]>
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" |
f6bed0ef SL |
27 | |
28 | /* | |
29 | * metadata/data stored in disk with 4k size unit (a block) regardless | |
30 | * underneath hardware sector size. only works with PAGE_SIZE == 4096 | |
31 | */ | |
32 | #define BLOCK_SECTORS (8) | |
effe6ee7 | 33 | #define BLOCK_SECTOR_SHIFT (3) |
f6bed0ef | 34 | |
0576b1c6 | 35 | /* |
a39f7afd SL |
36 | * log->max_free_space is min(1/4 disk size, 10G reclaimable space). |
37 | * | |
38 | * In write through mode, the reclaim runs every log->max_free_space. | |
39 | * This can prevent the recovery scans for too long | |
0576b1c6 SL |
40 | */ |
41 | #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */ | |
42 | #define RECLAIM_MAX_FREE_SPACE_SHIFT (2) | |
43 | ||
a39f7afd SL |
44 | /* wake up reclaim thread periodically */ |
45 | #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ) | |
46 | /* start flush with these full stripes */ | |
84890c03 | 47 | #define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4) |
a39f7afd SL |
48 | /* reclaim stripes in groups */ |
49 | #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) | |
50 | ||
c38d29b3 CH |
51 | /* |
52 | * We only need 2 bios per I/O unit to make progress, but ensure we | |
53 | * have a few more available to not get too tight. | |
54 | */ | |
55 | #define R5L_POOL_SIZE 4 | |
56 | ||
2ded3703 SL |
57 | /* |
58 | * r5c journal modes of the array: write-back or write-through. | |
59 | * write-through mode has identical behavior as existing log only | |
60 | * implementation. | |
61 | */ | |
62 | enum r5c_journal_mode { | |
63 | R5C_JOURNAL_MODE_WRITE_THROUGH = 0, | |
64 | R5C_JOURNAL_MODE_WRITE_BACK = 1, | |
65 | }; | |
66 | ||
2c7da14b SL |
67 | static char *r5c_journal_mode_str[] = {"write-through", |
68 | "write-back"}; | |
2ded3703 SL |
69 | /* |
70 | * raid5 cache state machine | |
71 | * | |
9b69173e | 72 | * With the RAID cache, each stripe works in two phases: |
2ded3703 SL |
73 | * - caching phase |
74 | * - writing-out phase | |
75 | * | |
76 | * These two phases are controlled by bit STRIPE_R5C_CACHING: | |
77 | * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase | |
78 | * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase | |
79 | * | |
80 | * When there is no journal, or the journal is in write-through mode, | |
81 | * the stripe is always in writing-out phase. | |
82 | * | |
83 | * For write-back journal, the stripe is sent to caching phase on write | |
84 | * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off | |
85 | * the write-out phase by clearing STRIPE_R5C_CACHING. | |
86 | * | |
87 | * Stripes in caching phase do not write the raid disks. Instead, all | |
88 | * writes are committed from the log device. Therefore, a stripe in | |
89 | * caching phase handles writes as: | |
90 | * - write to log device | |
91 | * - return IO | |
92 | * | |
93 | * Stripes in writing-out phase handle writes as: | |
94 | * - calculate parity | |
95 | * - write pending data and parity to journal | |
96 | * - write data and parity to raid disks | |
97 | * - return IO for pending writes | |
98 | */ | |
99 | ||
f6bed0ef SL |
100 | struct r5l_log { |
101 | struct md_rdev *rdev; | |
102 | ||
103 | u32 uuid_checksum; | |
104 | ||
105 | sector_t device_size; /* log device size, round to | |
106 | * BLOCK_SECTORS */ | |
0576b1c6 SL |
107 | sector_t max_free_space; /* reclaim run if free space is at |
108 | * this size */ | |
f6bed0ef SL |
109 | |
110 | sector_t last_checkpoint; /* log tail. where recovery scan | |
111 | * starts from */ | |
112 | u64 last_cp_seq; /* log tail sequence */ | |
113 | ||
114 | sector_t log_start; /* log head. where new data appends */ | |
115 | u64 seq; /* log head sequence */ | |
116 | ||
17036461 | 117 | sector_t next_checkpoint; |
17036461 | 118 | |
f6bed0ef SL |
119 | struct mutex io_mutex; |
120 | struct r5l_io_unit *current_io; /* current io_unit accepting new data */ | |
121 | ||
122 | spinlock_t io_list_lock; | |
123 | struct list_head running_ios; /* io_units which are still running, | |
124 | * and have not yet been completely | |
125 | * written to the log */ | |
126 | struct list_head io_end_ios; /* io_units which have been completely | |
127 | * written to the log but not yet written | |
128 | * to the RAID */ | |
a8c34f91 SL |
129 | struct list_head flushing_ios; /* io_units which are waiting for log |
130 | * cache flush */ | |
04732f74 | 131 | struct list_head finished_ios; /* io_units which settle down in log disk */ |
a8c34f91 | 132 | struct bio flush_bio; |
f6bed0ef | 133 | |
5036c390 CH |
134 | struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */ |
135 | ||
f6bed0ef | 136 | struct kmem_cache *io_kc; |
5036c390 | 137 | mempool_t *io_pool; |
c38d29b3 | 138 | struct bio_set *bs; |
e8deb638 | 139 | mempool_t *meta_pool; |
f6bed0ef | 140 | |
0576b1c6 SL |
141 | struct md_thread *reclaim_thread; |
142 | unsigned long reclaim_target; /* number of space that need to be | |
143 | * reclaimed. if it's 0, reclaim spaces | |
144 | * used by io_units which are in | |
145 | * IO_UNIT_STRIPE_END state (eg, reclaim | |
146 | * dones't wait for specific io_unit | |
147 | * switching to IO_UNIT_STRIPE_END | |
148 | * state) */ | |
0fd22b45 | 149 | wait_queue_head_t iounit_wait; |
0576b1c6 | 150 | |
f6bed0ef SL |
151 | struct list_head no_space_stripes; /* pending stripes, log has no space */ |
152 | spinlock_t no_space_stripes_lock; | |
56fef7c6 CH |
153 | |
154 | bool need_cache_flush; | |
2ded3703 SL |
155 | |
156 | /* for r5c_cache */ | |
157 | enum r5c_journal_mode r5c_journal_mode; | |
a39f7afd SL |
158 | |
159 | /* all stripes in r5cache, in the order of seq at sh->log_start */ | |
160 | struct list_head stripe_in_journal_list; | |
161 | ||
162 | spinlock_t stripe_in_journal_lock; | |
163 | atomic_t stripe_in_journal_count; | |
3bddb7f8 SL |
164 | |
165 | /* to submit async io_units, to fulfill ordering of flush */ | |
166 | struct work_struct deferred_io_work; | |
2e38a37f SL |
167 | /* to disable write back during in degraded mode */ |
168 | struct work_struct disable_writeback_work; | |
03b047f4 SL |
169 | |
170 | /* to for chunk_aligned_read in writeback mode, details below */ | |
171 | spinlock_t tree_lock; | |
172 | struct radix_tree_root big_stripe_tree; | |
f6bed0ef SL |
173 | }; |
174 | ||
03b047f4 SL |
175 | /* |
176 | * Enable chunk_aligned_read() with write back cache. | |
177 | * | |
178 | * Each chunk may contain more than one stripe (for example, a 256kB | |
179 | * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For | |
180 | * chunk_aligned_read, these stripes are grouped into one "big_stripe". | |
181 | * For each big_stripe, we count how many stripes of this big_stripe | |
182 | * are in the write back cache. These data are tracked in a radix tree | |
183 | * (big_stripe_tree). We use radix_tree item pointer as the counter. | |
184 | * r5c_tree_index() is used to calculate keys for the radix tree. | |
185 | * | |
186 | * chunk_aligned_read() calls r5c_big_stripe_cached() to look up | |
187 | * big_stripe of each chunk in the tree. If this big_stripe is in the | |
188 | * tree, chunk_aligned_read() aborts. This look up is protected by | |
189 | * rcu_read_lock(). | |
190 | * | |
191 | * It is necessary to remember whether a stripe is counted in | |
192 | * big_stripe_tree. Instead of adding new flag, we reuses existing flags: | |
193 | * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these | |
194 | * two flags are set, the stripe is counted in big_stripe_tree. This | |
195 | * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to | |
196 | * r5c_try_caching_write(); and moving clear_bit of | |
197 | * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to | |
198 | * r5c_finish_stripe_write_out(). | |
199 | */ | |
200 | ||
201 | /* | |
202 | * radix tree requests lowest 2 bits of data pointer to be 2b'00. | |
203 | * So it is necessary to left shift the counter by 2 bits before using it | |
204 | * as data pointer of the tree. | |
205 | */ | |
206 | #define R5C_RADIX_COUNT_SHIFT 2 | |
207 | ||
208 | /* | |
209 | * calculate key for big_stripe_tree | |
210 | * | |
211 | * sect: align_bi->bi_iter.bi_sector or sh->sector | |
212 | */ | |
213 | static inline sector_t r5c_tree_index(struct r5conf *conf, | |
214 | sector_t sect) | |
215 | { | |
216 | sector_t offset; | |
217 | ||
218 | offset = sector_div(sect, conf->chunk_sectors); | |
219 | return sect; | |
220 | } | |
221 | ||
f6bed0ef SL |
222 | /* |
223 | * an IO range starts from a meta data block and end at the next meta data | |
224 | * block. The io unit's the meta data block tracks data/parity followed it. io | |
225 | * unit is written to log disk with normal write, as we always flush log disk | |
226 | * first and then start move data to raid disks, there is no requirement to | |
227 | * write io unit with FLUSH/FUA | |
228 | */ | |
229 | struct r5l_io_unit { | |
230 | struct r5l_log *log; | |
231 | ||
232 | struct page *meta_page; /* store meta block */ | |
233 | int meta_offset; /* current offset in meta_page */ | |
234 | ||
f6bed0ef SL |
235 | struct bio *current_bio;/* current_bio accepting new data */ |
236 | ||
237 | atomic_t pending_stripe;/* how many stripes not flushed to raid */ | |
238 | u64 seq; /* seq number of the metablock */ | |
239 | sector_t log_start; /* where the io_unit starts */ | |
240 | sector_t log_end; /* where the io_unit ends */ | |
241 | struct list_head log_sibling; /* log->running_ios */ | |
242 | struct list_head stripe_list; /* stripes added to the io_unit */ | |
243 | ||
244 | int state; | |
6143e2ce | 245 | bool need_split_bio; |
3bddb7f8 SL |
246 | struct bio *split_bio; |
247 | ||
248 | unsigned int has_flush:1; /* include flush request */ | |
249 | unsigned int has_fua:1; /* include fua request */ | |
250 | unsigned int has_null_flush:1; /* include empty flush request */ | |
251 | /* | |
252 | * io isn't sent yet, flush/fua request can only be submitted till it's | |
253 | * the first IO in running_ios list | |
254 | */ | |
255 | unsigned int io_deferred:1; | |
256 | ||
257 | struct bio_list flush_barriers; /* size == 0 flush bios */ | |
f6bed0ef SL |
258 | }; |
259 | ||
260 | /* r5l_io_unit state */ | |
261 | enum r5l_io_unit_state { | |
262 | IO_UNIT_RUNNING = 0, /* accepting new IO */ | |
263 | IO_UNIT_IO_START = 1, /* io_unit bio start writing to log, | |
264 | * don't accepting new bio */ | |
265 | IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */ | |
a8c34f91 | 266 | IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */ |
f6bed0ef SL |
267 | }; |
268 | ||
2ded3703 SL |
269 | bool r5c_is_writeback(struct r5l_log *log) |
270 | { | |
271 | return (log != NULL && | |
272 | log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK); | |
273 | } | |
274 | ||
f6bed0ef SL |
275 | static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc) |
276 | { | |
277 | start += inc; | |
278 | if (start >= log->device_size) | |
279 | start = start - log->device_size; | |
280 | return start; | |
281 | } | |
282 | ||
283 | static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start, | |
284 | sector_t end) | |
285 | { | |
286 | if (end >= start) | |
287 | return end - start; | |
288 | else | |
289 | return end + log->device_size - start; | |
290 | } | |
291 | ||
292 | static bool r5l_has_free_space(struct r5l_log *log, sector_t size) | |
293 | { | |
294 | sector_t used_size; | |
295 | ||
296 | used_size = r5l_ring_distance(log, log->last_checkpoint, | |
297 | log->log_start); | |
298 | ||
299 | return log->device_size > used_size + size; | |
300 | } | |
301 | ||
f6bed0ef SL |
302 | static void __r5l_set_io_unit_state(struct r5l_io_unit *io, |
303 | enum r5l_io_unit_state state) | |
304 | { | |
f6bed0ef SL |
305 | if (WARN_ON(io->state >= state)) |
306 | return; | |
307 | io->state = state; | |
f6bed0ef SL |
308 | } |
309 | ||
1e6d690b | 310 | static void |
bd83d0a2 | 311 | r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev) |
1e6d690b SL |
312 | { |
313 | struct bio *wbi, *wbi2; | |
314 | ||
315 | wbi = dev->written; | |
316 | dev->written = NULL; | |
317 | while (wbi && wbi->bi_iter.bi_sector < | |
318 | dev->sector + STRIPE_SECTORS) { | |
319 | wbi2 = r5_next_bio(wbi, dev->sector); | |
49728050 | 320 | md_write_end(conf->mddev); |
bd83d0a2 N |
321 | if (!raid5_dec_bi_active_stripes(wbi)) |
322 | bio_endio(wbi); | |
1e6d690b SL |
323 | wbi = wbi2; |
324 | } | |
325 | } | |
326 | ||
327 | void r5c_handle_cached_data_endio(struct r5conf *conf, | |
bd83d0a2 | 328 | struct stripe_head *sh, int disks) |
1e6d690b SL |
329 | { |
330 | int i; | |
331 | ||
332 | for (i = sh->disks; i--; ) { | |
333 | if (sh->dev[i].written) { | |
334 | set_bit(R5_UPTODATE, &sh->dev[i].flags); | |
bd83d0a2 | 335 | r5c_return_dev_pending_writes(conf, &sh->dev[i]); |
1e6d690b SL |
336 | bitmap_endwrite(conf->mddev->bitmap, sh->sector, |
337 | STRIPE_SECTORS, | |
338 | !test_bit(STRIPE_DEGRADED, &sh->state), | |
339 | 0); | |
340 | } | |
341 | } | |
342 | } | |
343 | ||
ff875738 AP |
344 | void r5l_wake_reclaim(struct r5l_log *log, sector_t space); |
345 | ||
a39f7afd SL |
346 | /* Check whether we should flush some stripes to free up stripe cache */ |
347 | void r5c_check_stripe_cache_usage(struct r5conf *conf) | |
348 | { | |
349 | int total_cached; | |
350 | ||
351 | if (!r5c_is_writeback(conf->log)) | |
352 | return; | |
353 | ||
354 | total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + | |
355 | atomic_read(&conf->r5c_cached_full_stripes); | |
356 | ||
357 | /* | |
358 | * The following condition is true for either of the following: | |
359 | * - stripe cache pressure high: | |
360 | * total_cached > 3/4 min_nr_stripes || | |
361 | * empty_inactive_list_nr > 0 | |
362 | * - stripe cache pressure moderate: | |
363 | * total_cached > 1/2 min_nr_stripes | |
364 | */ | |
365 | if (total_cached > conf->min_nr_stripes * 1 / 2 || | |
366 | atomic_read(&conf->empty_inactive_list_nr) > 0) | |
367 | r5l_wake_reclaim(conf->log, 0); | |
368 | } | |
369 | ||
370 | /* | |
371 | * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full | |
372 | * stripes in the cache | |
373 | */ | |
374 | void r5c_check_cached_full_stripe(struct r5conf *conf) | |
375 | { | |
376 | if (!r5c_is_writeback(conf->log)) | |
377 | return; | |
378 | ||
379 | /* | |
380 | * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes | |
381 | * or a full stripe (chunk size / 4k stripes). | |
382 | */ | |
383 | if (atomic_read(&conf->r5c_cached_full_stripes) >= | |
84890c03 | 384 | min(R5C_FULL_STRIPE_FLUSH_BATCH(conf), |
a39f7afd SL |
385 | conf->chunk_sectors >> STRIPE_SHIFT)) |
386 | r5l_wake_reclaim(conf->log, 0); | |
387 | } | |
388 | ||
389 | /* | |
390 | * Total log space (in sectors) needed to flush all data in cache | |
391 | * | |
39b99586 SL |
392 | * To avoid deadlock due to log space, it is necessary to reserve log |
393 | * space to flush critical stripes (stripes that occupying log space near | |
394 | * last_checkpoint). This function helps check how much log space is | |
395 | * required to flush all cached stripes. | |
a39f7afd | 396 | * |
39b99586 SL |
397 | * To reduce log space requirements, two mechanisms are used to give cache |
398 | * flush higher priorities: | |
399 | * 1. In handle_stripe_dirtying() and schedule_reconstruction(), | |
400 | * stripes ALREADY in journal can be flushed w/o pending writes; | |
401 | * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal | |
402 | * can be delayed (r5l_add_no_space_stripe). | |
a39f7afd | 403 | * |
39b99586 SL |
404 | * In cache flush, the stripe goes through 1 and then 2. For a stripe that |
405 | * already passed 1, flushing it requires at most (conf->max_degraded + 1) | |
406 | * pages of journal space. For stripes that has not passed 1, flushing it | |
407 | * requires (conf->raid_disks + 1) pages of journal space. There are at | |
408 | * most (conf->group_cnt + 1) stripe that passed 1. So total journal space | |
409 | * required to flush all cached stripes (in pages) is: | |
410 | * | |
411 | * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) + | |
412 | * (group_cnt + 1) * (raid_disks + 1) | |
413 | * or | |
414 | * (stripe_in_journal_count) * (max_degraded + 1) + | |
415 | * (group_cnt + 1) * (raid_disks - max_degraded) | |
a39f7afd SL |
416 | */ |
417 | static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf) | |
418 | { | |
419 | struct r5l_log *log = conf->log; | |
420 | ||
421 | if (!r5c_is_writeback(log)) | |
422 | return 0; | |
423 | ||
39b99586 SL |
424 | return BLOCK_SECTORS * |
425 | ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) + | |
426 | (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1)); | |
a39f7afd SL |
427 | } |
428 | ||
429 | /* | |
430 | * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL | |
431 | * | |
432 | * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of | |
433 | * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log | |
434 | * device is less than 2x of reclaim_required_space. | |
435 | */ | |
436 | static inline void r5c_update_log_state(struct r5l_log *log) | |
437 | { | |
438 | struct r5conf *conf = log->rdev->mddev->private; | |
439 | sector_t free_space; | |
440 | sector_t reclaim_space; | |
f687a33e | 441 | bool wake_reclaim = false; |
a39f7afd SL |
442 | |
443 | if (!r5c_is_writeback(log)) | |
444 | return; | |
445 | ||
446 | free_space = r5l_ring_distance(log, log->log_start, | |
447 | log->last_checkpoint); | |
448 | reclaim_space = r5c_log_required_to_flush_cache(conf); | |
449 | if (free_space < 2 * reclaim_space) | |
450 | set_bit(R5C_LOG_CRITICAL, &conf->cache_state); | |
f687a33e SL |
451 | else { |
452 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state)) | |
453 | wake_reclaim = true; | |
a39f7afd | 454 | clear_bit(R5C_LOG_CRITICAL, &conf->cache_state); |
f687a33e | 455 | } |
a39f7afd SL |
456 | if (free_space < 3 * reclaim_space) |
457 | set_bit(R5C_LOG_TIGHT, &conf->cache_state); | |
458 | else | |
459 | clear_bit(R5C_LOG_TIGHT, &conf->cache_state); | |
f687a33e SL |
460 | |
461 | if (wake_reclaim) | |
462 | r5l_wake_reclaim(log, 0); | |
a39f7afd SL |
463 | } |
464 | ||
2ded3703 SL |
465 | /* |
466 | * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING. | |
467 | * This function should only be called in write-back mode. | |
468 | */ | |
a39f7afd | 469 | void r5c_make_stripe_write_out(struct stripe_head *sh) |
2ded3703 SL |
470 | { |
471 | struct r5conf *conf = sh->raid_conf; | |
472 | struct r5l_log *log = conf->log; | |
473 | ||
474 | BUG_ON(!r5c_is_writeback(log)); | |
475 | ||
476 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
477 | clear_bit(STRIPE_R5C_CACHING, &sh->state); | |
1e6d690b SL |
478 | |
479 | if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) | |
480 | atomic_inc(&conf->preread_active_stripes); | |
1e6d690b SL |
481 | } |
482 | ||
483 | static void r5c_handle_data_cached(struct stripe_head *sh) | |
484 | { | |
485 | int i; | |
486 | ||
487 | for (i = sh->disks; i--; ) | |
488 | if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) { | |
489 | set_bit(R5_InJournal, &sh->dev[i].flags); | |
490 | clear_bit(R5_LOCKED, &sh->dev[i].flags); | |
491 | } | |
492 | clear_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
493 | } | |
494 | ||
495 | /* | |
496 | * this journal write must contain full parity, | |
497 | * it may also contain some data pages | |
498 | */ | |
499 | static void r5c_handle_parity_cached(struct stripe_head *sh) | |
500 | { | |
501 | int i; | |
502 | ||
503 | for (i = sh->disks; i--; ) | |
504 | if (test_bit(R5_InJournal, &sh->dev[i].flags)) | |
505 | set_bit(R5_Wantwrite, &sh->dev[i].flags); | |
2ded3703 SL |
506 | } |
507 | ||
508 | /* | |
509 | * Setting proper flags after writing (or flushing) data and/or parity to the | |
510 | * log device. This is called from r5l_log_endio() or r5l_log_flush_endio(). | |
511 | */ | |
512 | static void r5c_finish_cache_stripe(struct stripe_head *sh) | |
513 | { | |
514 | struct r5l_log *log = sh->raid_conf->log; | |
515 | ||
516 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { | |
517 | BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
518 | /* | |
519 | * Set R5_InJournal for parity dev[pd_idx]. This means | |
520 | * all data AND parity in the journal. For RAID 6, it is | |
521 | * NOT necessary to set the flag for dev[qd_idx], as the | |
522 | * two parities are written out together. | |
523 | */ | |
524 | set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
1e6d690b SL |
525 | } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) { |
526 | r5c_handle_data_cached(sh); | |
527 | } else { | |
528 | r5c_handle_parity_cached(sh); | |
529 | set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
530 | } | |
2ded3703 SL |
531 | } |
532 | ||
d8858f43 CH |
533 | static void r5l_io_run_stripes(struct r5l_io_unit *io) |
534 | { | |
535 | struct stripe_head *sh, *next; | |
536 | ||
537 | list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) { | |
538 | list_del_init(&sh->log_list); | |
2ded3703 SL |
539 | |
540 | r5c_finish_cache_stripe(sh); | |
541 | ||
d8858f43 CH |
542 | set_bit(STRIPE_HANDLE, &sh->state); |
543 | raid5_release_stripe(sh); | |
544 | } | |
545 | } | |
546 | ||
56fef7c6 CH |
547 | static void r5l_log_run_stripes(struct r5l_log *log) |
548 | { | |
549 | struct r5l_io_unit *io, *next; | |
550 | ||
551 | assert_spin_locked(&log->io_list_lock); | |
552 | ||
553 | list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { | |
554 | /* don't change list order */ | |
555 | if (io->state < IO_UNIT_IO_END) | |
556 | break; | |
557 | ||
558 | list_move_tail(&io->log_sibling, &log->finished_ios); | |
559 | r5l_io_run_stripes(io); | |
560 | } | |
561 | } | |
562 | ||
3848c0bc CH |
563 | static void r5l_move_to_end_ios(struct r5l_log *log) |
564 | { | |
565 | struct r5l_io_unit *io, *next; | |
566 | ||
567 | assert_spin_locked(&log->io_list_lock); | |
568 | ||
569 | list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { | |
570 | /* don't change list order */ | |
571 | if (io->state < IO_UNIT_IO_END) | |
572 | break; | |
573 | list_move_tail(&io->log_sibling, &log->io_end_ios); | |
574 | } | |
575 | } | |
576 | ||
3bddb7f8 | 577 | static void __r5l_stripe_write_finished(struct r5l_io_unit *io); |
f6bed0ef SL |
578 | static void r5l_log_endio(struct bio *bio) |
579 | { | |
580 | struct r5l_io_unit *io = bio->bi_private; | |
3bddb7f8 | 581 | struct r5l_io_unit *io_deferred; |
f6bed0ef | 582 | struct r5l_log *log = io->log; |
509ffec7 | 583 | unsigned long flags; |
f6bed0ef | 584 | |
6e74a9cf SL |
585 | if (bio->bi_error) |
586 | md_error(log->rdev->mddev, log->rdev); | |
587 | ||
f6bed0ef | 588 | bio_put(bio); |
e8deb638 | 589 | mempool_free(io->meta_page, log->meta_pool); |
f6bed0ef | 590 | |
509ffec7 CH |
591 | spin_lock_irqsave(&log->io_list_lock, flags); |
592 | __r5l_set_io_unit_state(io, IO_UNIT_IO_END); | |
ea17481f | 593 | if (log->need_cache_flush && !list_empty(&io->stripe_list)) |
3848c0bc | 594 | r5l_move_to_end_ios(log); |
56fef7c6 CH |
595 | else |
596 | r5l_log_run_stripes(log); | |
3bddb7f8 SL |
597 | if (!list_empty(&log->running_ios)) { |
598 | /* | |
599 | * FLUSH/FUA io_unit is deferred because of ordering, now we | |
600 | * can dispatch it | |
601 | */ | |
602 | io_deferred = list_first_entry(&log->running_ios, | |
603 | struct r5l_io_unit, log_sibling); | |
604 | if (io_deferred->io_deferred) | |
605 | schedule_work(&log->deferred_io_work); | |
606 | } | |
607 | ||
509ffec7 CH |
608 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
609 | ||
56fef7c6 CH |
610 | if (log->need_cache_flush) |
611 | md_wakeup_thread(log->rdev->mddev->thread); | |
3bddb7f8 SL |
612 | |
613 | if (io->has_null_flush) { | |
614 | struct bio *bi; | |
615 | ||
616 | WARN_ON(bio_list_empty(&io->flush_barriers)); | |
617 | while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) { | |
618 | bio_endio(bi); | |
619 | atomic_dec(&io->pending_stripe); | |
620 | } | |
3bddb7f8 | 621 | } |
ea17481f SL |
622 | |
623 | /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */ | |
624 | if (atomic_read(&io->pending_stripe) == 0) | |
625 | __r5l_stripe_write_finished(io); | |
3bddb7f8 SL |
626 | } |
627 | ||
628 | static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io) | |
629 | { | |
630 | unsigned long flags; | |
631 | ||
632 | spin_lock_irqsave(&log->io_list_lock, flags); | |
633 | __r5l_set_io_unit_state(io, IO_UNIT_IO_START); | |
634 | spin_unlock_irqrestore(&log->io_list_lock, flags); | |
635 | ||
636 | if (io->has_flush) | |
20737738 | 637 | io->current_bio->bi_opf |= REQ_PREFLUSH; |
3bddb7f8 | 638 | if (io->has_fua) |
20737738 | 639 | io->current_bio->bi_opf |= REQ_FUA; |
3bddb7f8 SL |
640 | submit_bio(io->current_bio); |
641 | ||
642 | if (!io->split_bio) | |
643 | return; | |
644 | ||
645 | if (io->has_flush) | |
20737738 | 646 | io->split_bio->bi_opf |= REQ_PREFLUSH; |
3bddb7f8 | 647 | if (io->has_fua) |
20737738 | 648 | io->split_bio->bi_opf |= REQ_FUA; |
3bddb7f8 SL |
649 | submit_bio(io->split_bio); |
650 | } | |
651 | ||
652 | /* deferred io_unit will be dispatched here */ | |
653 | static void r5l_submit_io_async(struct work_struct *work) | |
654 | { | |
655 | struct r5l_log *log = container_of(work, struct r5l_log, | |
656 | deferred_io_work); | |
657 | struct r5l_io_unit *io = NULL; | |
658 | unsigned long flags; | |
659 | ||
660 | spin_lock_irqsave(&log->io_list_lock, flags); | |
661 | if (!list_empty(&log->running_ios)) { | |
662 | io = list_first_entry(&log->running_ios, struct r5l_io_unit, | |
663 | log_sibling); | |
664 | if (!io->io_deferred) | |
665 | io = NULL; | |
666 | else | |
667 | io->io_deferred = 0; | |
668 | } | |
669 | spin_unlock_irqrestore(&log->io_list_lock, flags); | |
670 | if (io) | |
671 | r5l_do_submit_io(log, io); | |
f6bed0ef SL |
672 | } |
673 | ||
2e38a37f SL |
674 | static void r5c_disable_writeback_async(struct work_struct *work) |
675 | { | |
676 | struct r5l_log *log = container_of(work, struct r5l_log, | |
677 | disable_writeback_work); | |
678 | struct mddev *mddev = log->rdev->mddev; | |
679 | ||
680 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) | |
681 | return; | |
682 | pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n", | |
683 | mdname(mddev)); | |
684 | mddev_suspend(mddev); | |
685 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; | |
686 | mddev_resume(mddev); | |
687 | } | |
688 | ||
f6bed0ef SL |
689 | static void r5l_submit_current_io(struct r5l_log *log) |
690 | { | |
691 | struct r5l_io_unit *io = log->current_io; | |
3bddb7f8 | 692 | struct bio *bio; |
f6bed0ef | 693 | struct r5l_meta_block *block; |
509ffec7 | 694 | unsigned long flags; |
f6bed0ef | 695 | u32 crc; |
3bddb7f8 | 696 | bool do_submit = true; |
f6bed0ef SL |
697 | |
698 | if (!io) | |
699 | return; | |
700 | ||
701 | block = page_address(io->meta_page); | |
702 | block->meta_size = cpu_to_le32(io->meta_offset); | |
5cb2fbd6 | 703 | crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE); |
f6bed0ef | 704 | block->checksum = cpu_to_le32(crc); |
3bddb7f8 | 705 | bio = io->current_bio; |
f6bed0ef SL |
706 | |
707 | log->current_io = NULL; | |
509ffec7 | 708 | spin_lock_irqsave(&log->io_list_lock, flags); |
3bddb7f8 SL |
709 | if (io->has_flush || io->has_fua) { |
710 | if (io != list_first_entry(&log->running_ios, | |
711 | struct r5l_io_unit, log_sibling)) { | |
712 | io->io_deferred = 1; | |
713 | do_submit = false; | |
714 | } | |
715 | } | |
509ffec7 | 716 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
3bddb7f8 SL |
717 | if (do_submit) |
718 | r5l_do_submit_io(log, io); | |
f6bed0ef SL |
719 | } |
720 | ||
6143e2ce | 721 | static struct bio *r5l_bio_alloc(struct r5l_log *log) |
b349feb3 | 722 | { |
c38d29b3 | 723 | struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs); |
b349feb3 | 724 | |
796a5cf0 | 725 | bio_set_op_attrs(bio, REQ_OP_WRITE, 0); |
b349feb3 | 726 | bio->bi_bdev = log->rdev->bdev; |
1e932a37 | 727 | bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start; |
b349feb3 | 728 | |
b349feb3 CH |
729 | return bio; |
730 | } | |
731 | ||
c1b99198 CH |
732 | static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io) |
733 | { | |
734 | log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS); | |
735 | ||
a39f7afd | 736 | r5c_update_log_state(log); |
c1b99198 CH |
737 | /* |
738 | * If we filled up the log device start from the beginning again, | |
739 | * which will require a new bio. | |
740 | * | |
741 | * Note: for this to work properly the log size needs to me a multiple | |
742 | * of BLOCK_SECTORS. | |
743 | */ | |
744 | if (log->log_start == 0) | |
6143e2ce | 745 | io->need_split_bio = true; |
c1b99198 CH |
746 | |
747 | io->log_end = log->log_start; | |
748 | } | |
749 | ||
f6bed0ef SL |
750 | static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log) |
751 | { | |
752 | struct r5l_io_unit *io; | |
753 | struct r5l_meta_block *block; | |
f6bed0ef | 754 | |
5036c390 CH |
755 | io = mempool_alloc(log->io_pool, GFP_ATOMIC); |
756 | if (!io) | |
757 | return NULL; | |
758 | memset(io, 0, sizeof(*io)); | |
759 | ||
51039cd0 | 760 | io->log = log; |
51039cd0 CH |
761 | INIT_LIST_HEAD(&io->log_sibling); |
762 | INIT_LIST_HEAD(&io->stripe_list); | |
3bddb7f8 | 763 | bio_list_init(&io->flush_barriers); |
51039cd0 | 764 | io->state = IO_UNIT_RUNNING; |
f6bed0ef | 765 | |
e8deb638 | 766 | io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO); |
f6bed0ef | 767 | block = page_address(io->meta_page); |
e8deb638 | 768 | clear_page(block); |
f6bed0ef SL |
769 | block->magic = cpu_to_le32(R5LOG_MAGIC); |
770 | block->version = R5LOG_VERSION; | |
771 | block->seq = cpu_to_le64(log->seq); | |
772 | block->position = cpu_to_le64(log->log_start); | |
773 | ||
774 | io->log_start = log->log_start; | |
775 | io->meta_offset = sizeof(struct r5l_meta_block); | |
2b8ef16e | 776 | io->seq = log->seq++; |
f6bed0ef | 777 | |
6143e2ce CH |
778 | io->current_bio = r5l_bio_alloc(log); |
779 | io->current_bio->bi_end_io = r5l_log_endio; | |
780 | io->current_bio->bi_private = io; | |
b349feb3 | 781 | bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0); |
f6bed0ef | 782 | |
c1b99198 | 783 | r5_reserve_log_entry(log, io); |
f6bed0ef SL |
784 | |
785 | spin_lock_irq(&log->io_list_lock); | |
786 | list_add_tail(&io->log_sibling, &log->running_ios); | |
787 | spin_unlock_irq(&log->io_list_lock); | |
788 | ||
789 | return io; | |
790 | } | |
791 | ||
792 | static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size) | |
793 | { | |
22581f58 CH |
794 | if (log->current_io && |
795 | log->current_io->meta_offset + payload_size > PAGE_SIZE) | |
f6bed0ef | 796 | r5l_submit_current_io(log); |
f6bed0ef | 797 | |
5036c390 | 798 | if (!log->current_io) { |
22581f58 | 799 | log->current_io = r5l_new_meta(log); |
5036c390 CH |
800 | if (!log->current_io) |
801 | return -ENOMEM; | |
802 | } | |
803 | ||
f6bed0ef SL |
804 | return 0; |
805 | } | |
806 | ||
807 | static void r5l_append_payload_meta(struct r5l_log *log, u16 type, | |
808 | sector_t location, | |
809 | u32 checksum1, u32 checksum2, | |
810 | bool checksum2_valid) | |
811 | { | |
812 | struct r5l_io_unit *io = log->current_io; | |
813 | struct r5l_payload_data_parity *payload; | |
814 | ||
815 | payload = page_address(io->meta_page) + io->meta_offset; | |
816 | payload->header.type = cpu_to_le16(type); | |
817 | payload->header.flags = cpu_to_le16(0); | |
818 | payload->size = cpu_to_le32((1 + !!checksum2_valid) << | |
819 | (PAGE_SHIFT - 9)); | |
820 | payload->location = cpu_to_le64(location); | |
821 | payload->checksum[0] = cpu_to_le32(checksum1); | |
822 | if (checksum2_valid) | |
823 | payload->checksum[1] = cpu_to_le32(checksum2); | |
824 | ||
825 | io->meta_offset += sizeof(struct r5l_payload_data_parity) + | |
826 | sizeof(__le32) * (1 + !!checksum2_valid); | |
827 | } | |
828 | ||
829 | static void r5l_append_payload_page(struct r5l_log *log, struct page *page) | |
830 | { | |
831 | struct r5l_io_unit *io = log->current_io; | |
832 | ||
6143e2ce | 833 | if (io->need_split_bio) { |
3bddb7f8 SL |
834 | BUG_ON(io->split_bio); |
835 | io->split_bio = io->current_bio; | |
6143e2ce | 836 | io->current_bio = r5l_bio_alloc(log); |
3bddb7f8 SL |
837 | bio_chain(io->current_bio, io->split_bio); |
838 | io->need_split_bio = false; | |
f6bed0ef | 839 | } |
f6bed0ef | 840 | |
6143e2ce CH |
841 | if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) |
842 | BUG(); | |
843 | ||
c1b99198 | 844 | r5_reserve_log_entry(log, io); |
f6bed0ef SL |
845 | } |
846 | ||
ea17481f SL |
847 | static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect) |
848 | { | |
849 | struct mddev *mddev = log->rdev->mddev; | |
850 | struct r5conf *conf = mddev->private; | |
851 | struct r5l_io_unit *io; | |
852 | struct r5l_payload_flush *payload; | |
853 | int meta_size; | |
854 | ||
855 | /* | |
856 | * payload_flush requires extra writes to the journal. | |
857 | * To avoid handling the extra IO in quiesce, just skip | |
858 | * flush_payload | |
859 | */ | |
860 | if (conf->quiesce) | |
861 | return; | |
862 | ||
863 | mutex_lock(&log->io_mutex); | |
864 | meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64); | |
865 | ||
866 | if (r5l_get_meta(log, meta_size)) { | |
867 | mutex_unlock(&log->io_mutex); | |
868 | return; | |
869 | } | |
870 | ||
871 | /* current implementation is one stripe per flush payload */ | |
872 | io = log->current_io; | |
873 | payload = page_address(io->meta_page) + io->meta_offset; | |
874 | payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH); | |
875 | payload->header.flags = cpu_to_le16(0); | |
876 | payload->size = cpu_to_le32(sizeof(__le64)); | |
877 | payload->flush_stripes[0] = cpu_to_le64(sect); | |
878 | io->meta_offset += meta_size; | |
879 | mutex_unlock(&log->io_mutex); | |
880 | } | |
881 | ||
5036c390 | 882 | static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh, |
f6bed0ef SL |
883 | int data_pages, int parity_pages) |
884 | { | |
885 | int i; | |
886 | int meta_size; | |
5036c390 | 887 | int ret; |
f6bed0ef SL |
888 | struct r5l_io_unit *io; |
889 | ||
890 | meta_size = | |
891 | ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) | |
892 | * data_pages) + | |
893 | sizeof(struct r5l_payload_data_parity) + | |
894 | sizeof(__le32) * parity_pages; | |
895 | ||
5036c390 CH |
896 | ret = r5l_get_meta(log, meta_size); |
897 | if (ret) | |
898 | return ret; | |
899 | ||
f6bed0ef SL |
900 | io = log->current_io; |
901 | ||
3bddb7f8 SL |
902 | if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state)) |
903 | io->has_flush = 1; | |
904 | ||
f6bed0ef | 905 | for (i = 0; i < sh->disks; i++) { |
1e6d690b SL |
906 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || |
907 | test_bit(R5_InJournal, &sh->dev[i].flags)) | |
f6bed0ef SL |
908 | continue; |
909 | if (i == sh->pd_idx || i == sh->qd_idx) | |
910 | continue; | |
3bddb7f8 SL |
911 | if (test_bit(R5_WantFUA, &sh->dev[i].flags) && |
912 | log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) { | |
913 | io->has_fua = 1; | |
914 | /* | |
915 | * we need to flush journal to make sure recovery can | |
916 | * reach the data with fua flag | |
917 | */ | |
918 | io->has_flush = 1; | |
919 | } | |
f6bed0ef SL |
920 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA, |
921 | raid5_compute_blocknr(sh, i, 0), | |
922 | sh->dev[i].log_checksum, 0, false); | |
923 | r5l_append_payload_page(log, sh->dev[i].page); | |
924 | } | |
925 | ||
2ded3703 | 926 | if (parity_pages == 2) { |
f6bed0ef SL |
927 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, |
928 | sh->sector, sh->dev[sh->pd_idx].log_checksum, | |
929 | sh->dev[sh->qd_idx].log_checksum, true); | |
930 | r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); | |
931 | r5l_append_payload_page(log, sh->dev[sh->qd_idx].page); | |
2ded3703 | 932 | } else if (parity_pages == 1) { |
f6bed0ef SL |
933 | r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, |
934 | sh->sector, sh->dev[sh->pd_idx].log_checksum, | |
935 | 0, false); | |
936 | r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); | |
2ded3703 SL |
937 | } else /* Just writing data, not parity, in caching phase */ |
938 | BUG_ON(parity_pages != 0); | |
f6bed0ef SL |
939 | |
940 | list_add_tail(&sh->log_list, &io->stripe_list); | |
941 | atomic_inc(&io->pending_stripe); | |
942 | sh->log_io = io; | |
5036c390 | 943 | |
a39f7afd SL |
944 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) |
945 | return 0; | |
946 | ||
947 | if (sh->log_start == MaxSector) { | |
948 | BUG_ON(!list_empty(&sh->r5c)); | |
949 | sh->log_start = io->log_start; | |
950 | spin_lock_irq(&log->stripe_in_journal_lock); | |
951 | list_add_tail(&sh->r5c, | |
952 | &log->stripe_in_journal_list); | |
953 | spin_unlock_irq(&log->stripe_in_journal_lock); | |
954 | atomic_inc(&log->stripe_in_journal_count); | |
955 | } | |
5036c390 | 956 | return 0; |
f6bed0ef SL |
957 | } |
958 | ||
a39f7afd SL |
959 | /* add stripe to no_space_stripes, and then wake up reclaim */ |
960 | static inline void r5l_add_no_space_stripe(struct r5l_log *log, | |
961 | struct stripe_head *sh) | |
962 | { | |
963 | spin_lock(&log->no_space_stripes_lock); | |
964 | list_add_tail(&sh->log_list, &log->no_space_stripes); | |
965 | spin_unlock(&log->no_space_stripes_lock); | |
966 | } | |
967 | ||
f6bed0ef SL |
968 | /* |
969 | * running in raid5d, where reclaim could wait for raid5d too (when it flushes | |
970 | * data from log to raid disks), so we shouldn't wait for reclaim here | |
971 | */ | |
972 | int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh) | |
973 | { | |
a39f7afd | 974 | struct r5conf *conf = sh->raid_conf; |
f6bed0ef SL |
975 | int write_disks = 0; |
976 | int data_pages, parity_pages; | |
f6bed0ef SL |
977 | int reserve; |
978 | int i; | |
5036c390 | 979 | int ret = 0; |
a39f7afd | 980 | bool wake_reclaim = false; |
f6bed0ef SL |
981 | |
982 | if (!log) | |
983 | return -EAGAIN; | |
984 | /* Don't support stripe batch */ | |
985 | if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) || | |
986 | test_bit(STRIPE_SYNCING, &sh->state)) { | |
987 | /* the stripe is written to log, we start writing it to raid */ | |
988 | clear_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
989 | return -EAGAIN; | |
990 | } | |
991 | ||
2ded3703 SL |
992 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); |
993 | ||
f6bed0ef SL |
994 | for (i = 0; i < sh->disks; i++) { |
995 | void *addr; | |
996 | ||
1e6d690b SL |
997 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || |
998 | test_bit(R5_InJournal, &sh->dev[i].flags)) | |
f6bed0ef | 999 | continue; |
1e6d690b | 1000 | |
f6bed0ef SL |
1001 | write_disks++; |
1002 | /* checksum is already calculated in last run */ | |
1003 | if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) | |
1004 | continue; | |
1005 | addr = kmap_atomic(sh->dev[i].page); | |
5cb2fbd6 SL |
1006 | sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, |
1007 | addr, PAGE_SIZE); | |
f6bed0ef SL |
1008 | kunmap_atomic(addr); |
1009 | } | |
1010 | parity_pages = 1 + !!(sh->qd_idx >= 0); | |
1011 | data_pages = write_disks - parity_pages; | |
1012 | ||
f6bed0ef | 1013 | set_bit(STRIPE_LOG_TRAPPED, &sh->state); |
253f9fd4 SL |
1014 | /* |
1015 | * The stripe must enter state machine again to finish the write, so | |
1016 | * don't delay. | |
1017 | */ | |
1018 | clear_bit(STRIPE_DELAYED, &sh->state); | |
f6bed0ef SL |
1019 | atomic_inc(&sh->count); |
1020 | ||
1021 | mutex_lock(&log->io_mutex); | |
1022 | /* meta + data */ | |
1023 | reserve = (1 + write_disks) << (PAGE_SHIFT - 9); | |
f6bed0ef | 1024 | |
a39f7afd SL |
1025 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { |
1026 | if (!r5l_has_free_space(log, reserve)) { | |
1027 | r5l_add_no_space_stripe(log, sh); | |
1028 | wake_reclaim = true; | |
1029 | } else { | |
1030 | ret = r5l_log_stripe(log, sh, data_pages, parity_pages); | |
1031 | if (ret) { | |
1032 | spin_lock_irq(&log->io_list_lock); | |
1033 | list_add_tail(&sh->log_list, | |
1034 | &log->no_mem_stripes); | |
1035 | spin_unlock_irq(&log->io_list_lock); | |
1036 | } | |
1037 | } | |
1038 | } else { /* R5C_JOURNAL_MODE_WRITE_BACK */ | |
1039 | /* | |
1040 | * log space critical, do not process stripes that are | |
1041 | * not in cache yet (sh->log_start == MaxSector). | |
1042 | */ | |
1043 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && | |
1044 | sh->log_start == MaxSector) { | |
1045 | r5l_add_no_space_stripe(log, sh); | |
1046 | wake_reclaim = true; | |
1047 | reserve = 0; | |
1048 | } else if (!r5l_has_free_space(log, reserve)) { | |
1049 | if (sh->log_start == log->last_checkpoint) | |
1050 | BUG(); | |
1051 | else | |
1052 | r5l_add_no_space_stripe(log, sh); | |
1053 | } else { | |
1054 | ret = r5l_log_stripe(log, sh, data_pages, parity_pages); | |
1055 | if (ret) { | |
1056 | spin_lock_irq(&log->io_list_lock); | |
1057 | list_add_tail(&sh->log_list, | |
1058 | &log->no_mem_stripes); | |
1059 | spin_unlock_irq(&log->io_list_lock); | |
1060 | } | |
5036c390 | 1061 | } |
f6bed0ef | 1062 | } |
f6bed0ef | 1063 | |
5036c390 | 1064 | mutex_unlock(&log->io_mutex); |
a39f7afd SL |
1065 | if (wake_reclaim) |
1066 | r5l_wake_reclaim(log, reserve); | |
f6bed0ef SL |
1067 | return 0; |
1068 | } | |
1069 | ||
1070 | void r5l_write_stripe_run(struct r5l_log *log) | |
1071 | { | |
1072 | if (!log) | |
1073 | return; | |
1074 | mutex_lock(&log->io_mutex); | |
1075 | r5l_submit_current_io(log); | |
1076 | mutex_unlock(&log->io_mutex); | |
1077 | } | |
1078 | ||
828cbe98 SL |
1079 | int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio) |
1080 | { | |
1081 | if (!log) | |
1082 | return -ENODEV; | |
3bddb7f8 SL |
1083 | |
1084 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { | |
1085 | /* | |
1086 | * in write through (journal only) | |
1087 | * we flush log disk cache first, then write stripe data to | |
1088 | * raid disks. So if bio is finished, the log disk cache is | |
1089 | * flushed already. The recovery guarantees we can recovery | |
1090 | * the bio from log disk, so we don't need to flush again | |
1091 | */ | |
1092 | if (bio->bi_iter.bi_size == 0) { | |
1093 | bio_endio(bio); | |
1094 | return 0; | |
1095 | } | |
1096 | bio->bi_opf &= ~REQ_PREFLUSH; | |
1097 | } else { | |
1098 | /* write back (with cache) */ | |
1099 | if (bio->bi_iter.bi_size == 0) { | |
1100 | mutex_lock(&log->io_mutex); | |
1101 | r5l_get_meta(log, 0); | |
1102 | bio_list_add(&log->current_io->flush_barriers, bio); | |
1103 | log->current_io->has_flush = 1; | |
1104 | log->current_io->has_null_flush = 1; | |
1105 | atomic_inc(&log->current_io->pending_stripe); | |
1106 | r5l_submit_current_io(log); | |
1107 | mutex_unlock(&log->io_mutex); | |
1108 | return 0; | |
1109 | } | |
828cbe98 | 1110 | } |
828cbe98 SL |
1111 | return -EAGAIN; |
1112 | } | |
1113 | ||
f6bed0ef SL |
1114 | /* This will run after log space is reclaimed */ |
1115 | static void r5l_run_no_space_stripes(struct r5l_log *log) | |
1116 | { | |
1117 | struct stripe_head *sh; | |
1118 | ||
1119 | spin_lock(&log->no_space_stripes_lock); | |
1120 | while (!list_empty(&log->no_space_stripes)) { | |
1121 | sh = list_first_entry(&log->no_space_stripes, | |
1122 | struct stripe_head, log_list); | |
1123 | list_del_init(&sh->log_list); | |
1124 | set_bit(STRIPE_HANDLE, &sh->state); | |
1125 | raid5_release_stripe(sh); | |
1126 | } | |
1127 | spin_unlock(&log->no_space_stripes_lock); | |
1128 | } | |
1129 | ||
a39f7afd SL |
1130 | /* |
1131 | * calculate new last_checkpoint | |
1132 | * for write through mode, returns log->next_checkpoint | |
1133 | * for write back, returns log_start of first sh in stripe_in_journal_list | |
1134 | */ | |
1135 | static sector_t r5c_calculate_new_cp(struct r5conf *conf) | |
1136 | { | |
1137 | struct stripe_head *sh; | |
1138 | struct r5l_log *log = conf->log; | |
1139 | sector_t new_cp; | |
1140 | unsigned long flags; | |
1141 | ||
1142 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) | |
1143 | return log->next_checkpoint; | |
1144 | ||
1145 | spin_lock_irqsave(&log->stripe_in_journal_lock, flags); | |
1146 | if (list_empty(&conf->log->stripe_in_journal_list)) { | |
1147 | /* all stripes flushed */ | |
d3014e21 | 1148 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); |
a39f7afd SL |
1149 | return log->next_checkpoint; |
1150 | } | |
1151 | sh = list_first_entry(&conf->log->stripe_in_journal_list, | |
1152 | struct stripe_head, r5c); | |
1153 | new_cp = sh->log_start; | |
1154 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); | |
1155 | return new_cp; | |
1156 | } | |
1157 | ||
17036461 CH |
1158 | static sector_t r5l_reclaimable_space(struct r5l_log *log) |
1159 | { | |
a39f7afd SL |
1160 | struct r5conf *conf = log->rdev->mddev->private; |
1161 | ||
17036461 | 1162 | return r5l_ring_distance(log, log->last_checkpoint, |
a39f7afd | 1163 | r5c_calculate_new_cp(conf)); |
17036461 CH |
1164 | } |
1165 | ||
5036c390 CH |
1166 | static void r5l_run_no_mem_stripe(struct r5l_log *log) |
1167 | { | |
1168 | struct stripe_head *sh; | |
1169 | ||
1170 | assert_spin_locked(&log->io_list_lock); | |
1171 | ||
1172 | if (!list_empty(&log->no_mem_stripes)) { | |
1173 | sh = list_first_entry(&log->no_mem_stripes, | |
1174 | struct stripe_head, log_list); | |
1175 | list_del_init(&sh->log_list); | |
1176 | set_bit(STRIPE_HANDLE, &sh->state); | |
1177 | raid5_release_stripe(sh); | |
1178 | } | |
1179 | } | |
1180 | ||
04732f74 | 1181 | static bool r5l_complete_finished_ios(struct r5l_log *log) |
17036461 CH |
1182 | { |
1183 | struct r5l_io_unit *io, *next; | |
1184 | bool found = false; | |
1185 | ||
1186 | assert_spin_locked(&log->io_list_lock); | |
1187 | ||
04732f74 | 1188 | list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) { |
17036461 CH |
1189 | /* don't change list order */ |
1190 | if (io->state < IO_UNIT_STRIPE_END) | |
1191 | break; | |
1192 | ||
1193 | log->next_checkpoint = io->log_start; | |
17036461 CH |
1194 | |
1195 | list_del(&io->log_sibling); | |
5036c390 CH |
1196 | mempool_free(io, log->io_pool); |
1197 | r5l_run_no_mem_stripe(log); | |
17036461 CH |
1198 | |
1199 | found = true; | |
1200 | } | |
1201 | ||
1202 | return found; | |
1203 | } | |
1204 | ||
509ffec7 CH |
1205 | static void __r5l_stripe_write_finished(struct r5l_io_unit *io) |
1206 | { | |
1207 | struct r5l_log *log = io->log; | |
a39f7afd | 1208 | struct r5conf *conf = log->rdev->mddev->private; |
509ffec7 CH |
1209 | unsigned long flags; |
1210 | ||
1211 | spin_lock_irqsave(&log->io_list_lock, flags); | |
1212 | __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END); | |
17036461 | 1213 | |
04732f74 | 1214 | if (!r5l_complete_finished_ios(log)) { |
85f2f9a4 SL |
1215 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
1216 | return; | |
1217 | } | |
509ffec7 | 1218 | |
a39f7afd SL |
1219 | if (r5l_reclaimable_space(log) > log->max_free_space || |
1220 | test_bit(R5C_LOG_TIGHT, &conf->cache_state)) | |
509ffec7 CH |
1221 | r5l_wake_reclaim(log, 0); |
1222 | ||
509ffec7 CH |
1223 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
1224 | wake_up(&log->iounit_wait); | |
1225 | } | |
1226 | ||
0576b1c6 SL |
1227 | void r5l_stripe_write_finished(struct stripe_head *sh) |
1228 | { | |
1229 | struct r5l_io_unit *io; | |
1230 | ||
0576b1c6 | 1231 | io = sh->log_io; |
0576b1c6 SL |
1232 | sh->log_io = NULL; |
1233 | ||
509ffec7 CH |
1234 | if (io && atomic_dec_and_test(&io->pending_stripe)) |
1235 | __r5l_stripe_write_finished(io); | |
0576b1c6 SL |
1236 | } |
1237 | ||
a8c34f91 SL |
1238 | static void r5l_log_flush_endio(struct bio *bio) |
1239 | { | |
1240 | struct r5l_log *log = container_of(bio, struct r5l_log, | |
1241 | flush_bio); | |
1242 | unsigned long flags; | |
1243 | struct r5l_io_unit *io; | |
a8c34f91 | 1244 | |
6e74a9cf SL |
1245 | if (bio->bi_error) |
1246 | md_error(log->rdev->mddev, log->rdev); | |
1247 | ||
a8c34f91 | 1248 | spin_lock_irqsave(&log->io_list_lock, flags); |
d8858f43 CH |
1249 | list_for_each_entry(io, &log->flushing_ios, log_sibling) |
1250 | r5l_io_run_stripes(io); | |
04732f74 | 1251 | list_splice_tail_init(&log->flushing_ios, &log->finished_ios); |
a8c34f91 SL |
1252 | spin_unlock_irqrestore(&log->io_list_lock, flags); |
1253 | } | |
1254 | ||
0576b1c6 SL |
1255 | /* |
1256 | * Starting dispatch IO to raid. | |
1257 | * io_unit(meta) consists of a log. There is one situation we want to avoid. A | |
1258 | * broken meta in the middle of a log causes recovery can't find meta at the | |
1259 | * head of log. If operations require meta at the head persistent in log, we | |
1260 | * must make sure meta before it persistent in log too. A case is: | |
1261 | * | |
1262 | * stripe data/parity is in log, we start write stripe to raid disks. stripe | |
1263 | * data/parity must be persistent in log before we do the write to raid disks. | |
1264 | * | |
1265 | * The solution is we restrictly maintain io_unit list order. In this case, we | |
1266 | * only write stripes of an io_unit to raid disks till the io_unit is the first | |
1267 | * one whose data/parity is in log. | |
1268 | */ | |
1269 | void r5l_flush_stripe_to_raid(struct r5l_log *log) | |
1270 | { | |
a8c34f91 | 1271 | bool do_flush; |
56fef7c6 CH |
1272 | |
1273 | if (!log || !log->need_cache_flush) | |
0576b1c6 | 1274 | return; |
0576b1c6 SL |
1275 | |
1276 | spin_lock_irq(&log->io_list_lock); | |
a8c34f91 SL |
1277 | /* flush bio is running */ |
1278 | if (!list_empty(&log->flushing_ios)) { | |
1279 | spin_unlock_irq(&log->io_list_lock); | |
1280 | return; | |
0576b1c6 | 1281 | } |
a8c34f91 SL |
1282 | list_splice_tail_init(&log->io_end_ios, &log->flushing_ios); |
1283 | do_flush = !list_empty(&log->flushing_ios); | |
0576b1c6 | 1284 | spin_unlock_irq(&log->io_list_lock); |
a8c34f91 SL |
1285 | |
1286 | if (!do_flush) | |
1287 | return; | |
1288 | bio_reset(&log->flush_bio); | |
1289 | log->flush_bio.bi_bdev = log->rdev->bdev; | |
1290 | log->flush_bio.bi_end_io = r5l_log_flush_endio; | |
70fd7614 | 1291 | log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH; |
4e49ea4a | 1292 | submit_bio(&log->flush_bio); |
0576b1c6 SL |
1293 | } |
1294 | ||
0576b1c6 | 1295 | static void r5l_write_super(struct r5l_log *log, sector_t cp); |
4b482044 SL |
1296 | static void r5l_write_super_and_discard_space(struct r5l_log *log, |
1297 | sector_t end) | |
1298 | { | |
1299 | struct block_device *bdev = log->rdev->bdev; | |
1300 | struct mddev *mddev; | |
1301 | ||
1302 | r5l_write_super(log, end); | |
1303 | ||
1304 | if (!blk_queue_discard(bdev_get_queue(bdev))) | |
1305 | return; | |
1306 | ||
1307 | mddev = log->rdev->mddev; | |
1308 | /* | |
8e018c21 SL |
1309 | * Discard could zero data, so before discard we must make sure |
1310 | * superblock is updated to new log tail. Updating superblock (either | |
1311 | * directly call md_update_sb() or depend on md thread) must hold | |
1312 | * reconfig mutex. On the other hand, raid5_quiesce is called with | |
1313 | * reconfig_mutex hold. The first step of raid5_quiesce() is waitting | |
1314 | * for all IO finish, hence waitting for reclaim thread, while reclaim | |
1315 | * thread is calling this function and waitting for reconfig mutex. So | |
1316 | * there is a deadlock. We workaround this issue with a trylock. | |
1317 | * FIXME: we could miss discard if we can't take reconfig mutex | |
4b482044 | 1318 | */ |
2953079c SL |
1319 | set_mask_bits(&mddev->sb_flags, 0, |
1320 | BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING)); | |
8e018c21 SL |
1321 | if (!mddev_trylock(mddev)) |
1322 | return; | |
1323 | md_update_sb(mddev, 1); | |
1324 | mddev_unlock(mddev); | |
4b482044 | 1325 | |
6e74a9cf | 1326 | /* discard IO error really doesn't matter, ignore it */ |
4b482044 SL |
1327 | if (log->last_checkpoint < end) { |
1328 | blkdev_issue_discard(bdev, | |
1329 | log->last_checkpoint + log->rdev->data_offset, | |
1330 | end - log->last_checkpoint, GFP_NOIO, 0); | |
1331 | } else { | |
1332 | blkdev_issue_discard(bdev, | |
1333 | log->last_checkpoint + log->rdev->data_offset, | |
1334 | log->device_size - log->last_checkpoint, | |
1335 | GFP_NOIO, 0); | |
1336 | blkdev_issue_discard(bdev, log->rdev->data_offset, end, | |
1337 | GFP_NOIO, 0); | |
1338 | } | |
1339 | } | |
1340 | ||
a39f7afd SL |
1341 | /* |
1342 | * r5c_flush_stripe moves stripe from cached list to handle_list. When called, | |
1343 | * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes. | |
1344 | * | |
1345 | * must hold conf->device_lock | |
1346 | */ | |
1347 | static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh) | |
0576b1c6 | 1348 | { |
a39f7afd SL |
1349 | BUG_ON(list_empty(&sh->lru)); |
1350 | BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
1351 | BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); | |
0576b1c6 | 1352 | |
0576b1c6 | 1353 | /* |
a39f7afd SL |
1354 | * The stripe is not ON_RELEASE_LIST, so it is safe to call |
1355 | * raid5_release_stripe() while holding conf->device_lock | |
0576b1c6 | 1356 | */ |
a39f7afd SL |
1357 | BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state)); |
1358 | assert_spin_locked(&conf->device_lock); | |
0576b1c6 | 1359 | |
a39f7afd SL |
1360 | list_del_init(&sh->lru); |
1361 | atomic_inc(&sh->count); | |
17036461 | 1362 | |
a39f7afd SL |
1363 | set_bit(STRIPE_HANDLE, &sh->state); |
1364 | atomic_inc(&conf->active_stripes); | |
1365 | r5c_make_stripe_write_out(sh); | |
0576b1c6 | 1366 | |
e33fbb9c SL |
1367 | if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) |
1368 | atomic_inc(&conf->r5c_flushing_partial_stripes); | |
1369 | else | |
1370 | atomic_inc(&conf->r5c_flushing_full_stripes); | |
a39f7afd SL |
1371 | raid5_release_stripe(sh); |
1372 | } | |
1373 | ||
1374 | /* | |
1375 | * if num == 0, flush all full stripes | |
1376 | * if num > 0, flush all full stripes. If less than num full stripes are | |
1377 | * flushed, flush some partial stripes until totally num stripes are | |
1378 | * flushed or there is no more cached stripes. | |
1379 | */ | |
1380 | void r5c_flush_cache(struct r5conf *conf, int num) | |
1381 | { | |
1382 | int count; | |
1383 | struct stripe_head *sh, *next; | |
1384 | ||
1385 | assert_spin_locked(&conf->device_lock); | |
1386 | if (!conf->log) | |
0576b1c6 SL |
1387 | return; |
1388 | ||
a39f7afd SL |
1389 | count = 0; |
1390 | list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) { | |
1391 | r5c_flush_stripe(conf, sh); | |
1392 | count++; | |
1393 | } | |
1394 | ||
1395 | if (count >= num) | |
1396 | return; | |
1397 | list_for_each_entry_safe(sh, next, | |
1398 | &conf->r5c_partial_stripe_list, lru) { | |
1399 | r5c_flush_stripe(conf, sh); | |
1400 | if (++count >= num) | |
1401 | break; | |
1402 | } | |
1403 | } | |
1404 | ||
1405 | static void r5c_do_reclaim(struct r5conf *conf) | |
1406 | { | |
1407 | struct r5l_log *log = conf->log; | |
1408 | struct stripe_head *sh; | |
1409 | int count = 0; | |
1410 | unsigned long flags; | |
1411 | int total_cached; | |
1412 | int stripes_to_flush; | |
e33fbb9c | 1413 | int flushing_partial, flushing_full; |
a39f7afd SL |
1414 | |
1415 | if (!r5c_is_writeback(log)) | |
1416 | return; | |
1417 | ||
e33fbb9c SL |
1418 | flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes); |
1419 | flushing_full = atomic_read(&conf->r5c_flushing_full_stripes); | |
a39f7afd | 1420 | total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + |
e33fbb9c SL |
1421 | atomic_read(&conf->r5c_cached_full_stripes) - |
1422 | flushing_full - flushing_partial; | |
a39f7afd SL |
1423 | |
1424 | if (total_cached > conf->min_nr_stripes * 3 / 4 || | |
1425 | atomic_read(&conf->empty_inactive_list_nr) > 0) | |
1426 | /* | |
1427 | * if stripe cache pressure high, flush all full stripes and | |
1428 | * some partial stripes | |
1429 | */ | |
1430 | stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP; | |
1431 | else if (total_cached > conf->min_nr_stripes * 1 / 2 || | |
e33fbb9c | 1432 | atomic_read(&conf->r5c_cached_full_stripes) - flushing_full > |
84890c03 | 1433 | R5C_FULL_STRIPE_FLUSH_BATCH(conf)) |
a39f7afd SL |
1434 | /* |
1435 | * if stripe cache pressure moderate, or if there is many full | |
1436 | * stripes,flush all full stripes | |
1437 | */ | |
1438 | stripes_to_flush = 0; | |
1439 | else | |
1440 | /* no need to flush */ | |
1441 | stripes_to_flush = -1; | |
1442 | ||
1443 | if (stripes_to_flush >= 0) { | |
1444 | spin_lock_irqsave(&conf->device_lock, flags); | |
1445 | r5c_flush_cache(conf, stripes_to_flush); | |
1446 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1447 | } | |
1448 | ||
1449 | /* if log space is tight, flush stripes on stripe_in_journal_list */ | |
1450 | if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) { | |
1451 | spin_lock_irqsave(&log->stripe_in_journal_lock, flags); | |
1452 | spin_lock(&conf->device_lock); | |
1453 | list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) { | |
1454 | /* | |
1455 | * stripes on stripe_in_journal_list could be in any | |
1456 | * state of the stripe_cache state machine. In this | |
1457 | * case, we only want to flush stripe on | |
1458 | * r5c_cached_full/partial_stripes. The following | |
1459 | * condition makes sure the stripe is on one of the | |
1460 | * two lists. | |
1461 | */ | |
1462 | if (!list_empty(&sh->lru) && | |
1463 | !test_bit(STRIPE_HANDLE, &sh->state) && | |
1464 | atomic_read(&sh->count) == 0) { | |
1465 | r5c_flush_stripe(conf, sh); | |
e8fd52ee SL |
1466 | if (count++ >= R5C_RECLAIM_STRIPE_GROUP) |
1467 | break; | |
a39f7afd | 1468 | } |
a39f7afd SL |
1469 | } |
1470 | spin_unlock(&conf->device_lock); | |
1471 | spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); | |
1472 | } | |
f687a33e SL |
1473 | |
1474 | if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state)) | |
1475 | r5l_run_no_space_stripes(log); | |
1476 | ||
a39f7afd SL |
1477 | md_wakeup_thread(conf->mddev->thread); |
1478 | } | |
1479 | ||
0576b1c6 SL |
1480 | static void r5l_do_reclaim(struct r5l_log *log) |
1481 | { | |
a39f7afd | 1482 | struct r5conf *conf = log->rdev->mddev->private; |
0576b1c6 | 1483 | sector_t reclaim_target = xchg(&log->reclaim_target, 0); |
17036461 CH |
1484 | sector_t reclaimable; |
1485 | sector_t next_checkpoint; | |
a39f7afd | 1486 | bool write_super; |
0576b1c6 SL |
1487 | |
1488 | spin_lock_irq(&log->io_list_lock); | |
a39f7afd SL |
1489 | write_super = r5l_reclaimable_space(log) > log->max_free_space || |
1490 | reclaim_target != 0 || !list_empty(&log->no_space_stripes); | |
0576b1c6 SL |
1491 | /* |
1492 | * move proper io_unit to reclaim list. We should not change the order. | |
1493 | * reclaimable/unreclaimable io_unit can be mixed in the list, we | |
1494 | * shouldn't reuse space of an unreclaimable io_unit | |
1495 | */ | |
1496 | while (1) { | |
17036461 CH |
1497 | reclaimable = r5l_reclaimable_space(log); |
1498 | if (reclaimable >= reclaim_target || | |
0576b1c6 SL |
1499 | (list_empty(&log->running_ios) && |
1500 | list_empty(&log->io_end_ios) && | |
a8c34f91 | 1501 | list_empty(&log->flushing_ios) && |
04732f74 | 1502 | list_empty(&log->finished_ios))) |
0576b1c6 SL |
1503 | break; |
1504 | ||
17036461 CH |
1505 | md_wakeup_thread(log->rdev->mddev->thread); |
1506 | wait_event_lock_irq(log->iounit_wait, | |
1507 | r5l_reclaimable_space(log) > reclaimable, | |
1508 | log->io_list_lock); | |
0576b1c6 | 1509 | } |
17036461 | 1510 | |
a39f7afd | 1511 | next_checkpoint = r5c_calculate_new_cp(conf); |
0576b1c6 SL |
1512 | spin_unlock_irq(&log->io_list_lock); |
1513 | ||
a39f7afd | 1514 | if (reclaimable == 0 || !write_super) |
0576b1c6 SL |
1515 | return; |
1516 | ||
0576b1c6 SL |
1517 | /* |
1518 | * write_super will flush cache of each raid disk. We must write super | |
1519 | * here, because the log area might be reused soon and we don't want to | |
1520 | * confuse recovery | |
1521 | */ | |
4b482044 | 1522 | r5l_write_super_and_discard_space(log, next_checkpoint); |
0576b1c6 SL |
1523 | |
1524 | mutex_lock(&log->io_mutex); | |
17036461 | 1525 | log->last_checkpoint = next_checkpoint; |
a39f7afd | 1526 | r5c_update_log_state(log); |
0576b1c6 | 1527 | mutex_unlock(&log->io_mutex); |
0576b1c6 | 1528 | |
17036461 | 1529 | r5l_run_no_space_stripes(log); |
0576b1c6 SL |
1530 | } |
1531 | ||
1532 | static void r5l_reclaim_thread(struct md_thread *thread) | |
1533 | { | |
1534 | struct mddev *mddev = thread->mddev; | |
1535 | struct r5conf *conf = mddev->private; | |
1536 | struct r5l_log *log = conf->log; | |
1537 | ||
1538 | if (!log) | |
1539 | return; | |
a39f7afd | 1540 | r5c_do_reclaim(conf); |
0576b1c6 SL |
1541 | r5l_do_reclaim(log); |
1542 | } | |
1543 | ||
a39f7afd | 1544 | void r5l_wake_reclaim(struct r5l_log *log, sector_t space) |
f6bed0ef | 1545 | { |
0576b1c6 SL |
1546 | unsigned long target; |
1547 | unsigned long new = (unsigned long)space; /* overflow in theory */ | |
1548 | ||
a39f7afd SL |
1549 | if (!log) |
1550 | return; | |
0576b1c6 SL |
1551 | do { |
1552 | target = log->reclaim_target; | |
1553 | if (new < target) | |
1554 | return; | |
1555 | } while (cmpxchg(&log->reclaim_target, target, new) != target); | |
1556 | md_wakeup_thread(log->reclaim_thread); | |
f6bed0ef SL |
1557 | } |
1558 | ||
e6c033f7 SL |
1559 | void r5l_quiesce(struct r5l_log *log, int state) |
1560 | { | |
4b482044 | 1561 | struct mddev *mddev; |
e6c033f7 SL |
1562 | if (!log || state == 2) |
1563 | return; | |
ce1ccd07 SL |
1564 | if (state == 0) |
1565 | kthread_unpark(log->reclaim_thread->tsk); | |
1566 | else if (state == 1) { | |
4b482044 SL |
1567 | /* make sure r5l_write_super_and_discard_space exits */ |
1568 | mddev = log->rdev->mddev; | |
1569 | wake_up(&mddev->sb_wait); | |
ce1ccd07 | 1570 | kthread_park(log->reclaim_thread->tsk); |
a39f7afd | 1571 | r5l_wake_reclaim(log, MaxSector); |
e6c033f7 SL |
1572 | r5l_do_reclaim(log); |
1573 | } | |
1574 | } | |
1575 | ||
6e74a9cf SL |
1576 | bool r5l_log_disk_error(struct r5conf *conf) |
1577 | { | |
f6b6ec5c SL |
1578 | struct r5l_log *log; |
1579 | bool ret; | |
7dde2ad3 | 1580 | /* don't allow write if journal disk is missing */ |
f6b6ec5c SL |
1581 | rcu_read_lock(); |
1582 | log = rcu_dereference(conf->log); | |
1583 | ||
1584 | if (!log) | |
1585 | ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags); | |
1586 | else | |
1587 | ret = test_bit(Faulty, &log->rdev->flags); | |
1588 | rcu_read_unlock(); | |
1589 | return ret; | |
6e74a9cf SL |
1590 | } |
1591 | ||
effe6ee7 SL |
1592 | #define R5L_RECOVERY_PAGE_POOL_SIZE 256 |
1593 | ||
355810d1 SL |
1594 | struct r5l_recovery_ctx { |
1595 | struct page *meta_page; /* current meta */ | |
1596 | sector_t meta_total_blocks; /* total size of current meta and data */ | |
1597 | sector_t pos; /* recovery position */ | |
1598 | u64 seq; /* recovery position seq */ | |
b4c625c6 SL |
1599 | int data_parity_stripes; /* number of data_parity stripes */ |
1600 | int data_only_stripes; /* number of data_only stripes */ | |
1601 | struct list_head cached_list; | |
effe6ee7 SL |
1602 | |
1603 | /* | |
1604 | * read ahead page pool (ra_pool) | |
1605 | * in recovery, log is read sequentially. It is not efficient to | |
1606 | * read every page with sync_page_io(). The read ahead page pool | |
1607 | * reads multiple pages with one IO, so further log read can | |
1608 | * just copy data from the pool. | |
1609 | */ | |
1610 | struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE]; | |
1611 | sector_t pool_offset; /* offset of first page in the pool */ | |
1612 | int total_pages; /* total allocated pages */ | |
1613 | int valid_pages; /* pages with valid data */ | |
1614 | struct bio *ra_bio; /* bio to do the read ahead */ | |
355810d1 SL |
1615 | }; |
1616 | ||
effe6ee7 SL |
1617 | static int r5l_recovery_allocate_ra_pool(struct r5l_log *log, |
1618 | struct r5l_recovery_ctx *ctx) | |
1619 | { | |
1620 | struct page *page; | |
1621 | ||
1622 | ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs); | |
1623 | if (!ctx->ra_bio) | |
1624 | return -ENOMEM; | |
1625 | ||
1626 | ctx->valid_pages = 0; | |
1627 | ctx->total_pages = 0; | |
1628 | while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) { | |
1629 | page = alloc_page(GFP_KERNEL); | |
1630 | ||
1631 | if (!page) | |
1632 | break; | |
1633 | ctx->ra_pool[ctx->total_pages] = page; | |
1634 | ctx->total_pages += 1; | |
1635 | } | |
1636 | ||
1637 | if (ctx->total_pages == 0) { | |
1638 | bio_put(ctx->ra_bio); | |
1639 | return -ENOMEM; | |
1640 | } | |
1641 | ||
1642 | ctx->pool_offset = 0; | |
1643 | return 0; | |
1644 | } | |
1645 | ||
1646 | static void r5l_recovery_free_ra_pool(struct r5l_log *log, | |
1647 | struct r5l_recovery_ctx *ctx) | |
1648 | { | |
1649 | int i; | |
1650 | ||
1651 | for (i = 0; i < ctx->total_pages; ++i) | |
1652 | put_page(ctx->ra_pool[i]); | |
1653 | bio_put(ctx->ra_bio); | |
1654 | } | |
1655 | ||
1656 | /* | |
1657 | * fetch ctx->valid_pages pages from offset | |
1658 | * In normal cases, ctx->valid_pages == ctx->total_pages after the call. | |
1659 | * However, if the offset is close to the end of the journal device, | |
1660 | * ctx->valid_pages could be smaller than ctx->total_pages | |
1661 | */ | |
1662 | static int r5l_recovery_fetch_ra_pool(struct r5l_log *log, | |
1663 | struct r5l_recovery_ctx *ctx, | |
1664 | sector_t offset) | |
1665 | { | |
1666 | bio_reset(ctx->ra_bio); | |
1667 | ctx->ra_bio->bi_bdev = log->rdev->bdev; | |
1668 | bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0); | |
1669 | ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset; | |
1670 | ||
1671 | ctx->valid_pages = 0; | |
1672 | ctx->pool_offset = offset; | |
1673 | ||
1674 | while (ctx->valid_pages < ctx->total_pages) { | |
1675 | bio_add_page(ctx->ra_bio, | |
1676 | ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0); | |
1677 | ctx->valid_pages += 1; | |
1678 | ||
1679 | offset = r5l_ring_add(log, offset, BLOCK_SECTORS); | |
1680 | ||
1681 | if (offset == 0) /* reached end of the device */ | |
1682 | break; | |
1683 | } | |
1684 | ||
1685 | return submit_bio_wait(ctx->ra_bio); | |
1686 | } | |
1687 | ||
1688 | /* | |
1689 | * try read a page from the read ahead page pool, if the page is not in the | |
1690 | * pool, call r5l_recovery_fetch_ra_pool | |
1691 | */ | |
1692 | static int r5l_recovery_read_page(struct r5l_log *log, | |
1693 | struct r5l_recovery_ctx *ctx, | |
1694 | struct page *page, | |
1695 | sector_t offset) | |
1696 | { | |
1697 | int ret; | |
1698 | ||
1699 | if (offset < ctx->pool_offset || | |
1700 | offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) { | |
1701 | ret = r5l_recovery_fetch_ra_pool(log, ctx, offset); | |
1702 | if (ret) | |
1703 | return ret; | |
1704 | } | |
1705 | ||
1706 | BUG_ON(offset < ctx->pool_offset || | |
1707 | offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS); | |
1708 | ||
1709 | memcpy(page_address(page), | |
1710 | page_address(ctx->ra_pool[(offset - ctx->pool_offset) >> | |
1711 | BLOCK_SECTOR_SHIFT]), | |
1712 | PAGE_SIZE); | |
1713 | return 0; | |
1714 | } | |
1715 | ||
9ed988f5 SL |
1716 | static int r5l_recovery_read_meta_block(struct r5l_log *log, |
1717 | struct r5l_recovery_ctx *ctx) | |
355810d1 SL |
1718 | { |
1719 | struct page *page = ctx->meta_page; | |
1720 | struct r5l_meta_block *mb; | |
1721 | u32 crc, stored_crc; | |
effe6ee7 | 1722 | int ret; |
355810d1 | 1723 | |
effe6ee7 SL |
1724 | ret = r5l_recovery_read_page(log, ctx, page, ctx->pos); |
1725 | if (ret != 0) | |
1726 | return ret; | |
355810d1 SL |
1727 | |
1728 | mb = page_address(page); | |
1729 | stored_crc = le32_to_cpu(mb->checksum); | |
1730 | mb->checksum = 0; | |
1731 | ||
1732 | if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || | |
1733 | le64_to_cpu(mb->seq) != ctx->seq || | |
1734 | mb->version != R5LOG_VERSION || | |
1735 | le64_to_cpu(mb->position) != ctx->pos) | |
1736 | return -EINVAL; | |
1737 | ||
5cb2fbd6 | 1738 | crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); |
355810d1 SL |
1739 | if (stored_crc != crc) |
1740 | return -EINVAL; | |
1741 | ||
1742 | if (le32_to_cpu(mb->meta_size) > PAGE_SIZE) | |
1743 | return -EINVAL; | |
1744 | ||
1745 | ctx->meta_total_blocks = BLOCK_SECTORS; | |
1746 | ||
1747 | return 0; | |
1748 | } | |
1749 | ||
9ed988f5 SL |
1750 | static void |
1751 | r5l_recovery_create_empty_meta_block(struct r5l_log *log, | |
1752 | struct page *page, | |
1753 | sector_t pos, u64 seq) | |
355810d1 | 1754 | { |
355810d1 | 1755 | struct r5l_meta_block *mb; |
355810d1 | 1756 | |
355810d1 | 1757 | mb = page_address(page); |
9ed988f5 | 1758 | clear_page(mb); |
355810d1 SL |
1759 | mb->magic = cpu_to_le32(R5LOG_MAGIC); |
1760 | mb->version = R5LOG_VERSION; | |
1761 | mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block)); | |
1762 | mb->seq = cpu_to_le64(seq); | |
1763 | mb->position = cpu_to_le64(pos); | |
9ed988f5 | 1764 | } |
355810d1 | 1765 | |
9ed988f5 SL |
1766 | static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos, |
1767 | u64 seq) | |
1768 | { | |
1769 | struct page *page; | |
5c88f403 | 1770 | struct r5l_meta_block *mb; |
355810d1 | 1771 | |
9ed988f5 SL |
1772 | page = alloc_page(GFP_KERNEL); |
1773 | if (!page) | |
1774 | return -ENOMEM; | |
1775 | r5l_recovery_create_empty_meta_block(log, page, pos, seq); | |
5c88f403 SL |
1776 | mb = page_address(page); |
1777 | mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum, | |
1778 | mb, PAGE_SIZE)); | |
796a5cf0 | 1779 | if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE, |
20737738 | 1780 | REQ_FUA, false)) { |
355810d1 SL |
1781 | __free_page(page); |
1782 | return -EIO; | |
1783 | } | |
1784 | __free_page(page); | |
1785 | return 0; | |
1786 | } | |
355810d1 | 1787 | |
b4c625c6 SL |
1788 | /* |
1789 | * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite | |
1790 | * to mark valid (potentially not flushed) data in the journal. | |
1791 | * | |
1792 | * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb, | |
1793 | * so there should not be any mismatch here. | |
1794 | */ | |
1795 | static void r5l_recovery_load_data(struct r5l_log *log, | |
1796 | struct stripe_head *sh, | |
1797 | struct r5l_recovery_ctx *ctx, | |
1798 | struct r5l_payload_data_parity *payload, | |
1799 | sector_t log_offset) | |
1800 | { | |
1801 | struct mddev *mddev = log->rdev->mddev; | |
1802 | struct r5conf *conf = mddev->private; | |
1803 | int dd_idx; | |
1804 | ||
1805 | raid5_compute_sector(conf, | |
1806 | le64_to_cpu(payload->location), 0, | |
1807 | &dd_idx, sh); | |
effe6ee7 | 1808 | r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset); |
b4c625c6 SL |
1809 | sh->dev[dd_idx].log_checksum = |
1810 | le32_to_cpu(payload->checksum[0]); | |
1811 | ctx->meta_total_blocks += BLOCK_SECTORS; | |
1812 | ||
1813 | set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags); | |
1814 | set_bit(STRIPE_R5C_CACHING, &sh->state); | |
1815 | } | |
1816 | ||
1817 | static void r5l_recovery_load_parity(struct r5l_log *log, | |
1818 | struct stripe_head *sh, | |
1819 | struct r5l_recovery_ctx *ctx, | |
1820 | struct r5l_payload_data_parity *payload, | |
1821 | sector_t log_offset) | |
1822 | { | |
1823 | struct mddev *mddev = log->rdev->mddev; | |
1824 | struct r5conf *conf = mddev->private; | |
1825 | ||
1826 | ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded; | |
effe6ee7 | 1827 | r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset); |
b4c625c6 SL |
1828 | sh->dev[sh->pd_idx].log_checksum = |
1829 | le32_to_cpu(payload->checksum[0]); | |
1830 | set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags); | |
1831 | ||
1832 | if (sh->qd_idx >= 0) { | |
effe6ee7 SL |
1833 | r5l_recovery_read_page( |
1834 | log, ctx, sh->dev[sh->qd_idx].page, | |
1835 | r5l_ring_add(log, log_offset, BLOCK_SECTORS)); | |
b4c625c6 SL |
1836 | sh->dev[sh->qd_idx].log_checksum = |
1837 | le32_to_cpu(payload->checksum[1]); | |
1838 | set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags); | |
355810d1 | 1839 | } |
b4c625c6 SL |
1840 | clear_bit(STRIPE_R5C_CACHING, &sh->state); |
1841 | } | |
355810d1 | 1842 | |
b4c625c6 SL |
1843 | static void r5l_recovery_reset_stripe(struct stripe_head *sh) |
1844 | { | |
1845 | int i; | |
1846 | ||
1847 | sh->state = 0; | |
1848 | sh->log_start = MaxSector; | |
1849 | for (i = sh->disks; i--; ) | |
1850 | sh->dev[i].flags = 0; | |
1851 | } | |
1852 | ||
1853 | static void | |
1854 | r5l_recovery_replay_one_stripe(struct r5conf *conf, | |
1855 | struct stripe_head *sh, | |
1856 | struct r5l_recovery_ctx *ctx) | |
1857 | { | |
1858 | struct md_rdev *rdev, *rrdev; | |
1859 | int disk_index; | |
1860 | int data_count = 0; | |
355810d1 | 1861 | |
b4c625c6 | 1862 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { |
355810d1 SL |
1863 | if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) |
1864 | continue; | |
b4c625c6 SL |
1865 | if (disk_index == sh->qd_idx || disk_index == sh->pd_idx) |
1866 | continue; | |
1867 | data_count++; | |
355810d1 SL |
1868 | } |
1869 | ||
b4c625c6 SL |
1870 | /* |
1871 | * stripes that only have parity must have been flushed | |
1872 | * before the crash that we are now recovering from, so | |
1873 | * there is nothing more to recovery. | |
1874 | */ | |
1875 | if (data_count == 0) | |
1876 | goto out; | |
355810d1 | 1877 | |
b4c625c6 SL |
1878 | for (disk_index = 0; disk_index < sh->disks; disk_index++) { |
1879 | if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) | |
355810d1 SL |
1880 | continue; |
1881 | ||
1882 | /* in case device is broken */ | |
b4c625c6 | 1883 | rcu_read_lock(); |
355810d1 | 1884 | rdev = rcu_dereference(conf->disks[disk_index].rdev); |
b4c625c6 SL |
1885 | if (rdev) { |
1886 | atomic_inc(&rdev->nr_pending); | |
1887 | rcu_read_unlock(); | |
1888 | sync_page_io(rdev, sh->sector, PAGE_SIZE, | |
796a5cf0 MC |
1889 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, |
1890 | false); | |
b4c625c6 SL |
1891 | rdev_dec_pending(rdev, rdev->mddev); |
1892 | rcu_read_lock(); | |
1893 | } | |
355810d1 | 1894 | rrdev = rcu_dereference(conf->disks[disk_index].replacement); |
b4c625c6 SL |
1895 | if (rrdev) { |
1896 | atomic_inc(&rrdev->nr_pending); | |
1897 | rcu_read_unlock(); | |
1898 | sync_page_io(rrdev, sh->sector, PAGE_SIZE, | |
796a5cf0 MC |
1899 | sh->dev[disk_index].page, REQ_OP_WRITE, 0, |
1900 | false); | |
b4c625c6 SL |
1901 | rdev_dec_pending(rrdev, rrdev->mddev); |
1902 | rcu_read_lock(); | |
1903 | } | |
1904 | rcu_read_unlock(); | |
355810d1 | 1905 | } |
b4c625c6 SL |
1906 | ctx->data_parity_stripes++; |
1907 | out: | |
1908 | r5l_recovery_reset_stripe(sh); | |
1909 | } | |
1910 | ||
1911 | static struct stripe_head * | |
1912 | r5c_recovery_alloc_stripe(struct r5conf *conf, | |
3c66abba | 1913 | sector_t stripe_sect) |
b4c625c6 SL |
1914 | { |
1915 | struct stripe_head *sh; | |
1916 | ||
1917 | sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0); | |
1918 | if (!sh) | |
1919 | return NULL; /* no more stripe available */ | |
1920 | ||
1921 | r5l_recovery_reset_stripe(sh); | |
b4c625c6 SL |
1922 | |
1923 | return sh; | |
1924 | } | |
1925 | ||
1926 | static struct stripe_head * | |
1927 | r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect) | |
1928 | { | |
1929 | struct stripe_head *sh; | |
1930 | ||
1931 | list_for_each_entry(sh, list, lru) | |
1932 | if (sh->sector == sect) | |
1933 | return sh; | |
1934 | return NULL; | |
1935 | } | |
1936 | ||
1937 | static void | |
1938 | r5c_recovery_drop_stripes(struct list_head *cached_stripe_list, | |
1939 | struct r5l_recovery_ctx *ctx) | |
1940 | { | |
1941 | struct stripe_head *sh, *next; | |
1942 | ||
1943 | list_for_each_entry_safe(sh, next, cached_stripe_list, lru) { | |
1944 | r5l_recovery_reset_stripe(sh); | |
1945 | list_del_init(&sh->lru); | |
1946 | raid5_release_stripe(sh); | |
1947 | } | |
1948 | } | |
1949 | ||
1950 | static void | |
1951 | r5c_recovery_replay_stripes(struct list_head *cached_stripe_list, | |
1952 | struct r5l_recovery_ctx *ctx) | |
1953 | { | |
1954 | struct stripe_head *sh, *next; | |
1955 | ||
1956 | list_for_each_entry_safe(sh, next, cached_stripe_list, lru) | |
1957 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { | |
1958 | r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx); | |
1959 | list_del_init(&sh->lru); | |
1960 | raid5_release_stripe(sh); | |
1961 | } | |
1962 | } | |
1963 | ||
1964 | /* if matches return 0; otherwise return -EINVAL */ | |
1965 | static int | |
effe6ee7 SL |
1966 | r5l_recovery_verify_data_checksum(struct r5l_log *log, |
1967 | struct r5l_recovery_ctx *ctx, | |
1968 | struct page *page, | |
b4c625c6 SL |
1969 | sector_t log_offset, __le32 log_checksum) |
1970 | { | |
1971 | void *addr; | |
1972 | u32 checksum; | |
1973 | ||
effe6ee7 | 1974 | r5l_recovery_read_page(log, ctx, page, log_offset); |
b4c625c6 SL |
1975 | addr = kmap_atomic(page); |
1976 | checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE); | |
1977 | kunmap_atomic(addr); | |
1978 | return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL; | |
1979 | } | |
1980 | ||
1981 | /* | |
1982 | * before loading data to stripe cache, we need verify checksum for all data, | |
1983 | * if there is mismatch for any data page, we drop all data in the mata block | |
1984 | */ | |
1985 | static int | |
1986 | r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log, | |
1987 | struct r5l_recovery_ctx *ctx) | |
1988 | { | |
1989 | struct mddev *mddev = log->rdev->mddev; | |
1990 | struct r5conf *conf = mddev->private; | |
1991 | struct r5l_meta_block *mb = page_address(ctx->meta_page); | |
1992 | sector_t mb_offset = sizeof(struct r5l_meta_block); | |
1993 | sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); | |
1994 | struct page *page; | |
1995 | struct r5l_payload_data_parity *payload; | |
2d4f4687 | 1996 | struct r5l_payload_flush *payload_flush; |
b4c625c6 SL |
1997 | |
1998 | page = alloc_page(GFP_KERNEL); | |
1999 | if (!page) | |
2000 | return -ENOMEM; | |
2001 | ||
2002 | while (mb_offset < le32_to_cpu(mb->meta_size)) { | |
2003 | payload = (void *)mb + mb_offset; | |
2d4f4687 | 2004 | payload_flush = (void *)mb + mb_offset; |
b4c625c6 SL |
2005 | |
2006 | if (payload->header.type == R5LOG_PAYLOAD_DATA) { | |
2007 | if (r5l_recovery_verify_data_checksum( | |
effe6ee7 | 2008 | log, ctx, page, log_offset, |
b4c625c6 SL |
2009 | payload->checksum[0]) < 0) |
2010 | goto mismatch; | |
2011 | } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) { | |
2012 | if (r5l_recovery_verify_data_checksum( | |
effe6ee7 | 2013 | log, ctx, page, log_offset, |
b4c625c6 SL |
2014 | payload->checksum[0]) < 0) |
2015 | goto mismatch; | |
2016 | if (conf->max_degraded == 2 && /* q for RAID 6 */ | |
2017 | r5l_recovery_verify_data_checksum( | |
effe6ee7 | 2018 | log, ctx, page, |
b4c625c6 SL |
2019 | r5l_ring_add(log, log_offset, |
2020 | BLOCK_SECTORS), | |
2021 | payload->checksum[1]) < 0) | |
2022 | goto mismatch; | |
2d4f4687 SL |
2023 | } else if (payload->header.type == R5LOG_PAYLOAD_FLUSH) { |
2024 | /* nothing to do for R5LOG_PAYLOAD_FLUSH here */ | |
2025 | } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */ | |
b4c625c6 SL |
2026 | goto mismatch; |
2027 | ||
2d4f4687 SL |
2028 | if (payload->header.type == R5LOG_PAYLOAD_FLUSH) { |
2029 | mb_offset += sizeof(struct r5l_payload_flush) + | |
2030 | le32_to_cpu(payload_flush->size); | |
2031 | } else { | |
2032 | /* DATA or PARITY payload */ | |
2033 | log_offset = r5l_ring_add(log, log_offset, | |
2034 | le32_to_cpu(payload->size)); | |
2035 | mb_offset += sizeof(struct r5l_payload_data_parity) + | |
2036 | sizeof(__le32) * | |
2037 | (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); | |
2038 | } | |
b4c625c6 | 2039 | |
b4c625c6 SL |
2040 | } |
2041 | ||
2042 | put_page(page); | |
355810d1 SL |
2043 | return 0; |
2044 | ||
b4c625c6 SL |
2045 | mismatch: |
2046 | put_page(page); | |
355810d1 SL |
2047 | return -EINVAL; |
2048 | } | |
2049 | ||
b4c625c6 SL |
2050 | /* |
2051 | * Analyze all data/parity pages in one meta block | |
2052 | * Returns: | |
2053 | * 0 for success | |
2054 | * -EINVAL for unknown playload type | |
2055 | * -EAGAIN for checksum mismatch of data page | |
2056 | * -ENOMEM for run out of memory (alloc_page failed or run out of stripes) | |
2057 | */ | |
2058 | static int | |
2059 | r5c_recovery_analyze_meta_block(struct r5l_log *log, | |
2060 | struct r5l_recovery_ctx *ctx, | |
2061 | struct list_head *cached_stripe_list) | |
355810d1 | 2062 | { |
b4c625c6 SL |
2063 | struct mddev *mddev = log->rdev->mddev; |
2064 | struct r5conf *conf = mddev->private; | |
355810d1 | 2065 | struct r5l_meta_block *mb; |
b4c625c6 | 2066 | struct r5l_payload_data_parity *payload; |
2d4f4687 | 2067 | struct r5l_payload_flush *payload_flush; |
b4c625c6 | 2068 | int mb_offset; |
355810d1 | 2069 | sector_t log_offset; |
b4c625c6 SL |
2070 | sector_t stripe_sect; |
2071 | struct stripe_head *sh; | |
2072 | int ret; | |
2073 | ||
2074 | /* | |
2075 | * for mismatch in data blocks, we will drop all data in this mb, but | |
2076 | * we will still read next mb for other data with FLUSH flag, as | |
2077 | * io_unit could finish out of order. | |
2078 | */ | |
2079 | ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx); | |
2080 | if (ret == -EINVAL) | |
2081 | return -EAGAIN; | |
2082 | else if (ret) | |
2083 | return ret; /* -ENOMEM duo to alloc_page() failed */ | |
355810d1 SL |
2084 | |
2085 | mb = page_address(ctx->meta_page); | |
b4c625c6 | 2086 | mb_offset = sizeof(struct r5l_meta_block); |
355810d1 SL |
2087 | log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); |
2088 | ||
b4c625c6 | 2089 | while (mb_offset < le32_to_cpu(mb->meta_size)) { |
355810d1 SL |
2090 | int dd; |
2091 | ||
b4c625c6 | 2092 | payload = (void *)mb + mb_offset; |
2d4f4687 SL |
2093 | payload_flush = (void *)mb + mb_offset; |
2094 | ||
2095 | if (payload->header.type == R5LOG_PAYLOAD_FLUSH) { | |
2096 | int i, count; | |
2097 | ||
2098 | count = le32_to_cpu(payload_flush->size) / sizeof(__le64); | |
2099 | for (i = 0; i < count; ++i) { | |
2100 | stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]); | |
2101 | sh = r5c_recovery_lookup_stripe(cached_stripe_list, | |
2102 | stripe_sect); | |
2103 | if (sh) { | |
2104 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2105 | r5l_recovery_reset_stripe(sh); | |
2106 | list_del_init(&sh->lru); | |
2107 | raid5_release_stripe(sh); | |
2108 | } | |
2109 | } | |
2110 | ||
2111 | mb_offset += sizeof(struct r5l_payload_flush) + | |
2112 | le32_to_cpu(payload_flush->size); | |
2113 | continue; | |
2114 | } | |
2115 | ||
2116 | /* DATA or PARITY payload */ | |
b4c625c6 SL |
2117 | stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ? |
2118 | raid5_compute_sector( | |
2119 | conf, le64_to_cpu(payload->location), 0, &dd, | |
2120 | NULL) | |
2121 | : le64_to_cpu(payload->location); | |
2122 | ||
2123 | sh = r5c_recovery_lookup_stripe(cached_stripe_list, | |
2124 | stripe_sect); | |
2125 | ||
2126 | if (!sh) { | |
3c66abba | 2127 | sh = r5c_recovery_alloc_stripe(conf, stripe_sect); |
b4c625c6 SL |
2128 | /* |
2129 | * cannot get stripe from raid5_get_active_stripe | |
2130 | * try replay some stripes | |
2131 | */ | |
2132 | if (!sh) { | |
2133 | r5c_recovery_replay_stripes( | |
2134 | cached_stripe_list, ctx); | |
2135 | sh = r5c_recovery_alloc_stripe( | |
3c66abba | 2136 | conf, stripe_sect); |
b4c625c6 SL |
2137 | } |
2138 | if (!sh) { | |
2139 | pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n", | |
2140 | mdname(mddev), | |
2141 | conf->min_nr_stripes * 2); | |
2142 | raid5_set_cache_size(mddev, | |
2143 | conf->min_nr_stripes * 2); | |
3c66abba SL |
2144 | sh = r5c_recovery_alloc_stripe(conf, |
2145 | stripe_sect); | |
b4c625c6 SL |
2146 | } |
2147 | if (!sh) { | |
2148 | pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n", | |
2149 | mdname(mddev)); | |
2150 | return -ENOMEM; | |
2151 | } | |
2152 | list_add_tail(&sh->lru, cached_stripe_list); | |
2153 | } | |
2154 | ||
2155 | if (payload->header.type == R5LOG_PAYLOAD_DATA) { | |
f7b7bee7 ZL |
2156 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state) && |
2157 | test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) { | |
b4c625c6 | 2158 | r5l_recovery_replay_one_stripe(conf, sh, ctx); |
b4c625c6 SL |
2159 | list_move_tail(&sh->lru, cached_stripe_list); |
2160 | } | |
2161 | r5l_recovery_load_data(log, sh, ctx, payload, | |
2162 | log_offset); | |
2163 | } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) | |
2164 | r5l_recovery_load_parity(log, sh, ctx, payload, | |
2165 | log_offset); | |
2166 | else | |
355810d1 | 2167 | return -EINVAL; |
b4c625c6 SL |
2168 | |
2169 | log_offset = r5l_ring_add(log, log_offset, | |
2170 | le32_to_cpu(payload->size)); | |
2171 | ||
2172 | mb_offset += sizeof(struct r5l_payload_data_parity) + | |
2173 | sizeof(__le32) * | |
2174 | (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); | |
355810d1 | 2175 | } |
b4c625c6 | 2176 | |
355810d1 SL |
2177 | return 0; |
2178 | } | |
2179 | ||
b4c625c6 SL |
2180 | /* |
2181 | * Load the stripe into cache. The stripe will be written out later by | |
2182 | * the stripe cache state machine. | |
2183 | */ | |
2184 | static void r5c_recovery_load_one_stripe(struct r5l_log *log, | |
2185 | struct stripe_head *sh) | |
355810d1 | 2186 | { |
b4c625c6 SL |
2187 | struct r5dev *dev; |
2188 | int i; | |
2189 | ||
2190 | for (i = sh->disks; i--; ) { | |
2191 | dev = sh->dev + i; | |
2192 | if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) { | |
2193 | set_bit(R5_InJournal, &dev->flags); | |
2194 | set_bit(R5_UPTODATE, &dev->flags); | |
2195 | } | |
2196 | } | |
b4c625c6 SL |
2197 | } |
2198 | ||
2199 | /* | |
2200 | * Scan through the log for all to-be-flushed data | |
2201 | * | |
2202 | * For stripes with data and parity, namely Data-Parity stripe | |
2203 | * (STRIPE_R5C_CACHING == 0), we simply replay all the writes. | |
2204 | * | |
2205 | * For stripes with only data, namely Data-Only stripe | |
2206 | * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine. | |
2207 | * | |
2208 | * For a stripe, if we see data after parity, we should discard all previous | |
2209 | * data and parity for this stripe, as these data are already flushed to | |
2210 | * the array. | |
2211 | * | |
2212 | * At the end of the scan, we return the new journal_tail, which points to | |
2213 | * first data-only stripe on the journal device, or next invalid meta block. | |
2214 | */ | |
2215 | static int r5c_recovery_flush_log(struct r5l_log *log, | |
2216 | struct r5l_recovery_ctx *ctx) | |
2217 | { | |
bc8f167f | 2218 | struct stripe_head *sh; |
b4c625c6 SL |
2219 | int ret = 0; |
2220 | ||
2221 | /* scan through the log */ | |
355810d1 | 2222 | while (1) { |
b4c625c6 SL |
2223 | if (r5l_recovery_read_meta_block(log, ctx)) |
2224 | break; | |
2225 | ||
2226 | ret = r5c_recovery_analyze_meta_block(log, ctx, | |
2227 | &ctx->cached_list); | |
2228 | /* | |
2229 | * -EAGAIN means mismatch in data block, in this case, we still | |
2230 | * try scan the next metablock | |
2231 | */ | |
2232 | if (ret && ret != -EAGAIN) | |
2233 | break; /* ret == -EINVAL or -ENOMEM */ | |
355810d1 SL |
2234 | ctx->seq++; |
2235 | ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks); | |
2236 | } | |
b4c625c6 SL |
2237 | |
2238 | if (ret == -ENOMEM) { | |
2239 | r5c_recovery_drop_stripes(&ctx->cached_list, ctx); | |
2240 | return ret; | |
2241 | } | |
2242 | ||
2243 | /* replay data-parity stripes */ | |
2244 | r5c_recovery_replay_stripes(&ctx->cached_list, ctx); | |
2245 | ||
2246 | /* load data-only stripes to stripe cache */ | |
bc8f167f | 2247 | list_for_each_entry(sh, &ctx->cached_list, lru) { |
b4c625c6 SL |
2248 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); |
2249 | r5c_recovery_load_one_stripe(log, sh); | |
b4c625c6 SL |
2250 | ctx->data_only_stripes++; |
2251 | } | |
2252 | ||
2253 | return 0; | |
355810d1 SL |
2254 | } |
2255 | ||
b4c625c6 SL |
2256 | /* |
2257 | * we did a recovery. Now ctx.pos points to an invalid meta block. New | |
2258 | * log will start here. but we can't let superblock point to last valid | |
2259 | * meta block. The log might looks like: | |
2260 | * | meta 1| meta 2| meta 3| | |
2261 | * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If | |
2262 | * superblock points to meta 1, we write a new valid meta 2n. if crash | |
2263 | * happens again, new recovery will start from meta 1. Since meta 2n is | |
2264 | * valid now, recovery will think meta 3 is valid, which is wrong. | |
2265 | * The solution is we create a new meta in meta2 with its seq == meta | |
3c6edc66 SL |
2266 | * 1's seq + 10000 and let superblock points to meta2. The same recovery |
2267 | * will not think meta 3 is a valid meta, because its seq doesn't match | |
b4c625c6 SL |
2268 | */ |
2269 | ||
2270 | /* | |
2271 | * Before recovery, the log looks like the following | |
2272 | * | |
2273 | * --------------------------------------------- | |
2274 | * | valid log | invalid log | | |
2275 | * --------------------------------------------- | |
2276 | * ^ | |
2277 | * |- log->last_checkpoint | |
2278 | * |- log->last_cp_seq | |
2279 | * | |
2280 | * Now we scan through the log until we see invalid entry | |
2281 | * | |
2282 | * --------------------------------------------- | |
2283 | * | valid log | invalid log | | |
2284 | * --------------------------------------------- | |
2285 | * ^ ^ | |
2286 | * |- log->last_checkpoint |- ctx->pos | |
2287 | * |- log->last_cp_seq |- ctx->seq | |
2288 | * | |
2289 | * From this point, we need to increase seq number by 10 to avoid | |
2290 | * confusing next recovery. | |
2291 | * | |
2292 | * --------------------------------------------- | |
2293 | * | valid log | invalid log | | |
2294 | * --------------------------------------------- | |
2295 | * ^ ^ | |
2296 | * |- log->last_checkpoint |- ctx->pos+1 | |
3c6edc66 | 2297 | * |- log->last_cp_seq |- ctx->seq+10001 |
b4c625c6 SL |
2298 | * |
2299 | * However, it is not safe to start the state machine yet, because data only | |
2300 | * parities are not yet secured in RAID. To save these data only parities, we | |
2301 | * rewrite them from seq+11. | |
2302 | * | |
2303 | * ----------------------------------------------------------------- | |
2304 | * | valid log | data only stripes | invalid log | | |
2305 | * ----------------------------------------------------------------- | |
2306 | * ^ ^ | |
2307 | * |- log->last_checkpoint |- ctx->pos+n | |
3c6edc66 | 2308 | * |- log->last_cp_seq |- ctx->seq+10000+n |
b4c625c6 SL |
2309 | * |
2310 | * If failure happens again during this process, the recovery can safe start | |
2311 | * again from log->last_checkpoint. | |
2312 | * | |
2313 | * Once data only stripes are rewritten to journal, we move log_tail | |
2314 | * | |
2315 | * ----------------------------------------------------------------- | |
2316 | * | old log | data only stripes | invalid log | | |
2317 | * ----------------------------------------------------------------- | |
2318 | * ^ ^ | |
2319 | * |- log->last_checkpoint |- ctx->pos+n | |
3c6edc66 | 2320 | * |- log->last_cp_seq |- ctx->seq+10000+n |
b4c625c6 SL |
2321 | * |
2322 | * Then we can safely start the state machine. If failure happens from this | |
2323 | * point on, the recovery will start from new log->last_checkpoint. | |
2324 | */ | |
2325 | static int | |
2326 | r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log, | |
2327 | struct r5l_recovery_ctx *ctx) | |
355810d1 | 2328 | { |
a85dd7b8 | 2329 | struct stripe_head *sh; |
b4c625c6 | 2330 | struct mddev *mddev = log->rdev->mddev; |
355810d1 | 2331 | struct page *page; |
3c66abba | 2332 | sector_t next_checkpoint = MaxSector; |
355810d1 | 2333 | |
b4c625c6 SL |
2334 | page = alloc_page(GFP_KERNEL); |
2335 | if (!page) { | |
2336 | pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n", | |
2337 | mdname(mddev)); | |
355810d1 | 2338 | return -ENOMEM; |
b4c625c6 | 2339 | } |
355810d1 | 2340 | |
3c66abba SL |
2341 | WARN_ON(list_empty(&ctx->cached_list)); |
2342 | ||
a85dd7b8 | 2343 | list_for_each_entry(sh, &ctx->cached_list, lru) { |
b4c625c6 SL |
2344 | struct r5l_meta_block *mb; |
2345 | int i; | |
2346 | int offset; | |
2347 | sector_t write_pos; | |
2348 | ||
2349 | WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2350 | r5l_recovery_create_empty_meta_block(log, page, | |
2351 | ctx->pos, ctx->seq); | |
2352 | mb = page_address(page); | |
2353 | offset = le32_to_cpu(mb->meta_size); | |
fc833c2a | 2354 | write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); |
b4c625c6 SL |
2355 | |
2356 | for (i = sh->disks; i--; ) { | |
2357 | struct r5dev *dev = &sh->dev[i]; | |
2358 | struct r5l_payload_data_parity *payload; | |
2359 | void *addr; | |
2360 | ||
2361 | if (test_bit(R5_InJournal, &dev->flags)) { | |
2362 | payload = (void *)mb + offset; | |
2363 | payload->header.type = cpu_to_le16( | |
2364 | R5LOG_PAYLOAD_DATA); | |
2365 | payload->size = BLOCK_SECTORS; | |
2366 | payload->location = cpu_to_le64( | |
2367 | raid5_compute_blocknr(sh, i, 0)); | |
2368 | addr = kmap_atomic(dev->page); | |
2369 | payload->checksum[0] = cpu_to_le32( | |
2370 | crc32c_le(log->uuid_checksum, addr, | |
2371 | PAGE_SIZE)); | |
2372 | kunmap_atomic(addr); | |
2373 | sync_page_io(log->rdev, write_pos, PAGE_SIZE, | |
2374 | dev->page, REQ_OP_WRITE, 0, false); | |
2375 | write_pos = r5l_ring_add(log, write_pos, | |
2376 | BLOCK_SECTORS); | |
2377 | offset += sizeof(__le32) + | |
2378 | sizeof(struct r5l_payload_data_parity); | |
2379 | ||
2380 | } | |
2381 | } | |
2382 | mb->meta_size = cpu_to_le32(offset); | |
5c88f403 SL |
2383 | mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum, |
2384 | mb, PAGE_SIZE)); | |
b4c625c6 | 2385 | sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, |
20737738 | 2386 | REQ_OP_WRITE, REQ_FUA, false); |
b4c625c6 | 2387 | sh->log_start = ctx->pos; |
3c66abba SL |
2388 | list_add_tail(&sh->r5c, &log->stripe_in_journal_list); |
2389 | atomic_inc(&log->stripe_in_journal_count); | |
b4c625c6 SL |
2390 | ctx->pos = write_pos; |
2391 | ctx->seq += 1; | |
3c66abba | 2392 | next_checkpoint = sh->log_start; |
355810d1 | 2393 | } |
3c66abba | 2394 | log->next_checkpoint = next_checkpoint; |
355810d1 SL |
2395 | __free_page(page); |
2396 | return 0; | |
2397 | } | |
2398 | ||
a85dd7b8 SL |
2399 | static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log, |
2400 | struct r5l_recovery_ctx *ctx) | |
2401 | { | |
2402 | struct mddev *mddev = log->rdev->mddev; | |
2403 | struct r5conf *conf = mddev->private; | |
2404 | struct stripe_head *sh, *next; | |
2405 | ||
2406 | if (ctx->data_only_stripes == 0) | |
2407 | return; | |
2408 | ||
2409 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK; | |
2410 | ||
2411 | list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) { | |
2412 | r5c_make_stripe_write_out(sh); | |
2413 | set_bit(STRIPE_HANDLE, &sh->state); | |
2414 | list_del_init(&sh->lru); | |
2415 | raid5_release_stripe(sh); | |
2416 | } | |
2417 | ||
2418 | md_wakeup_thread(conf->mddev->thread); | |
2419 | /* reuse conf->wait_for_quiescent in recovery */ | |
2420 | wait_event(conf->wait_for_quiescent, | |
2421 | atomic_read(&conf->active_stripes) == 0); | |
2422 | ||
2423 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; | |
2424 | } | |
2425 | ||
f6bed0ef SL |
2426 | static int r5l_recovery_log(struct r5l_log *log) |
2427 | { | |
5aabf7c4 | 2428 | struct mddev *mddev = log->rdev->mddev; |
effe6ee7 | 2429 | struct r5l_recovery_ctx *ctx; |
5aabf7c4 | 2430 | int ret; |
43b96748 | 2431 | sector_t pos; |
355810d1 | 2432 | |
effe6ee7 SL |
2433 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); |
2434 | if (!ctx) | |
355810d1 SL |
2435 | return -ENOMEM; |
2436 | ||
effe6ee7 SL |
2437 | ctx->pos = log->last_checkpoint; |
2438 | ctx->seq = log->last_cp_seq; | |
2439 | INIT_LIST_HEAD(&ctx->cached_list); | |
2440 | ctx->meta_page = alloc_page(GFP_KERNEL); | |
355810d1 | 2441 | |
effe6ee7 SL |
2442 | if (!ctx->meta_page) { |
2443 | ret = -ENOMEM; | |
2444 | goto meta_page; | |
2445 | } | |
b4c625c6 | 2446 | |
effe6ee7 SL |
2447 | if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) { |
2448 | ret = -ENOMEM; | |
2449 | goto ra_pool; | |
2450 | } | |
2451 | ||
2452 | ret = r5c_recovery_flush_log(log, ctx); | |
2453 | ||
2454 | if (ret) | |
2455 | goto error; | |
43b96748 | 2456 | |
effe6ee7 SL |
2457 | pos = ctx->pos; |
2458 | ctx->seq += 10000; | |
43b96748 | 2459 | |
effe6ee7 | 2460 | if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0)) |
5aabf7c4 SL |
2461 | pr_debug("md/raid:%s: starting from clean shutdown\n", |
2462 | mdname(mddev)); | |
a85dd7b8 | 2463 | else |
99f17890 | 2464 | pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n", |
effe6ee7 SL |
2465 | mdname(mddev), ctx->data_only_stripes, |
2466 | ctx->data_parity_stripes); | |
2467 | ||
2468 | if (ctx->data_only_stripes == 0) { | |
2469 | log->next_checkpoint = ctx->pos; | |
2470 | r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++); | |
2471 | ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); | |
2472 | } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) { | |
a85dd7b8 SL |
2473 | pr_err("md/raid:%s: failed to rewrite stripes to journal\n", |
2474 | mdname(mddev)); | |
effe6ee7 SL |
2475 | ret = -EIO; |
2476 | goto error; | |
b4c625c6 SL |
2477 | } |
2478 | ||
effe6ee7 SL |
2479 | log->log_start = ctx->pos; |
2480 | log->seq = ctx->seq; | |
43b96748 J |
2481 | log->last_checkpoint = pos; |
2482 | r5l_write_super(log, pos); | |
a85dd7b8 | 2483 | |
effe6ee7 SL |
2484 | r5c_recovery_flush_data_only_stripes(log, ctx); |
2485 | ret = 0; | |
2486 | error: | |
2487 | r5l_recovery_free_ra_pool(log, ctx); | |
2488 | ra_pool: | |
2489 | __free_page(ctx->meta_page); | |
2490 | meta_page: | |
2491 | kfree(ctx); | |
2492 | return ret; | |
f6bed0ef SL |
2493 | } |
2494 | ||
2495 | static void r5l_write_super(struct r5l_log *log, sector_t cp) | |
2496 | { | |
2497 | struct mddev *mddev = log->rdev->mddev; | |
2498 | ||
2499 | log->rdev->journal_tail = cp; | |
2953079c | 2500 | set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags); |
f6bed0ef SL |
2501 | } |
2502 | ||
2c7da14b SL |
2503 | static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page) |
2504 | { | |
2505 | struct r5conf *conf = mddev->private; | |
2506 | int ret; | |
2507 | ||
2508 | if (!conf->log) | |
2509 | return 0; | |
2510 | ||
2511 | switch (conf->log->r5c_journal_mode) { | |
2512 | case R5C_JOURNAL_MODE_WRITE_THROUGH: | |
2513 | ret = snprintf( | |
2514 | page, PAGE_SIZE, "[%s] %s\n", | |
2515 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], | |
2516 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); | |
2517 | break; | |
2518 | case R5C_JOURNAL_MODE_WRITE_BACK: | |
2519 | ret = snprintf( | |
2520 | page, PAGE_SIZE, "%s [%s]\n", | |
2521 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], | |
2522 | r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); | |
2523 | break; | |
2524 | default: | |
2525 | ret = 0; | |
2526 | } | |
2527 | return ret; | |
2528 | } | |
2529 | ||
2530 | static ssize_t r5c_journal_mode_store(struct mddev *mddev, | |
2531 | const char *page, size_t length) | |
2532 | { | |
2533 | struct r5conf *conf = mddev->private; | |
2534 | struct r5l_log *log = conf->log; | |
2535 | int val = -1, i; | |
2536 | int len = length; | |
2537 | ||
2538 | if (!log) | |
2539 | return -ENODEV; | |
2540 | ||
2541 | if (len && page[len - 1] == '\n') | |
2542 | len -= 1; | |
2543 | for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++) | |
2544 | if (strlen(r5c_journal_mode_str[i]) == len && | |
2545 | strncmp(page, r5c_journal_mode_str[i], len) == 0) { | |
2546 | val = i; | |
2547 | break; | |
2548 | } | |
2549 | if (val < R5C_JOURNAL_MODE_WRITE_THROUGH || | |
2550 | val > R5C_JOURNAL_MODE_WRITE_BACK) | |
2551 | return -EINVAL; | |
2552 | ||
2e38a37f SL |
2553 | if (raid5_calc_degraded(conf) > 0 && |
2554 | val == R5C_JOURNAL_MODE_WRITE_BACK) | |
2555 | return -EINVAL; | |
2556 | ||
2c7da14b SL |
2557 | mddev_suspend(mddev); |
2558 | conf->log->r5c_journal_mode = val; | |
2559 | mddev_resume(mddev); | |
2560 | ||
2561 | pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n", | |
2562 | mdname(mddev), val, r5c_journal_mode_str[val]); | |
2563 | return length; | |
2564 | } | |
2565 | ||
2566 | struct md_sysfs_entry | |
2567 | r5c_journal_mode = __ATTR(journal_mode, 0644, | |
2568 | r5c_journal_mode_show, r5c_journal_mode_store); | |
2569 | ||
2ded3703 SL |
2570 | /* |
2571 | * Try handle write operation in caching phase. This function should only | |
2572 | * be called in write-back mode. | |
2573 | * | |
2574 | * If all outstanding writes can be handled in caching phase, returns 0 | |
2575 | * If writes requires write-out phase, call r5c_make_stripe_write_out() | |
2576 | * and returns -EAGAIN | |
2577 | */ | |
2578 | int r5c_try_caching_write(struct r5conf *conf, | |
2579 | struct stripe_head *sh, | |
2580 | struct stripe_head_state *s, | |
2581 | int disks) | |
2582 | { | |
2583 | struct r5l_log *log = conf->log; | |
1e6d690b SL |
2584 | int i; |
2585 | struct r5dev *dev; | |
2586 | int to_cache = 0; | |
03b047f4 SL |
2587 | void **pslot; |
2588 | sector_t tree_index; | |
2589 | int ret; | |
2590 | uintptr_t refcount; | |
2ded3703 SL |
2591 | |
2592 | BUG_ON(!r5c_is_writeback(log)); | |
2593 | ||
1e6d690b SL |
2594 | if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { |
2595 | /* | |
2596 | * There are two different scenarios here: | |
2597 | * 1. The stripe has some data cached, and it is sent to | |
2598 | * write-out phase for reclaim | |
2599 | * 2. The stripe is clean, and this is the first write | |
2600 | * | |
2601 | * For 1, return -EAGAIN, so we continue with | |
2602 | * handle_stripe_dirtying(). | |
2603 | * | |
2604 | * For 2, set STRIPE_R5C_CACHING and continue with caching | |
2605 | * write. | |
2606 | */ | |
2607 | ||
2608 | /* case 1: anything injournal or anything in written */ | |
2609 | if (s->injournal > 0 || s->written > 0) | |
2610 | return -EAGAIN; | |
2611 | /* case 2 */ | |
2612 | set_bit(STRIPE_R5C_CACHING, &sh->state); | |
2613 | } | |
2614 | ||
2e38a37f SL |
2615 | /* |
2616 | * When run in degraded mode, array is set to write-through mode. | |
2617 | * This check helps drain pending write safely in the transition to | |
2618 | * write-through mode. | |
2619 | */ | |
2620 | if (s->failed) { | |
2621 | r5c_make_stripe_write_out(sh); | |
2622 | return -EAGAIN; | |
2623 | } | |
2624 | ||
1e6d690b SL |
2625 | for (i = disks; i--; ) { |
2626 | dev = &sh->dev[i]; | |
2627 | /* if non-overwrite, use writing-out phase */ | |
2628 | if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) && | |
2629 | !test_bit(R5_InJournal, &dev->flags)) { | |
2630 | r5c_make_stripe_write_out(sh); | |
2631 | return -EAGAIN; | |
2632 | } | |
2633 | } | |
2634 | ||
03b047f4 SL |
2635 | /* if the stripe is not counted in big_stripe_tree, add it now */ |
2636 | if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) && | |
2637 | !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { | |
2638 | tree_index = r5c_tree_index(conf, sh->sector); | |
2639 | spin_lock(&log->tree_lock); | |
2640 | pslot = radix_tree_lookup_slot(&log->big_stripe_tree, | |
2641 | tree_index); | |
2642 | if (pslot) { | |
2643 | refcount = (uintptr_t)radix_tree_deref_slot_protected( | |
2644 | pslot, &log->tree_lock) >> | |
2645 | R5C_RADIX_COUNT_SHIFT; | |
2646 | radix_tree_replace_slot( | |
2647 | &log->big_stripe_tree, pslot, | |
2648 | (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT)); | |
2649 | } else { | |
2650 | /* | |
2651 | * this radix_tree_insert can fail safely, so no | |
2652 | * need to call radix_tree_preload() | |
2653 | */ | |
2654 | ret = radix_tree_insert( | |
2655 | &log->big_stripe_tree, tree_index, | |
2656 | (void *)(1 << R5C_RADIX_COUNT_SHIFT)); | |
2657 | if (ret) { | |
2658 | spin_unlock(&log->tree_lock); | |
2659 | r5c_make_stripe_write_out(sh); | |
2660 | return -EAGAIN; | |
2661 | } | |
2662 | } | |
2663 | spin_unlock(&log->tree_lock); | |
2664 | ||
2665 | /* | |
2666 | * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is | |
2667 | * counted in the radix tree | |
2668 | */ | |
2669 | set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state); | |
2670 | atomic_inc(&conf->r5c_cached_partial_stripes); | |
2671 | } | |
2672 | ||
1e6d690b SL |
2673 | for (i = disks; i--; ) { |
2674 | dev = &sh->dev[i]; | |
2675 | if (dev->towrite) { | |
2676 | set_bit(R5_Wantwrite, &dev->flags); | |
2677 | set_bit(R5_Wantdrain, &dev->flags); | |
2678 | set_bit(R5_LOCKED, &dev->flags); | |
2679 | to_cache++; | |
2680 | } | |
2681 | } | |
2682 | ||
2683 | if (to_cache) { | |
2684 | set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); | |
2685 | /* | |
2686 | * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data() | |
2687 | * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in | |
2688 | * r5c_handle_data_cached() | |
2689 | */ | |
2690 | set_bit(STRIPE_LOG_TRAPPED, &sh->state); | |
2691 | } | |
2692 | ||
2693 | return 0; | |
2694 | } | |
2695 | ||
2696 | /* | |
2697 | * free extra pages (orig_page) we allocated for prexor | |
2698 | */ | |
2699 | void r5c_release_extra_page(struct stripe_head *sh) | |
2700 | { | |
d7bd398e | 2701 | struct r5conf *conf = sh->raid_conf; |
1e6d690b | 2702 | int i; |
d7bd398e SL |
2703 | bool using_disk_info_extra_page; |
2704 | ||
2705 | using_disk_info_extra_page = | |
2706 | sh->dev[0].orig_page == conf->disks[0].extra_page; | |
1e6d690b SL |
2707 | |
2708 | for (i = sh->disks; i--; ) | |
2709 | if (sh->dev[i].page != sh->dev[i].orig_page) { | |
2710 | struct page *p = sh->dev[i].orig_page; | |
2711 | ||
2712 | sh->dev[i].orig_page = sh->dev[i].page; | |
86aa1397 SL |
2713 | clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags); |
2714 | ||
d7bd398e SL |
2715 | if (!using_disk_info_extra_page) |
2716 | put_page(p); | |
1e6d690b | 2717 | } |
d7bd398e SL |
2718 | |
2719 | if (using_disk_info_extra_page) { | |
2720 | clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state); | |
2721 | md_wakeup_thread(conf->mddev->thread); | |
2722 | } | |
2723 | } | |
2724 | ||
2725 | void r5c_use_extra_page(struct stripe_head *sh) | |
2726 | { | |
2727 | struct r5conf *conf = sh->raid_conf; | |
2728 | int i; | |
2729 | struct r5dev *dev; | |
2730 | ||
2731 | for (i = sh->disks; i--; ) { | |
2732 | dev = &sh->dev[i]; | |
2733 | if (dev->orig_page != dev->page) | |
2734 | put_page(dev->orig_page); | |
2735 | dev->orig_page = conf->disks[i].extra_page; | |
2736 | } | |
2ded3703 SL |
2737 | } |
2738 | ||
2739 | /* | |
2740 | * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the | |
2741 | * stripe is committed to RAID disks. | |
2742 | */ | |
2743 | void r5c_finish_stripe_write_out(struct r5conf *conf, | |
2744 | struct stripe_head *sh, | |
2745 | struct stripe_head_state *s) | |
2746 | { | |
03b047f4 | 2747 | struct r5l_log *log = conf->log; |
1e6d690b SL |
2748 | int i; |
2749 | int do_wakeup = 0; | |
03b047f4 SL |
2750 | sector_t tree_index; |
2751 | void **pslot; | |
2752 | uintptr_t refcount; | |
1e6d690b | 2753 | |
03b047f4 | 2754 | if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags)) |
2ded3703 SL |
2755 | return; |
2756 | ||
2757 | WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); | |
2758 | clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); | |
2759 | ||
03b047f4 | 2760 | if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) |
2ded3703 | 2761 | return; |
1e6d690b SL |
2762 | |
2763 | for (i = sh->disks; i--; ) { | |
2764 | clear_bit(R5_InJournal, &sh->dev[i].flags); | |
2765 | if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) | |
2766 | do_wakeup = 1; | |
2767 | } | |
2768 | ||
2769 | /* | |
2770 | * analyse_stripe() runs before r5c_finish_stripe_write_out(), | |
2771 | * We updated R5_InJournal, so we also update s->injournal. | |
2772 | */ | |
2773 | s->injournal = 0; | |
2774 | ||
2775 | if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) | |
2776 | if (atomic_dec_and_test(&conf->pending_full_writes)) | |
2777 | md_wakeup_thread(conf->mddev->thread); | |
2778 | ||
2779 | if (do_wakeup) | |
2780 | wake_up(&conf->wait_for_overlap); | |
a39f7afd | 2781 | |
03b047f4 | 2782 | spin_lock_irq(&log->stripe_in_journal_lock); |
a39f7afd | 2783 | list_del_init(&sh->r5c); |
03b047f4 | 2784 | spin_unlock_irq(&log->stripe_in_journal_lock); |
a39f7afd | 2785 | sh->log_start = MaxSector; |
03b047f4 SL |
2786 | |
2787 | atomic_dec(&log->stripe_in_journal_count); | |
2788 | r5c_update_log_state(log); | |
2789 | ||
2790 | /* stop counting this stripe in big_stripe_tree */ | |
2791 | if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) || | |
2792 | test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { | |
2793 | tree_index = r5c_tree_index(conf, sh->sector); | |
2794 | spin_lock(&log->tree_lock); | |
2795 | pslot = radix_tree_lookup_slot(&log->big_stripe_tree, | |
2796 | tree_index); | |
2797 | BUG_ON(pslot == NULL); | |
2798 | refcount = (uintptr_t)radix_tree_deref_slot_protected( | |
2799 | pslot, &log->tree_lock) >> | |
2800 | R5C_RADIX_COUNT_SHIFT; | |
2801 | if (refcount == 1) | |
2802 | radix_tree_delete(&log->big_stripe_tree, tree_index); | |
2803 | else | |
2804 | radix_tree_replace_slot( | |
2805 | &log->big_stripe_tree, pslot, | |
2806 | (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT)); | |
2807 | spin_unlock(&log->tree_lock); | |
2808 | } | |
2809 | ||
2810 | if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) { | |
2811 | BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0); | |
e33fbb9c | 2812 | atomic_dec(&conf->r5c_flushing_partial_stripes); |
03b047f4 SL |
2813 | atomic_dec(&conf->r5c_cached_partial_stripes); |
2814 | } | |
2815 | ||
2816 | if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { | |
2817 | BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0); | |
e33fbb9c | 2818 | atomic_dec(&conf->r5c_flushing_full_stripes); |
03b047f4 SL |
2819 | atomic_dec(&conf->r5c_cached_full_stripes); |
2820 | } | |
ea17481f SL |
2821 | |
2822 | r5l_append_flush_payload(log, sh->sector); | |
1e6d690b SL |
2823 | } |
2824 | ||
ff875738 | 2825 | int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh) |
1e6d690b | 2826 | { |
a39f7afd | 2827 | struct r5conf *conf = sh->raid_conf; |
1e6d690b SL |
2828 | int pages = 0; |
2829 | int reserve; | |
2830 | int i; | |
2831 | int ret = 0; | |
2832 | ||
2833 | BUG_ON(!log); | |
2834 | ||
2835 | for (i = 0; i < sh->disks; i++) { | |
2836 | void *addr; | |
2837 | ||
2838 | if (!test_bit(R5_Wantwrite, &sh->dev[i].flags)) | |
2839 | continue; | |
2840 | addr = kmap_atomic(sh->dev[i].page); | |
2841 | sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, | |
2842 | addr, PAGE_SIZE); | |
2843 | kunmap_atomic(addr); | |
2844 | pages++; | |
2845 | } | |
2846 | WARN_ON(pages == 0); | |
2847 | ||
2848 | /* | |
2849 | * The stripe must enter state machine again to call endio, so | |
2850 | * don't delay. | |
2851 | */ | |
2852 | clear_bit(STRIPE_DELAYED, &sh->state); | |
2853 | atomic_inc(&sh->count); | |
2854 | ||
2855 | mutex_lock(&log->io_mutex); | |
2856 | /* meta + data */ | |
2857 | reserve = (1 + pages) << (PAGE_SHIFT - 9); | |
1e6d690b | 2858 | |
a39f7afd SL |
2859 | if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && |
2860 | sh->log_start == MaxSector) | |
2861 | r5l_add_no_space_stripe(log, sh); | |
2862 | else if (!r5l_has_free_space(log, reserve)) { | |
2863 | if (sh->log_start == log->last_checkpoint) | |
2864 | BUG(); | |
2865 | else | |
2866 | r5l_add_no_space_stripe(log, sh); | |
1e6d690b SL |
2867 | } else { |
2868 | ret = r5l_log_stripe(log, sh, pages, 0); | |
2869 | if (ret) { | |
2870 | spin_lock_irq(&log->io_list_lock); | |
2871 | list_add_tail(&sh->log_list, &log->no_mem_stripes); | |
2872 | spin_unlock_irq(&log->io_list_lock); | |
2873 | } | |
2874 | } | |
2875 | ||
2876 | mutex_unlock(&log->io_mutex); | |
2877 | return 0; | |
f6bed0ef SL |
2878 | } |
2879 | ||
03b047f4 SL |
2880 | /* check whether this big stripe is in write back cache. */ |
2881 | bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect) | |
2882 | { | |
2883 | struct r5l_log *log = conf->log; | |
2884 | sector_t tree_index; | |
2885 | void *slot; | |
2886 | ||
2887 | if (!log) | |
2888 | return false; | |
2889 | ||
2890 | WARN_ON_ONCE(!rcu_read_lock_held()); | |
2891 | tree_index = r5c_tree_index(conf, sect); | |
2892 | slot = radix_tree_lookup(&log->big_stripe_tree, tree_index); | |
2893 | return slot != NULL; | |
2894 | } | |
2895 | ||
f6bed0ef SL |
2896 | static int r5l_load_log(struct r5l_log *log) |
2897 | { | |
2898 | struct md_rdev *rdev = log->rdev; | |
2899 | struct page *page; | |
2900 | struct r5l_meta_block *mb; | |
2901 | sector_t cp = log->rdev->journal_tail; | |
2902 | u32 stored_crc, expected_crc; | |
2903 | bool create_super = false; | |
d30dfeb9 | 2904 | int ret = 0; |
f6bed0ef SL |
2905 | |
2906 | /* Make sure it's valid */ | |
2907 | if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp) | |
2908 | cp = 0; | |
2909 | page = alloc_page(GFP_KERNEL); | |
2910 | if (!page) | |
2911 | return -ENOMEM; | |
2912 | ||
796a5cf0 | 2913 | if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) { |
f6bed0ef SL |
2914 | ret = -EIO; |
2915 | goto ioerr; | |
2916 | } | |
2917 | mb = page_address(page); | |
2918 | ||
2919 | if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || | |
2920 | mb->version != R5LOG_VERSION) { | |
2921 | create_super = true; | |
2922 | goto create; | |
2923 | } | |
2924 | stored_crc = le32_to_cpu(mb->checksum); | |
2925 | mb->checksum = 0; | |
5cb2fbd6 | 2926 | expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); |
f6bed0ef SL |
2927 | if (stored_crc != expected_crc) { |
2928 | create_super = true; | |
2929 | goto create; | |
2930 | } | |
2931 | if (le64_to_cpu(mb->position) != cp) { | |
2932 | create_super = true; | |
2933 | goto create; | |
2934 | } | |
2935 | create: | |
2936 | if (create_super) { | |
2937 | log->last_cp_seq = prandom_u32(); | |
2938 | cp = 0; | |
56056c2e | 2939 | r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq); |
f6bed0ef SL |
2940 | /* |
2941 | * Make sure super points to correct address. Log might have | |
2942 | * data very soon. If super hasn't correct log tail address, | |
2943 | * recovery can't find the log | |
2944 | */ | |
2945 | r5l_write_super(log, cp); | |
2946 | } else | |
2947 | log->last_cp_seq = le64_to_cpu(mb->seq); | |
2948 | ||
2949 | log->device_size = round_down(rdev->sectors, BLOCK_SECTORS); | |
0576b1c6 SL |
2950 | log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT; |
2951 | if (log->max_free_space > RECLAIM_MAX_FREE_SPACE) | |
2952 | log->max_free_space = RECLAIM_MAX_FREE_SPACE; | |
f6bed0ef SL |
2953 | log->last_checkpoint = cp; |
2954 | ||
2955 | __free_page(page); | |
2956 | ||
d30dfeb9 J |
2957 | if (create_super) { |
2958 | log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS); | |
2959 | log->seq = log->last_cp_seq + 1; | |
2960 | log->next_checkpoint = cp; | |
2961 | } else | |
2962 | ret = r5l_recovery_log(log); | |
2963 | ||
3d7e7e1d ZL |
2964 | r5c_update_log_state(log); |
2965 | return ret; | |
f6bed0ef SL |
2966 | ioerr: |
2967 | __free_page(page); | |
2968 | return ret; | |
2969 | } | |
2970 | ||
2e38a37f SL |
2971 | void r5c_update_on_rdev_error(struct mddev *mddev) |
2972 | { | |
2973 | struct r5conf *conf = mddev->private; | |
2974 | struct r5l_log *log = conf->log; | |
2975 | ||
2976 | if (!log) | |
2977 | return; | |
2978 | ||
2979 | if (raid5_calc_degraded(conf) > 0 && | |
2980 | conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) | |
2981 | schedule_work(&log->disable_writeback_work); | |
2982 | } | |
2983 | ||
f6bed0ef SL |
2984 | int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev) |
2985 | { | |
c888a8f9 | 2986 | struct request_queue *q = bdev_get_queue(rdev->bdev); |
f6bed0ef | 2987 | struct r5l_log *log; |
ff875738 AP |
2988 | char b[BDEVNAME_SIZE]; |
2989 | ||
2990 | pr_debug("md/raid:%s: using device %s as journal\n", | |
2991 | mdname(conf->mddev), bdevname(rdev->bdev, b)); | |
f6bed0ef SL |
2992 | |
2993 | if (PAGE_SIZE != 4096) | |
2994 | return -EINVAL; | |
c757ec95 SL |
2995 | |
2996 | /* | |
2997 | * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and | |
2998 | * raid_disks r5l_payload_data_parity. | |
2999 | * | |
3000 | * Write journal and cache does not work for very big array | |
3001 | * (raid_disks > 203) | |
3002 | */ | |
3003 | if (sizeof(struct r5l_meta_block) + | |
3004 | ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) * | |
3005 | conf->raid_disks) > PAGE_SIZE) { | |
3006 | pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n", | |
3007 | mdname(conf->mddev), conf->raid_disks); | |
3008 | return -EINVAL; | |
3009 | } | |
3010 | ||
f6bed0ef SL |
3011 | log = kzalloc(sizeof(*log), GFP_KERNEL); |
3012 | if (!log) | |
3013 | return -ENOMEM; | |
3014 | log->rdev = rdev; | |
3015 | ||
c888a8f9 | 3016 | log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0; |
56fef7c6 | 3017 | |
5cb2fbd6 SL |
3018 | log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid, |
3019 | sizeof(rdev->mddev->uuid)); | |
f6bed0ef SL |
3020 | |
3021 | mutex_init(&log->io_mutex); | |
3022 | ||
3023 | spin_lock_init(&log->io_list_lock); | |
3024 | INIT_LIST_HEAD(&log->running_ios); | |
0576b1c6 | 3025 | INIT_LIST_HEAD(&log->io_end_ios); |
a8c34f91 | 3026 | INIT_LIST_HEAD(&log->flushing_ios); |
04732f74 | 3027 | INIT_LIST_HEAD(&log->finished_ios); |
3a83f467 | 3028 | bio_init(&log->flush_bio, NULL, 0); |
f6bed0ef SL |
3029 | |
3030 | log->io_kc = KMEM_CACHE(r5l_io_unit, 0); | |
3031 | if (!log->io_kc) | |
3032 | goto io_kc; | |
3033 | ||
5036c390 CH |
3034 | log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc); |
3035 | if (!log->io_pool) | |
3036 | goto io_pool; | |
3037 | ||
c38d29b3 CH |
3038 | log->bs = bioset_create(R5L_POOL_SIZE, 0); |
3039 | if (!log->bs) | |
3040 | goto io_bs; | |
3041 | ||
e8deb638 CH |
3042 | log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0); |
3043 | if (!log->meta_pool) | |
3044 | goto out_mempool; | |
3045 | ||
03b047f4 SL |
3046 | spin_lock_init(&log->tree_lock); |
3047 | INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN); | |
3048 | ||
0576b1c6 SL |
3049 | log->reclaim_thread = md_register_thread(r5l_reclaim_thread, |
3050 | log->rdev->mddev, "reclaim"); | |
3051 | if (!log->reclaim_thread) | |
3052 | goto reclaim_thread; | |
a39f7afd SL |
3053 | log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL; |
3054 | ||
0fd22b45 | 3055 | init_waitqueue_head(&log->iounit_wait); |
0576b1c6 | 3056 | |
5036c390 CH |
3057 | INIT_LIST_HEAD(&log->no_mem_stripes); |
3058 | ||
f6bed0ef SL |
3059 | INIT_LIST_HEAD(&log->no_space_stripes); |
3060 | spin_lock_init(&log->no_space_stripes_lock); | |
3061 | ||
3bddb7f8 | 3062 | INIT_WORK(&log->deferred_io_work, r5l_submit_io_async); |
2e38a37f | 3063 | INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async); |
3bddb7f8 | 3064 | |
2ded3703 | 3065 | log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; |
a39f7afd SL |
3066 | INIT_LIST_HEAD(&log->stripe_in_journal_list); |
3067 | spin_lock_init(&log->stripe_in_journal_lock); | |
3068 | atomic_set(&log->stripe_in_journal_count, 0); | |
2ded3703 | 3069 | |
d2250f10 SL |
3070 | rcu_assign_pointer(conf->log, log); |
3071 | ||
f6bed0ef SL |
3072 | if (r5l_load_log(log)) |
3073 | goto error; | |
3074 | ||
a62ab49e | 3075 | set_bit(MD_HAS_JOURNAL, &conf->mddev->flags); |
f6bed0ef | 3076 | return 0; |
e8deb638 | 3077 | |
f6bed0ef | 3078 | error: |
d2250f10 | 3079 | rcu_assign_pointer(conf->log, NULL); |
0576b1c6 SL |
3080 | md_unregister_thread(&log->reclaim_thread); |
3081 | reclaim_thread: | |
e8deb638 CH |
3082 | mempool_destroy(log->meta_pool); |
3083 | out_mempool: | |
c38d29b3 CH |
3084 | bioset_free(log->bs); |
3085 | io_bs: | |
5036c390 CH |
3086 | mempool_destroy(log->io_pool); |
3087 | io_pool: | |
f6bed0ef SL |
3088 | kmem_cache_destroy(log->io_kc); |
3089 | io_kc: | |
3090 | kfree(log); | |
3091 | return -EINVAL; | |
3092 | } | |
3093 | ||
ff875738 | 3094 | void r5l_exit_log(struct r5conf *conf) |
f6bed0ef | 3095 | { |
ff875738 AP |
3096 | struct r5l_log *log = conf->log; |
3097 | ||
3098 | conf->log = NULL; | |
3099 | synchronize_rcu(); | |
3100 | ||
2e38a37f | 3101 | flush_work(&log->disable_writeback_work); |
0576b1c6 | 3102 | md_unregister_thread(&log->reclaim_thread); |
e8deb638 | 3103 | mempool_destroy(log->meta_pool); |
c38d29b3 | 3104 | bioset_free(log->bs); |
5036c390 | 3105 | mempool_destroy(log->io_pool); |
f6bed0ef SL |
3106 | kmem_cache_destroy(log->io_kc); |
3107 | kfree(log); | |
3108 | } |