(rrdev, rrdev->mddev); | |
1700 | rcu_read_lock(); | |
1701 | } | |
1702 | rcu_read_unlock(); | |
1703 | } | |
1704 | ctx->data_parity_stripes++; | |
1705 | out: | |
1706 | r5l_recovery_reset_stripe(sh); | |
1707 | } | |
1708 | ||
1709 | static struct stripe_head * | |
1710 | r5c_recovery_alloc_stripe(struct r5conf *conf, | |
1711 | struct list_head *recovery_list, | |
1712 | sector_t stripe_sect, | |
1713 | sector_t log_start) | |
1714 | { | |
1715 | struct stripe_head *sh; | |
1716 | ||
1717 | sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0); | |
1718 | if (!sh) | |
1719 | return NULL; /* no more stripe available */ | |
1720 | ||
1721 | r5l_recovery_reset_stripe(sh); | |
1722 | sh->log_start = log_start; | |
1723 | ||
1724 | return sh; | |
1725 | } | |
1726 | ||
1727 | static struct stripe_head * | |
1728 | r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect) | |
1729 | { | |
1730 | struct stripe_head *sh; | |
1731 | ||
1732 | list_for_each_entry(sh, list, lru) | |
1733 | if (sh->sector == sect) | |
1734 | return sh; | |
1735 | return NULL; | |
1736 | } | |
1737 | ||
1738 | static void | |
1739 | r5c_recovery_drop_stripes(struct list_head *cached_stripe_list, | |
1740 | struct r5l_recovery_ctx *ctx) | |
1741 | { | |
1742 | struct stripe_head *sh, *next; | |
1743 | ||
1744 | list_for_each_entry_safe(sh, next, cached_stripe_list, lru) { | |
1745 | r5l_recovery_reset_stripe(sh); | |
1746 | list_del_init(&sh->lru); | |
1747 | raid5_release_stripe(sh); | |
1748 | } | |
1749 | } | |
1750 | ||
1751 | static void | |
1752 | r5c_recovery_replay_stripes(struct list_head *cached_stripe_list, | |
1753 | struct r5l_recovery_ctx *ctx) | |
1754 | { | |
1755 | struct stripe_head *sh, *next; | |
1756 | ||
1757 | list_for_each_entry_safe(sh, next, cached_stripe_list, lru) | |
1758 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { | |
1759 | r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx); | |
1760 | list_del_init(&sh->lru); | |
1761 | raid5_release_stripe(sh); | |
1762 | } | |
1763 | } | |
1764 | ||
1765 | /* if matches return 0; otherwise return -EINVAL */ | |
1766 | static int | |
1767 | r5l_recovery_verify_data_checksum(struct r5l_log *log, struct page *page, | |
1768 | sector_t log_offset, __le32 log_checksum) | |
1769 | { | |
1770 | void *addr; | |
1771 | u32 checksum; | |
1772 | ||
1773 | sync_page_io(log->rdev, log_offset, PAGE_SIZE, | |
1774 | page, REQ_OP_READ, 0, false); | |
1775 | addr = kmap_atomic(page); | |
1776 | checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE); | |
1777 | kunmap_atomic(addr); | |
1778 | return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL; | |
1779 | } | |
1780 | ||
1781 | /* | |
1782 | * before loading data to stripe cache, we need verify checksum for all data, | |
1783 | * if there is mismatch for any data page, we drop all data in the mata block | |
1784 | */ | |
1785 | static int | |
1786 | r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log, | |
1787 | struct r5l_recovery_ctx *ctx) | |
1788 | { | |
1789 | struct mddev *mddev = log->rdev->mddev; | |
1790 | struct r5conf *conf = mddev->private; | |
1791 | struct r5l_meta_block *mb = page_address(ctx->meta_page); | |
1792 | sector_t mb_offset = sizeof(struct r5l_meta_block); | |
1793 | sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); | |
1794 | struct page *page; | |
1795 | struct r5l_payload_data_parity *payload; | |
1796 | ||
1797 | page = alloc_page(GFP_KERNEL); | |
1798 | if (!page) | |
1799 | return -ENOMEM; | |
1800 | ||
1801 | while (mb_offset < le32_to_cpu(mb->meta_size)) { | |
1802 | payload = (void *)mb + mb_offset; | |
1803 | ||
1804 | if (payload->header.type == R5LOG_PAYLOAD_DATA) { | |
1805 | if (r5l_recovery_verify_data_checksum( | |
1806 | log, page, log_offset, | |
1807 | payload->checksum[0]) < 0) | |
1808 | goto mismatch; | |
1809 | } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) { | |
1810 | if (r5l_recovery_verify_data_checksum( | |
1811 | log, page, log_offset, | |
1812 | payload->checksum[0]) < 0) | |
1813 | goto mismatch; | |
1814 | if (conf->max_degraded == 2 && /* q for RAID 6 */ | |
1815 | r5l_recovery_verify_data_checksum( | |
1816 | log, page, | |
1817 | r5l_ring_add(log, log_offset, | |
1818 | BLOCK_SECTORS), | |
1819 | payload->checksum[1]) < 0) | |
1820 | goto mismatch; | |
1821 | } else /* not R5LOG_PAYLOAD_DATA or R5LOG_PAYLOAD_PARITY */ | |
1822 | goto mismatch; | |
1823 | ||
1824 | log_offset = r5l_ring_add(log, log_offset, | |
1825 | le32_to_cpu(payload->size)); | |
1826 | ||
1827 | mb_offset += sizeof(struct r5l_payload_data_parity) + | |
1828 | sizeof(__le32) * | |
1829 | (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); | |
1830 | } | |
1831 | ||
1832 | put_page(page); | |
1833 | return 0; | |
1834 | ||
1835 | mismatch: | |
1836 | put_page(page); | |
1837 | return -EINVAL; | |
1838 | } | |
1839 | ||
1840 | /* | |
1841 | * Analyze all data/parity pages in one meta block | |
1842 | * Returns: | |
1843 | * 0 for success | |
1844 | * -EINVAL for unknown playload type | |
1845 | * -EAGAIN for checksum mismatch of data page | |
1846 | * -ENOMEM for run out of memory (alloc_page failed or run out of stripes) | |
1847 | */ | |
1848 | static int | |
1849 | r5c_recovery_analyze_meta_block(struct r5l_log *log, | |
1850 | struct r5l_recovery_ctx *ctx, | |
1851 | struct list_head *cached_stripe_list) | |
1852 | { | |
1853 | struct mddev *mddev = log->rdev->mddev; | |
1854 | struct r5conf *conf = mddev->private; | |
1855 | struct r5l_meta_block *mb; | |
1856 | struct r5l_payload_data_parity *payload; | |
1857 | int mb_offset; | |
1858 | sector_t log_offset; | |
1859 | sector_t stripe_sect; | |
1860 | struct stripe_head *sh; | |
1861 | int ret; | |
1862 | ||
1863 | /* | |
1864 | * for mismatch in data blocks, we will drop all data in this mb, but | |
1865 | * we will still read next mb for other data with FLUSH flag, as | |
1866 | * io_unit could finish out of order. | |
1867 | */ | |
1868 | ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx); | |
1869 | if (ret == -EINVAL) | |
1870 | return -EAGAIN; | |
1871 | else if (ret) | |
1872 | return ret; /* -ENOMEM duo to alloc_page() failed */ | |
1873 | ||
1874 | mb = page_address(ctx->meta_page); | |
1875 | mb_offset = sizeof(struct r5l_meta_block); | |
1876 | log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); | |
1877 | ||
1878 | while (mb_offset < le32_to_cpu(mb->meta_size)) { | |
1879 | int dd; | |
1880 | ||
1881 | payload = (void *)mb + mb_offset; | |
1882 | stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ? | |
1883 | raid5_compute_sector( | |
1884 | conf, le64_to_cpu(payload->location), 0, &dd, | |
1885 | NULL) | |
1886 | : le64_to_cpu(payload->location); | |
1887 | ||
1888 | sh = r5c_recovery_lookup_stripe(cached_stripe_list, | |
1889 | stripe_sect); | |
1890 | ||
1891 | if (!sh) { | |
1892 | sh = r5c_recovery_alloc_stripe(conf, cached_stripe_list, | |
1893 | stripe_sect, ctx->pos); | |
1894 | /* | |
1895 | * cannot get stripe from raid5_get_active_stripe | |
1896 | * try replay some stripes | |
1897 | */ | |
1898 | if (!sh) { | |
1899 | r5c_recovery_replay_stripes( | |
1900 | cached_stripe_list, ctx); | |
1901 | sh = r5c_recovery_alloc_stripe( | |
1902 | conf, cached_stripe_list, | |
1903 | stripe_sect, ctx->pos); | |
1904 | } | |
1905 | if (!sh) { | |
1906 | pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n", | |
1907 | mdname(mddev), | |
1908 | conf->min_nr_stripes * 2); | |
1909 | raid5_set_cache_size(mddev, | |
1910 | conf->min_nr_stripes * 2); | |
1911 | sh = r5c_recovery_alloc_stripe( | |
1912 | conf, cached_stripe_list, stripe_sect, | |
1913 | ctx->pos); | |
1914 | } | |
1915 | if (!sh) { | |
1916 | pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n", | |
1917 | mdname(mddev)); | |
1918 | return -ENOMEM; | |
1919 | } | |
1920 | list_add_tail(&sh->lru, cached_stripe_list); | |
1921 | } | |
1922 | ||
1923 | if (payload->header.type == R5LOG_PAYLOAD_DATA) { | |
1924 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { | |
1925 | r5l_recovery_replay_one_stripe(conf, sh, ctx); | |
1926 | r5l_recovery_reset_stripe(sh); | |
1927 | sh->log_start = ctx->pos; | |
1928 | list_move_tail(&sh->lru, cached_stripe_list); | |
1929 | } | |
1930 | r5l_recovery_load_data(log, sh, ctx, payload, | |
1931 | log_offset); | |
1932 | } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) | |
1933 | r5l_recovery_load_parity(log, sh, ctx, payload, | |
1934 | log_offset); | |
1935 | else | |
1936 | return -EINVAL; | |
1937 | ||
1938 | log_offset = r5l_ring_add(log, log_offset, | |
1939 | le32_to_cpu(payload->size)); | |
1940 | ||
1941 | mb_offset += sizeof(struct r5l_payload_data_parity) + | |
1942 | sizeof(__le32) * | |
1943 | (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); | |
1944 | } | |
1945 | ||
1946 | return 0; | |
1947 | } | |
1948 | ||
1949 | /* | |
1950 | * Load the stripe into cache. The stripe will be written out later by | |
1951 | * the stripe cache state machine. | |
1952 | */ | |
1953 | static void r5c_recovery_load_one_stripe(struct r5l_log *log, | |
1954 | struct stripe_head *sh) | |
1955 | { | |
1956 | struct r5conf *conf = sh->raid_conf; | |
1957 | struct r5dev *dev; | |
1958 | int i; | |
1959 | ||
1960 | for (i = sh->disks; i--; ) { | |
1961 | dev = sh->dev + i; | |
1962 | if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) { | |
1963 | set_bit(R5_InJournal, &dev->flags); | |
1964 | set_bit(R5_UPTODATE, &dev->flags); | |
1965 | } | |
1966 | } | |
1967 | set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state); | |
1968 | atomic_inc(&conf->r5c_cached_partial_stripes); | |
1969 | list_add_tail(&sh->r5c, &log->stripe_in_journal_list); | |
1970 | } | |
1971 | ||
1972 | /* | |
1973 | * Scan through the log for all to-be-flushed data | |
1974 | * | |
1975 | * For stripes with data and parity, namely Data-Parity stripe | |
1976 | * (STRIPE_R5C_CACHING == 0), we simply replay all the writes. | |
1977 | * | |
1978 | * For stripes with only data, namely Data-Only stripe | |
1979 | * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine. | |
1980 | * | |
1981 | * For a stripe, if we see data after parity, we should discard all previous | |
1982 | * data and parity for this stripe, as these data are already flushed to | |
1983 | * the array. | |
1984 | * | |
1985 | * At the end of the scan, we return the new journal_tail, which points to | |
1986 | * first data-only stripe on the journal device, or next invalid meta block. | |
1987 | */ | |
1988 | static int r5c_recovery_flush_log(struct r5l_log *log, | |
1989 | struct r5l_recovery_ctx *ctx) | |
1990 | { | |
1991 | struct stripe_head *sh, *next; | |
1992 | int ret = 0; | |
1993 | ||
1994 | /* scan through the log */ | |
1995 | while (1) { | |
1996 | if (r5l_recovery_read_meta_block(log, ctx)) | |
1997 | break; | |
1998 | ||
1999 | ret = r5c_recovery_analyze_meta_block(log, ctx, | |
2000 | &ctx->cached_list); | |
2001 | /* | |
2002 | * -EAGAIN means mismatch in data block, in this case, we still | |
2003 | * try scan the next metablock | |
2004 | */ | |
2005 | if (ret && ret != -EAGAIN) | |
2006 | break; /* ret == -EINVAL or -ENOMEM */ | |
2007 | ctx->seq++; | |
2008 | ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks); | |
2009 | } | |
2010 | ||
2011 | if (ret == -ENOMEM) { | |
2012 | r5c_recovery_drop_stripes(&ctx->cached_list, ctx); | |
2013 | return ret; | |
2014 | } | |
2015 | ||
2016 | /* replay data-parity stripes */ | |
2017 | r5c_recovery_replay_stripes(&ctx->cached_list, ctx); | |
2018 | ||
2019 | /* load data-only stripes to stripe cache */ | |
2020 | list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) { | |
2021 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2022 | r5c_recovery_load_one_stripe(log, sh); | |
2023 | list_del_init(&sh->lru); | |
2024 | raid5_release_stripe(sh); | |
2025 | ctx->data_only_stripes++; | |
2026 | } | |
2027 | ||
2028 | return 0; | |
2029 | } | |
2030 | ||
2031 | /* | |
2032 | * we did a recovery. Now ctx.pos points to an invalid meta block. New | |
2033 | * log will start here. but we can't let superblock point to last valid | |
2034 | * meta block. The log might looks like: | |
2035 | * | meta 1| meta 2| meta 3| | |
2036 | * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If | |
2037 | * superblock points to meta 1, we write a new valid meta 2n. if crash | |
2038 | * happens again, new recovery will start from meta 1. Since meta 2n is | |
2039 | * valid now, recovery will think meta 3 is valid, which is wrong. | |
2040 | * The solution is we create a new meta in meta2 with its seq == meta | |
2041 | * 1's seq + 10 and let superblock points to meta2. The same recovery will | |
2042 | * not think meta 3 is a valid meta, because its seq doesn't match | |
2043 | */ | |
2044 | ||
2045 | /* | |
2046 | * Before recovery, the log looks like the following | |
2047 | * | |
2048 | * --------------------------------------------- | |
2049 | * | valid log | invalid log | | |
2050 | * --------------------------------------------- | |
2051 | * ^ | |
2052 | * |- log->last_checkpoint | |
2053 | * |- log->last_cp_seq | |
2054 | * | |
2055 | * Now we scan through the log until we see invalid entry | |
2056 | * | |
2057 | * --------------------------------------------- | |
2058 | * | valid log | invalid log | | |
2059 | * --------------------------------------------- | |
2060 | * ^ ^ | |
2061 | * |- log->last_checkpoint |- ctx->pos | |
2062 | * |- log->last_cp_seq |- ctx->seq | |
2063 | * | |
2064 | * From this point, we need to increase seq number by 10 to avoid | |
2065 | * confusing next recovery. | |
2066 | * | |
2067 | * --------------------------------------------- | |
2068 | * | valid log | invalid log | | |
2069 | * --------------------------------------------- | |
2070 | * ^ ^ | |
2071 | * |- log->last_checkpoint |- ctx->pos+1 | |
2072 | * |- log->last_cp_seq |- ctx->seq+11 | |
2073 | * | |
2074 | * However, it is not safe to start the state machine yet, because data only | |
2075 | * parities are not yet secured in RAID. To save these data only parities, we | |
2076 | * rewrite them from seq+11. | |
2077 | * | |
2078 | * ----------------------------------------------------------------- | |
2079 | * | valid log | data only stripes | invalid log | | |
2080 | * ----------------------------------------------------------------- | |
2081 | * ^ ^ | |
2082 | * |- log->last_checkpoint |- ctx->pos+n | |
2083 | * |- log->last_cp_seq |- ctx->seq+10+n | |
2084 | * | |
2085 | * If failure happens again during this process, the recovery can safe start | |
2086 | * again from log->last_checkpoint. | |
2087 | * | |
2088 | * Once data only stripes are rewritten to journal, we move log_tail | |
2089 | * | |
2090 | * ----------------------------------------------------------------- | |
2091 | * | old log | data only stripes | invalid log | | |
2092 | * ----------------------------------------------------------------- | |
2093 | * ^ ^ | |
2094 | * |- log->last_checkpoint |- ctx->pos+n | |
2095 | * |- log->last_cp_seq |- ctx->seq+10+n | |
2096 | * | |
2097 | * Then we can safely start the state machine. If failure happens from this | |
2098 | * point on, the recovery will start from new log->last_checkpoint. | |
2099 | */ | |
2100 | static int | |
2101 | r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log, | |
2102 | struct r5l_recovery_ctx *ctx) | |
2103 | { | |
2104 | struct stripe_head *sh; | |
2105 | struct mddev *mddev = log->rdev->mddev; | |
2106 | struct page *page; | |
2107 | ||
2108 | page = alloc_page(GFP_KERNEL); | |
2109 | if (!page) { | |
2110 | pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n", | |
2111 | mdname(mddev)); | |
2112 | return -ENOMEM; | |
2113 | } | |
2114 | ||
2115 | ctx->seq += 10; | |
2116 | list_for_each_entry(sh, &ctx->cached_list, lru) { | |
2117 | struct r5l_meta_block *mb; | |
2118 | int i; | |
2119 | int offset; | |
2120 | sector_t write_pos; | |
2121 | ||
2122 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2123 | r5l_recovery_create_empty_meta_block(log, page, | |
2124 | ctx->pos, ctx->seq); | |
2125 | mb = page_address(page); | |
2126 | offset = le32_to_cpu(mb->meta_size); | |
2127 | write_pos = ctx->pos + BLOCK_SECTORS; | |
2128 | ||
2129 | for (i = sh->disks; i--; ) { | |
2130 | struct r5dev *dev = &sh->dev[i]; | |
2131 | struct r5l_payload_data_parity *payload; | |
2132 | void *addr; | |
2133 | ||
2134 | if (test_bit(R5_InJournal, &dev->flags)) { | |
2135 | payload = (void *)mb + offset; | |
2136 | payload->header.type = cpu_to_le16( | |
2137 | R5LOG_PAYLOAD_DATA); | |
2138 | payload->size = BLOCK_SECTORS; | |
2139 | payload->location = cpu_to_le64( | |
2140 | raid5_compute_blocknr(sh, i, 0)); | |
2141 | addr = kmap_atomic(dev->page); | |
2142 | payload->checksum[0] = cpu_to_le32( | |
2143 | crc32c_le(log->uuid_checksum, addr, | |
2144 | PAGE_SIZE)); | |
2145 | kunmap_atomic(addr); | |
2146 | sync_page_io(log->rdev, write_pos, PAGE_SIZE, | |
2147 | dev->page, REQ_OP_WRITE, 0, false); | |
2148 | write_pos = r5l_ring_add(log, write_pos, | |
2149 | BLOCK_SECTORS); | |
2150 | offset += sizeof(__le32) + | |
2151 | sizeof(struct r5l_payload_data_parity); | |
2152 | ||
2153 | } | |
2154 | } | |
2155 | mb->meta_size = cpu_to_le32(offset); | |
2156 | mb->checksum = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); | |
2157 | sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, | |
2158 | REQ_OP_WRITE, WRITE_FUA, false); | |
2159 | sh->log_start = ctx->pos; | |
2160 | ctx->pos = write_pos; | |
2161 | ctx->seq += 1; | |
2162 | } | |
2163 | __free_page(page); | |
2164 | return 0; | |
2165 | } | |
2166 | ||
f6bed0ef SL |
2167 | static int r5l_recovery_log(struct r5l_log *log) |
2168 | { | |
355810d1 SL |
2169 | struct r5l_recovery_ctx ctx; |
2170 | ||
2171 | ctx.pos = log->last_checkpoint; | |
2172 | ctx.seq = log->last_cp_seq; | |
2173 | ctx.meta_page = alloc_page(GFP_KERNEL); | |
b4c625c6 SL |
2174 | ctx.data_only_stripes = 0; |
2175 | ctx.data_parity_stripes = 0; | |
2176 | INIT_LIST_HEAD(&ctx.cached_list); | |
2177 | ||
355810d1 SL |
2178 | if (!ctx.meta_page) |
2179 | return -ENOMEM; | |
2180 | ||
2181 | r5l_recovery_flush_log(log, &ctx); | |
2182 | __free_page(ctx.meta_page); | |
2183 | ||
2184 | /* | |
2185 | * we did a recovery. Now ctx.pos points to an invalid meta block. New | |
2186 | * log will start here. but we can't let superblock point to last valid | |
2187 | * meta block. The log might looks like: | |
2188 | * | meta 1| meta 2| meta 3| | |
2189 | * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If | |
2190 | * superblock points to meta 1, we write a new valid meta 2n. if crash | |
2191 | * happens again, new recovery will start from meta 1. Since meta 2n is | |
2192 | * valid now, recovery will think meta 3 is valid, which is wrong. | |
2193 | * The solution is we create a new meta in meta2 with its seq == meta | |
2194 | * 1's seq + 10 and let superblock points to meta2. The same recovery will | |
2195 | * not think meta 3 is a valid meta, because its seq doesn't match | |
2196 | */ | |
9a8b27fa | 2197 | if (ctx.seq > log->last_cp_seq) { |
355810d1 SL |
2198 | int ret; |
2199 | ||
2200 | ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10); | |
2201 | if (ret) | |
2202 | return ret; | |
2203 | log->seq = ctx.seq + 11; | |
2204 | log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS); | |
2205 | r5l_write_super(log, ctx.pos); | |
28cd88e2 ZL |
2206 | log->last_checkpoint = ctx.pos; |
2207 | log->next_checkpoint = ctx.pos; | |
355810d1 SL |
2208 | } else { |
2209 | log->log_start = ctx.pos; | |
2210 | log->seq = ctx.seq; | |
2211 | } | |
b4c625c6 SL |
2212 | |
2213 | /* | |
2214 | * This is to suppress "function defined but not used" warning. | |
2215 | * It will be removed when the two functions are used (next patch). | |
2216 | */ | |
2217 | if (!log) { | |
2218 | r5c_recovery_flush_log(log, &ctx); | |
2219 | r5c_recovery_rewrite_data_only_stripes(log, &ctx); | |
2220 | } | |
2221 | ||
f6bed0ef SL |
2222 | return 0; |
2223 | } | |
2224 | ||
2225 | static void r5l_write_super(struct r5l_log *log, sector_t cp) | |
2226 | { | |
2227 | struct mddev *mddev = log->rdev->mddev; | |
2228 | ||
2229 | log->rdev->journal_tail = cp; | |
2230 | set_bit(MD_CHANGE_DEVS, &mddev->flags); | |
2231 | } | |
2232 | ||
2c7da14b SL |
2233 | static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page) |
2234 | { | |
2235 | struct r5conf *conf = mddev->private; | |
2236 | int ret; | |
2237 | ||
2238 | if (!conf->log) | |
2239 | return 0; | |
2240 | ||
2241 | switch (conf->log->r5c_journal_mode) { | |
2242 | case R5C_JOURNAL_MODE_WRITE_THROUGH: | |
2243 | ret = snprintf( | |
2244 | page, PAGE_SIZE, "[%s] %s\n", | |
2245 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], | |
2246 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); | |
2247 | break; | |
2248 | case R5C_JOURNAL_MODE_WRITE_BACK: | |
2249 | ret = snprintf( | |
2250 | page, PAGE_SIZE, "%s [%s]\n", | |
2251 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], | |
2252 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); | |
2253 | break; | |
2254 | default: | |
2255 | ret = 0; | |
2256 | } | |
2257 | return ret; | |
2258 | } | |
2259 | ||
2260 | static ssize_t r5c_journal_mode_store(struct mddev *mddev, | |
2261 | const char *page, size_t length) | |
2262 | { | |
2263 | struct r5conf *conf = mddev->private; | |
2264 | struct r5l_log *log = conf->log; | |
2265 | int val = -1, i; | |
2266 | int len = length; | |
2267 | ||
2268 | if (!log) | |
2269 | return -ENODEV; | |
2270 | ||
2271 | if (len && page[len - 1] == '\n') | |
2272 | len -= 1; | |
2273 | for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++) | |
2274 | if (strlen(r5c_journal_mode_str[i]) == len && | |
2275 | strncmp(page, r5c_journal_mode_str[i], len) == 0) { | |
2276 | val = i; | |
2277 | break; | |
2278 | } | |
2279 | if (val < R5C_JOURNAL_MODE_WRITE_THROUGH || | |
2280 | val > R5C_JOURNAL_MODE_WRITE_BACK) | |
2281 | return -EINVAL; | |
2282 | ||
2283 | mddev_suspend(mddev); | |
2284 | conf->log->r5c_journal_mode = val; | |
2285 | mddev_resume(mddev); | |
2286 | ||
2287 | pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n", | |
2288 | mdname(mddev), val, r5c_journal_mode_str[val]); | |
2289 | return length; | |
2290 | } | |
2291 | ||
2292 | struct md_sysfs_entry | |
2293 | r5c_journal_mode = __ATTR(journal_mode, 0644, | |
2294 | r5c_journal_mode_show, r5c_journal_mode_store); | |
2295 | ||
2ded3703 SL |
2296 | /* |
2297 | * Try handle write operation in caching phase. This function should only | |
2298 | * be called in write-back mode. | |
2299 | * | |
2300 | * If all outstanding writes can be handled in caching phase, returns 0 | |
2301 | * If writes requires write-out phase, call r5c_make_stripe_write_out() | |
2302 | * and returns -EAGAIN | |
2303 | */ | |
2304 | int r5c_try_caching_write(struct r5conf *conf, | |
2305 | struct stripe_head *sh, | |
2306 | struct stripe_head_state *s, | |
2307 | int disks) | |
2308 | { | |
2309 | struct r5l_log *log = conf->log; | |
1e6d690b SL |
2310 | int i; |
2311 | struct r5dev *dev; | |
2312 | int to_cache = 0; | |
2ded3703 SL |
2313 | |
2314 | BUG_ON(!r5c_is_writeback(log)); | |
2315 | ||
1e6d690b SL |
2316 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { |
2317 | /* | |
2318 | * There are two different scenarios here: | |
2319 | * 1. The stripe has some data cached, and it is sent to | |
2320 | * write-out phase for reclaim | |
2321 | * 2. The stripe is clean, and this is the first write | |
2322 | * | |
2323 | * For 1, return -EAGAIN, so we continue with | |
2324 | * handle_stripe_dirtying(). | |
2325 | * | |
2326 | * For 2, set STRIPE_R5C_CACHING and continue with caching | |
2327 | * write. | |
2328 | */ | |
2329 | ||
2330 | /* case 1: anything injournal or anything in written */ | |
2331 | if (s->injournal > 0 || s->written > 0) | |
2332 | return -EAGAIN; | |
2333 | /* case 2 */ | |
2334 | set_bit(STRIPE_R5C_CACHING, &sh->state); | |
2335 | } | |
2336 | ||
2337 | for (i = disks; i--; ) { | |
2338 | dev = &sh->dev[i]; | |
2339 | /* if non-overwrite, use writing-out phase */ | |
2340 | if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) && | |
2341 | !test_bit(R5_InJournal, &dev->flags)) { | |
2342 | r5c_make_stripe_write_out(sh); | |
2343 | return -EAGAIN; | |
2344 | } | |
2345 | } | |
2346 | ||
2347 | for (i = disks; i--; ) { | |
2348 | dev = &sh->dev[i]; | |
2349 | if (dev->towrite) { | |
2350 | set_bit(R5_Wantwrite, &dev->flags); | |
2351 | set_bit(R5_Wantdrain, &dev->flags); | |
2352 | set_bit(R5_LOCKED, &dev->flags); | |
2353 | to_cache++; | |
2354 | } | |
2355 | } | |
2356 | ||
2357 | if (to_cache) { | |
2358 | set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); | |
2359 | /* | |
2360 | * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data() | |
2361 | * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in | |
2362 | * r5c_handle_data_cached() | |
2363 | */ | |
2364 | set_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
2365 | } | |
2366 | ||
2367 | return 0; | |
2368 | } | |
2369 | ||
2370 | /* | |
2371 | * free extra pages (orig_page) we allocated for prexor | |
2372 | */ | |
2373 | void r5c_release_extra_page(struct stripe_head *sh) | |
2374 | { | |
2375 | int i; | |
2376 | ||
2377 | for (i = sh->disks; i--; ) | |
2378 | if (sh->dev[i].page != sh->dev[i].orig_page) { | |
2379 | struct page *p = sh->dev[i].orig_page; | |
2380 | ||
2381 | sh->dev[i].orig_page = sh->dev[i].page; | |
2382 | put_page(p); | |
2383 | } | |
2ded3703 SL |
2384 | } |
2385 | ||
2386 | /* | |
2387 | * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the | |
2388 | * stripe is committed to RAID disks. | |
2389 | */ | |
2390 | void r5c_finish_stripe_write_out(struct r5conf *conf, | |
2391 | struct stripe_head *sh, | |
2392 | struct stripe_head_state *s) | |
2393 | { | |
1e6d690b SL |
2394 | int i; |
2395 | int do_wakeup = 0; | |
2396 | ||
2ded3703 SL |
2397 | if (!conf->log || |
2398 | !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags)) | |
2399 | return; | |
2400 | ||
2401 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2402 | clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
2403 | ||
2404 | if (conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) | |
2405 | return; | |
1e6d690b SL |
2406 | |
2407 | for (i = sh->disks; i--; ) { | |
2408 | clear_bit(R5_InJournal, &sh->dev[i].flags); | |
2409 | if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) | |
2410 | do_wakeup = 1; | |
2411 | } | |
2412 | ||
2413 | /* | |
2414 | * analyse_stripe() runs before r5c_finish_stripe_write_out(), | |
2415 | * We updated R5_InJournal, so we also update s->injournal. | |
2416 | */ | |
2417 | s->injournal = 0; | |
2418 | ||
2419 | if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) | |
2420 | if (atomic_dec_and_test(&conf->pending_full_writes)) | |
2421 | md_wakeup_thread(conf->mddev->thread); | |
2422 | ||
2423 | if (do_wakeup) | |
2424 | wake_up(&conf->wait_for_overlap); | |
a39f7afd SL |
2425 | |
2426 | if (conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) | |
2427 | return; | |
2428 | ||
2429 | spin_lock_irq(&conf->log->stripe_in_journal_lock); | |
2430 | list_del_init(&sh->r5c); | |
2431 | spin_unlock_irq(&conf->log->stripe_in_journal_lock); | |
2432 | sh->log_start = MaxSector; | |
2433 | atomic_dec(&conf->log->stripe_in_journal_count); | |
1e6d690b SL |
2434 | } |
2435 | ||
2436 | int | |
2437 | r5c_cache_data(struct r5l_log *log, struct stripe_head *sh, | |
2438 | struct stripe_head_state *s) | |
2439 | { | |
a39f7afd | 2440 | struct r5conf *conf = sh->raid_conf; |
1e6d690b SL |
2441 | int pages = 0; |
2442 | int reserve; | |
2443 | int i; | |
2444 | int ret = 0; | |
2445 | ||
2446 | BUG_ON(!log); | |
2447 | ||
2448 | for (i = 0; i < sh->disks; i++) { | |
2449 | void *addr; | |
2450 | ||
2451 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags)) | |
2452 | continue; | |
2453 | addr = kmap_atomic(sh->dev[i].page); | |
2454 | sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, | |
2455 | addr, PAGE_SIZE); | |
2456 | kunmap_atomic(addr); | |
2457 | pages++; | |
2458 | } | |
2459 | WARN_ON(pages == 0); | |
2460 | ||
2461 | /* | |
2462 | * The stripe must enter state machine again to call endio, so | |
2463 | * don't delay. | |
2464 | */ | |
2465 | clear_bit(STRIPE_DELAYED, &sh->state); | |
2466 | atomic_inc(&sh->count); | |
2467 | ||
2468 | mutex_lock(&log->io_mutex); | |
2469 | /* meta + data */ | |
2470 | reserve = (1 + pages) << (PAGE_SHIFT - 9); | |
1e6d690b | 2471 | |
a39f7afd SL |
2472 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && |
2473 | sh->log_start == MaxSector) | |
2474 | r5l_add_no_space_stripe(log, sh); | |
2475 | else if (!r5l_has_free_space(log, reserve)) { | |
2476 | if (sh->log_start == log->last_checkpoint) | |
2477 | BUG(); | |
2478 | else | |
2479 | r5l_add_no_space_stripe(log, sh); | |
1e6d690b SL |
2480 | } else { |
2481 | ret = r5l_log_stripe(log, sh, pages, 0); | |
2482 | if (ret) { | |
2483 | spin_lock_irq(&log->io_list_lock); | |
2484 | list_add_tail(&sh->log_list, &log->no_mem_stripes); | |
2485 | spin_unlock_irq(&log->io_list_lock); | |
2486 | } | |
2487 | } | |
2488 | ||
2489 | mutex_unlock(&log->io_mutex); | |
2490 | return 0; | |
2ded3703 SL |
2491 | } |
2492 | ||
f6bed0ef SL |
2493 | static int r5l_load_log(struct r5l_log *log) |
2494 | { | |
2495 | struct md_rdev *rdev = log->rdev; | |
2496 | struct page *page; | |
2497 | struct r5l_meta_block *mb; | |
2498 | sector_t cp = log->rdev->journal_tail; | |
2499 | u32 stored_crc, expected_crc; | |
2500 | bool create_super = false; | |
2501 | int ret; | |
2502 | ||
2503 | /* Make sure it's valid */ | |
2504 | if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp) | |
2505 | cp = 0; | |
2506 | page = alloc_page(GFP_KERNEL); | |
2507 | if (!page) | |
2508 | return -ENOMEM; | |
2509 | ||
796a5cf0 | 2510 | if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) { |
f6bed0ef SL |
2511 | ret = -EIO; |
2512 | goto ioerr; | |
2513 | } | |
2514 | mb = page_address(page); | |
2515 | ||
2516 | if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || | |
2517 | mb->version != R5LOG_VERSION) { | |
2518 | create_super = true; | |
2519 | goto create; | |
2520 | } | |
2521 | stored_crc = le32_to_cpu(mb->checksum); | |
2522 | mb->checksum = 0; | |
5cb2fbd6 | 2523 | expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); |
f6bed0ef SL |
2524 | if (stored_crc != expected_crc) { |
2525 | create_super = true; | |
2526 | goto create; | |
2527 | } | |
2528 | if (le64_to_cpu(mb->position) != cp) { | |
2529 | create_super = true; | |
2530 | goto create; | |
2531 | } | |
2532 | create: | |
2533 | if (create_super) { | |
2534 | log->last_cp_seq = prandom_u32(); | |
2535 | cp = 0; | |
56056c2e | 2536 | r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq); |
f6bed0ef SL |
2537 | /* |
2538 | * Make sure super points to correct address. Log might have | |
2539 | * data very soon. If super hasn't correct log tail address, | |
2540 | * recovery can't find the log | |
2541 | */ | |
2542 | r5l_write_super(log, cp); | |
2543 | } else | |
2544 | log->last_cp_seq = le64_to_cpu(mb->seq); | |
2545 | ||
2546 | log->device_size = round_down(rdev->sectors, BLOCK_SECTORS); | |
0576b1c6 SL |
2547 | log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT; |
2548 | if (log->max_free_space > RECLAIM_MAX_FREE_SPACE) | |
2549 | log->max_free_space = RECLAIM_MAX_FREE_SPACE; | |
f6bed0ef | 2550 | log->last_checkpoint = cp; |
28cd88e2 | 2551 | log->next_checkpoint = cp; |
a39f7afd SL |
2552 | mutex_lock(&log->io_mutex); |
2553 | r5c_update_log_state(log); | |
2554 | mutex_unlock(&log->io_mutex); | |
f6bed0ef SL |
2555 | |
2556 | __free_page(page); | |
2557 | ||
2558 | return r5l_recovery_log(log); | |
2559 | ioerr: | |
2560 | __free_page(page); | |
2561 | return ret; | |
2562 | } | |
2563 | ||
2564 | int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev) | |
2565 | { | |
c888a8f9 | 2566 | struct request_queue *q = bdev_get_queue(rdev->bdev); |
f6bed0ef SL |
2567 | struct r5l_log *log; |
2568 | ||
2569 | if (PAGE_SIZE != 4096) | |
2570 | return -EINVAL; | |
c757ec95 SL |
2571 | |
2572 | /* | |
2573 | * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and | |
2574 | * raid_disks r5l_payload_data_parity. | |
2575 | * | |
2576 | * Write journal and cache does not work for very big array | |
2577 | * (raid_disks > 203) | |
2578 | */ | |
2579 | if (sizeof(struct r5l_meta_block) + | |
2580 | ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) * | |
2581 | conf->raid_disks) > PAGE_SIZE) { | |
2582 | pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n", | |
2583 | mdname(conf->mddev), conf->raid_disks); | |
2584 | return -EINVAL; | |
2585 | } | |
2586 | ||
f6bed0ef SL |
2587 | log = kzalloc(sizeof(*log), GFP_KERNEL); |
2588 | if (!log) | |
2589 | return -ENOMEM; | |
2590 | log->rdev = rdev; | |
2591 | ||
c888a8f9 | 2592 | log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0; |
56fef7c6 | 2593 | |
5cb2fbd6 SL |
2594 | log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid, |
2595 | sizeof(rdev->mddev->uuid)); | |
f6bed0ef SL |
2596 | |
2597 | mutex_init(&log->io_mutex); | |
2598 | ||
2599 | spin_lock_init(&log->io_list_lock); | |
2600 | INIT_LIST_HEAD(&log->running_ios); | |
0576b1c6 | 2601 | INIT_LIST_HEAD(&log->io_end_ios); |
a8c34f91 | 2602 | INIT_LIST_HEAD(&log->flushing_ios); |
04732f74 | 2603 | INIT_LIST_HEAD(&log->finished_ios); |
a8c34f91 | 2604 | bio_init(&log->flush_bio); |
f6bed0ef SL |
2605 | |
2606 | log->io_kc = KMEM_CACHE(r5l_io_unit, 0); | |
2607 | if (!log->io_kc) | |
2608 | goto io_kc; | |
2609 | ||
5036c390 CH |
2610 | log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc); |
2611 | if (!log->io_pool) | |
2612 | goto io_pool; | |
2613 | ||
c38d29b3 CH |
2614 | log->bs = bioset_create(R5L_POOL_SIZE, 0); |
2615 | if (!log->bs) | |
2616 | goto io_bs; | |
2617 | ||
e8deb638 CH |
2618 | log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0); |
2619 | if (!log->meta_pool) | |
2620 | goto out_mempool; | |
2621 | ||
0576b1c6 SL |
2622 | log->reclaim_thread = md_register_thread(r5l_reclaim_thread, |
2623 | log->rdev->mddev, "reclaim"); | |
2624 | if (!log->reclaim_thread) | |
2625 | goto reclaim_thread; | |
a39f7afd SL |
2626 | log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL; |
2627 | ||
0fd22b45 | 2628 | init_waitqueue_head(&log->iounit_wait); |
0576b1c6 | 2629 | |
5036c390 CH |
2630 | INIT_LIST_HEAD(&log->no_mem_stripes); |
2631 | ||
f6bed0ef SL |
2632 | INIT_LIST_HEAD(&log->no_space_stripes); |
2633 | spin_lock_init(&log->no_space_stripes_lock); | |
2634 | ||
2ded3703 | 2635 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; |
a39f7afd SL |
2636 | INIT_LIST_HEAD(&log->stripe_in_journal_list); |
2637 | spin_lock_init(&log->stripe_in_journal_lock); | |
2638 | atomic_set(&log->stripe_in_journal_count, 0); | |
2ded3703 | 2639 | |
f6bed0ef SL |
2640 | if (r5l_load_log(log)) |
2641 | goto error; | |
2642 | ||
f6b6ec5c | 2643 | rcu_assign_pointer(conf->log, log); |
a62ab49e | 2644 | set_bit(MD_HAS_JOURNAL, &conf->mddev->flags); |
f6bed0ef | 2645 | return 0; |
e8deb638 | 2646 | |
f6bed0ef | 2647 | error: |
0576b1c6 SL |
2648 | md_unregister_thread(&log->reclaim_thread); |
2649 | reclaim_thread: | |
e8deb638 CH |
2650 | mempool_destroy(log->meta_pool); |
2651 | out_mempool: | |
c38d29b3 CH |
2652 | bioset_free(log->bs); |
2653 | io_bs: | |
5036c390 CH |
2654 | mempool_destroy(log->io_pool); |
2655 | io_pool: | |
f6bed0ef SL |
2656 | kmem_cache_destroy(log->io_kc); |
2657 | io_kc: | |
2658 | kfree(log); | |
2659 | return -EINVAL; | |
2660 | } | |
2661 | ||
2662 | void r5l_exit_log(struct r5l_log *log) | |
2663 | { | |
0576b1c6 | 2664 | md_unregister_thread(&log->reclaim_thread); |
e8deb638 | 2665 | mempool_destroy(log->meta_pool); |
c38d29b3 | 2666 | bioset_free(log->bs); |
5036c390 | 2667 | mempool_destroy(log->io_pool); |
f6bed0ef SL |
2668 | kmem_cache_destroy(log->io_kc); |
2669 | kfree(log); | |
2670 | } |