]>
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 | 21 | #include <linux/random.h> |
ce1ccd07 | 22 | #include <linux/kthread.h> |
03b047f4 | 23 | #include <linux/types.h> |
f6bed0ef SL |
24 | #include "md.h" |
25 | #include "raid5.h" | |
1e6d690b | 26 | #include "bitmap.h" |
70d466f7 | 27 | #include "raid5-log.h" |
f6bed0ef SL |
28 | |
29 | /* | |
30 | * metadata/data stored in disk with 4k size unit (a block) regardless | |
31 | * underneath hardware sector size. only works with PAGE_SIZE == 4096 | |
32 | */ | |
33 | #define BLOCK_SECTORS (8) | |
effe6ee7 | 34 | #define BLOCK_SECTOR_SHIFT (3) |
f6bed0ef | 35 | |
0576b1c6 | 36 | /* |
a39f7afd SL |
37 | * log->max_free_space is min(1/4 disk size, 10G reclaimable space). |
38 | * | |
39 | * In write through mode, the reclaim runs every log->max_free_space. | |
40 | * This can prevent the recovery scans for too long | |
0576b1c6 SL |
41 | */ |
42 | #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */ | |
43 | #define RECLAIM_MAX_FREE_SPACE_SHIFT (2) | |
44 | ||
a39f7afd SL |
45 | /* wake up reclaim thread periodically */ |
46 | #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ) | |
47 | /* start flush with these full stripes */ | |
84890c03 | 48 | #define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4) |
a39f7afd SL |
49 | /* reclaim stripes in groups */ |
50 | #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) | |
51 | ||
c38d29b3 CH |
52 | /* |
53 | * We only need 2 bios per I/O unit to make progress, but ensure we | |
54 | * have a few more available to not get too tight. | |
55 | */ | |
56 | #define R5L_POOL_SIZE 4 | |
57 | ||
2c7da14b SL |
58 | static char *r5c_journal_mode_str[] = {"write-through", |
59 | "write-back"}; | |
2ded3703 SL |
60 | /* |
61 | * raid5 cache state machine | |
62 | * | |
9b69173e | 63 | * With the RAID cache, each stripe works in two phases: |
2ded3703 SL |
64 | * - caching phase |
65 | * - writing-out phase | |
66 | * | |
67 | * These two phases are controlled by bit STRIPE_R5C_CACHING: | |
68 | * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase | |
69 | * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase | |
70 | * | |
71 | * When there is no journal, or the journal is in write-through mode, | |
72 | * the stripe is always in writing-out phase. | |
73 | * | |
74 | * For write-back journal, the stripe is sent to caching phase on write | |
75 | * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off | |
76 | * the write-out phase by clearing STRIPE_R5C_CACHING. | |
77 | * | |
78 | * Stripes in caching phase do not write the raid disks. Instead, all | |
79 | * writes are committed from the log device. Therefore, a stripe in | |
80 | * caching phase handles writes as: | |
81 | * - write to log device | |
82 | * - return IO | |
83 | * | |
84 | * Stripes in writing-out phase handle writes as: | |
85 | * - calculate parity | |
86 | * - write pending data and parity to journal | |
87 | * - write data and parity to raid disks | |
88 | * - return IO for pending writes | |
89 | */ | |
90 | ||
f6bed0ef SL |
91 | struct r5l_log { |
92 | struct md_rdev *rdev; | |
93 | ||
94 | u32 uuid_checksum; | |
95 | ||
96 | sector_t device_size; /* log device size, round to | |
97 | * BLOCK_SECTORS */ | |
0576b1c6 SL |
98 | sector_t max_free_space; /* reclaim run if free space is at |
99 | * this size */ | |
f6bed0ef SL |
100 | |
101 | sector_t last_checkpoint; /* log tail. where recovery scan | |
102 | * starts from */ | |
103 | u64 last_cp_seq; /* log tail sequence */ | |
104 | ||
105 | sector_t log_start; /* log head. where new data appends */ | |
106 | u64 seq; /* log head sequence */ | |
107 | ||
17036461 | 108 | sector_t next_checkpoint; |
17036461 | 109 | |
f6bed0ef SL |
110 | struct mutex io_mutex; |
111 | struct r5l_io_unit *current_io; /* current io_unit accepting new data */ | |
112 | ||
113 | spinlock_t io_list_lock; | |
114 | struct list_head running_ios; /* io_units which are still running, | |
115 | * and have not yet been completely | |
116 | * written to the log */ | |
117 | struct list_head io_end_ios; /* io_units which have been completely | |
118 | * written to the log but not yet written | |
119 | * to the RAID */ | |
a8c34f91 SL |
120 | struct list_head flushing_ios; /* io_units which are waiting for log |
121 | * cache flush */ | |
04732f74 | 122 | struct list_head finished_ios; /* io_units which settle down in log disk */ |
a8c34f91 | 123 | struct bio flush_bio; |
f6bed0ef | 124 | |
5036c390 CH |
125 | struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */ |
126 | ||
f6bed0ef | 127 | struct kmem_cache *io_kc; |
5036c390 | 128 | mempool_t *io_pool; |
c38d29b3 | 129 | struct bio_set *bs; |
e8deb638 | 130 | mempool_t *meta_pool; |
f6bed0ef | 131 | |
0576b1c6 SL |
132 | struct md_thread *reclaim_thread; |
133 | unsigned long reclaim_target; /* number of space that need to be | |
134 | * reclaimed. if it's 0, reclaim spaces | |
135 | * used by io_units which are in | |
136 | * IO_UNIT_STRIPE_END state (eg, reclaim | |
137 | * dones't wait for specific io_unit | |
138 | * switching to IO_UNIT_STRIPE_END | |
139 | * state) */ | |
0fd22b45 | 140 | wait_queue_head_t iounit_wait; |
0576b1c6 | 141 | |
f6bed0ef SL |
142 | struct list_head no_space_stripes; /* pending stripes, log has no space */ |
143 | spinlock_t no_space_stripes_lock; | |
56fef7c6 CH |
144 | |
145 | bool need_cache_flush; | |
2ded3703 SL |
146 | |
147 | /* for r5c_cache */ | |
148 | enum r5c_journal_mode r5c_journal_mode; | |
a39f7afd SL |
149 | |
150 | /* all stripes in r5cache, in the order of seq at sh->log_start */ | |
151 | struct list_head stripe_in_journal_list; | |
152 | ||
153 | spinlock_t stripe_in_journal_lock; | |
154 | atomic_t stripe_in_journal_count; | |
3bddb7f8 SL |
155 | |
156 | /* to submit async io_units, to fulfill ordering of flush */ | |
157 | struct work_struct deferred_io_work; | |
2e38a37f SL |
158 | /* to disable write back during in degraded mode */ |
159 | struct work_struct disable_writeback_work; | |
03b047f4 SL |
160 | |
161 | /* to for chunk_aligned_read in writeback mode, details below */ | |
162 | spinlock_t tree_lock; | |
163 | struct radix_tree_root big_stripe_tree; | |
f6bed0ef SL |
164 | }; |
165 | ||
03b047f4 SL |
166 | /* |
167 | * Enable chunk_aligned_read() with write back cache. | |
168 | * | |
169 | * Each chunk may contain more than one stripe (for example, a 256kB | |
170 | * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For | |
171 | * chunk_aligned_read, these stripes are grouped into one "big_stripe". | |
172 | * For each big_stripe, we count how many stripes of this big_stripe | |
173 | * are in the write back cache. These data are tracked in a radix tree | |
174 | * (big_stripe_tree). We use radix_tree item pointer as the counter. | |
175 | * r5c_tree_index() is used to calculate keys for the radix tree. | |
176 | * | |
177 | * chunk_aligned_read() calls r5c_big_stripe_cached() to look up | |
178 | * big_stripe of each chunk in the tree. If this big_stripe is in the | |
179 | * tree, chunk_aligned_read() aborts. This look up is protected by | |
180 | * rcu_read_lock(). | |
181 | * | |
182 | * It is necessary to remember whether a stripe is counted in | |
183 | * big_stripe_tree. Instead of adding new flag, we reuses existing flags: | |
184 | * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these | |
185 | * two flags are set, the stripe is counted in big_stripe_tree. This | |
186 | * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to | |
187 | * r5c_try_caching_write(); and moving clear_bit of | |
188 | * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to | |
189 | * r5c_finish_stripe_write_out(). | |
190 | */ | |
191 | ||
192 | /* | |
193 | * radix tree requests lowest 2 bits of data pointer to be 2b'00. | |
194 | * So it is necessary to left shift the counter by 2 bits before using it | |
195 | * as data pointer of the tree. | |
196 | */ | |
197 | #define R5C_RADIX_COUNT_SHIFT 2 | |
198 | ||
199 | /* | |
200 | * calculate key for big_stripe_tree | |
201 | * | |
202 | * sect: align_bi->bi_iter.bi_sector or sh->sector | |
203 | */ | |
204 | static inline sector_t r5c_tree_index(struct r5conf *conf, | |
205 | sector_t sect) | |
206 | { | |
207 | sector_t offset; | |
208 | ||
209 | offset = sector_div(sect, conf->chunk_sectors); | |
210 | return sect; | |
211 | } | |
212 | ||
f6bed0ef SL |
213 | /* |
214 | * an IO range starts from a meta data block and end at the next meta data | |
215 | * block. The io unit's the meta data block tracks data/parity followed it. io | |
216 | * unit is written to log disk with normal write, as we always flush log disk | |
217 | * first and then start move data to raid disks, there is no requirement to | |
218 | * write io unit with FLUSH/FUA | |
219 | */ | |
220 | struct r5l_io_unit { | |
221 | struct r5l_log *log; | |
222 | ||
223 | struct page *meta_page; /* store meta block */ | |
224 | int meta_offset; /* current offset in meta_page */ | |
225 | ||
f6bed0ef SL |
226 | struct bio *current_bio;/* current_bio accepting new data */ |
227 | ||
228 | atomic_t pending_stripe;/* how many stripes not flushed to raid */ | |
229 | u64 seq; /* seq number of the metablock */ | |
230 | sector_t log_start; /* where the io_unit starts */ | |
231 | sector_t log_end; /* where the io_unit ends */ | |
232 | struct list_head log_sibling; /* log->running_ios */ | |
233 | struct list_head stripe_list; /* stripes added to the io_unit */ | |
234 | ||
235 | int state; | |
6143e2ce | 236 | bool need_split_bio; |
3bddb7f8 SL |
237 | struct bio *split_bio; |
238 | ||
a9501d74 SL |
239 | unsigned int has_flush:1; /* include flush request */ |
240 | unsigned int has_fua:1; /* include fua request */ | |
241 | unsigned int has_null_flush:1; /* include null flush request */ | |
242 | unsigned int has_flush_payload:1; /* include flush payload */ | |
3bddb7f8 SL |
243 | /* |
244 | * io isn't sent yet, flush/fua request can only be submitted till it's | |
245 | * the first IO in running_ios list | |
246 | */ | |
247 | unsigned int io_deferred:1; | |
248 | ||
249 | struct bio_list flush_barriers; /* size == 0 flush bios */ | |
f6bed0ef SL |
250 | }; |
251 | ||
252 | /* r5l_io_unit state */ | |
253 | enum r5l_io_unit_state { | |
254 | IO_UNIT_RUNNING = 0, /* accepting new IO */ | |
255 | IO_UNIT_IO_START = 1, /* io_unit bio start writing to log, | |
256 | * don't accepting new bio */ | |
257 | IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */ | |
a8c34f91 | 258 | IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */ |
f6bed0ef SL |
259 | }; |
260 | ||
2ded3703 SL |
261 | bool r5c_is_writeback(struct r5l_log *log) |
262 | { | |
263 | return (log != NULL && | |
264 | log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK); | |
265 | } | |
266 | ||
f6bed0ef SL |
267 | static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc) |
268 | { | |
269 | start += inc; | |
270 | if (start >= log->device_size) | |
271 | start = start - log->device_size; | |
272 | return start; | |
273 | } | |
274 | ||
275 | static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start, | |
276 | sector_t end) | |
277 | { | |
278 | if (end >= start) | |
279 | return end - start; | |
280 | else | |
281 | return end + log->device_size - start; | |
282 | } | |
283 | ||
284 | static bool r5l_has_free_space(struct r5l_log *log, sector_t size) | |
285 | { | |
286 | sector_t used_size; | |
287 | ||
288 | used_size = r5l_ring_distance(log, log->last_checkpoint, | |
289 | log->log_start); | |
290 | ||
291 | return log->device_size > used_size + size; | |
292 | } | |
293 | ||
f6bed0ef SL |
294 | static void __r5l_set_io_unit_state(struct r5l_io_unit *io, |
295 | enum r5l_io_unit_state state) | |
296 | { | |
f6bed0ef SL |
297 | if (WARN_ON(io->state >= state)) |
298 | return; | |
299 | io->state = state; | |
f6bed0ef SL |
300 | } |
301 | ||
1e6d690b | 302 | static void |
bd83d0a2 | 303 | r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev) |
1e6d690b SL |
304 | { |
305 | struct bio *wbi, *wbi2; | |
306 | ||
307 | wbi = dev->written; | |
308 | dev->written = NULL; | |
309 | while (wbi && wbi->bi_iter.bi_sector < | |
310 | dev->sector + STRIPE_SECTORS) { | |
311 | wbi2 = r5_next_bio(wbi, dev->sector); | |
49728050 | 312 | md_write_end(conf->mddev); |
016c76ac | 313 | bio_endio(wbi); |
1e6d690b SL |
314 | wbi = wbi2; |
315 | } | |
316 | } | |
317 | ||
318 | void r5c_handle_cached_data_endio(struct r5conf *conf, | |
bd83d0a2 | 319 | struct stripe_head *sh, int disks) |
1e6d690b SL |
320 | { |
321 | int i; | |
322 | ||
323 | for (i = sh->disks; i--; ) { | |
324 | if (sh->dev[i].written) { | |
325 | set_bit(R5_UPTODATE, &sh->dev[i].flags); | |
bd83d0a2 | 326 | r5c_return_dev_pending_writes(conf, &sh->dev[i]); |
1e6d690b SL |
327 | bitmap_endwrite(conf->mddev->bitmap, sh->sector, |
328 | STRIPE_SECTORS, | |
329 | !test_bit(STRIPE_DEGRADED, &sh->state), | |
330 | 0); | |
331 | } | |
332 | } | |
333 | } | |
334 | ||
ff875738 AP |
335 | void r5l_wake_reclaim(struct r5l_log *log, sector_t space); |
336 | ||
a39f7afd SL |
337 | /* Check whether we should flush some stripes to free up stripe cache */ |
338 | void r5c_check_stripe_cache_usage(struct r5conf *conf) | |
339 | { | |
340 | int total_cached; | |
341 | ||
342 | if (!r5c_is_writeback(conf->log)) | |
343 | return; | |
344 | ||
345 | total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + | |
346 | atomic_read(&conf->r5c_cached_full_stripes); | |
347 | ||
348 | /* | |
349 | * The following condition is true for either of the following: | |
350 | * - stripe cache pressure high: | |
351 | * total_cached > 3/4 min_nr_stripes || | |
352 | * empty_inactive_list_nr > 0 | |
353 | * - stripe cache pressure moderate: | |
354 | * total_cached > 1/2 min_nr_stripes | |
355 | */ | |
356 | if (total_cached > conf->min_nr_stripes * 1 / 2 || | |
357 | atomic_read(&conf->empty_inactive_list_nr) > 0) | |
358 | r5l_wake_reclaim(conf->log, 0); | |
359 | } | |
360 | ||
361 | /* | |
362 | * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full | |
363 | * stripes in the cache | |
364 | */ | |
365 | void r5c_check_cached_full_stripe(struct r5conf *conf) | |
366 | { | |
367 | if (!r5c_is_writeback(conf->log)) | |
368 | return; | |
369 | ||
370 | /* | |
371 | * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes | |
372 | * or a full stripe (chunk size / 4k stripes). | |
373 | */ | |
374 | if (atomic_read(&conf->r5c_cached_full_stripes) >= | |
84890c03 | 375 | min(R5C_FULL_STRIPE_FLUSH_BATCH(conf), |
a39f7afd SL |
376 | conf->chunk_sectors >> STRIPE_SHIFT)) |
377 | r5l_wake_reclaim(conf->log, 0); | |
378 | } | |
379 | ||
380 | /* | |
381 | * Total log space (in sectors) needed to flush all data in cache | |
382 | * | |
39b99586 SL |
383 | * To avoid deadlock due to log space, it is necessary to reserve log |
384 | * space to flush critical stripes (stripes that occupying log space near | |
385 | * last_checkpoint). This function helps check how much log space is | |
386 | * required to flush all cached stripes. | |
a39f7afd | 387 | * |
39b99586 SL |
388 | * To reduce log space requirements, two mechanisms are used to give cache |
389 | * flush higher priorities: | |
390 | * 1. In handle_stripe_dirtying() and schedule_reconstruction(), | |
391 | * stripes ALREADY in journal can be flushed w/o pending writes; | |
392 | * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal | |
393 | * can be delayed (r5l_add_no_space_stripe). | |
a39f7afd | 394 | * |
39b99586 SL |
395 | * In cache flush, the stripe goes through 1 and then 2. For a stripe that |
396 | * already passed 1, flushing it requires at most (conf->max_degraded + 1) | |
397 | * pages of journal space. For stripes that has not passed 1, flushing it | |
398 | * requires (conf->raid_disks + 1) pages of journal space. There are at | |
399 | * most (conf->group_cnt + 1) stripe that passed 1. So total journal space | |
400 | * required to flush all cached stripes (in pages) is: | |
401 | * | |
402 | * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) + | |
403 | * (group_cnt + 1) * (raid_disks + 1) | |
404 | * or | |
405 | * (stripe_in_journal_count) * (max_degraded + 1) + | |
406 | * (group_cnt + 1) * (raid_disks - max_degraded) | |
a39f7afd SL |
407 | */ |
408 | static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf) | |
409 | { | |
410 | struct r5l_log *log = conf->log; | |
411 | ||
412 | if (!r5c_is_writeback(log)) | |
413 | return 0; | |
414 | ||
39b99586 SL |
415 | return BLOCK_SECTORS * |
416 | ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) + | |
417 | (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1)); | |
a39f7afd SL |
418 | } |
419 | ||
420 | /* | |
421 | * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL | |
422 | * | |
423 | * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of | |
424 | * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log | |
425 | * device is less than 2x of reclaim_required_space. | |
426 | */ | |
427 | static inline void r5c_update_log_state(struct r5l_log *log) | |
428 | { | |
429 | struct r5conf *conf = log->rdev->mddev->private; | |
430 | sector_t free_space; | |
431 | sector_t reclaim_space; | |
f687a33e | 432 | bool wake_reclaim = false; |
a39f7afd SL |
433 | |
434 | if (!r5c_is_writeback(log)) | |
435 | return; | |
436 | ||
437 | free_space = r5l_ring_distance(log, log->log_start, | |
438 | log->last_checkpoint); | |
439 | reclaim_space = r5c_log_required_to_flush_cache(conf); | |
440 | if (free_space < 2 * reclaim_space) | |
441 | set_bit(R5C_LOG_CRITICAL, &conf->cache_state); | |
f687a33e SL |
442 | else { |
443 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state)) | |
444 | wake_reclaim = true; | |
a39f7afd | 445 | clear_bit(R5C_LOG_CRITICAL, &conf->cache_state); |
f687a33e | 446 | } |
a39f7afd SL |
447 | if (free_space < 3 * reclaim_space) |
448 | set_bit(R5C_LOG_TIGHT, &conf->cache_state); | |
449 | else | |
450 | clear_bit(R5C_LOG_TIGHT, &conf->cache_state); | |
f687a33e SL |
451 | |
452 | if (wake_reclaim) | |
453 | r5l_wake_reclaim(log, 0); | |
a39f7afd SL |
454 | } |
455 | ||
2ded3703 SL |
456 | /* |
457 | * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING. | |
458 | * This function should only be called in write-back mode. | |
459 | */ | |
a39f7afd | 460 | void r5c_make_stripe_write_out(struct stripe_head *sh) |
2ded3703 SL |
461 | { |
462 | struct r5conf *conf = sh->raid_conf; | |
463 | struct r5l_log *log = conf->log; | |
464 | ||
465 | BUG_ON(!r5c_is_writeback(log)); | |
466 | ||
467 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
468 | clear_bit(STRIPE_R5C_CACHING, &sh->state); | |
1e6d690b SL |
469 | |
470 | if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) | |
471 | atomic_inc(&conf->preread_active_stripes); | |
1e6d690b SL |
472 | } |
473 | ||
474 | static void r5c_handle_data_cached(struct stripe_head *sh) | |
475 | { | |
476 | int i; | |
477 | ||
478 | for (i = sh->disks; i--; ) | |
479 | if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) { | |
480 | set_bit(R5_InJournal, &sh->dev[i].flags); | |
481 | clear_bit(R5_LOCKED, &sh->dev[i].flags); | |
482 | } | |
483 | clear_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
484 | } | |
485 | ||
486 | /* | |
487 | * this journal write must contain full parity, | |
488 | * it may also contain some data pages | |
489 | */ | |
490 | static void r5c_handle_parity_cached(struct stripe_head *sh) | |
491 | { | |
492 | int i; | |
493 | ||
494 | for (i = sh->disks; i--; ) | |
495 | if (test_bit(R5_InJournal, &sh->dev[i].flags)) | |
496 | set_bit(R5_Wantwrite, &sh->dev[i].flags); | |
2ded3703 SL |
497 | } |
498 | ||
499 | /* | |
500 | * Setting proper flags after writing (or flushing) data and/or parity to the | |
501 | * log device. This is called from r5l_log_endio() or r5l_log_flush_endio(). | |
502 | */ | |
503 | static void r5c_finish_cache_stripe(struct stripe_head *sh) | |
504 | { | |
505 | struct r5l_log *log = sh->raid_conf->log; | |
506 | ||
507 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { | |
508 | BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
509 | /* | |
510 | * Set R5_InJournal for parity dev[pd_idx]. This means | |
511 | * all data AND parity in the journal. For RAID 6, it is | |
512 | * NOT necessary to set the flag for dev[qd_idx], as the | |
513 | * two parities are written out together. | |
514 | */ | |
515 | set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
1e6d690b SL |
516 | } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) { |
517 | r5c_handle_data_cached(sh); | |
518 | } else { | |
519 | r5c_handle_parity_cached(sh); | |
520 | set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
521 | } | |
2ded3703 SL |
522 | } |
523 | ||
d8858f43 CH |
524 | static void r5l_io_run_stripes(struct r5l_io_unit *io) |
525 | { | |
526 | struct stripe_head *sh, *next; | |
527 | ||
528 | list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) { | |
529 | list_del_init(&sh->log_list); | |
2ded3703 SL |
530 | |
531 | r5c_finish_cache_stripe(sh); | |
532 | ||
d8858f43 CH |
533 | set_bit(STRIPE_HANDLE, &sh->state); |
534 | raid5_release_stripe(sh); | |
535 | } | |
536 | } | |
537 | ||
56fef7c6 CH |
538 | static void r5l_log_run_stripes(struct r5l_log *log) |
539 | { | |
540 | struct r5l_io_unit *io, *next; | |
541 | ||
542 | assert_spin_locked(&log->io_list_lock); | |
543 | ||
544 | list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { | |
545 | /* don't change list order */ | |
546 | if (io->state < IO_UNIT_IO_END) | |
547 | break; | |
548 | ||
549 | list_move_tail(&io->log_sibling, &log->finished_ios); | |
550 | r5l_io_run_stripes(io); | |
551 | } | |
552 | } | |
553 | ||
3848c0bc CH |
554 | static void r5l_move_to_end_ios(struct r5l_log *log) |
555 | { | |
556 | struct r5l_io_unit *io, *next; | |
557 | ||
558 | assert_spin_locked(&log->io_list_lock); | |
559 | ||
560 | list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { | |
561 | /* don't change list order */ | |
562 | if (io->state < IO_UNIT_IO_END) | |
563 | break; | |
564 | list_move_tail(&io->log_sibling, &log->io_end_ios); | |
565 | } | |
566 | } | |
567 | ||
3bddb7f8 | 568 | static void __r5l_stripe_write_finished(struct r5l_io_unit *io); |
f6bed0ef SL |
569 | static void r5l_log_endio(struct bio *bio) |
570 | { | |
571 | struct r5l_io_unit *io = bio->bi_private; | |
3bddb7f8 | 572 | struct r5l_io_unit *io_deferred; |
f6bed0ef | 573 | struct r5l_log *log = io->log; |
509ffec7 | 574 | unsigned long flags; |
a9501d74 SL |
575 | bool has_null_flush; |
576 | bool has_flush_payload; | |
f6bed0ef | 577 | |
4e4cbee9 | 578 | if (bio->bi_status) |
6e74a9cf SL |
579 | md_error(log->rdev->mddev, log->rdev); |
580 | ||
f6bed0ef | 581 | bio_put(bio); |
e8deb638 | 582 | mempool_free(io->meta_page, log->meta_pool); |
f6bed0ef | 583 | |
509ffec7 CH |
584 | spin_lock_irqsave(&log->io_list_lock, flags); |
585 | __r5l_set_io_unit_state(io, IO_UNIT_IO_END); | |
a9501d74 SL |
586 | |
587 | /* | |
588 | * if the io doesn't not have null_flush or flush payload, | |
589 | * it is not safe to access it after releasing io_list_lock. | |
590 | * Therefore, it is necessary to check the condition with | |
591 | * the lock held. | |
592 | */ | |
593 | has_null_flush = io->has_null_flush; | |
594 | has_flush_payload = io->has_flush_payload; | |
595 | ||
ea17481f | 596 | if (log->need_cache_flush && !list_empty(&io->stripe_list)) |
3848c0bc | 597 | r5l_move_to_end_ios(log); |
56fef7c6 CH |
598 | else |
599 | r5l_log_run_stripes(log); | |
3bddb7f8 SL |
600 | if (!list_empty(&log->running_ios)) { |
601 | /* | |
602 | * FLUSH/FUA io_unit is deferred because of ordering, now we | |
603 | * can dispatch it | |
604 | */ | |
605 | io_deferred = list_first_entry(&log->running_ios, | |
606 | struct r5l_io_unit, log_sibling); | |
607 | if (io_deferred->io_deferred) | |
608 | schedule_work(&log->deferred_io_work); | |
609 | } | |
610 | ||
509ffec7 CH |
611 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
612 | ||
56fef7c6 CH |
613 | if (log->need_cache_flush) |
614 | md_wakeup_thread(log->rdev->mddev->thread); | |
3bddb7f8 | 615 | |
a9501d74 SL |
616 | /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */ |
617 | if (has_null_flush) { | |
3bddb7f8 SL |
618 | struct bio *bi; |
619 | ||
620 | WARN_ON(bio_list_empty(&io->flush_barriers)); | |
621 | while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) { | |
622 | bio_endio(bi); | |
a9501d74 SL |
623 | if (atomic_dec_and_test(&io->pending_stripe)) { |
624 | __r5l_stripe_write_finished(io); | |
625 | return; | |
626 | } | |
3bddb7f8 | 627 | } |
3bddb7f8 | 628 | } |
a9501d74 SL |
629 | /* decrease pending_stripe for flush payload */ |
630 | if (has_flush_payload) | |
631 | if (atomic_dec_and_test(&io->pending_stripe)) | |
632 | __r5l_stripe_write_finished(io); | |
3bddb7f8 SL |
633 | } |
634 | ||
635 | static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io) | |
636 | { | |
637 | unsigned long flags; | |
638 | ||
639 | spin_lock_irqsave(&log->io_list_lock, flags); | |
640 | __r5l_set_io_unit_state(io, IO_UNIT_IO_START); | |
641 | spin_unlock_irqrestore(&log->io_list_lock, flags); | |
642 | ||
bb3338d3 SL |
643 | /* |
644 | * In case of journal device failures, submit_bio will get error | |
645 | * and calls endio, then active stripes will continue write | |
646 | * process. Therefore, it is not necessary to check Faulty bit | |
647 | * of journal device here. | |
648 | * | |
649 | * We can't check split_bio after current_bio is submitted. If | |
650 | * io->split_bio is null, after current_bio is submitted, current_bio | |
651 | * might already be completed and the io_unit is freed. We submit | |
652 | * split_bio first to avoid the issue. | |
653 | */ | |
654 | if (io->split_bio) { | |
655 | if (io->has_flush) | |
656 | io->split_bio->bi_opf |= REQ_PREFLUSH; | |
657 | if (io->has_fua) | |
658 | io->split_bio->bi_opf |= REQ_FUA; | |
659 | submit_bio(io->split_bio); | |
660 | } | |
661 | ||
3bddb7f8 | 662 | if (io->has_flush) |
20737738 | 663 | io->current_bio->bi_opf |= REQ_PREFLUSH; |
3bddb7f8 | 664 | if (io->has_fua) |
20737738 | 665 | io->current_bio->bi_opf |= REQ_FUA; |
3bddb7f8 | 666 | submit_bio(io->current_bio); |
3bddb7f8 SL |
667 | } |
668 | ||
669 | /* deferred io_unit will be dispatched here */ | |
670 | static void r5l_submit_io_async(struct work_struct *work) | |
671 | { | |
672 | struct r5l_log *log = container_of(work, struct r5l_log, | |
673 | deferred_io_work); | |
674 | struct r5l_io_unit *io = NULL; | |
675 | unsigned long flags; | |
676 | ||
677 | spin_lock_irqsave(&log->io_list_lock, flags); | |
678 | if (!list_empty(&log->running_ios)) { | |
679 | io = list_first_entry(&log->running_ios, struct r5l_io_unit, | |
680 | log_sibling); | |
681 | if (!io->io_deferred) | |
682 | io = NULL; | |
683 | else | |
684 | io->io_deferred = 0; | |
685 | } | |
686 | spin_unlock_irqrestore(&log->io_list_lock, flags); | |
687 | if (io) | |
688 | r5l_do_submit_io(log, io); | |
f6bed0ef SL |
689 | } |
690 | ||
2e38a37f SL |
691 | static void r5c_disable_writeback_async(struct work_struct *work) |
692 | { | |
693 | struct r5l_log *log = container_of(work, struct r5l_log, | |
694 | disable_writeback_work); | |
695 | struct mddev *mddev = log->rdev->mddev; | |
696 | ||
697 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) | |
698 | return; | |
699 | pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n", | |
700 | mdname(mddev)); | |
70d466f7 SL |
701 | |
702 | /* wait superblock change before suspend */ | |
703 | wait_event(mddev->sb_wait, | |
704 | !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)); | |
705 | ||
2e38a37f SL |
706 | mddev_suspend(mddev); |
707 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; | |
708 | mddev_resume(mddev); | |
709 | } | |
710 | ||
f6bed0ef SL |
711 | static void r5l_submit_current_io(struct r5l_log *log) |
712 | { | |
713 | struct r5l_io_unit *io = log->current_io; | |
3bddb7f8 | 714 | struct bio *bio; |
f6bed0ef | 715 | struct r5l_meta_block *block; |
509ffec7 | 716 | unsigned long flags; |
f6bed0ef | 717 | u32 crc; |
3bddb7f8 | 718 | bool do_submit = true; |
f6bed0ef SL |
719 | |
720 | if (!io) | |
721 | return; | |
722 | ||
723 | block = page_address(io->meta_page); | |
724 | block->meta_size = cpu_to_le32(io->meta_offset); | |
5cb2fbd6 | 725 | crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE); |
f6bed0ef | 726 | block->checksum = cpu_to_le32(crc); |
3bddb7f8 | 727 | bio = io->current_bio; |
f6bed0ef SL |
728 | |
729 | log->current_io = NULL; | |
509ffec7 | 730 | spin_lock_irqsave(&log->io_list_lock, flags); |
3bddb7f8 SL |
731 | if (io->has_flush || io->has_fua) { |
732 | if (io != list_first_entry(&log->running_ios, | |
733 | struct r5l_io_unit, log_sibling)) { | |
734 | io->io_deferred = 1; | |
735 | do_submit = false; | |
736 | } | |
737 | } | |
509ffec7 | 738 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
3bddb7f8 SL |
739 | if (do_submit) |
740 | r5l_do_submit_io(log, io); | |
f6bed0ef SL |
741 | } |
742 | ||
6143e2ce | 743 | static struct bio *r5l_bio_alloc(struct r5l_log *log) |
b349feb3 | 744 | { |
c38d29b3 | 745 | struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs); |
b349feb3 | 746 | |
796a5cf0 | 747 | bio_set_op_attrs(bio, REQ_OP_WRITE, 0); |
b349feb3 | 748 | bio->bi_bdev = log->rdev->bdev; |
1e932a37 | 749 | bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start; |
b349feb3 | 750 | |
b349feb3 CH |
751 | return bio; |
752 | } | |
753 | ||
c1b99198 CH |
754 | static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io) |
755 | { | |
756 | log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS); | |
757 | ||
a39f7afd | 758 | r5c_update_log_state(log); |
c1b99198 CH |
759 | /* |
760 | * If we filled up the log device start from the beginning again, | |
761 | * which will require a new bio. | |
762 | * | |
763 | * Note: for this to work properly the log size needs to me a multiple | |
764 | * of BLOCK_SECTORS. | |
765 | */ | |
766 | if (log->log_start == 0) | |
6143e2ce | 767 | io->need_split_bio = true; |
c1b99198 CH |
768 | |
769 | io->log_end = log->log_start; | |
770 | } | |
771 | ||
f6bed0ef SL |
772 | static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log) |
773 | { | |
774 | struct r5l_io_unit *io; | |
775 | struct r5l_meta_block *block; | |
f6bed0ef | 776 | |
5036c390 CH |
777 | io = mempool_alloc(log->io_pool, GFP_ATOMIC); |
778 | if (!io) | |
779 | return NULL; | |
780 | memset(io, 0, sizeof(*io)); | |
781 | ||
51039cd0 | 782 | io->log = log; |
51039cd0 CH |
783 | INIT_LIST_HEAD(&io->log_sibling); |
784 | INIT_LIST_HEAD(&io->stripe_list); | |
3bddb7f8 | 785 | bio_list_init(&io->flush_barriers); |
51039cd0 | 786 | io->state = IO_UNIT_RUNNING; |
f6bed0ef | 787 | |
e8deb638 | 788 | io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO); |
f6bed0ef | 789 | block = page_address(io->meta_page); |
e8deb638 | 790 | clear_page(block); |
f6bed0ef SL |
791 | block->magic = cpu_to_le32(R5LOG_MAGIC); |
792 | block->version = R5LOG_VERSION; | |
793 | block->seq = cpu_to_le64(log->seq); | |
794 | block->position = cpu_to_le64(log->log_start); | |
795 | ||
796 | io->log_start = log->log_start; | |
797 | io->meta_offset = sizeof(struct r5l_meta_block); | |
2b8ef16e | 798 | io->seq = log->seq++; |
f6bed0ef | 799 | |
6143e2ce CH |
800 | io->current_bio = r5l_bio_alloc(log); |
801 | io->current_bio->bi_end_io = r5l_log_endio; | |
802 | io->current_bio->bi_private = io; | |
b349feb3 | 803 | bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0); |
f6bed0ef | 804 | |
c1b99198 | 805 | r5_reserve_log_entry(log, io); |
f6bed0ef SL |
806 | |
807 | spin_lock_irq(&log->io_list_lock); | |
808 | list_add_tail(&io->log_sibling, &log->running_ios); | |
809 | spin_unlock_irq(&log->io_list_lock); | |
810 | ||
811 | return io; | |
812 | } | |
813 | ||
814 | static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size) | |
815 | { | |
22581f58 CH |
816 | if (log->current_io && |
817 | log->current_io->meta_offset + payload_size > PAGE_SIZE) | |
f6bed0ef | 818 | r5l_submit_current_io(log); |
f6bed0ef | 819 | |
5036c390 | 820 | if (!log->current_io) { |
22581f58 | 821 | log->current_io = r5l_new_meta(log); |
5036c390 CH |
822 | if (!log->current_io) |
823 | return -ENOMEM; | |
824 | } | |
825 | ||
f6bed0ef SL |
826 | return 0; |
827 | } | |
828 | ||
829 | static void r5l_append_payload_meta(struct r5l_log *log, u16 type, | |
830 | sector_t location, | |
831 | u32 checksum1, u32 checksum2, | |
832 | bool checksum2_valid) | |
833 | { | |
834 | struct r5l_io_unit *io = log->current_io; | |
835 | struct r5l_payload_data_parity *payload; | |
836 | ||
837 | payload = page_address(io->meta_page) + io->meta_offset; | |
838 | payload->header.type = cpu_to_le16(type); | |
839 | payload->header.flags = cpu_to_le16(0); | |
840 | payload->size = cpu_to_le32((1 + !!checksum2_valid) << | |
841 | (PAGE_SHIFT - 9)); | |
842 | payload->location = cpu_to_le64(location); | |
843 | payload->checksum[0] = cpu_to_le32(checksum1); | |
844 | if (checksum2_valid) | |
845 | payload->checksum[1] = cpu_to_le32(checksum2); | |
846 | ||
847 | io->meta_offset += sizeof(struct r5l_payload_data_parity) + | |
848 | sizeof(__le32) * (1 + !!checksum2_valid); | |
849 | } | |
850 | ||
851 | static void r5l_append_payload_page(struct r5l_log *log, struct page *page) | |
852 | { | |
853 | struct r5l_io_unit *io = log->current_io; | |
854 | ||
6143e2ce | 855 | if (io->need_split_bio) { |
3bddb7f8 SL |
856 | BUG_ON(io->split_bio); |
857 | io->split_bio = io->current_bio; | |
6143e2ce | 858 | io->current_bio = r5l_bio_alloc(log); |
3bddb7f8 SL |
859 | bio_chain(io->current_bio, io->split_bio); |
860 | io->need_split_bio = false; | |
f6bed0ef | 861 | } |
f6bed0ef | 862 | |
6143e2ce CH |
863 | if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) |
864 | BUG(); | |
865 | ||
c1b99198 | 866 | r5_reserve_log_entry(log, io); |
f6bed0ef SL |
867 | } |
868 | ||
ea17481f SL |
869 | static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect) |
870 | { | |
871 | struct mddev *mddev = log->rdev->mddev; | |
872 | struct r5conf *conf = mddev->private; | |
873 | struct r5l_io_unit *io; | |
874 | struct r5l_payload_flush *payload; | |
875 | int meta_size; | |
876 | ||
877 | /* | |
878 | * payload_flush requires extra writes to the journal. | |
879 | * To avoid handling the extra IO in quiesce, just skip | |
880 | * flush_payload | |
881 | */ | |
882 | if (conf->quiesce) | |
883 | return; | |
884 | ||
885 | mutex_lock(&log->io_mutex); | |
886 | meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64); | |
887 | ||
888 | if (r5l_get_meta(log, meta_size)) { | |
889 | mutex_unlock(&log->io_mutex); | |
890 | return; | |
891 | } | |
892 | ||
893 | /* current implementation is one stripe per flush payload */ | |
894 | io = log->current_io; | |
895 | payload = page_address(io->meta_page) + io->meta_offset; | |
896 | payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH); | |
897 | payload->header.flags = cpu_to_le16(0); | |
898 | payload->size = cpu_to_le32(sizeof(__le64)); | |
899 | payload->flush_stripes[0] = cpu_to_le64(sect); | |
900 | io->meta_offset += meta_size; | |
a9501d74 SL |
901 | /* multiple flush payloads count as one pending_stripe */ |
902 | if (!io->has_flush_payload) { | |
903 | io->has_flush_payload = 1; | |
904 | atomic_inc(&io->pending_stripe); | |
905 | } | |
ea17481f SL |
906 | mutex_unlock(&log->io_mutex); |
907 | } | |
908 | ||
5036c390 | 909 | static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh, |
f6bed0ef SL |
910 | int data_pages, int parity_pages) |
911 | { | |
912 | int i; | |
913 | int meta_size; | |
5036c390 | 914 | int ret; |
f6bed0ef SL |
915 | struct r5l_io_unit *io; |
916 | ||
917 | meta_size = | |
918 | ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) | |
919 | * data_pages) + | |
920 | sizeof(struct r5l_payload_data_parity) + | |
921 | sizeof(__le32) * parity_pages; | |
922 | ||
5036c390 CH |
923 | ret = r5l_get_meta(log, meta_size); |
924 | if (ret) | |
925 | return ret; | |
926 | ||
f6bed0ef SL |
927 | io = log->current_io; |
928 | ||
3bddb7f8 SL |
929 | if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state)) |
930 | io->has_flush = 1; | |
931 | ||
f6bed0ef | 932 | for (i = 0; i < sh->disks; i++) { |
1e6d690b SL |
933 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || |
934 | test_bit(R5_InJournal, &sh->dev[i].flags)) | |
f6bed0ef SL |
935 | continue; |
936 | if (i == sh->pd_idx || i == sh->qd_idx) | |
937 | continue; | |
3bddb7f8 SL |
938 | if (test_bit(R5_WantFUA, &sh->dev[i].flags) && |
939 | log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) { | |
940 | io->has_fua = 1; | |
941 | /* | |
942 | * we need to flush journal to make sure recovery can | |
943 | * reach the data with fua flag | |
944 | */ | |
945 | io->has_flush = 1; | |
946 | } | |
f6bed0ef SL |
947 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA, |
948 | raid5_compute_blocknr(sh, i, 0), | |
949 | sh->dev[i].log_checksum, 0, false); | |
950 | r5l_append_payload_page(log, sh->dev[i].page); | |
951 | } | |
952 | ||
2ded3703 | 953 | if (parity_pages == 2) { |
f6bed0ef SL |
954 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, |
955 | sh->sector, sh->dev[sh->pd_idx].log_checksum, | |
956 | sh->dev[sh->qd_idx].log_checksum, true); | |
957 | r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); | |
958 | r5l_append_payload_page(log, sh->dev[sh->qd_idx].page); | |
2ded3703 | 959 | } else if (parity_pages == 1) { |
f6bed0ef SL |
960 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, |
961 | sh->sector, sh->dev[sh->pd_idx].log_checksum, | |
962 | 0, false); | |
963 | r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); | |
2ded3703 SL |
964 | } else /* Just writing data, not parity, in caching phase */ |
965 | BUG_ON(parity_pages != 0); | |
f6bed0ef SL |
966 | |
967 | list_add_tail(&sh->log_list, &io->stripe_list); | |
968 | atomic_inc(&io->pending_stripe); | |
969 | sh->log_io = io; | |
5036c390 | 970 | |
a39f7afd SL |
971 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) |
972 | return 0; | |
973 | ||
974 | if (sh->log_start == MaxSector) { | |
975 | BUG_ON(!list_empty(&sh->r5c)); | |
976 | sh->log_start = io->log_start; | |
977 | spin_lock_irq(&log->stripe_in_journal_lock); | |
978 | list_add_tail(&sh->r5c, | |
979 | &log->stripe_in_journal_list); | |
980 | spin_unlock_irq(&log->stripe_in_journal_lock); | |
981 | atomic_inc(&log->stripe_in_journal_count); | |
982 | } | |
5036c390 | 983 | return 0; |
f6bed0ef SL |
984 | } |
985 | ||
a39f7afd SL |
986 | /* add stripe to no_space_stripes, and then wake up reclaim */ |
987 | static inline void r5l_add_no_space_stripe(struct r5l_log *log, | |
988 | struct stripe_head *sh) | |
989 | { | |
990 | spin_lock(&log->no_space_stripes_lock); | |
991 | list_add_tail(&sh->log_list, &log->no_space_stripes); | |
992 | spin_unlock(&log->no_space_stripes_lock); | |
993 | } | |
994 | ||
f6bed0ef SL |
995 | /* |
996 | * running in raid5d, where reclaim could wait for raid5d too (when it flushes | |
997 | * data from log to raid disks), so we shouldn't wait for reclaim here | |
998 | */ | |
999 | int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh) | |
1000 | { | |
a39f7afd | 1001 | struct r5conf *conf = sh->raid_conf; |
f6bed0ef SL |
1002 | int write_disks = 0; |
1003 | int data_pages, parity_pages; | |
f6bed0ef SL |
1004 | int reserve; |
1005 | int i; | |
5036c390 | 1006 | int ret = 0; |
a39f7afd | 1007 | bool wake_reclaim = false; |
f6bed0ef SL |
1008 | |
1009 | if (!log) | |
1010 | return -EAGAIN; | |
1011 | /* Don't support stripe batch */ | |
1012 | if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) || | |
1013 | test_bit(STRIPE_SYNCING, &sh->state)) { | |
1014 | /* the stripe is written to log, we start writing it to raid */ | |
1015 | clear_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
1016 | return -EAGAIN; | |
1017 | } | |
1018 | ||
2ded3703 SL |
1019 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); |
1020 | ||
f6bed0ef SL |
1021 | for (i = 0; i < sh->disks; i++) { |
1022 | void *addr; | |
1023 | ||
1e6d690b SL |
1024 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || |
1025 | test_bit(R5_InJournal, &sh->dev[i].flags)) | |
f6bed0ef | 1026 | continue; |
1e6d690b | 1027 | |
f6bed0ef SL |
1028 | write_disks++; |
1029 | /* checksum is already calculated in last run */ | |
1030 | if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) | |
1031 | continue; | |
1032 | addr = kmap_atomic(sh->dev[i].page); | |
5cb2fbd6 SL |
1033 | sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, |
1034 | addr, PAGE_SIZE); | |
f6bed0ef SL |
1035 | kunmap_atomic(addr); |
1036 | } | |
1037 | parity_pages = 1 + !!(sh->qd_idx >= 0); | |
1038 | data_pages = write_disks - parity_pages; | |
1039 | ||
f6bed0ef | 1040 | set_bit(STRIPE_LOG_TRAPPED, &sh->state); |
253f9fd4 SL |
1041 | /* |
1042 | * The stripe must enter state machine again to finish the write, so | |
1043 | * don't delay. | |
1044 | */ | |
1045 | clear_bit(STRIPE_DELAYED, &sh->state); | |
f6bed0ef SL |
1046 | atomic_inc(&sh->count); |
1047 | ||
1048 | mutex_lock(&log->io_mutex); | |
1049 | /* meta + data */ | |
1050 | reserve = (1 + write_disks) << (PAGE_SHIFT - 9); | |
f6bed0ef | 1051 | |
a39f7afd SL |
1052 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { |
1053 | if (!r5l_has_free_space(log, reserve)) { | |
1054 | r5l_add_no_space_stripe(log, sh); | |
1055 | wake_reclaim = true; | |
1056 | } else { | |
1057 | ret = r5l_log_stripe(log, sh, data_pages, parity_pages); | |
1058 | if (ret) { | |
1059 | spin_lock_irq(&log->io_list_lock); | |
1060 | list_add_tail(&sh->log_list, | |
1061 | &log->no_mem_stripes); | |
1062 | spin_unlock_irq(&log->io_list_lock); | |
1063 | } | |
1064 | } | |
1065 | } else { /* R5C_JOURNAL_MODE_WRITE_BACK */ | |
1066 | /* | |
1067 | * log space critical, do not process stripes that are | |
1068 | * not in cache yet (sh->log_start == MaxSector). | |
1069 | */ | |
1070 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && | |
1071 | sh->log_start == MaxSector) { | |
1072 | r5l_add_no_space_stripe(log, sh); | |
1073 | wake_reclaim = true; | |
1074 | reserve = 0; | |
1075 | } else if (!r5l_has_free_space(log, reserve)) { | |
1076 | if (sh->log_start == log->last_checkpoint) | |
1077 | BUG(); | |
1078 | else | |
1079 | r5l_add_no_space_stripe(log, sh); | |
1080 | } else { | |
1081 | ret = r5l_log_stripe(log, sh, data_pages, parity_pages); | |
1082 | if (ret) { | |
1083 | spin_lock_irq(&log->io_list_lock); | |
1084 | list_add_tail(&sh->log_list, | |
1085 | &log->no_mem_stripes); | |
1086 | spin_unlock_irq(&log->io_list_lock); | |
1087 | } | |
5036c390 | 1088 | } |
f6bed0ef | 1089 | } |
f6bed0ef | 1090 | |
5036c390 | 1091 | mutex_unlock(&log->io_mutex); |
a39f7afd SL |
1092 | if (wake_reclaim) |
1093 | r5l_wake_reclaim(log, reserve); | |
f6bed0ef SL |
1094 | return 0; |
1095 | } | |
1096 | ||
1097 | void r5l_write_stripe_run(struct r5l_log *log) | |
1098 | { | |
1099 | if (!log) | |
1100 | return; | |
1101 | mutex_lock(&log->io_mutex); | |
1102 | r5l_submit_current_io(log); | |
1103 | mutex_unlock(&log->io_mutex); | |
1104 | } | |
1105 | ||
828cbe98 SL |
1106 | int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio) |
1107 | { | |
1108 | if (!log) | |
1109 | return -ENODEV; | |
3bddb7f8 SL |
1110 | |
1111 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { | |
1112 | /* | |
1113 | * in write through (journal only) | |
1114 | * we flush log disk cache first, then write stripe data to | |
1115 | * raid disks. So if bio is finished, the log disk cache is | |
1116 | * flushed already. The recovery guarantees we can recovery | |
1117 | * the bio from log disk, so we don't need to flush again | |
1118 | */ | |
1119 | if (bio->bi_iter.bi_size == 0) { | |
1120 | bio_endio(bio); | |
1121 | return 0; | |
1122 | } | |
1123 | bio->bi_opf &= ~REQ_PREFLUSH; | |
1124 | } else { | |
1125 | /* write back (with cache) */ | |
1126 | if (bio->bi_iter.bi_size == 0) { | |
1127 | mutex_lock(&log->io_mutex); | |
1128 | r5l_get_meta(log, 0); | |
1129 | bio_list_add(&log->current_io->flush_barriers, bio); | |
1130 | log->current_io->has_flush = 1; | |
1131 | log->current_io->has_null_flush = 1; | |
1132 | atomic_inc(&log->current_io->pending_stripe); | |
1133 | r5l_submit_current_io(log); | |
1134 | mutex_unlock(&log->io_mutex); | |
1135 | return 0; | |
1136 | } | |
828cbe98 | 1137 | } |
828cbe98 SL |
1138 | return -EAGAIN; |
1139 | } | |
1140 | ||
f6bed0ef SL |
1141 | /* This will run after log space is reclaimed */ |
1142 | static void r5l_run_no_space_stripes(struct r5l_log *log) | |
1143 | { | |
1144 | struct stripe_head *sh; | |
1145 | ||
1146 | spin_lock(&log->no_space_stripes_lock); | |
1147 | while (!list_empty(&log->no_space_stripes)) { | |
1148 | sh = list_first_entry(&log->no_space_stripes, | |
1149 | struct stripe_head, log_list); | |
1150 | list_del_init(&sh->log_list); | |
1151 | set_bit(STRIPE_HANDLE, &sh->state); | |
1152 | raid5_release_stripe(sh); | |
1153 | } | |
1154 | spin_unlock(&log->no_space_stripes_lock); | |
1155 | } | |
1156 | ||
a39f7afd SL |
1157 | /* |
1158 | * calculate new last_checkpoint | |
1159 | * for write through mode, returns log->next_checkpoint | |
1160 | * for write back, returns log_start of first sh in stripe_in_journal_list | |
1161 | */ | |
1162 | static sector_t r5c_calculate_new_cp(struct r5conf *conf) | |
1163 | { | |
1164 | struct stripe_head *sh; | |
1165 | struct r5l_log *log = conf->log; | |
1166 | sector_t new_cp; | |
1167 | unsigned long flags; | |
1168 | ||
1169 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) | |
1170 | return log->next_checkpoint; | |
1171 | ||
1172 | spin_lock_irqsave(&log->stripe_in_journal_lock, flags); | |
1173 | if (list_empty(&conf->log->stripe_in_journal_list)) { | |
1174 | /* all stripes flushed */ | |
d3014e21 | 1175 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); |
a39f7afd SL |
1176 | return log->next_checkpoint; |
1177 | } | |
1178 | sh = list_first_entry(&conf->log->stripe_in_journal_list, | |
1179 | struct stripe_head, r5c); | |
1180 | new_cp = sh->log_start; | |
1181 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); | |
1182 | return new_cp; | |
1183 | } | |
1184 | ||
17036461 CH |
1185 | static sector_t r5l_reclaimable_space(struct r5l_log *log) |
1186 | { | |
a39f7afd SL |
1187 | struct r5conf *conf = log->rdev->mddev->private; |
1188 | ||
17036461 | 1189 | return r5l_ring_distance(log, log->last_checkpoint, |
a39f7afd | 1190 | r5c_calculate_new_cp(conf)); |
17036461 CH |
1191 | } |
1192 | ||
5036c390 CH |
1193 | static void r5l_run_no_mem_stripe(struct r5l_log *log) |
1194 | { | |
1195 | struct stripe_head *sh; | |
1196 | ||
1197 | assert_spin_locked(&log->io_list_lock); | |
1198 | ||
1199 | if (!list_empty(&log->no_mem_stripes)) { | |
1200 | sh = list_first_entry(&log->no_mem_stripes, | |
1201 | struct stripe_head, log_list); | |
1202 | list_del_init(&sh->log_list); | |
1203 | set_bit(STRIPE_HANDLE, &sh->state); | |
1204 | raid5_release_stripe(sh); | |
1205 | } | |
1206 | } | |
1207 | ||
04732f74 | 1208 | static bool r5l_complete_finished_ios(struct r5l_log *log) |
17036461 CH |
1209 | { |
1210 | struct r5l_io_unit *io, *next; | |
1211 | bool found = false; | |
1212 | ||
1213 | assert_spin_locked(&log->io_list_lock); | |
1214 | ||
04732f74 | 1215 | list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) { |
17036461 CH |
1216 | /* don't change list order */ |
1217 | if (io->state < IO_UNIT_STRIPE_END) | |
1218 | break; | |
1219 | ||
1220 | log->next_checkpoint = io->log_start; | |
17036461 CH |
1221 | |
1222 | list_del(&io->log_sibling); | |
5036c390 CH |
1223 | mempool_free(io, log->io_pool); |
1224 | r5l_run_no_mem_stripe(log); | |
17036461 CH |
1225 | |
1226 | found = true; | |
1227 | } | |
1228 | ||
1229 | return found; | |
1230 | } | |
1231 | ||
509ffec7 CH |
1232 | static void __r5l_stripe_write_finished(struct r5l_io_unit *io) |
1233 | { | |
1234 | struct r5l_log *log = io->log; | |
a39f7afd | 1235 | struct r5conf *conf = log->rdev->mddev->private; |
509ffec7 CH |
1236 | unsigned long flags; |
1237 | ||
1238 | spin_lock_irqsave(&log->io_list_lock, flags); | |
1239 | __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END); | |
17036461 | 1240 | |
04732f74 | 1241 | if (!r5l_complete_finished_ios(log)) { |
85f2f9a4 SL |
1242 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
1243 | return; | |
1244 | } | |
509ffec7 | 1245 | |
a39f7afd SL |
1246 | if (r5l_reclaimable_space(log) > log->max_free_space || |
1247 | test_bit(R5C_LOG_TIGHT, &conf->cache_state)) | |
509ffec7 CH |
1248 | r5l_wake_reclaim(log, 0); |
1249 | ||
509ffec7 CH |
1250 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
1251 | wake_up(&log->iounit_wait); | |
1252 | } | |
1253 | ||
0576b1c6 SL |
1254 | void r5l_stripe_write_finished(struct stripe_head *sh) |
1255 | { | |
1256 | struct r5l_io_unit *io; | |
1257 | ||
0576b1c6 | 1258 | io = sh->log_io; |
0576b1c6 SL |
1259 | sh->log_io = NULL; |
1260 | ||
509ffec7 CH |
1261 | if (io && atomic_dec_and_test(&io->pending_stripe)) |
1262 | __r5l_stripe_write_finished(io); | |
0576b1c6 SL |
1263 | } |
1264 | ||
a8c34f91 SL |
1265 | static void r5l_log_flush_endio(struct bio *bio) |
1266 | { | |
1267 | struct r5l_log *log = container_of(bio, struct r5l_log, | |
1268 | flush_bio); | |
1269 | unsigned long flags; | |
1270 | struct r5l_io_unit *io; | |
a8c34f91 | 1271 | |
4e4cbee9 | 1272 | if (bio->bi_status) |
6e74a9cf SL |
1273 | md_error(log->rdev->mddev, log->rdev); |
1274 | ||
a8c34f91 | 1275 | spin_lock_irqsave(&log->io_list_lock, flags); |
d8858f43 CH |
1276 | list_for_each_entry(io, &log->flushing_ios, log_sibling) |
1277 | r5l_io_run_stripes(io); | |
04732f74 | 1278 | list_splice_tail_init(&log->flushing_ios, &log->finished_ios); |
a8c34f91 SL |
1279 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
1280 | } | |
1281 | ||
0576b1c6 SL |
1282 | /* |
1283 | * Starting dispatch IO to raid. | |
1284 | * io_unit(meta) consists of a log. There is one situation we want to avoid. A | |
1285 | * broken meta in the middle of a log causes recovery can't find meta at the | |
1286 | * head of log. If operations require meta at the head persistent in log, we | |
1287 | * must make sure meta before it persistent in log too. A case is: | |
1288 | * | |
1289 | * stripe data/parity is in log, we start write stripe to raid disks. stripe | |
1290 | * data/parity must be persistent in log before we do the write to raid disks. | |
1291 | * | |
1292 | * The solution is we restrictly maintain io_unit list order. In this case, we | |
1293 | * only write stripes of an io_unit to raid disks till the io_unit is the first | |
1294 | * one whose data/parity is in log. | |
1295 | */ | |
1296 | void r5l_flush_stripe_to_raid(struct r5l_log *log) | |
1297 | { | |
a8c34f91 | 1298 | bool do_flush; |
56fef7c6 CH |
1299 | |
1300 | if (!log || !log->need_cache_flush) | |
0576b1c6 | 1301 | return; |
0576b1c6 SL |
1302 | |
1303 | spin_lock_irq(&log->io_list_lock); | |
a8c34f91 SL |
1304 | /* flush bio is running */ |
1305 | if (!list_empty(&log->flushing_ios)) { | |
1306 | spin_unlock_irq(&log->io_list_lock); | |
1307 | return; | |
0576b1c6 | 1308 | } |
a8c34f91 SL |
1309 | list_splice_tail_init(&log->io_end_ios, &log->flushing_ios); |
1310 | do_flush = !list_empty(&log->flushing_ios); | |
0576b1c6 | 1311 | spin_unlock_irq(&log->io_list_lock); |
a8c34f91 SL |
1312 | |
1313 | if (!do_flush) | |
1314 | return; | |
1315 | bio_reset(&log->flush_bio); | |
1316 | log->flush_bio.bi_bdev = log->rdev->bdev; | |
1317 | log->flush_bio.bi_end_io = r5l_log_flush_endio; | |
70fd7614 | 1318 | log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH; |
4e49ea4a | 1319 | submit_bio(&log->flush_bio); |
0576b1c6 SL |
1320 | } |
1321 | ||
0576b1c6 | 1322 | static void r5l_write_super(struct r5l_log *log, sector_t cp); |
4b482044 SL |
1323 | static void r5l_write_super_and_discard_space(struct r5l_log *log, |
1324 | sector_t end) | |
1325 | { | |
1326 | struct block_device *bdev = log->rdev->bdev; | |
1327 | struct mddev *mddev; | |
1328 | ||
1329 | r5l_write_super(log, end); | |
1330 | ||
1331 | if (!blk_queue_discard(bdev_get_queue(bdev))) | |
1332 | return; | |
1333 | ||
1334 | mddev = log->rdev->mddev; | |
1335 | /* | |
8e018c21 SL |
1336 | * Discard could zero data, so before discard we must make sure |
1337 | * superblock is updated to new log tail. Updating superblock (either | |
1338 | * directly call md_update_sb() or depend on md thread) must hold | |
1339 | * reconfig mutex. On the other hand, raid5_quiesce is called with | |
1340 | * reconfig_mutex hold. The first step of raid5_quiesce() is waitting | |
1341 | * for all IO finish, hence waitting for reclaim thread, while reclaim | |
1342 | * thread is calling this function and waitting for reconfig mutex. So | |
1343 | * there is a deadlock. We workaround this issue with a trylock. | |
1344 | * FIXME: we could miss discard if we can't take reconfig mutex | |
4b482044 | 1345 | */ |
2953079c SL |
1346 | set_mask_bits(&mddev->sb_flags, 0, |
1347 | BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING)); | |
8e018c21 SL |
1348 | if (!mddev_trylock(mddev)) |
1349 | return; | |
1350 | md_update_sb(mddev, 1); | |
1351 | mddev_unlock(mddev); | |
4b482044 | 1352 | |
6e74a9cf | 1353 | /* discard IO error really doesn't matter, ignore it */ |
4b482044 SL |
1354 | if (log->last_checkpoint < end) { |
1355 | blkdev_issue_discard(bdev, | |
1356 | log->last_checkpoint + log->rdev->data_offset, | |
1357 | end - log->last_checkpoint, GFP_NOIO, 0); | |
1358 | } else { | |
1359 | blkdev_issue_discard(bdev, | |
1360 | log->last_checkpoint + log->rdev->data_offset, | |
1361 | log->device_size - log->last_checkpoint, | |
1362 | GFP_NOIO, 0); | |
1363 | blkdev_issue_discard(bdev, log->rdev->data_offset, end, | |
1364 | GFP_NOIO, 0); | |
1365 | } | |
1366 | } | |
1367 | ||
a39f7afd SL |
1368 | /* |
1369 | * r5c_flush_stripe moves stripe from cached list to handle_list. When called, | |
1370 | * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes. | |
1371 | * | |
1372 | * must hold conf->device_lock | |
1373 | */ | |
1374 | static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh) | |
0576b1c6 | 1375 | { |
a39f7afd SL |
1376 | BUG_ON(list_empty(&sh->lru)); |
1377 | BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
1378 | BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); | |
0576b1c6 | 1379 | |
0576b1c6 | 1380 | /* |
a39f7afd SL |
1381 | * The stripe is not ON_RELEASE_LIST, so it is safe to call |
1382 | * raid5_release_stripe() while holding conf->device_lock | |
0576b1c6 | 1383 | */ |
a39f7afd SL |
1384 | BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state)); |
1385 | assert_spin_locked(&conf->device_lock); | |
0576b1c6 | 1386 | |
a39f7afd SL |
1387 | list_del_init(&sh->lru); |
1388 | atomic_inc(&sh->count); | |
17036461 | 1389 | |
a39f7afd SL |
1390 | set_bit(STRIPE_HANDLE, &sh->state); |
1391 | atomic_inc(&conf->active_stripes); | |
1392 | r5c_make_stripe_write_out(sh); | |
0576b1c6 | 1393 | |
e33fbb9c SL |
1394 | if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) |
1395 | atomic_inc(&conf->r5c_flushing_partial_stripes); | |
1396 | else | |
1397 | atomic_inc(&conf->r5c_flushing_full_stripes); | |
a39f7afd SL |
1398 | raid5_release_stripe(sh); |
1399 | } | |
1400 | ||
1401 | /* | |
1402 | * if num == 0, flush all full stripes | |
1403 | * if num > 0, flush all full stripes. If less than num full stripes are | |
1404 | * flushed, flush some partial stripes until totally num stripes are | |
1405 | * flushed or there is no more cached stripes. | |
1406 | */ | |
1407 | void r5c_flush_cache(struct r5conf *conf, int num) | |
1408 | { | |
1409 | int count; | |
1410 | struct stripe_head *sh, *next; | |
1411 | ||
1412 | assert_spin_locked(&conf->device_lock); | |
1413 | if (!conf->log) | |
0576b1c6 SL |
1414 | return; |
1415 | ||
a39f7afd SL |
1416 | count = 0; |
1417 | list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) { | |
1418 | r5c_flush_stripe(conf, sh); | |
1419 | count++; | |
1420 | } | |
1421 | ||
1422 | if (count >= num) | |
1423 | return; | |
1424 | list_for_each_entry_safe(sh, next, | |
1425 | &conf->r5c_partial_stripe_list, lru) { | |
1426 | r5c_flush_stripe(conf, sh); | |
1427 | if (++count >= num) | |
1428 | break; | |
1429 | } | |
1430 | } | |
1431 | ||
1432 | static void r5c_do_reclaim(struct r5conf *conf) | |
1433 | { | |
1434 | struct r5l_log *log = conf->log; | |
1435 | struct stripe_head *sh; | |
1436 | int count = 0; | |
1437 | unsigned long flags; | |
1438 | int total_cached; | |
1439 | int stripes_to_flush; | |
e33fbb9c | 1440 | int flushing_partial, flushing_full; |
a39f7afd SL |
1441 | |
1442 | if (!r5c_is_writeback(log)) | |
1443 | return; | |
1444 | ||
e33fbb9c SL |
1445 | flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes); |
1446 | flushing_full = atomic_read(&conf->r5c_flushing_full_stripes); | |
a39f7afd | 1447 | total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + |
e33fbb9c SL |
1448 | atomic_read(&conf->r5c_cached_full_stripes) - |
1449 | flushing_full - flushing_partial; | |
a39f7afd SL |
1450 | |
1451 | if (total_cached > conf->min_nr_stripes * 3 / 4 || | |
1452 | atomic_read(&conf->empty_inactive_list_nr) > 0) | |
1453 | /* | |
1454 | * if stripe cache pressure high, flush all full stripes and | |
1455 | * some partial stripes | |
1456 | */ | |
1457 | stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP; | |
1458 | else if (total_cached > conf->min_nr_stripes * 1 / 2 || | |
e33fbb9c | 1459 | atomic_read(&conf->r5c_cached_full_stripes) - flushing_full > |
84890c03 | 1460 | R5C_FULL_STRIPE_FLUSH_BATCH(conf)) |
a39f7afd SL |
1461 | /* |
1462 | * if stripe cache pressure moderate, or if there is many full | |
1463 | * stripes,flush all full stripes | |
1464 | */ | |
1465 | stripes_to_flush = 0; | |
1466 | else | |
1467 | /* no need to flush */ | |
1468 | stripes_to_flush = -1; | |
1469 | ||
1470 | if (stripes_to_flush >= 0) { | |
1471 | spin_lock_irqsave(&conf->device_lock, flags); | |
1472 | r5c_flush_cache(conf, stripes_to_flush); | |
1473 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1474 | } | |
1475 | ||
1476 | /* if log space is tight, flush stripes on stripe_in_journal_list */ | |
1477 | if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) { | |
1478 | spin_lock_irqsave(&log->stripe_in_journal_lock, flags); | |
1479 | spin_lock(&conf->device_lock); | |
1480 | list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) { | |
1481 | /* | |
1482 | * stripes on stripe_in_journal_list could be in any | |
1483 | * state of the stripe_cache state machine. In this | |
1484 | * case, we only want to flush stripe on | |
1485 | * r5c_cached_full/partial_stripes. The following | |
1486 | * condition makes sure the stripe is on one of the | |
1487 | * two lists. | |
1488 | */ | |
1489 | if (!list_empty(&sh->lru) && | |
1490 | !test_bit(STRIPE_HANDLE, &sh->state) && | |
1491 | atomic_read(&sh->count) == 0) { | |
1492 | r5c_flush_stripe(conf, sh); | |
e8fd52ee SL |
1493 | if (count++ >= R5C_RECLAIM_STRIPE_GROUP) |
1494 | break; | |
a39f7afd | 1495 | } |
a39f7afd SL |
1496 | } |
1497 | spin_unlock(&conf->device_lock); | |
1498 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); | |
1499 | } | |
f687a33e SL |
1500 | |
1501 | if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state)) | |
1502 | r5l_run_no_space_stripes(log); | |
1503 | ||
a39f7afd SL |
1504 | md_wakeup_thread(conf->mddev->thread); |
1505 | } | |
1506 | ||
0576b1c6 SL |
1507 | static void r5l_do_reclaim(struct r5l_log *log) |
1508 | { | |
a39f7afd | 1509 | struct r5conf *conf = log->rdev->mddev->private; |
0576b1c6 | 1510 | sector_t reclaim_target = xchg(&log->reclaim_target, 0); |
17036461 CH |
1511 | sector_t reclaimable; |
1512 | sector_t next_checkpoint; | |
a39f7afd | 1513 | bool write_super; |
0576b1c6 SL |
1514 | |
1515 | spin_lock_irq(&log->io_list_lock); | |
a39f7afd SL |
1516 | write_super = r5l_reclaimable_space(log) > log->max_free_space || |
1517 | reclaim_target != 0 || !list_empty(&log->no_space_stripes); | |
0576b1c6 SL |
1518 | /* |
1519 | * move proper io_unit to reclaim list. We should not change the order. | |
1520 | * reclaimable/unreclaimable io_unit can be mixed in the list, we | |
1521 | * shouldn't reuse space of an unreclaimable io_unit | |
1522 | */ | |
1523 | while (1) { | |
17036461 CH |
1524 | reclaimable = r5l_reclaimable_space(log); |
1525 | if (reclaimable >= reclaim_target || | |
0576b1c6 SL |
1526 | (list_empty(&log->running_ios) && |
1527 | list_empty(&log->io_end_ios) && | |
a8c34f91 | 1528 | list_empty(&log->flushing_ios) && |
04732f74 | 1529 | list_empty(&log->finished_ios))) |
0576b1c6 SL |
1530 | break; |
1531 | ||
17036461 CH |
1532 | md_wakeup_thread(log->rdev->mddev->thread); |
1533 | wait_event_lock_irq(log->iounit_wait, | |
1534 | r5l_reclaimable_space(log) > reclaimable, | |
1535 | log->io_list_lock); | |
0576b1c6 | 1536 | } |
17036461 | 1537 | |
a39f7afd | 1538 | next_checkpoint = r5c_calculate_new_cp(conf); |
0576b1c6 SL |
1539 | spin_unlock_irq(&log->io_list_lock); |
1540 | ||
a39f7afd | 1541 | if (reclaimable == 0 || !write_super) |
0576b1c6 SL |
1542 | return; |
1543 | ||
0576b1c6 SL |
1544 | /* |
1545 | * write_super will flush cache of each raid disk. We must write super | |
1546 | * here, because the log area might be reused soon and we don't want to | |
1547 | * confuse recovery | |
1548 | */ | |
4b482044 | 1549 | r5l_write_super_and_discard_space(log, next_checkpoint); |
0576b1c6 SL |
1550 | |
1551 | mutex_lock(&log->io_mutex); | |
17036461 | 1552 | log->last_checkpoint = next_checkpoint; |
a39f7afd | 1553 | r5c_update_log_state(log); |
0576b1c6 | 1554 | mutex_unlock(&log->io_mutex); |
0576b1c6 | 1555 | |
17036461 | 1556 | r5l_run_no_space_stripes(log); |
0576b1c6 SL |
1557 | } |
1558 | ||
1559 | static void r5l_reclaim_thread(struct md_thread *thread) | |
1560 | { | |
1561 | struct mddev *mddev = thread->mddev; | |
1562 | struct r5conf *conf = mddev->private; | |
1563 | struct r5l_log *log = conf->log; | |
1564 | ||
1565 | if (!log) | |
1566 | return; | |
a39f7afd | 1567 | r5c_do_reclaim(conf); |
0576b1c6 SL |
1568 | r5l_do_reclaim(log); |
1569 | } | |
1570 | ||
a39f7afd | 1571 | void r5l_wake_reclaim(struct r5l_log *log, sector_t space) |
f6bed0ef | 1572 | { |
0576b1c6 SL |
1573 | unsigned long target; |
1574 | unsigned long new = (unsigned long)space; /* overflow in theory */ | |
1575 | ||
a39f7afd SL |
1576 | if (!log) |
1577 | return; | |
0576b1c6 SL |
1578 | do { |
1579 | target = log->reclaim_target; | |
1580 | if (new < target) | |
1581 | return; | |
1582 | } while (cmpxchg(&log->reclaim_target, target, new) != target); | |
1583 | md_wakeup_thread(log->reclaim_thread); | |
f6bed0ef SL |
1584 | } |
1585 | ||
e6c033f7 SL |
1586 | void r5l_quiesce(struct r5l_log *log, int state) |
1587 | { | |
4b482044 | 1588 | struct mddev *mddev; |
e6c033f7 SL |
1589 | if (!log || state == 2) |
1590 | return; | |
ce1ccd07 SL |
1591 | if (state == 0) |
1592 | kthread_unpark(log->reclaim_thread->tsk); | |
1593 | else if (state == 1) { | |
4b482044 SL |
1594 | /* make sure r5l_write_super_and_discard_space exits */ |
1595 | mddev = log->rdev->mddev; | |
1596 | wake_up(&mddev->sb_wait); | |
ce1ccd07 | 1597 | kthread_park(log->reclaim_thread->tsk); |
a39f7afd | 1598 | r5l_wake_reclaim(log, MaxSector); |
e6c033f7 SL |
1599 | r5l_do_reclaim(log); |
1600 | } | |
1601 | } | |
1602 | ||
6e74a9cf SL |
1603 | bool r5l_log_disk_error(struct r5conf *conf) |
1604 | { | |
f6b6ec5c SL |
1605 | struct r5l_log *log; |
1606 | bool ret; | |
7dde2ad3 | 1607 | /* don't allow write if journal disk is missing */ |
f6b6ec5c SL |
1608 | rcu_read_lock(); |
1609 | log = rcu_dereference(conf->log); | |
1610 | ||
1611 | if (!log) | |
1612 | ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags); | |
1613 | else | |
1614 | ret = test_bit(Faulty, &log->rdev->flags); | |
1615 | rcu_read_unlock(); | |
1616 | return ret; | |
6e74a9cf SL |
1617 | } |
1618 | ||
effe6ee7 SL |
1619 | #define R5L_RECOVERY_PAGE_POOL_SIZE 256 |
1620 | ||
355810d1 SL |
1621 | struct r5l_recovery_ctx { |
1622 | struct page *meta_page; /* current meta */ | |
1623 | sector_t meta_total_blocks; /* total size of current meta and data */ | |
1624 | sector_t pos; /* recovery position */ | |
1625 | u64 seq; /* recovery position seq */ | |
b4c625c6 SL |
1626 | int data_parity_stripes; /* number of data_parity stripes */ |
1627 | int data_only_stripes; /* number of data_only stripes */ | |
1628 | struct list_head cached_list; | |
effe6ee7 SL |
1629 | |
1630 | /* | |
1631 | * read ahead page pool (ra_pool) | |
1632 | * in recovery, log is read sequentially. It is not efficient to | |
1633 | * read every page with sync_page_io(). The read ahead page pool | |
1634 | * reads multiple pages with one IO, so further log read can | |
1635 | * just copy data from the pool. | |
1636 | */ | |
1637 | struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE]; | |
1638 | sector_t pool_offset; /* offset of first page in the pool */ | |
1639 | int total_pages; /* total allocated pages */ | |
1640 | int valid_pages; /* pages with valid data */ | |
1641 | struct bio *ra_bio; /* bio to do the read ahead */ | |
355810d1 SL |
1642 | }; |
1643 | ||
effe6ee7 SL |
1644 | static int r5l_recovery_allocate_ra_pool(struct r5l_log *log, |
1645 | struct r5l_recovery_ctx *ctx) | |
1646 | { | |
1647 | struct page *page; | |
1648 | ||
1649 | ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs); | |
1650 | if (!ctx->ra_bio) | |
1651 | return -ENOMEM; | |
1652 | ||
1653 | ctx->valid_pages = 0; | |
1654 | ctx->total_pages = 0; | |
1655 | while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) { | |
1656 | page = alloc_page(GFP_KERNEL); | |
1657 | ||
1658 | if (!page) | |
1659 | break; | |
1660 | ctx->ra_pool[ctx->total_pages] = page; | |
1661 | ctx->total_pages += 1; | |
1662 | } | |
1663 | ||
1664 | if (ctx->total_pages == 0) { | |
1665 | bio_put(ctx->ra_bio); | |
1666 | return -ENOMEM; | |
1667 | } | |
1668 | ||
1669 | ctx->pool_offset = 0; | |
1670 | return 0; | |
1671 | } | |
1672 | ||
1673 | static void r5l_recovery_free_ra_pool(struct r5l_log *log, | |
1674 | struct r5l_recovery_ctx *ctx) | |
1675 | { | |
1676 | int i; | |
1677 | ||
1678 | for (i = 0; i < ctx->total_pages; ++i) | |
1679 | put_page(ctx->ra_pool[i]); | |
1680 | bio_put(ctx->ra_bio); | |
1681 | } | |
1682 | ||
1683 | /* | |
1684 | * fetch ctx->valid_pages pages from offset | |
1685 | * In normal cases, ctx->valid_pages == ctx->total_pages after the call. | |
1686 | * However, if the offset is close to the end of the journal device, | |
1687 | * ctx->valid_pages could be smaller than ctx->total_pages | |
1688 | */ | |
1689 | static int r5l_recovery_fetch_ra_pool(struct r5l_log *log, | |
1690 | struct r5l_recovery_ctx *ctx, | |
1691 | sector_t offset) | |
1692 | { | |
1693 | bio_reset(ctx->ra_bio); | |
1694 | ctx->ra_bio->bi_bdev = log->rdev->bdev; | |
1695 | bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0); | |
1696 | ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset; | |
1697 | ||
1698 | ctx->valid_pages = 0; | |
1699 | ctx->pool_offset = offset; | |
1700 | ||
1701 | while (ctx->valid_pages < ctx->total_pages) { | |
1702 | bio_add_page(ctx->ra_bio, | |
1703 | ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0); | |
1704 | ctx->valid_pages += 1; | |
1705 | ||
1706 | offset = r5l_ring_add(log, offset, BLOCK_SECTORS); | |
1707 | ||
1708 | if (offset == 0) /* reached end of the device */ | |
1709 | break; | |
1710 | } | |
1711 | ||
1712 | return submit_bio_wait(ctx->ra_bio); | |
1713 | } | |
1714 | ||
1715 | /* | |
1716 | * try read a page from the read ahead page pool, if the page is not in the | |
1717 | * pool, call r5l_recovery_fetch_ra_pool | |
1718 | */ | |
1719 | static int r5l_recovery_read_page(struct r5l_log *log, | |
1720 | struct r5l_recovery_ctx *ctx, | |
1721 | struct page *page, | |
1722 | sector_t offset) | |
1723 | { | |
1724 | int ret; | |
1725 | ||
1726 | if (offset < ctx->pool_offset || | |
1727 | offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) { | |
1728 | ret = r5l_recovery_fetch_ra_pool(log, ctx, offset); | |
1729 | if (ret) | |
1730 | return ret; | |
1731 | } | |
1732 | ||
1733 | BUG_ON(offset < ctx->pool_offset || | |
1734 | offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS); | |
1735 | ||
1736 | memcpy(page_address(page), | |
1737 | page_address(ctx->ra_pool[(offset - ctx->pool_offset) >> | |
1738 | BLOCK_SECTOR_SHIFT]), | |
1739 | PAGE_SIZE); | |
1740 | return 0; | |
1741 | } | |
1742 | ||
9ed988f5 SL |
1743 | static int r5l_recovery_read_meta_block(struct r5l_log *log, |
1744 | struct r5l_recovery_ctx *ctx) | |
355810d1 SL |
1745 | { |
1746 | struct page *page = ctx->meta_page; | |
1747 | struct r5l_meta_block *mb; | |
1748 | u32 crc, stored_crc; | |
effe6ee7 | 1749 | int ret; |
355810d1 | 1750 | |
effe6ee7 SL |
1751 | ret = r5l_recovery_read_page(log, ctx, page, ctx->pos); |
1752 | if (ret != 0) | |
1753 | return ret; | |
355810d1 SL |
1754 | |
1755 | mb = page_address(page); | |
1756 | stored_crc = le32_to_cpu(mb->checksum); | |
1757 | mb->checksum = 0; | |
1758 | ||
1759 | if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || | |
1760 | le64_to_cpu(mb->seq) != ctx->seq || | |
1761 | mb->version != R5LOG_VERSION || | |
1762 | le64_to_cpu(mb->position) != ctx->pos) | |
1763 | return -EINVAL; | |
1764 | ||
5cb2fbd6 | 1765 | crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); |
355810d1 SL |
1766 | if (stored_crc != crc) |
1767 | return -EINVAL; | |
1768 | ||
1769 | if (le32_to_cpu(mb->meta_size) > PAGE_SIZE) | |
1770 | return -EINVAL; | |
1771 | ||
1772 | ctx->meta_total_blocks = BLOCK_SECTORS; | |
1773 | ||
1774 | return 0; | |
1775 | } | |
1776 | ||
9ed988f5 SL |
1777 | static void |
1778 | r5l_recovery_create_empty_meta_block(struct r5l_log *log, | |
1779 | struct page *page, | |
1780 | sector_t pos, u64 seq) | |
355810d1 | 1781 | { |
355810d1 | 1782 | struct r5l_meta_block *mb; |
355810d1 | 1783 | |
355810d1 | 1784 | mb = page_address(page); |
9ed988f5 | 1785 | clear_page(mb); |
355810d1 SL |
1786 | mb->magic = cpu_to_le32(R5LOG_MAGIC); |
1787 | mb->version = R5LOG_VERSION; | |
1788 | mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block)); | |
1789 | mb->seq = cpu_to_le64(seq); | |
1790 | mb->position = cpu_to_le64(pos); | |
9ed988f5 | 1791 | } |
355810d1 | 1792 | |
9ed988f5 SL |
1793 | static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos, |
1794 | u64 seq) | |
1795 | { | |
1796 | struct page *page; | |
5c88f403 | 1797 | struct r5l_meta_block *mb; |
355810d1 | 1798 | |
9ed988f5 SL |
1799 | page = alloc_page(GFP_KERNEL); |
1800 | if (!page) | |
1801 | return -ENOMEM; | |
1802 | r5l_recovery_create_empty_meta_block(log, page, pos, seq); | |
5c88f403 SL |
1803 | mb = page_address(page); |
1804 | mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum, | |
1805 | mb, PAGE_SIZE)); | |
796a5cf0 | 1806 | if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE, |
5a8948f8 | 1807 | REQ_SYNC | REQ_FUA, false)) { |
355810d1 SL |
1808 | __free_page(page); |
1809 | return -EIO; | |
1810 | } | |
1811 | __free_page(page); | |
1812 | return 0; | |
1813 | } | |
355810d1 | 1814 | |
b4c625c6 SL |
1815 | /* |
1816 | * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite | |
1817 | * to mark valid (potentially not flushed) data in the journal. | |
1818 | * | |
1819 | * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb, | |
1820 | * so there should not be any mismatch here. | |
1821 | */ | |
1822 | static void r5l_recovery_load_data(struct r5l_log *log, | |
1823 | struct stripe_head *sh, | |
1824 | struct r5l_recovery_ctx *ctx, | |
1825 | struct r5l_payload_data_parity *payload, | |
1826 | sector_t log_offset) | |
1827 | { | |
1828 | struct mddev *mddev = log->rdev->mddev; | |
1829 | struct r5conf *conf = mddev->private; | |
1830 | int dd_idx; | |
1831 | ||
1832 | raid5_compute_sector(conf, | |
1833 | le64_to_cpu(payload->location), 0, | |
1834 | &dd_idx, sh); | |
effe6ee7 | 1835 | r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset); |
b4c625c6 SL |
1836 | sh->dev[dd_idx].log_checksum = |
1837 | le32_to_cpu(payload->checksum[0]); | |
1838 | ctx->meta_total_blocks += BLOCK_SECTORS; | |
1839 | ||
1840 | set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags); | |
1841 | set_bit(STRIPE_R5C_CACHING, &sh->state); | |
1842 | } | |
1843 | ||
1844 | static void r5l_recovery_load_parity(struct r5l_log *log, | |
1845 | struct stripe_head *sh, | |
1846 | struct r5l_recovery_ctx *ctx, | |
1847 | struct r5l_payload_data_parity *payload, | |
1848 | sector_t log_offset) | |
1849 | { | |
1850 | struct mddev *mddev = log->rdev->mddev; | |
1851 | struct r5conf *conf = mddev->private; | |
1852 | ||
1853 | ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded; | |
effe6ee7 | 1854 | r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset); |
b4c625c6 SL |
1855 | sh->dev[sh->pd_idx].log_checksum = |
1856 | le32_to_cpu(payload->checksum[0]); | |
1857 | set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags); | |
1858 | ||
1859 | if (sh->qd_idx >= 0) { | |
effe6ee7 SL |
1860 | r5l_recovery_read_page( |
1861 | log, ctx, sh->dev[sh->qd_idx].page, | |
1862 | r5l_ring_add(log, log_offset, BLOCK_SECTORS)); | |
b4c625c6 SL |
1863 | sh->dev[sh->qd_idx].log_checksum = |
1864 | le32_to_cpu(payload->checksum[1]); | |
1865 | set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags); | |
355810d1 | 1866 | } |
b4c625c6 SL |
1867 | clear_bit(STRIPE_R5C_CACHING, &sh->state); |
1868 | } | |
355810d1 | 1869 | |
b4c625c6 SL |
1870 | static void r5l_recovery_reset_stripe(struct stripe_head *sh) |
1871 | { | |
1872 | int i; | |
1873 | ||
1874 | sh->state = 0; | |
1875 | sh->log_start = MaxSector; | |
1876 | for (i = sh->disks; i--; ) | |
1877 | sh->dev[i].flags = 0; | |
1878 | } | |
1879 | ||
1880 | static void | |
1881 | r5l_recovery_replay_one_stripe(struct r5conf *conf, | |
1882 | struct stripe_head *sh, | |
1883 | struct r5l_recovery_ctx *ctx) | |
1884 | { | |
1885 | struct md_rdev *rdev, *rrdev; | |
1886 | int disk_index; | |
1887 | int data_count = 0; | |
355810d1 | 1888 | |
b4c625c6 | 1889 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { |
355810d1 SL |
1890 | if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) |
1891 | continue; | |
b4c625c6 SL |
1892 | if (disk_index == sh->qd_idx || disk_index == sh->pd_idx) |
1893 | continue; | |
1894 | data_count++; | |
355810d1 SL |
1895 | } |
1896 | ||
b4c625c6 SL |
1897 | /* |
1898 | * stripes that only have parity must have been flushed | |
1899 | * before the crash that we are now recovering from, so | |
1900 | * there is nothing more to recovery. | |
1901 | */ | |
1902 | if (data_count == 0) | |
1903 | goto out; | |
355810d1 | 1904 | |
b4c625c6 SL |
1905 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { |
1906 | if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) | |
355810d1 SL |
1907 | continue; |
1908 | ||
1909 | /* in case device is broken */ | |
b4c625c6 | 1910 | rcu_read_lock(); |
355810d1 | 1911 | rdev = rcu_dereference(conf->disks[disk_index].rdev); |
b4c625c6 SL |
1912 | if (rdev) { |
1913 | atomic_inc(&rdev->nr_pending); | |
1914 | rcu_read_unlock(); | |
1915 | sync_page_io(rdev, sh->sector, PAGE_SIZE, | |
796a5cf0 MC |
1916 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, |
1917 | false); | |
b4c625c6 SL |
1918 | rdev_dec_pending(rdev, rdev->mddev); |
1919 | rcu_read_lock(); | |
1920 | } | |
355810d1 | 1921 | rrdev = rcu_dereference(conf->disks[disk_index].replacement); |
b4c625c6 SL |
1922 | if (rrdev) { |
1923 | atomic_inc(&rrdev->nr_pending); | |
1924 | rcu_read_unlock(); | |
1925 | sync_page_io(rrdev, sh->sector, PAGE_SIZE, | |
796a5cf0 MC |
1926 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, |
1927 | false); | |
b4c625c6 SL |
1928 | rdev_dec_pending(rrdev, rrdev->mddev); |
1929 | rcu_read_lock(); | |
1930 | } | |
1931 | rcu_read_unlock(); | |
355810d1 | 1932 | } |
b4c625c6 SL |
1933 | ctx->data_parity_stripes++; |
1934 | out: | |
1935 | r5l_recovery_reset_stripe(sh); | |
1936 | } | |
1937 | ||
1938 | static struct stripe_head * | |
1939 | r5c_recovery_alloc_stripe(struct r5conf *conf, | |
3c66abba | 1940 | sector_t stripe_sect) |
b4c625c6 SL |
1941 | { |
1942 | struct stripe_head *sh; | |
1943 | ||
1944 | sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0); | |
1945 | if (!sh) | |
1946 | return NULL; /* no more stripe available */ | |
1947 | ||
1948 | r5l_recovery_reset_stripe(sh); | |
b4c625c6 SL |
1949 | |
1950 | return sh; | |
1951 | } | |
1952 | ||
1953 | static struct stripe_head * | |
1954 | r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect) | |
1955 | { | |
1956 | struct stripe_head *sh; | |
1957 | ||
1958 | list_for_each_entry(sh, list, lru) | |
1959 | if (sh->sector == sect) | |
1960 | return sh; | |
1961 | return NULL; | |
1962 | } | |
1963 | ||
1964 | static void | |
1965 | r5c_recovery_drop_stripes(struct list_head *cached_stripe_list, | |
1966 | struct r5l_recovery_ctx *ctx) | |
1967 | { | |
1968 | struct stripe_head *sh, *next; | |
1969 | ||
1970 | list_for_each_entry_safe(sh, next, cached_stripe_list, lru) { | |
1971 | r5l_recovery_reset_stripe(sh); | |
1972 | list_del_init(&sh->lru); | |
1973 | raid5_release_stripe(sh); | |
1974 | } | |
1975 | } | |
1976 | ||
1977 | static void | |
1978 | r5c_recovery_replay_stripes(struct list_head *cached_stripe_list, | |
1979 | struct r5l_recovery_ctx *ctx) | |
1980 | { | |
1981 | struct stripe_head *sh, *next; | |
1982 | ||
1983 | list_for_each_entry_safe(sh, next, cached_stripe_list, lru) | |
1984 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { | |
1985 | r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx); | |
1986 | list_del_init(&sh->lru); | |
1987 | raid5_release_stripe(sh); | |
1988 | } | |
1989 | } | |
1990 | ||
1991 | /* if matches return 0; otherwise return -EINVAL */ | |
1992 | static int | |
effe6ee7 SL |
1993 | r5l_recovery_verify_data_checksum(struct r5l_log *log, |
1994 | struct r5l_recovery_ctx *ctx, | |
1995 | struct page *page, | |
b4c625c6 SL |
1996 | sector_t log_offset, __le32 log_checksum) |
1997 | { | |
1998 | void *addr; | |
1999 | u32 checksum; | |
2000 | ||
effe6ee7 | 2001 | r5l_recovery_read_page(log, ctx, page, log_offset); |
b4c625c6 SL |
2002 | addr = kmap_atomic(page); |
2003 | checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE); | |
2004 | kunmap_atomic(addr); | |
2005 | return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL; | |
2006 | } | |
2007 | ||
2008 | /* | |
2009 | * before loading data to stripe cache, we need verify checksum for all data, | |
2010 | * if there is mismatch for any data page, we drop all data in the mata block | |
2011 | */ | |
2012 | static int | |
2013 | r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log, | |
2014 | struct r5l_recovery_ctx *ctx) | |
2015 | { | |
2016 | struct mddev *mddev = log->rdev->mddev; | |
2017 | struct r5conf *conf = mddev->private; | |
2018 | struct r5l_meta_block *mb = page_address(ctx->meta_page); | |
2019 | sector_t mb_offset = sizeof(struct r5l_meta_block); | |
2020 | sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); | |
2021 | struct page *page; | |
2022 | struct r5l_payload_data_parity *payload; | |
2d4f4687 | 2023 | struct r5l_payload_flush *payload_flush; |
b4c625c6 SL |
2024 | |
2025 | page = alloc_page(GFP_KERNEL); | |
2026 | if (!page) | |
2027 | return -ENOMEM; | |
2028 | ||
2029 | while (mb_offset < le32_to_cpu(mb->meta_size)) { | |
2030 | payload = (void *)mb + mb_offset; | |
2d4f4687 | 2031 | payload_flush = (void *)mb + mb_offset; |
b4c625c6 | 2032 | |
1ad45a9b | 2033 | if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) { |
b4c625c6 | 2034 | if (r5l_recovery_verify_data_checksum( |
effe6ee7 | 2035 | log, ctx, page, log_offset, |
b4c625c6 SL |
2036 | payload->checksum[0]) < 0) |
2037 | goto mismatch; | |
1ad45a9b | 2038 | } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) { |
b4c625c6 | 2039 | if (r5l_recovery_verify_data_checksum( |
effe6ee7 | 2040 | log, ctx, page, log_offset, |
b4c625c6 SL |
2041 | payload->checksum[0]) < 0) |
2042 | goto mismatch; | |
2043 | if (conf->max_degraded == 2 && /* q for RAID 6 */ | |
2044 | r5l_recovery_verify_data_checksum( | |
effe6ee7 | 2045 | log, ctx, page, |
b4c625c6 SL |
2046 | r5l_ring_add(log, log_offset, |
2047 | BLOCK_SECTORS), | |
2048 | payload->checksum[1]) < 0) | |
2049 | goto mismatch; | |
1ad45a9b | 2050 | } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) { |
2d4f4687 SL |
2051 | /* nothing to do for R5LOG_PAYLOAD_FLUSH here */ |
2052 | } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */ | |
b4c625c6 SL |
2053 | goto mismatch; |
2054 | ||
1ad45a9b | 2055 | if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) { |
2d4f4687 SL |
2056 | mb_offset += sizeof(struct r5l_payload_flush) + |
2057 | le32_to_cpu(payload_flush->size); | |
2058 | } else { | |
2059 | /* DATA or PARITY payload */ | |
2060 | log_offset = r5l_ring_add(log, log_offset, | |
2061 | le32_to_cpu(payload->size)); | |
2062 | mb_offset += sizeof(struct r5l_payload_data_parity) + | |
2063 | sizeof(__le32) * | |
2064 | (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); | |
2065 | } | |
b4c625c6 | 2066 | |
b4c625c6 SL |
2067 | } |
2068 | ||
2069 | put_page(page); | |
355810d1 SL |
2070 | return 0; |
2071 | ||
b4c625c6 SL |
2072 | mismatch: |
2073 | put_page(page); | |
355810d1 SL |
2074 | return -EINVAL; |
2075 | } | |
2076 | ||
b4c625c6 SL |
2077 | /* |
2078 | * Analyze all data/parity pages in one meta block | |
2079 | * Returns: | |
2080 | * 0 for success | |
2081 | * -EINVAL for unknown playload type | |
2082 | * -EAGAIN for checksum mismatch of data page | |
2083 | * -ENOMEM for run out of memory (alloc_page failed or run out of stripes) | |
2084 | */ | |
2085 | static int | |
2086 | r5c_recovery_analyze_meta_block(struct r5l_log *log, | |
2087 | struct r5l_recovery_ctx *ctx, | |
2088 | struct list_head *cached_stripe_list) | |
355810d1 | 2089 | { |
b4c625c6 SL |
2090 | struct mddev *mddev = log->rdev->mddev; |
2091 | struct r5conf *conf = mddev->private; | |
355810d1 | 2092 | struct r5l_meta_block *mb; |
b4c625c6 | 2093 | struct r5l_payload_data_parity *payload; |
2d4f4687 | 2094 | struct r5l_payload_flush *payload_flush; |
b4c625c6 | 2095 | int mb_offset; |
355810d1 | 2096 | sector_t log_offset; |
b4c625c6 SL |
2097 | sector_t stripe_sect; |
2098 | struct stripe_head *sh; | |
2099 | int ret; | |
2100 | ||
2101 | /* | |
2102 | * for mismatch in data blocks, we will drop all data in this mb, but | |
2103 | * we will still read next mb for other data with FLUSH flag, as | |
2104 | * io_unit could finish out of order. | |
2105 | */ | |
2106 | ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx); | |
2107 | if (ret == -EINVAL) | |
2108 | return -EAGAIN; | |
2109 | else if (ret) | |
2110 | return ret; /* -ENOMEM duo to alloc_page() failed */ | |
355810d1 SL |
2111 | |
2112 | mb = page_address(ctx->meta_page); | |
b4c625c6 | 2113 | mb_offset = sizeof(struct r5l_meta_block); |
355810d1 SL |
2114 | log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); |
2115 | ||
b4c625c6 | 2116 | while (mb_offset < le32_to_cpu(mb->meta_size)) { |
355810d1 SL |
2117 | int dd; |
2118 | ||
b4c625c6 | 2119 | payload = (void *)mb + mb_offset; |
2d4f4687 SL |
2120 | payload_flush = (void *)mb + mb_offset; |
2121 | ||
1ad45a9b | 2122 | if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) { |
2d4f4687 SL |
2123 | int i, count; |
2124 | ||
2125 | count = le32_to_cpu(payload_flush->size) / sizeof(__le64); | |
2126 | for (i = 0; i < count; ++i) { | |
2127 | stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]); | |
2128 | sh = r5c_recovery_lookup_stripe(cached_stripe_list, | |
2129 | stripe_sect); | |
2130 | if (sh) { | |
2131 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2132 | r5l_recovery_reset_stripe(sh); | |
2133 | list_del_init(&sh->lru); | |
2134 | raid5_release_stripe(sh); | |
2135 | } | |
2136 | } | |
2137 | ||
2138 | mb_offset += sizeof(struct r5l_payload_flush) + | |
2139 | le32_to_cpu(payload_flush->size); | |
2140 | continue; | |
2141 | } | |
2142 | ||
2143 | /* DATA or PARITY payload */ | |
1ad45a9b | 2144 | stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ? |
b4c625c6 SL |
2145 | raid5_compute_sector( |
2146 | conf, le64_to_cpu(payload->location), 0, &dd, | |
2147 | NULL) | |
2148 | : le64_to_cpu(payload->location); | |
2149 | ||
2150 | sh = r5c_recovery_lookup_stripe(cached_stripe_list, | |
2151 | stripe_sect); | |
2152 | ||
2153 | if (!sh) { | |
3c66abba | 2154 | sh = r5c_recovery_alloc_stripe(conf, stripe_sect); |
b4c625c6 SL |
2155 | /* |
2156 | * cannot get stripe from raid5_get_active_stripe | |
2157 | * try replay some stripes | |
2158 | */ | |
2159 | if (!sh) { | |
2160 | r5c_recovery_replay_stripes( | |
2161 | cached_stripe_list, ctx); | |
2162 | sh = r5c_recovery_alloc_stripe( | |
3c66abba | 2163 | conf, stripe_sect); |
b4c625c6 SL |
2164 | } |
2165 | if (!sh) { | |
2166 | pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n", | |
2167 | mdname(mddev), | |
2168 | conf->min_nr_stripes * 2); | |
2169 | raid5_set_cache_size(mddev, | |
2170 | conf->min_nr_stripes * 2); | |
3c66abba SL |
2171 | sh = r5c_recovery_alloc_stripe(conf, |
2172 | stripe_sect); | |
b4c625c6 SL |
2173 | } |
2174 | if (!sh) { | |
2175 | pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n", | |
2176 | mdname(mddev)); | |
2177 | return -ENOMEM; | |
2178 | } | |
2179 | list_add_tail(&sh->lru, cached_stripe_list); | |
2180 | } | |
2181 | ||
1ad45a9b | 2182 | if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) { |
f7b7bee7 ZL |
2183 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state) && |
2184 | test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) { | |
b4c625c6 | 2185 | r5l_recovery_replay_one_stripe(conf, sh, ctx); |
b4c625c6 SL |
2186 | list_move_tail(&sh->lru, cached_stripe_list); |
2187 | } | |
2188 | r5l_recovery_load_data(log, sh, ctx, payload, | |
2189 | log_offset); | |
1ad45a9b | 2190 | } else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) |
b4c625c6 SL |
2191 | r5l_recovery_load_parity(log, sh, ctx, payload, |
2192 | log_offset); | |
2193 | else | |
355810d1 | 2194 | return -EINVAL; |
b4c625c6 SL |
2195 | |
2196 | log_offset = r5l_ring_add(log, log_offset, | |
2197 | le32_to_cpu(payload->size)); | |
2198 | ||
2199 | mb_offset += sizeof(struct r5l_payload_data_parity) + | |
2200 | sizeof(__le32) * | |
2201 | (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); | |
355810d1 | 2202 | } |
b4c625c6 | 2203 | |
355810d1 SL |
2204 | return 0; |
2205 | } | |
2206 | ||
b4c625c6 SL |
2207 | /* |
2208 | * Load the stripe into cache. The stripe will be written out later by | |
2209 | * the stripe cache state machine. | |
2210 | */ | |
2211 | static void r5c_recovery_load_one_stripe(struct r5l_log *log, | |
2212 | struct stripe_head *sh) | |
355810d1 | 2213 | { |
b4c625c6 SL |
2214 | struct r5dev *dev; |
2215 | int i; | |
2216 | ||
2217 | for (i = sh->disks; i--; ) { | |
2218 | dev = sh->dev + i; | |
2219 | if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) { | |
2220 | set_bit(R5_InJournal, &dev->flags); | |
2221 | set_bit(R5_UPTODATE, &dev->flags); | |
2222 | } | |
2223 | } | |
b4c625c6 SL |
2224 | } |
2225 | ||
2226 | /* | |
2227 | * Scan through the log for all to-be-flushed data | |
2228 | * | |
2229 | * For stripes with data and parity, namely Data-Parity stripe | |
2230 | * (STRIPE_R5C_CACHING == 0), we simply replay all the writes. | |
2231 | * | |
2232 | * For stripes with only data, namely Data-Only stripe | |
2233 | * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine. | |
2234 | * | |
2235 | * For a stripe, if we see data after parity, we should discard all previous | |
2236 | * data and parity for this stripe, as these data are already flushed to | |
2237 | * the array. | |
2238 | * | |
2239 | * At the end of the scan, we return the new journal_tail, which points to | |
2240 | * first data-only stripe on the journal device, or next invalid meta block. | |
2241 | */ | |
2242 | static int r5c_recovery_flush_log(struct r5l_log *log, | |
2243 | struct r5l_recovery_ctx *ctx) | |
2244 | { | |
bc8f167f | 2245 | struct stripe_head *sh; |
b4c625c6 SL |
2246 | int ret = 0; |
2247 | ||
2248 | /* scan through the log */ | |
355810d1 | 2249 | while (1) { |
b4c625c6 SL |
2250 | if (r5l_recovery_read_meta_block(log, ctx)) |
2251 | break; | |
2252 | ||
2253 | ret = r5c_recovery_analyze_meta_block(log, ctx, | |
2254 | &ctx->cached_list); | |
2255 | /* | |
2256 | * -EAGAIN means mismatch in data block, in this case, we still | |
2257 | * try scan the next metablock | |
2258 | */ | |
2259 | if (ret && ret != -EAGAIN) | |
2260 | break; /* ret == -EINVAL or -ENOMEM */ | |
355810d1 SL |
2261 | ctx->seq++; |
2262 | ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks); | |
2263 | } | |
b4c625c6 SL |
2264 | |
2265 | if (ret == -ENOMEM) { | |
2266 | r5c_recovery_drop_stripes(&ctx->cached_list, ctx); | |
2267 | return ret; | |
2268 | } | |
2269 | ||
2270 | /* replay data-parity stripes */ | |
2271 | r5c_recovery_replay_stripes(&ctx->cached_list, ctx); | |
2272 | ||
2273 | /* load data-only stripes to stripe cache */ | |
bc8f167f | 2274 | list_for_each_entry(sh, &ctx->cached_list, lru) { |
b4c625c6 SL |
2275 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); |
2276 | r5c_recovery_load_one_stripe(log, sh); | |
b4c625c6 SL |
2277 | ctx->data_only_stripes++; |
2278 | } | |
2279 | ||
2280 | return 0; | |
355810d1 SL |
2281 | } |
2282 | ||
b4c625c6 SL |
2283 | /* |
2284 | * we did a recovery. Now ctx.pos points to an invalid meta block. New | |
2285 | * log will start here. but we can't let superblock point to last valid | |
2286 | * meta block. The log might looks like: | |
2287 | * | meta 1| meta 2| meta 3| | |
2288 | * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If | |
2289 | * superblock points to meta 1, we write a new valid meta 2n. if crash | |
2290 | * happens again, new recovery will start from meta 1. Since meta 2n is | |
2291 | * valid now, recovery will think meta 3 is valid, which is wrong. | |
2292 | * The solution is we create a new meta in meta2 with its seq == meta | |
3c6edc66 SL |
2293 | * 1's seq + 10000 and let superblock points to meta2. The same recovery |
2294 | * will not think meta 3 is a valid meta, because its seq doesn't match | |
b4c625c6 SL |
2295 | */ |
2296 | ||
2297 | /* | |
2298 | * Before recovery, the log looks like the following | |
2299 | * | |
2300 | * --------------------------------------------- | |
2301 | * | valid log | invalid log | | |
2302 | * --------------------------------------------- | |
2303 | * ^ | |
2304 | * |- log->last_checkpoint | |
2305 | * |- log->last_cp_seq | |
2306 | * | |
2307 | * Now we scan through the log until we see invalid entry | |
2308 | * | |
2309 | * --------------------------------------------- | |
2310 | * | valid log | invalid log | | |
2311 | * --------------------------------------------- | |
2312 | * ^ ^ | |
2313 | * |- log->last_checkpoint |- ctx->pos | |
2314 | * |- log->last_cp_seq |- ctx->seq | |
2315 | * | |
2316 | * From this point, we need to increase seq number by 10 to avoid | |
2317 | * confusing next recovery. | |
2318 | * | |
2319 | * --------------------------------------------- | |
2320 | * | valid log | invalid log | | |
2321 | * --------------------------------------------- | |
2322 | * ^ ^ | |
2323 | * |- log->last_checkpoint |- ctx->pos+1 | |
3c6edc66 | 2324 | * |- log->last_cp_seq |- ctx->seq+10001 |
b4c625c6 SL |
2325 | * |
2326 | * However, it is not safe to start the state machine yet, because data only | |
2327 | * parities are not yet secured in RAID. To save these data only parities, we | |
2328 | * rewrite them from seq+11. | |
2329 | * | |
2330 | * ----------------------------------------------------------------- | |
2331 | * | valid log | data only stripes | invalid log | | |
2332 | * ----------------------------------------------------------------- | |
2333 | * ^ ^ | |
2334 | * |- log->last_checkpoint |- ctx->pos+n | |
3c6edc66 | 2335 | * |- log->last_cp_seq |- ctx->seq+10000+n |
b4c625c6 SL |
2336 | * |
2337 | * If failure happens again during this process, the recovery can safe start | |
2338 | * again from log->last_checkpoint. | |
2339 | * | |
2340 | * Once data only stripes are rewritten to journal, we move log_tail | |
2341 | * | |
2342 | * ----------------------------------------------------------------- | |
2343 | * | old log | data only stripes | invalid log | | |
2344 | * ----------------------------------------------------------------- | |
2345 | * ^ ^ | |
2346 | * |- log->last_checkpoint |- ctx->pos+n | |
3c6edc66 | 2347 | * |- log->last_cp_seq |- ctx->seq+10000+n |
b4c625c6 SL |
2348 | * |
2349 | * Then we can safely start the state machine. If failure happens from this | |
2350 | * point on, the recovery will start from new log->last_checkpoint. | |
2351 | */ | |
2352 | static int | |
2353 | r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log, | |
2354 | struct r5l_recovery_ctx *ctx) | |
355810d1 | 2355 | { |
a85dd7b8 | 2356 | struct stripe_head *sh; |
b4c625c6 | 2357 | struct mddev *mddev = log->rdev->mddev; |
355810d1 | 2358 | struct page *page; |
3c66abba | 2359 | sector_t next_checkpoint = MaxSector; |
355810d1 | 2360 | |
b4c625c6 SL |
2361 | page = alloc_page(GFP_KERNEL); |
2362 | if (!page) { | |
2363 | pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n", | |
2364 | mdname(mddev)); | |
355810d1 | 2365 | return -ENOMEM; |
b4c625c6 | 2366 | } |
355810d1 | 2367 | |
3c66abba SL |
2368 | WARN_ON(list_empty(&ctx->cached_list)); |
2369 | ||
a85dd7b8 | 2370 | list_for_each_entry(sh, &ctx->cached_list, lru) { |
b4c625c6 SL |
2371 | struct r5l_meta_block *mb; |
2372 | int i; | |
2373 | int offset; | |
2374 | sector_t write_pos; | |
2375 | ||
2376 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2377 | r5l_recovery_create_empty_meta_block(log, page, | |
2378 | ctx->pos, ctx->seq); | |
2379 | mb = page_address(page); | |
2380 | offset = le32_to_cpu(mb->meta_size); | |
fc833c2a | 2381 | write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); |
b4c625c6 SL |
2382 | |
2383 | for (i = sh->disks; i--; ) { | |
2384 | struct r5dev *dev = &sh->dev[i]; | |
2385 | struct r5l_payload_data_parity *payload; | |
2386 | void *addr; | |
2387 | ||
2388 | if (test_bit(R5_InJournal, &dev->flags)) { | |
2389 | payload = (void *)mb + offset; | |
2390 | payload->header.type = cpu_to_le16( | |
2391 | R5LOG_PAYLOAD_DATA); | |
1ad45a9b | 2392 | payload->size = cpu_to_le32(BLOCK_SECTORS); |
b4c625c6 SL |
2393 | payload->location = cpu_to_le64( |
2394 | raid5_compute_blocknr(sh, i, 0)); | |
2395 | addr = kmap_atomic(dev->page); | |
2396 | payload->checksum[0] = cpu_to_le32( | |
2397 | crc32c_le(log->uuid_checksum, addr, | |
2398 | PAGE_SIZE)); | |
2399 | kunmap_atomic(addr); | |
2400 | sync_page_io(log->rdev, write_pos, PAGE_SIZE, | |
2401 | dev->page, REQ_OP_WRITE, 0, false); | |
2402 | write_pos = r5l_ring_add(log, write_pos, | |
2403 | BLOCK_SECTORS); | |
2404 | offset += sizeof(__le32) + | |
2405 | sizeof(struct r5l_payload_data_parity); | |
2406 | ||
2407 | } | |
2408 | } | |
2409 | mb->meta_size = cpu_to_le32(offset); | |
5c88f403 SL |
2410 | mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum, |
2411 | mb, PAGE_SIZE)); | |
b4c625c6 | 2412 | sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, |
5a8948f8 | 2413 | REQ_OP_WRITE, REQ_SYNC | REQ_FUA, false); |
b4c625c6 | 2414 | sh->log_start = ctx->pos; |
3c66abba SL |
2415 | list_add_tail(&sh->r5c, &log->stripe_in_journal_list); |
2416 | atomic_inc(&log->stripe_in_journal_count); | |
b4c625c6 SL |
2417 | ctx->pos = write_pos; |
2418 | ctx->seq += 1; | |
3c66abba | 2419 | next_checkpoint = sh->log_start; |
355810d1 | 2420 | } |
3c66abba | 2421 | log->next_checkpoint = next_checkpoint; |
355810d1 SL |
2422 | __free_page(page); |
2423 | return 0; | |
2424 | } | |
2425 | ||
a85dd7b8 SL |
2426 | static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log, |
2427 | struct r5l_recovery_ctx *ctx) | |
2428 | { | |
2429 | struct mddev *mddev = log->rdev->mddev; | |
2430 | struct r5conf *conf = mddev->private; | |
2431 | struct stripe_head *sh, *next; | |
2432 | ||
2433 | if (ctx->data_only_stripes == 0) | |
2434 | return; | |
2435 | ||
2436 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK; | |
2437 | ||
2438 | list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) { | |
2439 | r5c_make_stripe_write_out(sh); | |
2440 | set_bit(STRIPE_HANDLE, &sh->state); | |
2441 | list_del_init(&sh->lru); | |
2442 | raid5_release_stripe(sh); | |
2443 | } | |
2444 | ||
2445 | md_wakeup_thread(conf->mddev->thread); | |
2446 | /* reuse conf->wait_for_quiescent in recovery */ | |
2447 | wait_event(conf->wait_for_quiescent, | |
2448 | atomic_read(&conf->active_stripes) == 0); | |
2449 | ||
2450 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; | |
2451 | } | |
2452 | ||
f6bed0ef SL |
2453 | static int r5l_recovery_log(struct r5l_log *log) |
2454 | { | |
5aabf7c4 | 2455 | struct mddev *mddev = log->rdev->mddev; |
effe6ee7 | 2456 | struct r5l_recovery_ctx *ctx; |
5aabf7c4 | 2457 | int ret; |
43b96748 | 2458 | sector_t pos; |
355810d1 | 2459 | |
effe6ee7 SL |
2460 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); |
2461 | if (!ctx) | |
355810d1 SL |
2462 | return -ENOMEM; |
2463 | ||
effe6ee7 SL |
2464 | ctx->pos = log->last_checkpoint; |
2465 | ctx->seq = log->last_cp_seq; | |
2466 | INIT_LIST_HEAD(&ctx->cached_list); | |
2467 | ctx->meta_page = alloc_page(GFP_KERNEL); | |
355810d1 | 2468 | |
effe6ee7 SL |
2469 | if (!ctx->meta_page) { |
2470 | ret = -ENOMEM; | |
2471 | goto meta_page; | |
2472 | } | |
b4c625c6 | 2473 | |
effe6ee7 SL |
2474 | if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) { |
2475 | ret = -ENOMEM; | |
2476 | goto ra_pool; | |
2477 | } | |
2478 | ||
2479 | ret = r5c_recovery_flush_log(log, ctx); | |
2480 | ||
2481 | if (ret) | |
2482 | goto error; | |
43b96748 | 2483 | |
effe6ee7 SL |
2484 | pos = ctx->pos; |
2485 | ctx->seq += 10000; | |
43b96748 | 2486 | |
effe6ee7 | 2487 | if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0)) |
5aabf7c4 SL |
2488 | pr_debug("md/raid:%s: starting from clean shutdown\n", |
2489 | mdname(mddev)); | |
a85dd7b8 | 2490 | else |
99f17890 | 2491 | pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n", |
effe6ee7 SL |
2492 | mdname(mddev), ctx->data_only_stripes, |
2493 | ctx->data_parity_stripes); | |
2494 | ||
2495 | if (ctx->data_only_stripes == 0) { | |
2496 | log->next_checkpoint = ctx->pos; | |
2497 | r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++); | |
2498 | ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); | |
2499 | } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) { | |
a85dd7b8 SL |
2500 | pr_err("md/raid:%s: failed to rewrite stripes to journal\n", |
2501 | mdname(mddev)); | |
effe6ee7 SL |
2502 | ret = -EIO; |
2503 | goto error; | |
b4c625c6 SL |
2504 | } |
2505 | ||
effe6ee7 SL |
2506 | log->log_start = ctx->pos; |
2507 | log->seq = ctx->seq; | |
43b96748 J |
2508 | log->last_checkpoint = pos; |
2509 | r5l_write_super(log, pos); | |
a85dd7b8 | 2510 | |
effe6ee7 SL |
2511 | r5c_recovery_flush_data_only_stripes(log, ctx); |
2512 | ret = 0; | |
2513 | error: | |
2514 | r5l_recovery_free_ra_pool(log, ctx); | |
2515 | ra_pool: | |
2516 | __free_page(ctx->meta_page); | |
2517 | meta_page: | |
2518 | kfree(ctx); | |
2519 | return ret; | |
f6bed0ef SL |
2520 | } |
2521 | ||
2522 | static void r5l_write_super(struct r5l_log *log, sector_t cp) | |
2523 | { | |
2524 | struct mddev *mddev = log->rdev->mddev; | |
2525 | ||
2526 | log->rdev->journal_tail = cp; | |
2953079c | 2527 | set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags); |
f6bed0ef SL |
2528 | } |
2529 | ||
2c7da14b SL |
2530 | static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page) |
2531 | { | |
2532 | struct r5conf *conf = mddev->private; | |
2533 | int ret; | |
2534 | ||
2535 | if (!conf->log) | |
2536 | return 0; | |
2537 | ||
2538 | switch (conf->log->r5c_journal_mode) { | |
2539 | case R5C_JOURNAL_MODE_WRITE_THROUGH: | |
2540 | ret = snprintf( | |
2541 | page, PAGE_SIZE, "[%s] %s\n", | |
2542 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], | |
2543 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); | |
2544 | break; | |
2545 | case R5C_JOURNAL_MODE_WRITE_BACK: | |
2546 | ret = snprintf( | |
2547 | page, PAGE_SIZE, "%s [%s]\n", | |
2548 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], | |
2549 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); | |
2550 | break; | |
2551 | default: | |
2552 | ret = 0; | |
2553 | } | |
2554 | return ret; | |
2555 | } | |
2556 | ||
78e470c2 HM |
2557 | /* |
2558 | * Set journal cache mode on @mddev (external API initially needed by dm-raid). | |
2559 | * | |
2560 | * @mode as defined in 'enum r5c_journal_mode'. | |
2561 | * | |
2562 | */ | |
2563 | int r5c_journal_mode_set(struct mddev *mddev, int mode) | |
2c7da14b | 2564 | { |
b44886c5 SL |
2565 | struct r5conf *conf; |
2566 | int err; | |
2c7da14b | 2567 | |
78e470c2 HM |
2568 | if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH || |
2569 | mode > R5C_JOURNAL_MODE_WRITE_BACK) | |
2c7da14b SL |
2570 | return -EINVAL; |
2571 | ||
b44886c5 SL |
2572 | err = mddev_lock(mddev); |
2573 | if (err) | |
2574 | return err; | |
2575 | conf = mddev->private; | |
2576 | if (!conf || !conf->log) { | |
2577 | mddev_unlock(mddev); | |
2578 | return -ENODEV; | |
2579 | } | |
2580 | ||
2e38a37f | 2581 | if (raid5_calc_degraded(conf) > 0 && |
b44886c5 SL |
2582 | mode == R5C_JOURNAL_MODE_WRITE_BACK) { |
2583 | mddev_unlock(mddev); | |
2e38a37f | 2584 | return -EINVAL; |
b44886c5 | 2585 | } |
2e38a37f | 2586 | |
2c7da14b | 2587 | mddev_suspend(mddev); |
78e470c2 | 2588 | conf->log->r5c_journal_mode = mode; |
2c7da14b | 2589 | mddev_resume(mddev); |
b44886c5 | 2590 | mddev_unlock(mddev); |
2c7da14b SL |
2591 | |
2592 | pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n", | |
78e470c2 HM |
2593 | mdname(mddev), mode, r5c_journal_mode_str[mode]); |
2594 | return 0; | |
2595 | } | |
2596 | EXPORT_SYMBOL(r5c_journal_mode_set); | |
2597 | ||
2598 | static ssize_t r5c_journal_mode_store(struct mddev *mddev, | |
2599 | const char *page, size_t length) | |
2600 | { | |
2601 | int mode = ARRAY_SIZE(r5c_journal_mode_str); | |
2602 | size_t len = length; | |
2603 | ||
2604 | if (len < 2) | |
2605 | return -EINVAL; | |
2606 | ||
2607 | if (page[len - 1] == '\n') | |
2608 | len--; | |
2609 | ||
2610 | while (mode--) | |
2611 | if (strlen(r5c_journal_mode_str[mode]) == len && | |
2612 | !strncmp(page, r5c_journal_mode_str[mode], len)) | |
2613 | break; | |
2614 | ||
2615 | return r5c_journal_mode_set(mddev, mode) ?: length; | |
2c7da14b SL |
2616 | } |
2617 | ||
2618 | struct md_sysfs_entry | |
2619 | r5c_journal_mode = __ATTR(journal_mode, 0644, | |
2620 | r5c_journal_mode_show, r5c_journal_mode_store); | |
2621 | ||
2ded3703 SL |
2622 | /* |
2623 | * Try handle write operation in caching phase. This function should only | |
2624 | * be called in write-back mode. | |
2625 | * | |
2626 | * If all outstanding writes can be handled in caching phase, returns 0 | |
2627 | * If writes requires write-out phase, call r5c_make_stripe_write_out() | |
2628 | * and returns -EAGAIN | |
2629 | */ | |
2630 | int r5c_try_caching_write(struct r5conf *conf, | |
2631 | struct stripe_head *sh, | |
2632 | struct stripe_head_state *s, | |
2633 | int disks) | |
2634 | { | |
2635 | struct r5l_log *log = conf->log; | |
1e6d690b SL |
2636 | int i; |
2637 | struct r5dev *dev; | |
2638 | int to_cache = 0; | |
03b047f4 SL |
2639 | void **pslot; |
2640 | sector_t tree_index; | |
2641 | int ret; | |
2642 | uintptr_t refcount; | |
2ded3703 SL |
2643 | |
2644 | BUG_ON(!r5c_is_writeback(log)); | |
2645 | ||
1e6d690b SL |
2646 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { |
2647 | /* | |
2648 | * There are two different scenarios here: | |
2649 | * 1. The stripe has some data cached, and it is sent to | |
2650 | * write-out phase for reclaim | |
2651 | * 2. The stripe is clean, and this is the first write | |
2652 | * | |
2653 | * For 1, return -EAGAIN, so we continue with | |
2654 | * handle_stripe_dirtying(). | |
2655 | * | |
2656 | * For 2, set STRIPE_R5C_CACHING and continue with caching | |
2657 | * write. | |
2658 | */ | |
2659 | ||
2660 | /* case 1: anything injournal or anything in written */ | |
2661 | if (s->injournal > 0 || s->written > 0) | |
2662 | return -EAGAIN; | |
2663 | /* case 2 */ | |
2664 | set_bit(STRIPE_R5C_CACHING, &sh->state); | |
2665 | } | |
2666 | ||
2e38a37f SL |
2667 | /* |
2668 | * When run in degraded mode, array is set to write-through mode. | |
2669 | * This check helps drain pending write safely in the transition to | |
2670 | * write-through mode. | |
5ddf0440 SL |
2671 | * |
2672 | * When a stripe is syncing, the write is also handled in write | |
2673 | * through mode. | |
2e38a37f | 2674 | */ |
5ddf0440 | 2675 | if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) { |
2e38a37f SL |
2676 | r5c_make_stripe_write_out(sh); |
2677 | return -EAGAIN; | |
2678 | } | |
2679 | ||
1e6d690b SL |
2680 | for (i = disks; i--; ) { |
2681 | dev = &sh->dev[i]; | |
2682 | /* if non-overwrite, use writing-out phase */ | |
2683 | if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) && | |
2684 | !test_bit(R5_InJournal, &dev->flags)) { | |
2685 | r5c_make_stripe_write_out(sh); | |
2686 | return -EAGAIN; | |
2687 | } | |
2688 | } | |
2689 | ||
03b047f4 SL |
2690 | /* if the stripe is not counted in big_stripe_tree, add it now */ |
2691 | if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) && | |
2692 | !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { | |
2693 | tree_index = r5c_tree_index(conf, sh->sector); | |
2694 | spin_lock(&log->tree_lock); | |
2695 | pslot = radix_tree_lookup_slot(&log->big_stripe_tree, | |
2696 | tree_index); | |
2697 | if (pslot) { | |
2698 | refcount = (uintptr_t)radix_tree_deref_slot_protected( | |
2699 | pslot, &log->tree_lock) >> | |
2700 | R5C_RADIX_COUNT_SHIFT; | |
2701 | radix_tree_replace_slot( | |
2702 | &log->big_stripe_tree, pslot, | |
2703 | (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT)); | |
2704 | } else { | |
2705 | /* | |
2706 | * this radix_tree_insert can fail safely, so no | |
2707 | * need to call radix_tree_preload() | |
2708 | */ | |
2709 | ret = radix_tree_insert( | |
2710 | &log->big_stripe_tree, tree_index, | |
2711 | (void *)(1 << R5C_RADIX_COUNT_SHIFT)); | |
2712 | if (ret) { | |
2713 | spin_unlock(&log->tree_lock); | |
2714 | r5c_make_stripe_write_out(sh); | |
2715 | return -EAGAIN; | |
2716 | } | |
2717 | } | |
2718 | spin_unlock(&log->tree_lock); | |
2719 | ||
2720 | /* | |
2721 | * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is | |
2722 | * counted in the radix tree | |
2723 | */ | |
2724 | set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state); | |
2725 | atomic_inc(&conf->r5c_cached_partial_stripes); | |
2726 | } | |
2727 | ||
1e6d690b SL |
2728 | for (i = disks; i--; ) { |
2729 | dev = &sh->dev[i]; | |
2730 | if (dev->towrite) { | |
2731 | set_bit(R5_Wantwrite, &dev->flags); | |
2732 | set_bit(R5_Wantdrain, &dev->flags); | |
2733 | set_bit(R5_LOCKED, &dev->flags); | |
2734 | to_cache++; | |
2735 | } | |
2736 | } | |
2737 | ||
2738 | if (to_cache) { | |
2739 | set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); | |
2740 | /* | |
2741 | * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data() | |
2742 | * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in | |
2743 | * r5c_handle_data_cached() | |
2744 | */ | |
2745 | set_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
2746 | } | |
2747 | ||
2748 | return 0; | |
2749 | } | |
2750 | ||
2751 | /* | |
2752 | * free extra pages (orig_page) we allocated for prexor | |
2753 | */ | |
2754 | void r5c_release_extra_page(struct stripe_head *sh) | |
2755 | { | |
d7bd398e | 2756 | struct r5conf *conf = sh->raid_conf; |
1e6d690b | 2757 | int i; |
d7bd398e SL |
2758 | bool using_disk_info_extra_page; |
2759 | ||
2760 | using_disk_info_extra_page = | |
2761 | sh->dev[0].orig_page == conf->disks[0].extra_page; | |
1e6d690b SL |
2762 | |
2763 | for (i = sh->disks; i--; ) | |
2764 | if (sh->dev[i].page != sh->dev[i].orig_page) { | |
2765 | struct page *p = sh->dev[i].orig_page; | |
2766 | ||
2767 | sh->dev[i].orig_page = sh->dev[i].page; | |
86aa1397 SL |
2768 | clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags); |
2769 | ||
d7bd398e SL |
2770 | if (!using_disk_info_extra_page) |
2771 | put_page(p); | |
1e6d690b | 2772 | } |
d7bd398e SL |
2773 | |
2774 | if (using_disk_info_extra_page) { | |
2775 | clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state); | |
2776 | md_wakeup_thread(conf->mddev->thread); | |
2777 | } | |
2778 | } | |
2779 | ||
2780 | void r5c_use_extra_page(struct stripe_head *sh) | |
2781 | { | |
2782 | struct r5conf *conf = sh->raid_conf; | |
2783 | int i; | |
2784 | struct r5dev *dev; | |
2785 | ||
2786 | for (i = sh->disks; i--; ) { | |
2787 | dev = &sh->dev[i]; | |
2788 | if (dev->orig_page != dev->page) | |
2789 | put_page(dev->orig_page); | |
2790 | dev->orig_page = conf->disks[i].extra_page; | |
2791 | } | |
2ded3703 SL |
2792 | } |
2793 | ||
2794 | /* | |
2795 | * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the | |
2796 | * stripe is committed to RAID disks. | |
2797 | */ | |
2798 | void r5c_finish_stripe_write_out(struct r5conf *conf, | |
2799 | struct stripe_head *sh, | |
2800 | struct stripe_head_state *s) | |
2801 | { | |
03b047f4 | 2802 | struct r5l_log *log = conf->log; |
1e6d690b SL |
2803 | int i; |
2804 | int do_wakeup = 0; | |
03b047f4 SL |
2805 | sector_t tree_index; |
2806 | void **pslot; | |
2807 | uintptr_t refcount; | |
1e6d690b | 2808 | |
03b047f4 | 2809 | if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags)) |
2ded3703 SL |
2810 | return; |
2811 | ||
2812 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2813 | clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
2814 | ||
03b047f4 | 2815 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) |
2ded3703 | 2816 | return; |
1e6d690b SL |
2817 | |
2818 | for (i = sh->disks; i--; ) { | |
2819 | clear_bit(R5_InJournal, &sh->dev[i].flags); | |
2820 | if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) | |
2821 | do_wakeup = 1; | |
2822 | } | |
2823 | ||
2824 | /* | |
2825 | * analyse_stripe() runs before r5c_finish_stripe_write_out(), | |
2826 | * We updated R5_InJournal, so we also update s->injournal. | |
2827 | */ | |
2828 | s->injournal = 0; | |
2829 | ||
2830 | if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) | |
2831 | if (atomic_dec_and_test(&conf->pending_full_writes)) | |
2832 | md_wakeup_thread(conf->mddev->thread); | |
2833 | ||
2834 | if (do_wakeup) | |
2835 | wake_up(&conf->wait_for_overlap); | |
a39f7afd | 2836 | |
03b047f4 | 2837 | spin_lock_irq(&log->stripe_in_journal_lock); |
a39f7afd | 2838 | list_del_init(&sh->r5c); |
03b047f4 | 2839 | spin_unlock_irq(&log->stripe_in_journal_lock); |
a39f7afd | 2840 | sh->log_start = MaxSector; |
03b047f4 SL |
2841 | |
2842 | atomic_dec(&log->stripe_in_journal_count); | |
2843 | r5c_update_log_state(log); | |
2844 | ||
2845 | /* stop counting this stripe in big_stripe_tree */ | |
2846 | if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) || | |
2847 | test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { | |
2848 | tree_index = r5c_tree_index(conf, sh->sector); | |
2849 | spin_lock(&log->tree_lock); | |
2850 | pslot = radix_tree_lookup_slot(&log->big_stripe_tree, | |
2851 | tree_index); | |
2852 | BUG_ON(pslot == NULL); | |
2853 | refcount = (uintptr_t)radix_tree_deref_slot_protected( | |
2854 | pslot, &log->tree_lock) >> | |
2855 | R5C_RADIX_COUNT_SHIFT; | |
2856 | if (refcount == 1) | |
2857 | radix_tree_delete(&log->big_stripe_tree, tree_index); | |
2858 | else | |
2859 | radix_tree_replace_slot( | |
2860 | &log->big_stripe_tree, pslot, | |
2861 | (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT)); | |
2862 | spin_unlock(&log->tree_lock); | |
2863 | } | |
2864 | ||
2865 | if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) { | |
2866 | BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0); | |
e33fbb9c | 2867 | atomic_dec(&conf->r5c_flushing_partial_stripes); |
03b047f4 SL |
2868 | atomic_dec(&conf->r5c_cached_partial_stripes); |
2869 | } | |
2870 | ||
2871 | if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { | |
2872 | BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0); | |
e33fbb9c | 2873 | atomic_dec(&conf->r5c_flushing_full_stripes); |
03b047f4 SL |
2874 | atomic_dec(&conf->r5c_cached_full_stripes); |
2875 | } | |
ea17481f SL |
2876 | |
2877 | r5l_append_flush_payload(log, sh->sector); | |
5ddf0440 SL |
2878 | /* stripe is flused to raid disks, we can do resync now */ |
2879 | if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) | |
2880 | set_bit(STRIPE_HANDLE, &sh->state); | |
1e6d690b SL |
2881 | } |
2882 | ||
ff875738 | 2883 | int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh) |
1e6d690b | 2884 | { |
a39f7afd | 2885 | struct r5conf *conf = sh->raid_conf; |
1e6d690b SL |
2886 | int pages = 0; |
2887 | int reserve; | |
2888 | int i; | |
2889 | int ret = 0; | |
2890 | ||
2891 | BUG_ON(!log); | |
2892 | ||
2893 | for (i = 0; i < sh->disks; i++) { | |
2894 | void *addr; | |
2895 | ||
2896 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags)) | |
2897 | continue; | |
2898 | addr = kmap_atomic(sh->dev[i].page); | |
2899 | sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, | |
2900 | addr, PAGE_SIZE); | |
2901 | kunmap_atomic(addr); | |
2902 | pages++; | |
2903 | } | |
2904 | WARN_ON(pages == 0); | |
2905 | ||
2906 | /* | |
2907 | * The stripe must enter state machine again to call endio, so | |
2908 | * don't delay. | |
2909 | */ | |
2910 | clear_bit(STRIPE_DELAYED, &sh->state); | |
2911 | atomic_inc(&sh->count); | |
2912 | ||
2913 | mutex_lock(&log->io_mutex); | |
2914 | /* meta + data */ | |
2915 | reserve = (1 + pages) << (PAGE_SHIFT - 9); | |
1e6d690b | 2916 | |
a39f7afd SL |
2917 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && |
2918 | sh->log_start == MaxSector) | |
2919 | r5l_add_no_space_stripe(log, sh); | |
2920 | else if (!r5l_has_free_space(log, reserve)) { | |
2921 | if (sh->log_start == log->last_checkpoint) | |
2922 | BUG(); | |
2923 | else | |
2924 | r5l_add_no_space_stripe(log, sh); | |
1e6d690b SL |
2925 | } else { |
2926 | ret = r5l_log_stripe(log, sh, pages, 0); | |
2927 | if (ret) { | |
2928 | spin_lock_irq(&log->io_list_lock); | |
2929 | list_add_tail(&sh->log_list, &log->no_mem_stripes); | |
2930 | spin_unlock_irq(&log->io_list_lock); | |
2931 | } | |
2932 | } | |
2933 | ||
2934 | mutex_unlock(&log->io_mutex); | |
2935 | return 0; | |
f6bed0ef SL |
2936 | } |
2937 | ||
03b047f4 SL |
2938 | /* check whether this big stripe is in write back cache. */ |
2939 | bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect) | |
2940 | { | |
2941 | struct r5l_log *log = conf->log; | |
2942 | sector_t tree_index; | |
2943 | void *slot; | |
2944 | ||
2945 | if (!log) | |
2946 | return false; | |
2947 | ||
2948 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
2949 | tree_index = r5c_tree_index(conf, sect); | |
2950 | slot = radix_tree_lookup(&log->big_stripe_tree, tree_index); | |
2951 | return slot != NULL; | |
2952 | } | |
2953 | ||
f6bed0ef SL |
2954 | static int r5l_load_log(struct r5l_log *log) |
2955 | { | |
2956 | struct md_rdev *rdev = log->rdev; | |
2957 | struct page *page; | |
2958 | struct r5l_meta_block *mb; | |
2959 | sector_t cp = log->rdev->journal_tail; | |
2960 | u32 stored_crc, expected_crc; | |
2961 | bool create_super = false; | |
d30dfeb9 | 2962 | int ret = 0; |
f6bed0ef SL |
2963 | |
2964 | /* Make sure it's valid */ | |
2965 | if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp) | |
2966 | cp = 0; | |
2967 | page = alloc_page(GFP_KERNEL); | |
2968 | if (!page) | |
2969 | return -ENOMEM; | |
2970 | ||
796a5cf0 | 2971 | if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) { |
f6bed0ef SL |
2972 | ret = -EIO; |
2973 | goto ioerr; | |
2974 | } | |
2975 | mb = page_address(page); | |
2976 | ||
2977 | if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || | |
2978 | mb->version != R5LOG_VERSION) { | |
2979 | create_super = true; | |
2980 | goto create; | |
2981 | } | |
2982 | stored_crc = le32_to_cpu(mb->checksum); | |
2983 | mb->checksum = 0; | |
5cb2fbd6 | 2984 | expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); |
f6bed0ef SL |
2985 | if (stored_crc != expected_crc) { |
2986 | create_super = true; | |
2987 | goto create; | |
2988 | } | |
2989 | if (le64_to_cpu(mb->position) != cp) { | |
2990 | create_super = true; | |
2991 | goto create; | |
2992 | } | |
2993 | create: | |
2994 | if (create_super) { | |
2995 | log->last_cp_seq = prandom_u32(); | |
2996 | cp = 0; | |
56056c2e | 2997 | r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq); |
f6bed0ef SL |
2998 | /* |
2999 | * Make sure super points to correct address. Log might have | |
3000 | * data very soon. If super hasn't correct log tail address, | |
3001 | * recovery can't find the log | |
3002 | */ | |
3003 | r5l_write_super(log, cp); | |
3004 | } else | |
3005 | log->last_cp_seq = le64_to_cpu(mb->seq); | |
3006 | ||
3007 | log->device_size = round_down(rdev->sectors, BLOCK_SECTORS); | |
0576b1c6 SL |
3008 | log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT; |
3009 | if (log->max_free_space > RECLAIM_MAX_FREE_SPACE) | |
3010 | log->max_free_space = RECLAIM_MAX_FREE_SPACE; | |
f6bed0ef SL |
3011 | log->last_checkpoint = cp; |
3012 | ||
3013 | __free_page(page); | |
3014 | ||
d30dfeb9 J |
3015 | if (create_super) { |
3016 | log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS); | |
3017 | log->seq = log->last_cp_seq + 1; | |
3018 | log->next_checkpoint = cp; | |
3019 | } else | |
3020 | ret = r5l_recovery_log(log); | |
3021 | ||
3d7e7e1d ZL |
3022 | r5c_update_log_state(log); |
3023 | return ret; | |
f6bed0ef SL |
3024 | ioerr: |
3025 | __free_page(page); | |
3026 | return ret; | |
3027 | } | |
3028 | ||
70d466f7 | 3029 | void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev) |
2e38a37f SL |
3030 | { |
3031 | struct r5conf *conf = mddev->private; | |
3032 | struct r5l_log *log = conf->log; | |
3033 | ||
3034 | if (!log) | |
3035 | return; | |
3036 | ||
70d466f7 SL |
3037 | if ((raid5_calc_degraded(conf) > 0 || |
3038 | test_bit(Journal, &rdev->flags)) && | |
2e38a37f SL |
3039 | conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) |
3040 | schedule_work(&log->disable_writeback_work); | |
3041 | } | |
3042 | ||
f6bed0ef SL |
3043 | int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev) |
3044 | { | |
c888a8f9 | 3045 | struct request_queue *q = bdev_get_queue(rdev->bdev); |
f6bed0ef | 3046 | struct r5l_log *log; |
ff875738 AP |
3047 | char b[BDEVNAME_SIZE]; |
3048 | ||
3049 | pr_debug("md/raid:%s: using device %s as journal\n", | |
3050 | mdname(conf->mddev), bdevname(rdev->bdev, b)); | |
f6bed0ef SL |
3051 | |
3052 | if (PAGE_SIZE != 4096) | |
3053 | return -EINVAL; | |
c757ec95 SL |
3054 | |
3055 | /* | |
3056 | * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and | |
3057 | * raid_disks r5l_payload_data_parity. | |
3058 | * | |
3059 | * Write journal and cache does not work for very big array | |
3060 | * (raid_disks > 203) | |
3061 | */ | |
3062 | if (sizeof(struct r5l_meta_block) + | |
3063 | ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) * | |
3064 | conf->raid_disks) > PAGE_SIZE) { | |
3065 | pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n", | |
3066 | mdname(conf->mddev), conf->raid_disks); | |
3067 | return -EINVAL; | |
3068 | } | |
3069 | ||
f6bed0ef SL |
3070 | log = kzalloc(sizeof(*log), GFP_KERNEL); |
3071 | if (!log) | |
3072 | return -ENOMEM; | |
3073 | log->rdev = rdev; | |
3074 | ||
c888a8f9 | 3075 | log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0; |
56fef7c6 | 3076 | |
5cb2fbd6 SL |
3077 | log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid, |
3078 | sizeof(rdev->mddev->uuid)); | |
f6bed0ef SL |
3079 | |
3080 | mutex_init(&log->io_mutex); | |
3081 | ||
3082 | spin_lock_init(&log->io_list_lock); | |
3083 | INIT_LIST_HEAD(&log->running_ios); | |
0576b1c6 | 3084 | INIT_LIST_HEAD(&log->io_end_ios); |
a8c34f91 | 3085 | INIT_LIST_HEAD(&log->flushing_ios); |
04732f74 | 3086 | INIT_LIST_HEAD(&log->finished_ios); |
3a83f467 | 3087 | bio_init(&log->flush_bio, NULL, 0); |
f6bed0ef SL |
3088 | |
3089 | log->io_kc = KMEM_CACHE(r5l_io_unit, 0); | |
3090 | if (!log->io_kc) | |
3091 | goto io_kc; | |
3092 | ||
5036c390 CH |
3093 | log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc); |
3094 | if (!log->io_pool) | |
3095 | goto io_pool; | |
3096 | ||
011067b0 | 3097 | log->bs = bioset_create(R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS); |
c38d29b3 CH |
3098 | if (!log->bs) |
3099 | goto io_bs; | |
3100 | ||
e8deb638 CH |
3101 | log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0); |
3102 | if (!log->meta_pool) | |
3103 | goto out_mempool; | |
3104 | ||
03b047f4 SL |
3105 | spin_lock_init(&log->tree_lock); |
3106 | INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN); | |
3107 | ||
0576b1c6 SL |
3108 | log->reclaim_thread = md_register_thread(r5l_reclaim_thread, |
3109 | log->rdev->mddev, "reclaim"); | |
3110 | if (!log->reclaim_thread) | |
3111 | goto reclaim_thread; | |
a39f7afd SL |
3112 | log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL; |
3113 | ||
0fd22b45 | 3114 | init_waitqueue_head(&log->iounit_wait); |
0576b1c6 | 3115 | |
5036c390 CH |
3116 | INIT_LIST_HEAD(&log->no_mem_stripes); |
3117 | ||
f6bed0ef SL |
3118 | INIT_LIST_HEAD(&log->no_space_stripes); |
3119 | spin_lock_init(&log->no_space_stripes_lock); | |
3120 | ||
3bddb7f8 | 3121 | INIT_WORK(&log->deferred_io_work, r5l_submit_io_async); |
2e38a37f | 3122 | INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async); |
3bddb7f8 | 3123 | |
2ded3703 | 3124 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; |
a39f7afd SL |
3125 | INIT_LIST_HEAD(&log->stripe_in_journal_list); |
3126 | spin_lock_init(&log->stripe_in_journal_lock); | |
3127 | atomic_set(&log->stripe_in_journal_count, 0); | |
2ded3703 | 3128 | |
d2250f10 SL |
3129 | rcu_assign_pointer(conf->log, log); |
3130 | ||
f6bed0ef SL |
3131 | if (r5l_load_log(log)) |
3132 | goto error; | |
3133 | ||
a62ab49e | 3134 | set_bit(MD_HAS_JOURNAL, &conf->mddev->flags); |
f6bed0ef | 3135 | return 0; |
e8deb638 | 3136 | |
f6bed0ef | 3137 | error: |
d2250f10 | 3138 | rcu_assign_pointer(conf->log, NULL); |
0576b1c6 SL |
3139 | md_unregister_thread(&log->reclaim_thread); |
3140 | reclaim_thread: | |
e8deb638 CH |
3141 | mempool_destroy(log->meta_pool); |
3142 | out_mempool: | |
c38d29b3 CH |
3143 | bioset_free(log->bs); |
3144 | io_bs: | |
5036c390 CH |
3145 | mempool_destroy(log->io_pool); |
3146 | io_pool: | |
f6bed0ef SL |
3147 | kmem_cache_destroy(log->io_kc); |
3148 | io_kc: | |
3149 | kfree(log); | |
3150 | return -EINVAL; | |
3151 | } | |
3152 | ||
ff875738 | 3153 | void r5l_exit_log(struct r5conf *conf) |
f6bed0ef | 3154 | { |
ff875738 AP |
3155 | struct r5l_log *log = conf->log; |
3156 | ||
3157 | conf->log = NULL; | |
3158 | synchronize_rcu(); | |
3159 | ||
2e38a37f | 3160 | flush_work(&log->disable_writeback_work); |
0576b1c6 | 3161 | md_unregister_thread(&log->reclaim_thread); |
e8deb638 | 3162 | mempool_destroy(log->meta_pool); |
c38d29b3 | 3163 | bioset_free(log->bs); |
5036c390 | 3164 | mempool_destroy(log->io_pool); |
f6bed0ef SL |
3165 | kmem_cache_destroy(log->io_kc); |
3166 | kfree(log); | |
3167 | } |