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