]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * raid1.c : Multiple Devices driver for Linux | |
3 | * | |
4 | * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat | |
5 | * | |
6 | * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman | |
7 | * | |
8 | * RAID-1 management functions. | |
9 | * | |
10 | * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000 | |
11 | * | |
12 | * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk> | |
13 | * Various fixes by Neil Brown <neilb@cse.unsw.edu.au> | |
14 | * | |
15 | * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support | |
16 | * bitmapped intelligence in resync: | |
17 | * | |
18 | * - bitmap marked during normal i/o | |
19 | * - bitmap used to skip nondirty blocks during sync | |
20 | * | |
21 | * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology: | |
22 | * - persistent bitmap code | |
23 | * | |
24 | * This program is free software; you can redistribute it and/or modify | |
25 | * it under the terms of the GNU General Public License as published by | |
26 | * the Free Software Foundation; either version 2, or (at your option) | |
27 | * any later version. | |
28 | * | |
29 | * You should have received a copy of the GNU General Public License | |
30 | * (for example /usr/src/linux/COPYING); if not, write to the Free | |
31 | * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. | |
32 | */ | |
33 | ||
34 | #include <linux/slab.h> | |
35 | #include <linux/delay.h> | |
36 | #include <linux/blkdev.h> | |
37 | #include <linux/module.h> | |
38 | #include <linux/seq_file.h> | |
39 | #include <linux/ratelimit.h> | |
40 | #include <linux/sched/signal.h> | |
41 | ||
42 | #include <trace/events/block.h> | |
43 | ||
44 | #include "md.h" | |
45 | #include "raid1.h" | |
46 | #include "bitmap.h" | |
47 | ||
48 | #define UNSUPPORTED_MDDEV_FLAGS \ | |
49 | ((1L << MD_HAS_JOURNAL) | \ | |
50 | (1L << MD_JOURNAL_CLEAN) | \ | |
51 | (1L << MD_HAS_PPL)) | |
52 | ||
53 | /* | |
54 | * Number of guaranteed r1bios in case of extreme VM load: | |
55 | */ | |
56 | #define NR_RAID1_BIOS 256 | |
57 | ||
58 | /* when we get a read error on a read-only array, we redirect to another | |
59 | * device without failing the first device, or trying to over-write to | |
60 | * correct the read error. To keep track of bad blocks on a per-bio | |
61 | * level, we store IO_BLOCKED in the appropriate 'bios' pointer | |
62 | */ | |
63 | #define IO_BLOCKED ((struct bio *)1) | |
64 | /* When we successfully write to a known bad-block, we need to remove the | |
65 | * bad-block marking which must be done from process context. So we record | |
66 | * the success by setting devs[n].bio to IO_MADE_GOOD | |
67 | */ | |
68 | #define IO_MADE_GOOD ((struct bio *)2) | |
69 | ||
70 | #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2) | |
71 | ||
72 | /* When there are this many requests queue to be written by | |
73 | * the raid1 thread, we become 'congested' to provide back-pressure | |
74 | * for writeback. | |
75 | */ | |
76 | static int max_queued_requests = 1024; | |
77 | ||
78 | static void allow_barrier(struct r1conf *conf, sector_t sector_nr); | |
79 | static void lower_barrier(struct r1conf *conf, sector_t sector_nr); | |
80 | ||
81 | #define raid1_log(md, fmt, args...) \ | |
82 | do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0) | |
83 | ||
84 | #include "raid1-10.c" | |
85 | ||
86 | /* | |
87 | * for resync bio, r1bio pointer can be retrieved from the per-bio | |
88 | * 'struct resync_pages'. | |
89 | */ | |
90 | static inline struct r1bio *get_resync_r1bio(struct bio *bio) | |
91 | { | |
92 | return get_resync_pages(bio)->raid_bio; | |
93 | } | |
94 | ||
95 | static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data) | |
96 | { | |
97 | struct pool_info *pi = data; | |
98 | int size = offsetof(struct r1bio, bios[pi->raid_disks]); | |
99 | ||
100 | /* allocate a r1bio with room for raid_disks entries in the bios array */ | |
101 | return kzalloc(size, gfp_flags); | |
102 | } | |
103 | ||
104 | static void r1bio_pool_free(void *r1_bio, void *data) | |
105 | { | |
106 | kfree(r1_bio); | |
107 | } | |
108 | ||
109 | #define RESYNC_DEPTH 32 | |
110 | #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9) | |
111 | #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH) | |
112 | #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9) | |
113 | #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW) | |
114 | #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9) | |
115 | ||
116 | static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data) | |
117 | { | |
118 | struct pool_info *pi = data; | |
119 | struct r1bio *r1_bio; | |
120 | struct bio *bio; | |
121 | int need_pages; | |
122 | int j; | |
123 | struct resync_pages *rps; | |
124 | ||
125 | r1_bio = r1bio_pool_alloc(gfp_flags, pi); | |
126 | if (!r1_bio) | |
127 | return NULL; | |
128 | ||
129 | rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks, | |
130 | gfp_flags); | |
131 | if (!rps) | |
132 | goto out_free_r1bio; | |
133 | ||
134 | /* | |
135 | * Allocate bios : 1 for reading, n-1 for writing | |
136 | */ | |
137 | for (j = pi->raid_disks ; j-- ; ) { | |
138 | bio = bio_kmalloc(gfp_flags, RESYNC_PAGES); | |
139 | if (!bio) | |
140 | goto out_free_bio; | |
141 | r1_bio->bios[j] = bio; | |
142 | } | |
143 | /* | |
144 | * Allocate RESYNC_PAGES data pages and attach them to | |
145 | * the first bio. | |
146 | * If this is a user-requested check/repair, allocate | |
147 | * RESYNC_PAGES for each bio. | |
148 | */ | |
149 | if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) | |
150 | need_pages = pi->raid_disks; | |
151 | else | |
152 | need_pages = 1; | |
153 | for (j = 0; j < pi->raid_disks; j++) { | |
154 | struct resync_pages *rp = &rps[j]; | |
155 | ||
156 | bio = r1_bio->bios[j]; | |
157 | ||
158 | if (j < need_pages) { | |
159 | if (resync_alloc_pages(rp, gfp_flags)) | |
160 | goto out_free_pages; | |
161 | } else { | |
162 | memcpy(rp, &rps[0], sizeof(*rp)); | |
163 | resync_get_all_pages(rp); | |
164 | } | |
165 | ||
166 | rp->raid_bio = r1_bio; | |
167 | bio->bi_private = rp; | |
168 | } | |
169 | ||
170 | r1_bio->master_bio = NULL; | |
171 | ||
172 | return r1_bio; | |
173 | ||
174 | out_free_pages: | |
175 | while (--j >= 0) | |
176 | resync_free_pages(&rps[j]); | |
177 | ||
178 | out_free_bio: | |
179 | while (++j < pi->raid_disks) | |
180 | bio_put(r1_bio->bios[j]); | |
181 | kfree(rps); | |
182 | ||
183 | out_free_r1bio: | |
184 | r1bio_pool_free(r1_bio, data); | |
185 | return NULL; | |
186 | } | |
187 | ||
188 | static void r1buf_pool_free(void *__r1_bio, void *data) | |
189 | { | |
190 | struct pool_info *pi = data; | |
191 | int i; | |
192 | struct r1bio *r1bio = __r1_bio; | |
193 | struct resync_pages *rp = NULL; | |
194 | ||
195 | for (i = pi->raid_disks; i--; ) { | |
196 | rp = get_resync_pages(r1bio->bios[i]); | |
197 | resync_free_pages(rp); | |
198 | bio_put(r1bio->bios[i]); | |
199 | } | |
200 | ||
201 | /* resync pages array stored in the 1st bio's .bi_private */ | |
202 | kfree(rp); | |
203 | ||
204 | r1bio_pool_free(r1bio, data); | |
205 | } | |
206 | ||
207 | static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio) | |
208 | { | |
209 | int i; | |
210 | ||
211 | for (i = 0; i < conf->raid_disks * 2; i++) { | |
212 | struct bio **bio = r1_bio->bios + i; | |
213 | if (!BIO_SPECIAL(*bio)) | |
214 | bio_put(*bio); | |
215 | *bio = NULL; | |
216 | } | |
217 | } | |
218 | ||
219 | static void free_r1bio(struct r1bio *r1_bio) | |
220 | { | |
221 | struct r1conf *conf = r1_bio->mddev->private; | |
222 | ||
223 | put_all_bios(conf, r1_bio); | |
224 | mempool_free(r1_bio, conf->r1bio_pool); | |
225 | } | |
226 | ||
227 | static void put_buf(struct r1bio *r1_bio) | |
228 | { | |
229 | struct r1conf *conf = r1_bio->mddev->private; | |
230 | sector_t sect = r1_bio->sector; | |
231 | int i; | |
232 | ||
233 | for (i = 0; i < conf->raid_disks * 2; i++) { | |
234 | struct bio *bio = r1_bio->bios[i]; | |
235 | if (bio->bi_end_io) | |
236 | rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev); | |
237 | } | |
238 | ||
239 | mempool_free(r1_bio, conf->r1buf_pool); | |
240 | ||
241 | lower_barrier(conf, sect); | |
242 | } | |
243 | ||
244 | static void reschedule_retry(struct r1bio *r1_bio) | |
245 | { | |
246 | unsigned long flags; | |
247 | struct mddev *mddev = r1_bio->mddev; | |
248 | struct r1conf *conf = mddev->private; | |
249 | int idx; | |
250 | ||
251 | idx = sector_to_idx(r1_bio->sector); | |
252 | spin_lock_irqsave(&conf->device_lock, flags); | |
253 | list_add(&r1_bio->retry_list, &conf->retry_list); | |
254 | atomic_inc(&conf->nr_queued[idx]); | |
255 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
256 | ||
257 | wake_up(&conf->wait_barrier); | |
258 | md_wakeup_thread(mddev->thread); | |
259 | } | |
260 | ||
261 | /* | |
262 | * raid_end_bio_io() is called when we have finished servicing a mirrored | |
263 | * operation and are ready to return a success/failure code to the buffer | |
264 | * cache layer. | |
265 | */ | |
266 | static void call_bio_endio(struct r1bio *r1_bio) | |
267 | { | |
268 | struct bio *bio = r1_bio->master_bio; | |
269 | struct r1conf *conf = r1_bio->mddev->private; | |
270 | ||
271 | if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) | |
272 | bio->bi_status = BLK_STS_IOERR; | |
273 | ||
274 | bio_endio(bio); | |
275 | /* | |
276 | * Wake up any possible resync thread that waits for the device | |
277 | * to go idle. | |
278 | */ | |
279 | allow_barrier(conf, r1_bio->sector); | |
280 | } | |
281 | ||
282 | static void raid_end_bio_io(struct r1bio *r1_bio) | |
283 | { | |
284 | struct bio *bio = r1_bio->master_bio; | |
285 | ||
286 | /* if nobody has done the final endio yet, do it now */ | |
287 | if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { | |
288 | pr_debug("raid1: sync end %s on sectors %llu-%llu\n", | |
289 | (bio_data_dir(bio) == WRITE) ? "write" : "read", | |
290 | (unsigned long long) bio->bi_iter.bi_sector, | |
291 | (unsigned long long) bio_end_sector(bio) - 1); | |
292 | ||
293 | call_bio_endio(r1_bio); | |
294 | } | |
295 | free_r1bio(r1_bio); | |
296 | } | |
297 | ||
298 | /* | |
299 | * Update disk head position estimator based on IRQ completion info. | |
300 | */ | |
301 | static inline void update_head_pos(int disk, struct r1bio *r1_bio) | |
302 | { | |
303 | struct r1conf *conf = r1_bio->mddev->private; | |
304 | ||
305 | conf->mirrors[disk].head_position = | |
306 | r1_bio->sector + (r1_bio->sectors); | |
307 | } | |
308 | ||
309 | /* | |
310 | * Find the disk number which triggered given bio | |
311 | */ | |
312 | static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio) | |
313 | { | |
314 | int mirror; | |
315 | struct r1conf *conf = r1_bio->mddev->private; | |
316 | int raid_disks = conf->raid_disks; | |
317 | ||
318 | for (mirror = 0; mirror < raid_disks * 2; mirror++) | |
319 | if (r1_bio->bios[mirror] == bio) | |
320 | break; | |
321 | ||
322 | BUG_ON(mirror == raid_disks * 2); | |
323 | update_head_pos(mirror, r1_bio); | |
324 | ||
325 | return mirror; | |
326 | } | |
327 | ||
328 | static void raid1_end_read_request(struct bio *bio) | |
329 | { | |
330 | int uptodate = !bio->bi_status; | |
331 | struct r1bio *r1_bio = bio->bi_private; | |
332 | struct r1conf *conf = r1_bio->mddev->private; | |
333 | struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev; | |
334 | ||
335 | /* | |
336 | * this branch is our 'one mirror IO has finished' event handler: | |
337 | */ | |
338 | update_head_pos(r1_bio->read_disk, r1_bio); | |
339 | ||
340 | if (uptodate) | |
341 | set_bit(R1BIO_Uptodate, &r1_bio->state); | |
342 | else if (test_bit(FailFast, &rdev->flags) && | |
343 | test_bit(R1BIO_FailFast, &r1_bio->state)) | |
344 | /* This was a fail-fast read so we definitely | |
345 | * want to retry */ | |
346 | ; | |
347 | else { | |
348 | /* If all other devices have failed, we want to return | |
349 | * the error upwards rather than fail the last device. | |
350 | * Here we redefine "uptodate" to mean "Don't want to retry" | |
351 | */ | |
352 | unsigned long flags; | |
353 | spin_lock_irqsave(&conf->device_lock, flags); | |
354 | if (r1_bio->mddev->degraded == conf->raid_disks || | |
355 | (r1_bio->mddev->degraded == conf->raid_disks-1 && | |
356 | test_bit(In_sync, &rdev->flags))) | |
357 | uptodate = 1; | |
358 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
359 | } | |
360 | ||
361 | if (uptodate) { | |
362 | raid_end_bio_io(r1_bio); | |
363 | rdev_dec_pending(rdev, conf->mddev); | |
364 | } else { | |
365 | /* | |
366 | * oops, read error: | |
367 | */ | |
368 | char b[BDEVNAME_SIZE]; | |
369 | pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n", | |
370 | mdname(conf->mddev), | |
371 | bdevname(rdev->bdev, b), | |
372 | (unsigned long long)r1_bio->sector); | |
373 | set_bit(R1BIO_ReadError, &r1_bio->state); | |
374 | reschedule_retry(r1_bio); | |
375 | /* don't drop the reference on read_disk yet */ | |
376 | } | |
377 | } | |
378 | ||
379 | static void close_write(struct r1bio *r1_bio) | |
380 | { | |
381 | /* it really is the end of this request */ | |
382 | if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { | |
383 | bio_free_pages(r1_bio->behind_master_bio); | |
384 | bio_put(r1_bio->behind_master_bio); | |
385 | r1_bio->behind_master_bio = NULL; | |
386 | } | |
387 | /* clear the bitmap if all writes complete successfully */ | |
388 | bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector, | |
389 | r1_bio->sectors, | |
390 | !test_bit(R1BIO_Degraded, &r1_bio->state), | |
391 | test_bit(R1BIO_BehindIO, &r1_bio->state)); | |
392 | md_write_end(r1_bio->mddev); | |
393 | } | |
394 | ||
395 | static void r1_bio_write_done(struct r1bio *r1_bio) | |
396 | { | |
397 | if (!atomic_dec_and_test(&r1_bio->remaining)) | |
398 | return; | |
399 | ||
400 | if (test_bit(R1BIO_WriteError, &r1_bio->state)) | |
401 | reschedule_retry(r1_bio); | |
402 | else { | |
403 | close_write(r1_bio); | |
404 | if (test_bit(R1BIO_MadeGood, &r1_bio->state)) | |
405 | reschedule_retry(r1_bio); | |
406 | else | |
407 | raid_end_bio_io(r1_bio); | |
408 | } | |
409 | } | |
410 | ||
411 | static void raid1_end_write_request(struct bio *bio) | |
412 | { | |
413 | struct r1bio *r1_bio = bio->bi_private; | |
414 | int behind = test_bit(R1BIO_BehindIO, &r1_bio->state); | |
415 | struct r1conf *conf = r1_bio->mddev->private; | |
416 | struct bio *to_put = NULL; | |
417 | int mirror = find_bio_disk(r1_bio, bio); | |
418 | struct md_rdev *rdev = conf->mirrors[mirror].rdev; | |
419 | bool discard_error; | |
420 | ||
421 | discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD; | |
422 | ||
423 | /* | |
424 | * 'one mirror IO has finished' event handler: | |
425 | */ | |
426 | if (bio->bi_status && !discard_error) { | |
427 | set_bit(WriteErrorSeen, &rdev->flags); | |
428 | if (!test_and_set_bit(WantReplacement, &rdev->flags)) | |
429 | set_bit(MD_RECOVERY_NEEDED, & | |
430 | conf->mddev->recovery); | |
431 | ||
432 | if (test_bit(FailFast, &rdev->flags) && | |
433 | (bio->bi_opf & MD_FAILFAST) && | |
434 | /* We never try FailFast to WriteMostly devices */ | |
435 | !test_bit(WriteMostly, &rdev->flags)) { | |
436 | md_error(r1_bio->mddev, rdev); | |
437 | if (!test_bit(Faulty, &rdev->flags)) | |
438 | /* This is the only remaining device, | |
439 | * We need to retry the write without | |
440 | * FailFast | |
441 | */ | |
442 | set_bit(R1BIO_WriteError, &r1_bio->state); | |
443 | else { | |
444 | /* Finished with this branch */ | |
445 | r1_bio->bios[mirror] = NULL; | |
446 | to_put = bio; | |
447 | } | |
448 | } else | |
449 | set_bit(R1BIO_WriteError, &r1_bio->state); | |
450 | } else { | |
451 | /* | |
452 | * Set R1BIO_Uptodate in our master bio, so that we | |
453 | * will return a good error code for to the higher | |
454 | * levels even if IO on some other mirrored buffer | |
455 | * fails. | |
456 | * | |
457 | * The 'master' represents the composite IO operation | |
458 | * to user-side. So if something waits for IO, then it | |
459 | * will wait for the 'master' bio. | |
460 | */ | |
461 | sector_t first_bad; | |
462 | int bad_sectors; | |
463 | ||
464 | r1_bio->bios[mirror] = NULL; | |
465 | to_put = bio; | |
466 | /* | |
467 | * Do not set R1BIO_Uptodate if the current device is | |
468 | * rebuilding or Faulty. This is because we cannot use | |
469 | * such device for properly reading the data back (we could | |
470 | * potentially use it, if the current write would have felt | |
471 | * before rdev->recovery_offset, but for simplicity we don't | |
472 | * check this here. | |
473 | */ | |
474 | if (test_bit(In_sync, &rdev->flags) && | |
475 | !test_bit(Faulty, &rdev->flags)) | |
476 | set_bit(R1BIO_Uptodate, &r1_bio->state); | |
477 | ||
478 | /* Maybe we can clear some bad blocks. */ | |
479 | if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors, | |
480 | &first_bad, &bad_sectors) && !discard_error) { | |
481 | r1_bio->bios[mirror] = IO_MADE_GOOD; | |
482 | set_bit(R1BIO_MadeGood, &r1_bio->state); | |
483 | } | |
484 | } | |
485 | ||
486 | if (behind) { | |
487 | if (test_bit(WriteMostly, &rdev->flags)) | |
488 | atomic_dec(&r1_bio->behind_remaining); | |
489 | ||
490 | /* | |
491 | * In behind mode, we ACK the master bio once the I/O | |
492 | * has safely reached all non-writemostly | |
493 | * disks. Setting the Returned bit ensures that this | |
494 | * gets done only once -- we don't ever want to return | |
495 | * -EIO here, instead we'll wait | |
496 | */ | |
497 | if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) && | |
498 | test_bit(R1BIO_Uptodate, &r1_bio->state)) { | |
499 | /* Maybe we can return now */ | |
500 | if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) { | |
501 | struct bio *mbio = r1_bio->master_bio; | |
502 | pr_debug("raid1: behind end write sectors" | |
503 | " %llu-%llu\n", | |
504 | (unsigned long long) mbio->bi_iter.bi_sector, | |
505 | (unsigned long long) bio_end_sector(mbio) - 1); | |
506 | call_bio_endio(r1_bio); | |
507 | } | |
508 | } | |
509 | } | |
510 | if (r1_bio->bios[mirror] == NULL) | |
511 | rdev_dec_pending(rdev, conf->mddev); | |
512 | ||
513 | /* | |
514 | * Let's see if all mirrored write operations have finished | |
515 | * already. | |
516 | */ | |
517 | r1_bio_write_done(r1_bio); | |
518 | ||
519 | if (to_put) | |
520 | bio_put(to_put); | |
521 | } | |
522 | ||
523 | static sector_t align_to_barrier_unit_end(sector_t start_sector, | |
524 | sector_t sectors) | |
525 | { | |
526 | sector_t len; | |
527 | ||
528 | WARN_ON(sectors == 0); | |
529 | /* | |
530 | * len is the number of sectors from start_sector to end of the | |
531 | * barrier unit which start_sector belongs to. | |
532 | */ | |
533 | len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) - | |
534 | start_sector; | |
535 | ||
536 | if (len > sectors) | |
537 | len = sectors; | |
538 | ||
539 | return len; | |
540 | } | |
541 | ||
542 | /* | |
543 | * This routine returns the disk from which the requested read should | |
544 | * be done. There is a per-array 'next expected sequential IO' sector | |
545 | * number - if this matches on the next IO then we use the last disk. | |
546 | * There is also a per-disk 'last know head position' sector that is | |
547 | * maintained from IRQ contexts, both the normal and the resync IO | |
548 | * completion handlers update this position correctly. If there is no | |
549 | * perfect sequential match then we pick the disk whose head is closest. | |
550 | * | |
551 | * If there are 2 mirrors in the same 2 devices, performance degrades | |
552 | * because position is mirror, not device based. | |
553 | * | |
554 | * The rdev for the device selected will have nr_pending incremented. | |
555 | */ | |
556 | static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors) | |
557 | { | |
558 | const sector_t this_sector = r1_bio->sector; | |
559 | int sectors; | |
560 | int best_good_sectors; | |
561 | int best_disk, best_dist_disk, best_pending_disk; | |
562 | int has_nonrot_disk; | |
563 | int disk; | |
564 | sector_t best_dist; | |
565 | unsigned int min_pending; | |
566 | struct md_rdev *rdev; | |
567 | int choose_first; | |
568 | int choose_next_idle; | |
569 | ||
570 | rcu_read_lock(); | |
571 | /* | |
572 | * Check if we can balance. We can balance on the whole | |
573 | * device if no resync is going on, or below the resync window. | |
574 | * We take the first readable disk when above the resync window. | |
575 | */ | |
576 | retry: | |
577 | sectors = r1_bio->sectors; | |
578 | best_disk = -1; | |
579 | best_dist_disk = -1; | |
580 | best_dist = MaxSector; | |
581 | best_pending_disk = -1; | |
582 | min_pending = UINT_MAX; | |
583 | best_good_sectors = 0; | |
584 | has_nonrot_disk = 0; | |
585 | choose_next_idle = 0; | |
586 | clear_bit(R1BIO_FailFast, &r1_bio->state); | |
587 | ||
588 | if ((conf->mddev->recovery_cp < this_sector + sectors) || | |
589 | (mddev_is_clustered(conf->mddev) && | |
590 | md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector, | |
591 | this_sector + sectors))) | |
592 | choose_first = 1; | |
593 | else | |
594 | choose_first = 0; | |
595 | ||
596 | for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) { | |
597 | sector_t dist; | |
598 | sector_t first_bad; | |
599 | int bad_sectors; | |
600 | unsigned int pending; | |
601 | bool nonrot; | |
602 | ||
603 | rdev = rcu_dereference(conf->mirrors[disk].rdev); | |
604 | if (r1_bio->bios[disk] == IO_BLOCKED | |
605 | || rdev == NULL | |
606 | || test_bit(Faulty, &rdev->flags)) | |
607 | continue; | |
608 | if (!test_bit(In_sync, &rdev->flags) && | |
609 | rdev->recovery_offset < this_sector + sectors) | |
610 | continue; | |
611 | if (test_bit(WriteMostly, &rdev->flags)) { | |
612 | /* Don't balance among write-mostly, just | |
613 | * use the first as a last resort */ | |
614 | if (best_dist_disk < 0) { | |
615 | if (is_badblock(rdev, this_sector, sectors, | |
616 | &first_bad, &bad_sectors)) { | |
617 | if (first_bad <= this_sector) | |
618 | /* Cannot use this */ | |
619 | continue; | |
620 | best_good_sectors = first_bad - this_sector; | |
621 | } else | |
622 | best_good_sectors = sectors; | |
623 | best_dist_disk = disk; | |
624 | best_pending_disk = disk; | |
625 | } | |
626 | continue; | |
627 | } | |
628 | /* This is a reasonable device to use. It might | |
629 | * even be best. | |
630 | */ | |
631 | if (is_badblock(rdev, this_sector, sectors, | |
632 | &first_bad, &bad_sectors)) { | |
633 | if (best_dist < MaxSector) | |
634 | /* already have a better device */ | |
635 | continue; | |
636 | if (first_bad <= this_sector) { | |
637 | /* cannot read here. If this is the 'primary' | |
638 | * device, then we must not read beyond | |
639 | * bad_sectors from another device.. | |
640 | */ | |
641 | bad_sectors -= (this_sector - first_bad); | |
642 | if (choose_first && sectors > bad_sectors) | |
643 | sectors = bad_sectors; | |
644 | if (best_good_sectors > sectors) | |
645 | best_good_sectors = sectors; | |
646 | ||
647 | } else { | |
648 | sector_t good_sectors = first_bad - this_sector; | |
649 | if (good_sectors > best_good_sectors) { | |
650 | best_good_sectors = good_sectors; | |
651 | best_disk = disk; | |
652 | } | |
653 | if (choose_first) | |
654 | break; | |
655 | } | |
656 | continue; | |
657 | } else { | |
658 | if ((sectors > best_good_sectors) && (best_disk >= 0)) | |
659 | best_disk = -1; | |
660 | best_good_sectors = sectors; | |
661 | } | |
662 | ||
663 | if (best_disk >= 0) | |
664 | /* At least two disks to choose from so failfast is OK */ | |
665 | set_bit(R1BIO_FailFast, &r1_bio->state); | |
666 | ||
667 | nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev)); | |
668 | has_nonrot_disk |= nonrot; | |
669 | pending = atomic_read(&rdev->nr_pending); | |
670 | dist = abs(this_sector - conf->mirrors[disk].head_position); | |
671 | if (choose_first) { | |
672 | best_disk = disk; | |
673 | break; | |
674 | } | |
675 | /* Don't change to another disk for sequential reads */ | |
676 | if (conf->mirrors[disk].next_seq_sect == this_sector | |
677 | || dist == 0) { | |
678 | int opt_iosize = bdev_io_opt(rdev->bdev) >> 9; | |
679 | struct raid1_info *mirror = &conf->mirrors[disk]; | |
680 | ||
681 | best_disk = disk; | |
682 | /* | |
683 | * If buffered sequential IO size exceeds optimal | |
684 | * iosize, check if there is idle disk. If yes, choose | |
685 | * the idle disk. read_balance could already choose an | |
686 | * idle disk before noticing it's a sequential IO in | |
687 | * this disk. This doesn't matter because this disk | |
688 | * will idle, next time it will be utilized after the | |
689 | * first disk has IO size exceeds optimal iosize. In | |
690 | * this way, iosize of the first disk will be optimal | |
691 | * iosize at least. iosize of the second disk might be | |
692 | * small, but not a big deal since when the second disk | |
693 | * starts IO, the first disk is likely still busy. | |
694 | */ | |
695 | if (nonrot && opt_iosize > 0 && | |
696 | mirror->seq_start != MaxSector && | |
697 | mirror->next_seq_sect > opt_iosize && | |
698 | mirror->next_seq_sect - opt_iosize >= | |
699 | mirror->seq_start) { | |
700 | choose_next_idle = 1; | |
701 | continue; | |
702 | } | |
703 | break; | |
704 | } | |
705 | ||
706 | if (choose_next_idle) | |
707 | continue; | |
708 | ||
709 | if (min_pending > pending) { | |
710 | min_pending = pending; | |
711 | best_pending_disk = disk; | |
712 | } | |
713 | ||
714 | if (dist < best_dist) { | |
715 | best_dist = dist; | |
716 | best_dist_disk = disk; | |
717 | } | |
718 | } | |
719 | ||
720 | /* | |
721 | * If all disks are rotational, choose the closest disk. If any disk is | |
722 | * non-rotational, choose the disk with less pending request even the | |
723 | * disk is rotational, which might/might not be optimal for raids with | |
724 | * mixed ratation/non-rotational disks depending on workload. | |
725 | */ | |
726 | if (best_disk == -1) { | |
727 | if (has_nonrot_disk || min_pending == 0) | |
728 | best_disk = best_pending_disk; | |
729 | else | |
730 | best_disk = best_dist_disk; | |
731 | } | |
732 | ||
733 | if (best_disk >= 0) { | |
734 | rdev = rcu_dereference(conf->mirrors[best_disk].rdev); | |
735 | if (!rdev) | |
736 | goto retry; | |
737 | atomic_inc(&rdev->nr_pending); | |
738 | sectors = best_good_sectors; | |
739 | ||
740 | if (conf->mirrors[best_disk].next_seq_sect != this_sector) | |
741 | conf->mirrors[best_disk].seq_start = this_sector; | |
742 | ||
743 | conf->mirrors[best_disk].next_seq_sect = this_sector + sectors; | |
744 | } | |
745 | rcu_read_unlock(); | |
746 | *max_sectors = sectors; | |
747 | ||
748 | return best_disk; | |
749 | } | |
750 | ||
751 | static int raid1_congested(struct mddev *mddev, int bits) | |
752 | { | |
753 | struct r1conf *conf = mddev->private; | |
754 | int i, ret = 0; | |
755 | ||
756 | if ((bits & (1 << WB_async_congested)) && | |
757 | conf->pending_count >= max_queued_requests) | |
758 | return 1; | |
759 | ||
760 | rcu_read_lock(); | |
761 | for (i = 0; i < conf->raid_disks * 2; i++) { | |
762 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); | |
763 | if (rdev && !test_bit(Faulty, &rdev->flags)) { | |
764 | struct request_queue *q = bdev_get_queue(rdev->bdev); | |
765 | ||
766 | BUG_ON(!q); | |
767 | ||
768 | /* Note the '|| 1' - when read_balance prefers | |
769 | * non-congested targets, it can be removed | |
770 | */ | |
771 | if ((bits & (1 << WB_async_congested)) || 1) | |
772 | ret |= bdi_congested(q->backing_dev_info, bits); | |
773 | else | |
774 | ret &= bdi_congested(q->backing_dev_info, bits); | |
775 | } | |
776 | } | |
777 | rcu_read_unlock(); | |
778 | return ret; | |
779 | } | |
780 | ||
781 | static void flush_bio_list(struct r1conf *conf, struct bio *bio) | |
782 | { | |
783 | /* flush any pending bitmap writes to disk before proceeding w/ I/O */ | |
784 | bitmap_unplug(conf->mddev->bitmap); | |
785 | wake_up(&conf->wait_barrier); | |
786 | ||
787 | while (bio) { /* submit pending writes */ | |
788 | struct bio *next = bio->bi_next; | |
789 | struct md_rdev *rdev = (void *)bio->bi_disk; | |
790 | bio->bi_next = NULL; | |
791 | bio_set_dev(bio, rdev->bdev); | |
792 | if (test_bit(Faulty, &rdev->flags)) { | |
793 | bio_io_error(bio); | |
794 | } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) && | |
795 | !blk_queue_discard(bio->bi_disk->queue))) | |
796 | /* Just ignore it */ | |
797 | bio_endio(bio); | |
798 | else | |
799 | generic_make_request(bio); | |
800 | bio = next; | |
801 | } | |
802 | } | |
803 | ||
804 | static void flush_pending_writes(struct r1conf *conf) | |
805 | { | |
806 | /* Any writes that have been queued but are awaiting | |
807 | * bitmap updates get flushed here. | |
808 | */ | |
809 | spin_lock_irq(&conf->device_lock); | |
810 | ||
811 | if (conf->pending_bio_list.head) { | |
812 | struct bio *bio; | |
813 | bio = bio_list_get(&conf->pending_bio_list); | |
814 | conf->pending_count = 0; | |
815 | spin_unlock_irq(&conf->device_lock); | |
816 | flush_bio_list(conf, bio); | |
817 | } else | |
818 | spin_unlock_irq(&conf->device_lock); | |
819 | } | |
820 | ||
821 | /* Barriers.... | |
822 | * Sometimes we need to suspend IO while we do something else, | |
823 | * either some resync/recovery, or reconfigure the array. | |
824 | * To do this we raise a 'barrier'. | |
825 | * The 'barrier' is a counter that can be raised multiple times | |
826 | * to count how many activities are happening which preclude | |
827 | * normal IO. | |
828 | * We can only raise the barrier if there is no pending IO. | |
829 | * i.e. if nr_pending == 0. | |
830 | * We choose only to raise the barrier if no-one is waiting for the | |
831 | * barrier to go down. This means that as soon as an IO request | |
832 | * is ready, no other operations which require a barrier will start | |
833 | * until the IO request has had a chance. | |
834 | * | |
835 | * So: regular IO calls 'wait_barrier'. When that returns there | |
836 | * is no backgroup IO happening, It must arrange to call | |
837 | * allow_barrier when it has finished its IO. | |
838 | * backgroup IO calls must call raise_barrier. Once that returns | |
839 | * there is no normal IO happeing. It must arrange to call | |
840 | * lower_barrier when the particular background IO completes. | |
841 | */ | |
842 | static void raise_barrier(struct r1conf *conf, sector_t sector_nr) | |
843 | { | |
844 | int idx = sector_to_idx(sector_nr); | |
845 | ||
846 | spin_lock_irq(&conf->resync_lock); | |
847 | ||
848 | /* Wait until no block IO is waiting */ | |
849 | wait_event_lock_irq(conf->wait_barrier, | |
850 | !atomic_read(&conf->nr_waiting[idx]), | |
851 | conf->resync_lock); | |
852 | ||
853 | /* block any new IO from starting */ | |
854 | atomic_inc(&conf->barrier[idx]); | |
855 | /* | |
856 | * In raise_barrier() we firstly increase conf->barrier[idx] then | |
857 | * check conf->nr_pending[idx]. In _wait_barrier() we firstly | |
858 | * increase conf->nr_pending[idx] then check conf->barrier[idx]. | |
859 | * A memory barrier here to make sure conf->nr_pending[idx] won't | |
860 | * be fetched before conf->barrier[idx] is increased. Otherwise | |
861 | * there will be a race between raise_barrier() and _wait_barrier(). | |
862 | */ | |
863 | smp_mb__after_atomic(); | |
864 | ||
865 | /* For these conditions we must wait: | |
866 | * A: while the array is in frozen state | |
867 | * B: while conf->nr_pending[idx] is not 0, meaning regular I/O | |
868 | * existing in corresponding I/O barrier bucket. | |
869 | * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches | |
870 | * max resync count which allowed on current I/O barrier bucket. | |
871 | */ | |
872 | wait_event_lock_irq(conf->wait_barrier, | |
873 | !conf->array_frozen && | |
874 | !atomic_read(&conf->nr_pending[idx]) && | |
875 | atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH, | |
876 | conf->resync_lock); | |
877 | ||
878 | atomic_inc(&conf->nr_sync_pending); | |
879 | spin_unlock_irq(&conf->resync_lock); | |
880 | } | |
881 | ||
882 | static void lower_barrier(struct r1conf *conf, sector_t sector_nr) | |
883 | { | |
884 | int idx = sector_to_idx(sector_nr); | |
885 | ||
886 | BUG_ON(atomic_read(&conf->barrier[idx]) <= 0); | |
887 | ||
888 | atomic_dec(&conf->barrier[idx]); | |
889 | atomic_dec(&conf->nr_sync_pending); | |
890 | wake_up(&conf->wait_barrier); | |
891 | } | |
892 | ||
893 | static void _wait_barrier(struct r1conf *conf, int idx) | |
894 | { | |
895 | /* | |
896 | * We need to increase conf->nr_pending[idx] very early here, | |
897 | * then raise_barrier() can be blocked when it waits for | |
898 | * conf->nr_pending[idx] to be 0. Then we can avoid holding | |
899 | * conf->resync_lock when there is no barrier raised in same | |
900 | * barrier unit bucket. Also if the array is frozen, I/O | |
901 | * should be blocked until array is unfrozen. | |
902 | */ | |
903 | atomic_inc(&conf->nr_pending[idx]); | |
904 | /* | |
905 | * In _wait_barrier() we firstly increase conf->nr_pending[idx], then | |
906 | * check conf->barrier[idx]. In raise_barrier() we firstly increase | |
907 | * conf->barrier[idx], then check conf->nr_pending[idx]. A memory | |
908 | * barrier is necessary here to make sure conf->barrier[idx] won't be | |
909 | * fetched before conf->nr_pending[idx] is increased. Otherwise there | |
910 | * will be a race between _wait_barrier() and raise_barrier(). | |
911 | */ | |
912 | smp_mb__after_atomic(); | |
913 | ||
914 | /* | |
915 | * Don't worry about checking two atomic_t variables at same time | |
916 | * here. If during we check conf->barrier[idx], the array is | |
917 | * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is | |
918 | * 0, it is safe to return and make the I/O continue. Because the | |
919 | * array is frozen, all I/O returned here will eventually complete | |
920 | * or be queued, no race will happen. See code comment in | |
921 | * frozen_array(). | |
922 | */ | |
923 | if (!READ_ONCE(conf->array_frozen) && | |
924 | !atomic_read(&conf->barrier[idx])) | |
925 | return; | |
926 | ||
927 | /* | |
928 | * After holding conf->resync_lock, conf->nr_pending[idx] | |
929 | * should be decreased before waiting for barrier to drop. | |
930 | * Otherwise, we may encounter a race condition because | |
931 | * raise_barrer() might be waiting for conf->nr_pending[idx] | |
932 | * to be 0 at same time. | |
933 | */ | |
934 | spin_lock_irq(&conf->resync_lock); | |
935 | atomic_inc(&conf->nr_waiting[idx]); | |
936 | atomic_dec(&conf->nr_pending[idx]); | |
937 | /* | |
938 | * In case freeze_array() is waiting for | |
939 | * get_unqueued_pending() == extra | |
940 | */ | |
941 | wake_up(&conf->wait_barrier); | |
942 | /* Wait for the barrier in same barrier unit bucket to drop. */ | |
943 | wait_event_lock_irq(conf->wait_barrier, | |
944 | !conf->array_frozen && | |
945 | !atomic_read(&conf->barrier[idx]), | |
946 | conf->resync_lock); | |
947 | atomic_inc(&conf->nr_pending[idx]); | |
948 | atomic_dec(&conf->nr_waiting[idx]); | |
949 | spin_unlock_irq(&conf->resync_lock); | |
950 | } | |
951 | ||
952 | static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr) | |
953 | { | |
954 | int idx = sector_to_idx(sector_nr); | |
955 | ||
956 | /* | |
957 | * Very similar to _wait_barrier(). The difference is, for read | |
958 | * I/O we don't need wait for sync I/O, but if the whole array | |
959 | * is frozen, the read I/O still has to wait until the array is | |
960 | * unfrozen. Since there is no ordering requirement with | |
961 | * conf->barrier[idx] here, memory barrier is unnecessary as well. | |
962 | */ | |
963 | atomic_inc(&conf->nr_pending[idx]); | |
964 | ||
965 | if (!READ_ONCE(conf->array_frozen)) | |
966 | return; | |
967 | ||
968 | spin_lock_irq(&conf->resync_lock); | |
969 | atomic_inc(&conf->nr_waiting[idx]); | |
970 | atomic_dec(&conf->nr_pending[idx]); | |
971 | /* | |
972 | * In case freeze_array() is waiting for | |
973 | * get_unqueued_pending() == extra | |
974 | */ | |
975 | wake_up(&conf->wait_barrier); | |
976 | /* Wait for array to be unfrozen */ | |
977 | wait_event_lock_irq(conf->wait_barrier, | |
978 | !conf->array_frozen, | |
979 | conf->resync_lock); | |
980 | atomic_inc(&conf->nr_pending[idx]); | |
981 | atomic_dec(&conf->nr_waiting[idx]); | |
982 | spin_unlock_irq(&conf->resync_lock); | |
983 | } | |
984 | ||
985 | static void wait_barrier(struct r1conf *conf, sector_t sector_nr) | |
986 | { | |
987 | int idx = sector_to_idx(sector_nr); | |
988 | ||
989 | _wait_barrier(conf, idx); | |
990 | } | |
991 | ||
992 | static void wait_all_barriers(struct r1conf *conf) | |
993 | { | |
994 | int idx; | |
995 | ||
996 | for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) | |
997 | _wait_barrier(conf, idx); | |
998 | } | |
999 | ||
1000 | static void _allow_barrier(struct r1conf *conf, int idx) | |
1001 | { | |
1002 | atomic_dec(&conf->nr_pending[idx]); | |
1003 | wake_up(&conf->wait_barrier); | |
1004 | } | |
1005 | ||
1006 | static void allow_barrier(struct r1conf *conf, sector_t sector_nr) | |
1007 | { | |
1008 | int idx = sector_to_idx(sector_nr); | |
1009 | ||
1010 | _allow_barrier(conf, idx); | |
1011 | } | |
1012 | ||
1013 | static void allow_all_barriers(struct r1conf *conf) | |
1014 | { | |
1015 | int idx; | |
1016 | ||
1017 | for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) | |
1018 | _allow_barrier(conf, idx); | |
1019 | } | |
1020 | ||
1021 | /* conf->resync_lock should be held */ | |
1022 | static int get_unqueued_pending(struct r1conf *conf) | |
1023 | { | |
1024 | int idx, ret; | |
1025 | ||
1026 | ret = atomic_read(&conf->nr_sync_pending); | |
1027 | for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) | |
1028 | ret += atomic_read(&conf->nr_pending[idx]) - | |
1029 | atomic_read(&conf->nr_queued[idx]); | |
1030 | ||
1031 | return ret; | |
1032 | } | |
1033 | ||
1034 | static void freeze_array(struct r1conf *conf, int extra) | |
1035 | { | |
1036 | /* Stop sync I/O and normal I/O and wait for everything to | |
1037 | * go quiet. | |
1038 | * This is called in two situations: | |
1039 | * 1) management command handlers (reshape, remove disk, quiesce). | |
1040 | * 2) one normal I/O request failed. | |
1041 | ||
1042 | * After array_frozen is set to 1, new sync IO will be blocked at | |
1043 | * raise_barrier(), and new normal I/O will blocked at _wait_barrier() | |
1044 | * or wait_read_barrier(). The flying I/Os will either complete or be | |
1045 | * queued. When everything goes quite, there are only queued I/Os left. | |
1046 | ||
1047 | * Every flying I/O contributes to a conf->nr_pending[idx], idx is the | |
1048 | * barrier bucket index which this I/O request hits. When all sync and | |
1049 | * normal I/O are queued, sum of all conf->nr_pending[] will match sum | |
1050 | * of all conf->nr_queued[]. But normal I/O failure is an exception, | |
1051 | * in handle_read_error(), we may call freeze_array() before trying to | |
1052 | * fix the read error. In this case, the error read I/O is not queued, | |
1053 | * so get_unqueued_pending() == 1. | |
1054 | * | |
1055 | * Therefore before this function returns, we need to wait until | |
1056 | * get_unqueued_pendings(conf) gets equal to extra. For | |
1057 | * normal I/O context, extra is 1, in rested situations extra is 0. | |
1058 | */ | |
1059 | spin_lock_irq(&conf->resync_lock); | |
1060 | conf->array_frozen = 1; | |
1061 | raid1_log(conf->mddev, "wait freeze"); | |
1062 | wait_event_lock_irq_cmd( | |
1063 | conf->wait_barrier, | |
1064 | get_unqueued_pending(conf) == extra, | |
1065 | conf->resync_lock, | |
1066 | flush_pending_writes(conf)); | |
1067 | spin_unlock_irq(&conf->resync_lock); | |
1068 | } | |
1069 | static void unfreeze_array(struct r1conf *conf) | |
1070 | { | |
1071 | /* reverse the effect of the freeze */ | |
1072 | spin_lock_irq(&conf->resync_lock); | |
1073 | conf->array_frozen = 0; | |
1074 | spin_unlock_irq(&conf->resync_lock); | |
1075 | wake_up(&conf->wait_barrier); | |
1076 | } | |
1077 | ||
1078 | static void alloc_behind_master_bio(struct r1bio *r1_bio, | |
1079 | struct bio *bio) | |
1080 | { | |
1081 | int size = bio->bi_iter.bi_size; | |
1082 | unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
1083 | int i = 0; | |
1084 | struct bio *behind_bio = NULL; | |
1085 | ||
1086 | behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev); | |
1087 | if (!behind_bio) | |
1088 | return; | |
1089 | ||
1090 | /* discard op, we don't support writezero/writesame yet */ | |
1091 | if (!bio_has_data(bio)) { | |
1092 | behind_bio->bi_iter.bi_size = size; | |
1093 | goto skip_copy; | |
1094 | } | |
1095 | ||
1096 | while (i < vcnt && size) { | |
1097 | struct page *page; | |
1098 | int len = min_t(int, PAGE_SIZE, size); | |
1099 | ||
1100 | page = alloc_page(GFP_NOIO); | |
1101 | if (unlikely(!page)) | |
1102 | goto free_pages; | |
1103 | ||
1104 | bio_add_page(behind_bio, page, len, 0); | |
1105 | ||
1106 | size -= len; | |
1107 | i++; | |
1108 | } | |
1109 | ||
1110 | bio_copy_data(behind_bio, bio); | |
1111 | skip_copy: | |
1112 | r1_bio->behind_master_bio = behind_bio;; | |
1113 | set_bit(R1BIO_BehindIO, &r1_bio->state); | |
1114 | ||
1115 | return; | |
1116 | ||
1117 | free_pages: | |
1118 | pr_debug("%dB behind alloc failed, doing sync I/O\n", | |
1119 | bio->bi_iter.bi_size); | |
1120 | bio_free_pages(behind_bio); | |
1121 | bio_put(behind_bio); | |
1122 | } | |
1123 | ||
1124 | struct raid1_plug_cb { | |
1125 | struct blk_plug_cb cb; | |
1126 | struct bio_list pending; | |
1127 | int pending_cnt; | |
1128 | }; | |
1129 | ||
1130 | static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule) | |
1131 | { | |
1132 | struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, | |
1133 | cb); | |
1134 | struct mddev *mddev = plug->cb.data; | |
1135 | struct r1conf *conf = mddev->private; | |
1136 | struct bio *bio; | |
1137 | ||
1138 | if (from_schedule || current->bio_list) { | |
1139 | spin_lock_irq(&conf->device_lock); | |
1140 | bio_list_merge(&conf->pending_bio_list, &plug->pending); | |
1141 | conf->pending_count += plug->pending_cnt; | |
1142 | spin_unlock_irq(&conf->device_lock); | |
1143 | wake_up(&conf->wait_barrier); | |
1144 | md_wakeup_thread(mddev->thread); | |
1145 | kfree(plug); | |
1146 | return; | |
1147 | } | |
1148 | ||
1149 | /* we aren't scheduling, so we can do the write-out directly. */ | |
1150 | bio = bio_list_get(&plug->pending); | |
1151 | flush_bio_list(conf, bio); | |
1152 | kfree(plug); | |
1153 | } | |
1154 | ||
1155 | static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio) | |
1156 | { | |
1157 | r1_bio->master_bio = bio; | |
1158 | r1_bio->sectors = bio_sectors(bio); | |
1159 | r1_bio->state = 0; | |
1160 | r1_bio->mddev = mddev; | |
1161 | r1_bio->sector = bio->bi_iter.bi_sector; | |
1162 | } | |
1163 | ||
1164 | static inline struct r1bio * | |
1165 | alloc_r1bio(struct mddev *mddev, struct bio *bio) | |
1166 | { | |
1167 | struct r1conf *conf = mddev->private; | |
1168 | struct r1bio *r1_bio; | |
1169 | ||
1170 | r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO); | |
1171 | /* Ensure no bio records IO_BLOCKED */ | |
1172 | memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0])); | |
1173 | init_r1bio(r1_bio, mddev, bio); | |
1174 | return r1_bio; | |
1175 | } | |
1176 | ||
1177 | static void raid1_read_request(struct mddev *mddev, struct bio *bio, | |
1178 | int max_read_sectors, struct r1bio *r1_bio) | |
1179 | { | |
1180 | struct r1conf *conf = mddev->private; | |
1181 | struct raid1_info *mirror; | |
1182 | struct bio *read_bio; | |
1183 | struct bitmap *bitmap = mddev->bitmap; | |
1184 | const int op = bio_op(bio); | |
1185 | const unsigned long do_sync = (bio->bi_opf & REQ_SYNC); | |
1186 | int max_sectors; | |
1187 | int rdisk; | |
1188 | bool print_msg = !!r1_bio; | |
1189 | char b[BDEVNAME_SIZE]; | |
1190 | ||
1191 | /* | |
1192 | * If r1_bio is set, we are blocking the raid1d thread | |
1193 | * so there is a tiny risk of deadlock. So ask for | |
1194 | * emergency memory if needed. | |
1195 | */ | |
1196 | gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO; | |
1197 | ||
1198 | if (print_msg) { | |
1199 | /* Need to get the block device name carefully */ | |
1200 | struct md_rdev *rdev; | |
1201 | rcu_read_lock(); | |
1202 | rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev); | |
1203 | if (rdev) | |
1204 | bdevname(rdev->bdev, b); | |
1205 | else | |
1206 | strcpy(b, "???"); | |
1207 | rcu_read_unlock(); | |
1208 | } | |
1209 | ||
1210 | /* | |
1211 | * Still need barrier for READ in case that whole | |
1212 | * array is frozen. | |
1213 | */ | |
1214 | wait_read_barrier(conf, bio->bi_iter.bi_sector); | |
1215 | ||
1216 | if (!r1_bio) | |
1217 | r1_bio = alloc_r1bio(mddev, bio); | |
1218 | else | |
1219 | init_r1bio(r1_bio, mddev, bio); | |
1220 | r1_bio->sectors = max_read_sectors; | |
1221 | ||
1222 | /* | |
1223 | * make_request() can abort the operation when read-ahead is being | |
1224 | * used and no empty request is available. | |
1225 | */ | |
1226 | rdisk = read_balance(conf, r1_bio, &max_sectors); | |
1227 | ||
1228 | if (rdisk < 0) { | |
1229 | /* couldn't find anywhere to read from */ | |
1230 | if (print_msg) { | |
1231 | pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n", | |
1232 | mdname(mddev), | |
1233 | b, | |
1234 | (unsigned long long)r1_bio->sector); | |
1235 | } | |
1236 | raid_end_bio_io(r1_bio); | |
1237 | return; | |
1238 | } | |
1239 | mirror = conf->mirrors + rdisk; | |
1240 | ||
1241 | if (print_msg) | |
1242 | pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n", | |
1243 | mdname(mddev), | |
1244 | (unsigned long long)r1_bio->sector, | |
1245 | bdevname(mirror->rdev->bdev, b)); | |
1246 | ||
1247 | if (test_bit(WriteMostly, &mirror->rdev->flags) && | |
1248 | bitmap) { | |
1249 | /* | |
1250 | * Reading from a write-mostly device must take care not to | |
1251 | * over-take any writes that are 'behind' | |
1252 | */ | |
1253 | raid1_log(mddev, "wait behind writes"); | |
1254 | wait_event(bitmap->behind_wait, | |
1255 | atomic_read(&bitmap->behind_writes) == 0); | |
1256 | } | |
1257 | ||
1258 | if (max_sectors < bio_sectors(bio)) { | |
1259 | struct bio *split = bio_split(bio, max_sectors, | |
1260 | gfp, conf->bio_split); | |
1261 | bio_chain(split, bio); | |
1262 | generic_make_request(bio); | |
1263 | bio = split; | |
1264 | r1_bio->master_bio = bio; | |
1265 | r1_bio->sectors = max_sectors; | |
1266 | } | |
1267 | ||
1268 | r1_bio->read_disk = rdisk; | |
1269 | ||
1270 | read_bio = bio_clone_fast(bio, gfp, mddev->bio_set); | |
1271 | ||
1272 | r1_bio->bios[rdisk] = read_bio; | |
1273 | ||
1274 | read_bio->bi_iter.bi_sector = r1_bio->sector + | |
1275 | mirror->rdev->data_offset; | |
1276 | bio_set_dev(read_bio, mirror->rdev->bdev); | |
1277 | read_bio->bi_end_io = raid1_end_read_request; | |
1278 | bio_set_op_attrs(read_bio, op, do_sync); | |
1279 | if (test_bit(FailFast, &mirror->rdev->flags) && | |
1280 | test_bit(R1BIO_FailFast, &r1_bio->state)) | |
1281 | read_bio->bi_opf |= MD_FAILFAST; | |
1282 | read_bio->bi_private = r1_bio; | |
1283 | ||
1284 | if (mddev->gendisk) | |
1285 | trace_block_bio_remap(read_bio->bi_disk->queue, read_bio, | |
1286 | disk_devt(mddev->gendisk), r1_bio->sector); | |
1287 | ||
1288 | generic_make_request(read_bio); | |
1289 | } | |
1290 | ||
1291 | static void raid1_write_request(struct mddev *mddev, struct bio *bio, | |
1292 | int max_write_sectors) | |
1293 | { | |
1294 | struct r1conf *conf = mddev->private; | |
1295 | struct r1bio *r1_bio; | |
1296 | int i, disks; | |
1297 | struct bitmap *bitmap = mddev->bitmap; | |
1298 | unsigned long flags; | |
1299 | struct md_rdev *blocked_rdev; | |
1300 | struct blk_plug_cb *cb; | |
1301 | struct raid1_plug_cb *plug = NULL; | |
1302 | int first_clone; | |
1303 | int max_sectors; | |
1304 | ||
1305 | /* | |
1306 | * Register the new request and wait if the reconstruction | |
1307 | * thread has put up a bar for new requests. | |
1308 | * Continue immediately if no resync is active currently. | |
1309 | */ | |
1310 | ||
1311 | ||
1312 | if ((bio_end_sector(bio) > mddev->suspend_lo && | |
1313 | bio->bi_iter.bi_sector < mddev->suspend_hi) || | |
1314 | (mddev_is_clustered(mddev) && | |
1315 | md_cluster_ops->area_resyncing(mddev, WRITE, | |
1316 | bio->bi_iter.bi_sector, bio_end_sector(bio)))) { | |
1317 | ||
1318 | /* | |
1319 | * As the suspend_* range is controlled by userspace, we want | |
1320 | * an interruptible wait. | |
1321 | */ | |
1322 | DEFINE_WAIT(w); | |
1323 | for (;;) { | |
1324 | sigset_t full, old; | |
1325 | prepare_to_wait(&conf->wait_barrier, | |
1326 | &w, TASK_INTERRUPTIBLE); | |
1327 | if (bio_end_sector(bio) <= mddev->suspend_lo || | |
1328 | bio->bi_iter.bi_sector >= mddev->suspend_hi || | |
1329 | (mddev_is_clustered(mddev) && | |
1330 | !md_cluster_ops->area_resyncing(mddev, WRITE, | |
1331 | bio->bi_iter.bi_sector, | |
1332 | bio_end_sector(bio)))) | |
1333 | break; | |
1334 | sigfillset(&full); | |
1335 | sigprocmask(SIG_BLOCK, &full, &old); | |
1336 | schedule(); | |
1337 | sigprocmask(SIG_SETMASK, &old, NULL); | |
1338 | } | |
1339 | finish_wait(&conf->wait_barrier, &w); | |
1340 | } | |
1341 | wait_barrier(conf, bio->bi_iter.bi_sector); | |
1342 | ||
1343 | r1_bio = alloc_r1bio(mddev, bio); | |
1344 | r1_bio->sectors = max_write_sectors; | |
1345 | ||
1346 | if (conf->pending_count >= max_queued_requests) { | |
1347 | md_wakeup_thread(mddev->thread); | |
1348 | raid1_log(mddev, "wait queued"); | |
1349 | wait_event(conf->wait_barrier, | |
1350 | conf->pending_count < max_queued_requests); | |
1351 | } | |
1352 | /* first select target devices under rcu_lock and | |
1353 | * inc refcount on their rdev. Record them by setting | |
1354 | * bios[x] to bio | |
1355 | * If there are known/acknowledged bad blocks on any device on | |
1356 | * which we have seen a write error, we want to avoid writing those | |
1357 | * blocks. | |
1358 | * This potentially requires several writes to write around | |
1359 | * the bad blocks. Each set of writes gets it's own r1bio | |
1360 | * with a set of bios attached. | |
1361 | */ | |
1362 | ||
1363 | disks = conf->raid_disks * 2; | |
1364 | retry_write: | |
1365 | blocked_rdev = NULL; | |
1366 | rcu_read_lock(); | |
1367 | max_sectors = r1_bio->sectors; | |
1368 | for (i = 0; i < disks; i++) { | |
1369 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); | |
1370 | if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) { | |
1371 | atomic_inc(&rdev->nr_pending); | |
1372 | blocked_rdev = rdev; | |
1373 | break; | |
1374 | } | |
1375 | r1_bio->bios[i] = NULL; | |
1376 | if (!rdev || test_bit(Faulty, &rdev->flags)) { | |
1377 | if (i < conf->raid_disks) | |
1378 | set_bit(R1BIO_Degraded, &r1_bio->state); | |
1379 | continue; | |
1380 | } | |
1381 | ||
1382 | atomic_inc(&rdev->nr_pending); | |
1383 | if (test_bit(WriteErrorSeen, &rdev->flags)) { | |
1384 | sector_t first_bad; | |
1385 | int bad_sectors; | |
1386 | int is_bad; | |
1387 | ||
1388 | is_bad = is_badblock(rdev, r1_bio->sector, max_sectors, | |
1389 | &first_bad, &bad_sectors); | |
1390 | if (is_bad < 0) { | |
1391 | /* mustn't write here until the bad block is | |
1392 | * acknowledged*/ | |
1393 | set_bit(BlockedBadBlocks, &rdev->flags); | |
1394 | blocked_rdev = rdev; | |
1395 | break; | |
1396 | } | |
1397 | if (is_bad && first_bad <= r1_bio->sector) { | |
1398 | /* Cannot write here at all */ | |
1399 | bad_sectors -= (r1_bio->sector - first_bad); | |
1400 | if (bad_sectors < max_sectors) | |
1401 | /* mustn't write more than bad_sectors | |
1402 | * to other devices yet | |
1403 | */ | |
1404 | max_sectors = bad_sectors; | |
1405 | rdev_dec_pending(rdev, mddev); | |
1406 | /* We don't set R1BIO_Degraded as that | |
1407 | * only applies if the disk is | |
1408 | * missing, so it might be re-added, | |
1409 | * and we want to know to recover this | |
1410 | * chunk. | |
1411 | * In this case the device is here, | |
1412 | * and the fact that this chunk is not | |
1413 | * in-sync is recorded in the bad | |
1414 | * block log | |
1415 | */ | |
1416 | continue; | |
1417 | } | |
1418 | if (is_bad) { | |
1419 | int good_sectors = first_bad - r1_bio->sector; | |
1420 | if (good_sectors < max_sectors) | |
1421 | max_sectors = good_sectors; | |
1422 | } | |
1423 | } | |
1424 | r1_bio->bios[i] = bio; | |
1425 | } | |
1426 | rcu_read_unlock(); | |
1427 | ||
1428 | if (unlikely(blocked_rdev)) { | |
1429 | /* Wait for this device to become unblocked */ | |
1430 | int j; | |
1431 | ||
1432 | for (j = 0; j < i; j++) | |
1433 | if (r1_bio->bios[j]) | |
1434 | rdev_dec_pending(conf->mirrors[j].rdev, mddev); | |
1435 | r1_bio->state = 0; | |
1436 | allow_barrier(conf, bio->bi_iter.bi_sector); | |
1437 | raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk); | |
1438 | md_wait_for_blocked_rdev(blocked_rdev, mddev); | |
1439 | wait_barrier(conf, bio->bi_iter.bi_sector); | |
1440 | goto retry_write; | |
1441 | } | |
1442 | ||
1443 | if (max_sectors < bio_sectors(bio)) { | |
1444 | struct bio *split = bio_split(bio, max_sectors, | |
1445 | GFP_NOIO, conf->bio_split); | |
1446 | bio_chain(split, bio); | |
1447 | generic_make_request(bio); | |
1448 | bio = split; | |
1449 | r1_bio->master_bio = bio; | |
1450 | r1_bio->sectors = max_sectors; | |
1451 | } | |
1452 | ||
1453 | atomic_set(&r1_bio->remaining, 1); | |
1454 | atomic_set(&r1_bio->behind_remaining, 0); | |
1455 | ||
1456 | first_clone = 1; | |
1457 | ||
1458 | for (i = 0; i < disks; i++) { | |
1459 | struct bio *mbio = NULL; | |
1460 | if (!r1_bio->bios[i]) | |
1461 | continue; | |
1462 | ||
1463 | ||
1464 | if (first_clone) { | |
1465 | /* do behind I/O ? | |
1466 | * Not if there are too many, or cannot | |
1467 | * allocate memory, or a reader on WriteMostly | |
1468 | * is waiting for behind writes to flush */ | |
1469 | if (bitmap && | |
1470 | (atomic_read(&bitmap->behind_writes) | |
1471 | < mddev->bitmap_info.max_write_behind) && | |
1472 | !waitqueue_active(&bitmap->behind_wait)) { | |
1473 | alloc_behind_master_bio(r1_bio, bio); | |
1474 | } | |
1475 | ||
1476 | bitmap_startwrite(bitmap, r1_bio->sector, | |
1477 | r1_bio->sectors, | |
1478 | test_bit(R1BIO_BehindIO, | |
1479 | &r1_bio->state)); | |
1480 | first_clone = 0; | |
1481 | } | |
1482 | ||
1483 | if (r1_bio->behind_master_bio) | |
1484 | mbio = bio_clone_fast(r1_bio->behind_master_bio, | |
1485 | GFP_NOIO, mddev->bio_set); | |
1486 | else | |
1487 | mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set); | |
1488 | ||
1489 | if (r1_bio->behind_master_bio) { | |
1490 | if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags)) | |
1491 | atomic_inc(&r1_bio->behind_remaining); | |
1492 | } | |
1493 | ||
1494 | r1_bio->bios[i] = mbio; | |
1495 | ||
1496 | mbio->bi_iter.bi_sector = (r1_bio->sector + | |
1497 | conf->mirrors[i].rdev->data_offset); | |
1498 | bio_set_dev(mbio, conf->mirrors[i].rdev->bdev); | |
1499 | mbio->bi_end_io = raid1_end_write_request; | |
1500 | mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA)); | |
1501 | if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) && | |
1502 | !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) && | |
1503 | conf->raid_disks - mddev->degraded > 1) | |
1504 | mbio->bi_opf |= MD_FAILFAST; | |
1505 | mbio->bi_private = r1_bio; | |
1506 | ||
1507 | atomic_inc(&r1_bio->remaining); | |
1508 | ||
1509 | if (mddev->gendisk) | |
1510 | trace_block_bio_remap(mbio->bi_disk->queue, | |
1511 | mbio, disk_devt(mddev->gendisk), | |
1512 | r1_bio->sector); | |
1513 | /* flush_pending_writes() needs access to the rdev so...*/ | |
1514 | mbio->bi_disk = (void *)conf->mirrors[i].rdev; | |
1515 | ||
1516 | cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug)); | |
1517 | if (cb) | |
1518 | plug = container_of(cb, struct raid1_plug_cb, cb); | |
1519 | else | |
1520 | plug = NULL; | |
1521 | if (plug) { | |
1522 | bio_list_add(&plug->pending, mbio); | |
1523 | plug->pending_cnt++; | |
1524 | } else { | |
1525 | spin_lock_irqsave(&conf->device_lock, flags); | |
1526 | bio_list_add(&conf->pending_bio_list, mbio); | |
1527 | conf->pending_count++; | |
1528 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1529 | md_wakeup_thread(mddev->thread); | |
1530 | } | |
1531 | } | |
1532 | ||
1533 | r1_bio_write_done(r1_bio); | |
1534 | ||
1535 | /* In case raid1d snuck in to freeze_array */ | |
1536 | wake_up(&conf->wait_barrier); | |
1537 | } | |
1538 | ||
1539 | static bool raid1_make_request(struct mddev *mddev, struct bio *bio) | |
1540 | { | |
1541 | sector_t sectors; | |
1542 | ||
1543 | if (unlikely(bio->bi_opf & REQ_PREFLUSH)) { | |
1544 | md_flush_request(mddev, bio); | |
1545 | return true; | |
1546 | } | |
1547 | ||
1548 | /* | |
1549 | * There is a limit to the maximum size, but | |
1550 | * the read/write handler might find a lower limit | |
1551 | * due to bad blocks. To avoid multiple splits, | |
1552 | * we pass the maximum number of sectors down | |
1553 | * and let the lower level perform the split. | |
1554 | */ | |
1555 | sectors = align_to_barrier_unit_end( | |
1556 | bio->bi_iter.bi_sector, bio_sectors(bio)); | |
1557 | ||
1558 | if (bio_data_dir(bio) == READ) | |
1559 | raid1_read_request(mddev, bio, sectors, NULL); | |
1560 | else { | |
1561 | if (!md_write_start(mddev,bio)) | |
1562 | return false; | |
1563 | raid1_write_request(mddev, bio, sectors); | |
1564 | } | |
1565 | return true; | |
1566 | } | |
1567 | ||
1568 | static void raid1_status(struct seq_file *seq, struct mddev *mddev) | |
1569 | { | |
1570 | struct r1conf *conf = mddev->private; | |
1571 | int i; | |
1572 | ||
1573 | seq_printf(seq, " [%d/%d] [", conf->raid_disks, | |
1574 | conf->raid_disks - mddev->degraded); | |
1575 | rcu_read_lock(); | |
1576 | for (i = 0; i < conf->raid_disks; i++) { | |
1577 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); | |
1578 | seq_printf(seq, "%s", | |
1579 | rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_"); | |
1580 | } | |
1581 | rcu_read_unlock(); | |
1582 | seq_printf(seq, "]"); | |
1583 | } | |
1584 | ||
1585 | static void raid1_error(struct mddev *mddev, struct md_rdev *rdev) | |
1586 | { | |
1587 | char b[BDEVNAME_SIZE]; | |
1588 | struct r1conf *conf = mddev->private; | |
1589 | unsigned long flags; | |
1590 | ||
1591 | /* | |
1592 | * If it is not operational, then we have already marked it as dead | |
1593 | * else if it is the last working disks, ignore the error, let the | |
1594 | * next level up know. | |
1595 | * else mark the drive as failed | |
1596 | */ | |
1597 | spin_lock_irqsave(&conf->device_lock, flags); | |
1598 | if (test_bit(In_sync, &rdev->flags) | |
1599 | && (conf->raid_disks - mddev->degraded) == 1) { | |
1600 | /* | |
1601 | * Don't fail the drive, act as though we were just a | |
1602 | * normal single drive. | |
1603 | * However don't try a recovery from this drive as | |
1604 | * it is very likely to fail. | |
1605 | */ | |
1606 | conf->recovery_disabled = mddev->recovery_disabled; | |
1607 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1608 | return; | |
1609 | } | |
1610 | set_bit(Blocked, &rdev->flags); | |
1611 | if (test_and_clear_bit(In_sync, &rdev->flags)) { | |
1612 | mddev->degraded++; | |
1613 | set_bit(Faulty, &rdev->flags); | |
1614 | } else | |
1615 | set_bit(Faulty, &rdev->flags); | |
1616 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1617 | /* | |
1618 | * if recovery is running, make sure it aborts. | |
1619 | */ | |
1620 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); | |
1621 | set_mask_bits(&mddev->sb_flags, 0, | |
1622 | BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING)); | |
1623 | pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n" | |
1624 | "md/raid1:%s: Operation continuing on %d devices.\n", | |
1625 | mdname(mddev), bdevname(rdev->bdev, b), | |
1626 | mdname(mddev), conf->raid_disks - mddev->degraded); | |
1627 | } | |
1628 | ||
1629 | static void print_conf(struct r1conf *conf) | |
1630 | { | |
1631 | int i; | |
1632 | ||
1633 | pr_debug("RAID1 conf printout:\n"); | |
1634 | if (!conf) { | |
1635 | pr_debug("(!conf)\n"); | |
1636 | return; | |
1637 | } | |
1638 | pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded, | |
1639 | conf->raid_disks); | |
1640 | ||
1641 | rcu_read_lock(); | |
1642 | for (i = 0; i < conf->raid_disks; i++) { | |
1643 | char b[BDEVNAME_SIZE]; | |
1644 | struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev); | |
1645 | if (rdev) | |
1646 | pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n", | |
1647 | i, !test_bit(In_sync, &rdev->flags), | |
1648 | !test_bit(Faulty, &rdev->flags), | |
1649 | bdevname(rdev->bdev,b)); | |
1650 | } | |
1651 | rcu_read_unlock(); | |
1652 | } | |
1653 | ||
1654 | static void close_sync(struct r1conf *conf) | |
1655 | { | |
1656 | wait_all_barriers(conf); | |
1657 | allow_all_barriers(conf); | |
1658 | ||
1659 | mempool_destroy(conf->r1buf_pool); | |
1660 | conf->r1buf_pool = NULL; | |
1661 | } | |
1662 | ||
1663 | static int raid1_spare_active(struct mddev *mddev) | |
1664 | { | |
1665 | int i; | |
1666 | struct r1conf *conf = mddev->private; | |
1667 | int count = 0; | |
1668 | unsigned long flags; | |
1669 | ||
1670 | /* | |
1671 | * Find all failed disks within the RAID1 configuration | |
1672 | * and mark them readable. | |
1673 | * Called under mddev lock, so rcu protection not needed. | |
1674 | * device_lock used to avoid races with raid1_end_read_request | |
1675 | * which expects 'In_sync' flags and ->degraded to be consistent. | |
1676 | */ | |
1677 | spin_lock_irqsave(&conf->device_lock, flags); | |
1678 | for (i = 0; i < conf->raid_disks; i++) { | |
1679 | struct md_rdev *rdev = conf->mirrors[i].rdev; | |
1680 | struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev; | |
1681 | if (repl | |
1682 | && !test_bit(Candidate, &repl->flags) | |
1683 | && repl->recovery_offset == MaxSector | |
1684 | && !test_bit(Faulty, &repl->flags) | |
1685 | && !test_and_set_bit(In_sync, &repl->flags)) { | |
1686 | /* replacement has just become active */ | |
1687 | if (!rdev || | |
1688 | !test_and_clear_bit(In_sync, &rdev->flags)) | |
1689 | count++; | |
1690 | if (rdev) { | |
1691 | /* Replaced device not technically | |
1692 | * faulty, but we need to be sure | |
1693 | * it gets removed and never re-added | |
1694 | */ | |
1695 | set_bit(Faulty, &rdev->flags); | |
1696 | sysfs_notify_dirent_safe( | |
1697 | rdev->sysfs_state); | |
1698 | } | |
1699 | } | |
1700 | if (rdev | |
1701 | && rdev->recovery_offset == MaxSector | |
1702 | && !test_bit(Faulty, &rdev->flags) | |
1703 | && !test_and_set_bit(In_sync, &rdev->flags)) { | |
1704 | count++; | |
1705 | sysfs_notify_dirent_safe(rdev->sysfs_state); | |
1706 | } | |
1707 | } | |
1708 | mddev->degraded -= count; | |
1709 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
1710 | ||
1711 | print_conf(conf); | |
1712 | return count; | |
1713 | } | |
1714 | ||
1715 | static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev) | |
1716 | { | |
1717 | struct r1conf *conf = mddev->private; | |
1718 | int err = -EEXIST; | |
1719 | int mirror = 0; | |
1720 | struct raid1_info *p; | |
1721 | int first = 0; | |
1722 | int last = conf->raid_disks - 1; | |
1723 | ||
1724 | if (mddev->recovery_disabled == conf->recovery_disabled) | |
1725 | return -EBUSY; | |
1726 | ||
1727 | if (md_integrity_add_rdev(rdev, mddev)) | |
1728 | return -ENXIO; | |
1729 | ||
1730 | if (rdev->raid_disk >= 0) | |
1731 | first = last = rdev->raid_disk; | |
1732 | ||
1733 | /* | |
1734 | * find the disk ... but prefer rdev->saved_raid_disk | |
1735 | * if possible. | |
1736 | */ | |
1737 | if (rdev->saved_raid_disk >= 0 && | |
1738 | rdev->saved_raid_disk >= first && | |
1739 | conf->mirrors[rdev->saved_raid_disk].rdev == NULL) | |
1740 | first = last = rdev->saved_raid_disk; | |
1741 | ||
1742 | for (mirror = first; mirror <= last; mirror++) { | |
1743 | p = conf->mirrors+mirror; | |
1744 | if (!p->rdev) { | |
1745 | ||
1746 | if (mddev->gendisk) | |
1747 | disk_stack_limits(mddev->gendisk, rdev->bdev, | |
1748 | rdev->data_offset << 9); | |
1749 | ||
1750 | p->head_position = 0; | |
1751 | rdev->raid_disk = mirror; | |
1752 | err = 0; | |
1753 | /* As all devices are equivalent, we don't need a full recovery | |
1754 | * if this was recently any drive of the array | |
1755 | */ | |
1756 | if (rdev->saved_raid_disk < 0) | |
1757 | conf->fullsync = 1; | |
1758 | rcu_assign_pointer(p->rdev, rdev); | |
1759 | break; | |
1760 | } | |
1761 | if (test_bit(WantReplacement, &p->rdev->flags) && | |
1762 | p[conf->raid_disks].rdev == NULL) { | |
1763 | /* Add this device as a replacement */ | |
1764 | clear_bit(In_sync, &rdev->flags); | |
1765 | set_bit(Replacement, &rdev->flags); | |
1766 | rdev->raid_disk = mirror; | |
1767 | err = 0; | |
1768 | conf->fullsync = 1; | |
1769 | rcu_assign_pointer(p[conf->raid_disks].rdev, rdev); | |
1770 | break; | |
1771 | } | |
1772 | } | |
1773 | if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev))) | |
1774 | queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue); | |
1775 | print_conf(conf); | |
1776 | return err; | |
1777 | } | |
1778 | ||
1779 | static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev) | |
1780 | { | |
1781 | struct r1conf *conf = mddev->private; | |
1782 | int err = 0; | |
1783 | int number = rdev->raid_disk; | |
1784 | struct raid1_info *p = conf->mirrors + number; | |
1785 | ||
1786 | if (rdev != p->rdev) | |
1787 | p = conf->mirrors + conf->raid_disks + number; | |
1788 | ||
1789 | print_conf(conf); | |
1790 | if (rdev == p->rdev) { | |
1791 | if (test_bit(In_sync, &rdev->flags) || | |
1792 | atomic_read(&rdev->nr_pending)) { | |
1793 | err = -EBUSY; | |
1794 | goto abort; | |
1795 | } | |
1796 | /* Only remove non-faulty devices if recovery | |
1797 | * is not possible. | |
1798 | */ | |
1799 | if (!test_bit(Faulty, &rdev->flags) && | |
1800 | mddev->recovery_disabled != conf->recovery_disabled && | |
1801 | mddev->degraded < conf->raid_disks) { | |
1802 | err = -EBUSY; | |
1803 | goto abort; | |
1804 | } | |
1805 | p->rdev = NULL; | |
1806 | if (!test_bit(RemoveSynchronized, &rdev->flags)) { | |
1807 | synchronize_rcu(); | |
1808 | if (atomic_read(&rdev->nr_pending)) { | |
1809 | /* lost the race, try later */ | |
1810 | err = -EBUSY; | |
1811 | p->rdev = rdev; | |
1812 | goto abort; | |
1813 | } | |
1814 | } | |
1815 | if (conf->mirrors[conf->raid_disks + number].rdev) { | |
1816 | /* We just removed a device that is being replaced. | |
1817 | * Move down the replacement. We drain all IO before | |
1818 | * doing this to avoid confusion. | |
1819 | */ | |
1820 | struct md_rdev *repl = | |
1821 | conf->mirrors[conf->raid_disks + number].rdev; | |
1822 | freeze_array(conf, 0); | |
1823 | clear_bit(Replacement, &repl->flags); | |
1824 | p->rdev = repl; | |
1825 | conf->mirrors[conf->raid_disks + number].rdev = NULL; | |
1826 | unfreeze_array(conf); | |
1827 | } | |
1828 | ||
1829 | clear_bit(WantReplacement, &rdev->flags); | |
1830 | err = md_integrity_register(mddev); | |
1831 | } | |
1832 | abort: | |
1833 | ||
1834 | print_conf(conf); | |
1835 | return err; | |
1836 | } | |
1837 | ||
1838 | static void end_sync_read(struct bio *bio) | |
1839 | { | |
1840 | struct r1bio *r1_bio = get_resync_r1bio(bio); | |
1841 | ||
1842 | update_head_pos(r1_bio->read_disk, r1_bio); | |
1843 | ||
1844 | /* | |
1845 | * we have read a block, now it needs to be re-written, | |
1846 | * or re-read if the read failed. | |
1847 | * We don't do much here, just schedule handling by raid1d | |
1848 | */ | |
1849 | if (!bio->bi_status) | |
1850 | set_bit(R1BIO_Uptodate, &r1_bio->state); | |
1851 | ||
1852 | if (atomic_dec_and_test(&r1_bio->remaining)) | |
1853 | reschedule_retry(r1_bio); | |
1854 | } | |
1855 | ||
1856 | static void end_sync_write(struct bio *bio) | |
1857 | { | |
1858 | int uptodate = !bio->bi_status; | |
1859 | struct r1bio *r1_bio = get_resync_r1bio(bio); | |
1860 | struct mddev *mddev = r1_bio->mddev; | |
1861 | struct r1conf *conf = mddev->private; | |
1862 | sector_t first_bad; | |
1863 | int bad_sectors; | |
1864 | struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev; | |
1865 | ||
1866 | if (!uptodate) { | |
1867 | sector_t sync_blocks = 0; | |
1868 | sector_t s = r1_bio->sector; | |
1869 | long sectors_to_go = r1_bio->sectors; | |
1870 | /* make sure these bits doesn't get cleared. */ | |
1871 | do { | |
1872 | bitmap_end_sync(mddev->bitmap, s, | |
1873 | &sync_blocks, 1); | |
1874 | s += sync_blocks; | |
1875 | sectors_to_go -= sync_blocks; | |
1876 | } while (sectors_to_go > 0); | |
1877 | set_bit(WriteErrorSeen, &rdev->flags); | |
1878 | if (!test_and_set_bit(WantReplacement, &rdev->flags)) | |
1879 | set_bit(MD_RECOVERY_NEEDED, & | |
1880 | mddev->recovery); | |
1881 | set_bit(R1BIO_WriteError, &r1_bio->state); | |
1882 | } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors, | |
1883 | &first_bad, &bad_sectors) && | |
1884 | !is_badblock(conf->mirrors[r1_bio->read_disk].rdev, | |
1885 | r1_bio->sector, | |
1886 | r1_bio->sectors, | |
1887 | &first_bad, &bad_sectors) | |
1888 | ) | |
1889 | set_bit(R1BIO_MadeGood, &r1_bio->state); | |
1890 | ||
1891 | if (atomic_dec_and_test(&r1_bio->remaining)) { | |
1892 | int s = r1_bio->sectors; | |
1893 | if (test_bit(R1BIO_MadeGood, &r1_bio->state) || | |
1894 | test_bit(R1BIO_WriteError, &r1_bio->state)) | |
1895 | reschedule_retry(r1_bio); | |
1896 | else { | |
1897 | put_buf(r1_bio); | |
1898 | md_done_sync(mddev, s, uptodate); | |
1899 | } | |
1900 | } | |
1901 | } | |
1902 | ||
1903 | static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector, | |
1904 | int sectors, struct page *page, int rw) | |
1905 | { | |
1906 | if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false)) | |
1907 | /* success */ | |
1908 | return 1; | |
1909 | if (rw == WRITE) { | |
1910 | set_bit(WriteErrorSeen, &rdev->flags); | |
1911 | if (!test_and_set_bit(WantReplacement, | |
1912 | &rdev->flags)) | |
1913 | set_bit(MD_RECOVERY_NEEDED, & | |
1914 | rdev->mddev->recovery); | |
1915 | } | |
1916 | /* need to record an error - either for the block or the device */ | |
1917 | if (!rdev_set_badblocks(rdev, sector, sectors, 0)) | |
1918 | md_error(rdev->mddev, rdev); | |
1919 | return 0; | |
1920 | } | |
1921 | ||
1922 | static int fix_sync_read_error(struct r1bio *r1_bio) | |
1923 | { | |
1924 | /* Try some synchronous reads of other devices to get | |
1925 | * good data, much like with normal read errors. Only | |
1926 | * read into the pages we already have so we don't | |
1927 | * need to re-issue the read request. | |
1928 | * We don't need to freeze the array, because being in an | |
1929 | * active sync request, there is no normal IO, and | |
1930 | * no overlapping syncs. | |
1931 | * We don't need to check is_badblock() again as we | |
1932 | * made sure that anything with a bad block in range | |
1933 | * will have bi_end_io clear. | |
1934 | */ | |
1935 | struct mddev *mddev = r1_bio->mddev; | |
1936 | struct r1conf *conf = mddev->private; | |
1937 | struct bio *bio = r1_bio->bios[r1_bio->read_disk]; | |
1938 | struct page **pages = get_resync_pages(bio)->pages; | |
1939 | sector_t sect = r1_bio->sector; | |
1940 | int sectors = r1_bio->sectors; | |
1941 | int idx = 0; | |
1942 | struct md_rdev *rdev; | |
1943 | ||
1944 | rdev = conf->mirrors[r1_bio->read_disk].rdev; | |
1945 | if (test_bit(FailFast, &rdev->flags)) { | |
1946 | /* Don't try recovering from here - just fail it | |
1947 | * ... unless it is the last working device of course */ | |
1948 | md_error(mddev, rdev); | |
1949 | if (test_bit(Faulty, &rdev->flags)) | |
1950 | /* Don't try to read from here, but make sure | |
1951 | * put_buf does it's thing | |
1952 | */ | |
1953 | bio->bi_end_io = end_sync_write; | |
1954 | } | |
1955 | ||
1956 | while(sectors) { | |
1957 | int s = sectors; | |
1958 | int d = r1_bio->read_disk; | |
1959 | int success = 0; | |
1960 | int start; | |
1961 | ||
1962 | if (s > (PAGE_SIZE>>9)) | |
1963 | s = PAGE_SIZE >> 9; | |
1964 | do { | |
1965 | if (r1_bio->bios[d]->bi_end_io == end_sync_read) { | |
1966 | /* No rcu protection needed here devices | |
1967 | * can only be removed when no resync is | |
1968 | * active, and resync is currently active | |
1969 | */ | |
1970 | rdev = conf->mirrors[d].rdev; | |
1971 | if (sync_page_io(rdev, sect, s<<9, | |
1972 | pages[idx], | |
1973 | REQ_OP_READ, 0, false)) { | |
1974 | success = 1; | |
1975 | break; | |
1976 | } | |
1977 | } | |
1978 | d++; | |
1979 | if (d == conf->raid_disks * 2) | |
1980 | d = 0; | |
1981 | } while (!success && d != r1_bio->read_disk); | |
1982 | ||
1983 | if (!success) { | |
1984 | char b[BDEVNAME_SIZE]; | |
1985 | int abort = 0; | |
1986 | /* Cannot read from anywhere, this block is lost. | |
1987 | * Record a bad block on each device. If that doesn't | |
1988 | * work just disable and interrupt the recovery. | |
1989 | * Don't fail devices as that won't really help. | |
1990 | */ | |
1991 | pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n", | |
1992 | mdname(mddev), bio_devname(bio, b), | |
1993 | (unsigned long long)r1_bio->sector); | |
1994 | for (d = 0; d < conf->raid_disks * 2; d++) { | |
1995 | rdev = conf->mirrors[d].rdev; | |
1996 | if (!rdev || test_bit(Faulty, &rdev->flags)) | |
1997 | continue; | |
1998 | if (!rdev_set_badblocks(rdev, sect, s, 0)) | |
1999 | abort = 1; | |
2000 | } | |
2001 | if (abort) { | |
2002 | conf->recovery_disabled = | |
2003 | mddev->recovery_disabled; | |
2004 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); | |
2005 | md_done_sync(mddev, r1_bio->sectors, 0); | |
2006 | put_buf(r1_bio); | |
2007 | return 0; | |
2008 | } | |
2009 | /* Try next page */ | |
2010 | sectors -= s; | |
2011 | sect += s; | |
2012 | idx++; | |
2013 | continue; | |
2014 | } | |
2015 | ||
2016 | start = d; | |
2017 | /* write it back and re-read */ | |
2018 | while (d != r1_bio->read_disk) { | |
2019 | if (d == 0) | |
2020 | d = conf->raid_disks * 2; | |
2021 | d--; | |
2022 | if (r1_bio->bios[d]->bi_end_io != end_sync_read) | |
2023 | continue; | |
2024 | rdev = conf->mirrors[d].rdev; | |
2025 | if (r1_sync_page_io(rdev, sect, s, | |
2026 | pages[idx], | |
2027 | WRITE) == 0) { | |
2028 | r1_bio->bios[d]->bi_end_io = NULL; | |
2029 | rdev_dec_pending(rdev, mddev); | |
2030 | } | |
2031 | } | |
2032 | d = start; | |
2033 | while (d != r1_bio->read_disk) { | |
2034 | if (d == 0) | |
2035 | d = conf->raid_disks * 2; | |
2036 | d--; | |
2037 | if (r1_bio->bios[d]->bi_end_io != end_sync_read) | |
2038 | continue; | |
2039 | rdev = conf->mirrors[d].rdev; | |
2040 | if (r1_sync_page_io(rdev, sect, s, | |
2041 | pages[idx], | |
2042 | READ) != 0) | |
2043 | atomic_add(s, &rdev->corrected_errors); | |
2044 | } | |
2045 | sectors -= s; | |
2046 | sect += s; | |
2047 | idx ++; | |
2048 | } | |
2049 | set_bit(R1BIO_Uptodate, &r1_bio->state); | |
2050 | bio->bi_status = 0; | |
2051 | return 1; | |
2052 | } | |
2053 | ||
2054 | static void process_checks(struct r1bio *r1_bio) | |
2055 | { | |
2056 | /* We have read all readable devices. If we haven't | |
2057 | * got the block, then there is no hope left. | |
2058 | * If we have, then we want to do a comparison | |
2059 | * and skip the write if everything is the same. | |
2060 | * If any blocks failed to read, then we need to | |
2061 | * attempt an over-write | |
2062 | */ | |
2063 | struct mddev *mddev = r1_bio->mddev; | |
2064 | struct r1conf *conf = mddev->private; | |
2065 | int primary; | |
2066 | int i; | |
2067 | int vcnt; | |
2068 | ||
2069 | /* Fix variable parts of all bios */ | |
2070 | vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9); | |
2071 | for (i = 0; i < conf->raid_disks * 2; i++) { | |
2072 | blk_status_t status; | |
2073 | struct bio *b = r1_bio->bios[i]; | |
2074 | struct resync_pages *rp = get_resync_pages(b); | |
2075 | if (b->bi_end_io != end_sync_read) | |
2076 | continue; | |
2077 | /* fixup the bio for reuse, but preserve errno */ | |
2078 | status = b->bi_status; | |
2079 | bio_reset(b); | |
2080 | b->bi_status = status; | |
2081 | b->bi_iter.bi_sector = r1_bio->sector + | |
2082 | conf->mirrors[i].rdev->data_offset; | |
2083 | bio_set_dev(b, conf->mirrors[i].rdev->bdev); | |
2084 | b->bi_end_io = end_sync_read; | |
2085 | rp->raid_bio = r1_bio; | |
2086 | b->bi_private = rp; | |
2087 | ||
2088 | /* initialize bvec table again */ | |
2089 | md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9); | |
2090 | } | |
2091 | for (primary = 0; primary < conf->raid_disks * 2; primary++) | |
2092 | if (r1_bio->bios[primary]->bi_end_io == end_sync_read && | |
2093 | !r1_bio->bios[primary]->bi_status) { | |
2094 | r1_bio->bios[primary]->bi_end_io = NULL; | |
2095 | rdev_dec_pending(conf->mirrors[primary].rdev, mddev); | |
2096 | break; | |
2097 | } | |
2098 | r1_bio->read_disk = primary; | |
2099 | for (i = 0; i < conf->raid_disks * 2; i++) { | |
2100 | int j; | |
2101 | struct bio *pbio = r1_bio->bios[primary]; | |
2102 | struct bio *sbio = r1_bio->bios[i]; | |
2103 | blk_status_t status = sbio->bi_status; | |
2104 | struct page **ppages = get_resync_pages(pbio)->pages; | |
2105 | struct page **spages = get_resync_pages(sbio)->pages; | |
2106 | struct bio_vec *bi; | |
2107 | int page_len[RESYNC_PAGES] = { 0 }; | |
2108 | ||
2109 | if (sbio->bi_end_io != end_sync_read) | |
2110 | continue; | |
2111 | /* Now we can 'fixup' the error value */ | |
2112 | sbio->bi_status = 0; | |
2113 | ||
2114 | bio_for_each_segment_all(bi, sbio, j) | |
2115 | page_len[j] = bi->bv_len; | |
2116 | ||
2117 | if (!status) { | |
2118 | for (j = vcnt; j-- ; ) { | |
2119 | if (memcmp(page_address(ppages[j]), | |
2120 | page_address(spages[j]), | |
2121 | page_len[j])) | |
2122 | break; | |
2123 | } | |
2124 | } else | |
2125 | j = 0; | |
2126 | if (j >= 0) | |
2127 | atomic64_add(r1_bio->sectors, &mddev->resync_mismatches); | |
2128 | if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery) | |
2129 | && !status)) { | |
2130 | /* No need to write to this device. */ | |
2131 | sbio->bi_end_io = NULL; | |
2132 | rdev_dec_pending(conf->mirrors[i].rdev, mddev); | |
2133 | continue; | |
2134 | } | |
2135 | ||
2136 | bio_copy_data(sbio, pbio); | |
2137 | } | |
2138 | } | |
2139 | ||
2140 | static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio) | |
2141 | { | |
2142 | struct r1conf *conf = mddev->private; | |
2143 | int i; | |
2144 | int disks = conf->raid_disks * 2; | |
2145 | struct bio *wbio; | |
2146 | ||
2147 | if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) | |
2148 | /* ouch - failed to read all of that. */ | |
2149 | if (!fix_sync_read_error(r1_bio)) | |
2150 | return; | |
2151 | ||
2152 | if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) | |
2153 | process_checks(r1_bio); | |
2154 | ||
2155 | /* | |
2156 | * schedule writes | |
2157 | */ | |
2158 | atomic_set(&r1_bio->remaining, 1); | |
2159 | for (i = 0; i < disks ; i++) { | |
2160 | wbio = r1_bio->bios[i]; | |
2161 | if (wbio->bi_end_io == NULL || | |
2162 | (wbio->bi_end_io == end_sync_read && | |
2163 | (i == r1_bio->read_disk || | |
2164 | !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)))) | |
2165 | continue; | |
2166 | if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) | |
2167 | continue; | |
2168 | ||
2169 | bio_set_op_attrs(wbio, REQ_OP_WRITE, 0); | |
2170 | if (test_bit(FailFast, &conf->mirrors[i].rdev->flags)) | |
2171 | wbio->bi_opf |= MD_FAILFAST; | |
2172 | ||
2173 | wbio->bi_end_io = end_sync_write; | |
2174 | atomic_inc(&r1_bio->remaining); | |
2175 | md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio)); | |
2176 | ||
2177 | generic_make_request(wbio); | |
2178 | } | |
2179 | ||
2180 | if (atomic_dec_and_test(&r1_bio->remaining)) { | |
2181 | /* if we're here, all write(s) have completed, so clean up */ | |
2182 | int s = r1_bio->sectors; | |
2183 | if (test_bit(R1BIO_MadeGood, &r1_bio->state) || | |
2184 | test_bit(R1BIO_WriteError, &r1_bio->state)) | |
2185 | reschedule_retry(r1_bio); | |
2186 | else { | |
2187 | put_buf(r1_bio); | |
2188 | md_done_sync(mddev, s, 1); | |
2189 | } | |
2190 | } | |
2191 | } | |
2192 | ||
2193 | /* | |
2194 | * This is a kernel thread which: | |
2195 | * | |
2196 | * 1. Retries failed read operations on working mirrors. | |
2197 | * 2. Updates the raid superblock when problems encounter. | |
2198 | * 3. Performs writes following reads for array synchronising. | |
2199 | */ | |
2200 | ||
2201 | static void fix_read_error(struct r1conf *conf, int read_disk, | |
2202 | sector_t sect, int sectors) | |
2203 | { | |
2204 | struct mddev *mddev = conf->mddev; | |
2205 | while(sectors) { | |
2206 | int s = sectors; | |
2207 | int d = read_disk; | |
2208 | int success = 0; | |
2209 | int start; | |
2210 | struct md_rdev *rdev; | |
2211 | ||
2212 | if (s > (PAGE_SIZE>>9)) | |
2213 | s = PAGE_SIZE >> 9; | |
2214 | ||
2215 | do { | |
2216 | sector_t first_bad; | |
2217 | int bad_sectors; | |
2218 | ||
2219 | rcu_read_lock(); | |
2220 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
2221 | if (rdev && | |
2222 | (test_bit(In_sync, &rdev->flags) || | |
2223 | (!test_bit(Faulty, &rdev->flags) && | |
2224 | rdev->recovery_offset >= sect + s)) && | |
2225 | is_badblock(rdev, sect, s, | |
2226 | &first_bad, &bad_sectors) == 0) { | |
2227 | atomic_inc(&rdev->nr_pending); | |
2228 | rcu_read_unlock(); | |
2229 | if (sync_page_io(rdev, sect, s<<9, | |
2230 | conf->tmppage, REQ_OP_READ, 0, false)) | |
2231 | success = 1; | |
2232 | rdev_dec_pending(rdev, mddev); | |
2233 | if (success) | |
2234 | break; | |
2235 | } else | |
2236 | rcu_read_unlock(); | |
2237 | d++; | |
2238 | if (d == conf->raid_disks * 2) | |
2239 | d = 0; | |
2240 | } while (!success && d != read_disk); | |
2241 | ||
2242 | if (!success) { | |
2243 | /* Cannot read from anywhere - mark it bad */ | |
2244 | struct md_rdev *rdev = conf->mirrors[read_disk].rdev; | |
2245 | if (!rdev_set_badblocks(rdev, sect, s, 0)) | |
2246 | md_error(mddev, rdev); | |
2247 | break; | |
2248 | } | |
2249 | /* write it back and re-read */ | |
2250 | start = d; | |
2251 | while (d != read_disk) { | |
2252 | if (d==0) | |
2253 | d = conf->raid_disks * 2; | |
2254 | d--; | |
2255 | rcu_read_lock(); | |
2256 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
2257 | if (rdev && | |
2258 | !test_bit(Faulty, &rdev->flags)) { | |
2259 | atomic_inc(&rdev->nr_pending); | |
2260 | rcu_read_unlock(); | |
2261 | r1_sync_page_io(rdev, sect, s, | |
2262 | conf->tmppage, WRITE); | |
2263 | rdev_dec_pending(rdev, mddev); | |
2264 | } else | |
2265 | rcu_read_unlock(); | |
2266 | } | |
2267 | d = start; | |
2268 | while (d != read_disk) { | |
2269 | char b[BDEVNAME_SIZE]; | |
2270 | if (d==0) | |
2271 | d = conf->raid_disks * 2; | |
2272 | d--; | |
2273 | rcu_read_lock(); | |
2274 | rdev = rcu_dereference(conf->mirrors[d].rdev); | |
2275 | if (rdev && | |
2276 | !test_bit(Faulty, &rdev->flags)) { | |
2277 | atomic_inc(&rdev->nr_pending); | |
2278 | rcu_read_unlock(); | |
2279 | if (r1_sync_page_io(rdev, sect, s, | |
2280 | conf->tmppage, READ)) { | |
2281 | atomic_add(s, &rdev->corrected_errors); | |
2282 | pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n", | |
2283 | mdname(mddev), s, | |
2284 | (unsigned long long)(sect + | |
2285 | rdev->data_offset), | |
2286 | bdevname(rdev->bdev, b)); | |
2287 | } | |
2288 | rdev_dec_pending(rdev, mddev); | |
2289 | } else | |
2290 | rcu_read_unlock(); | |
2291 | } | |
2292 | sectors -= s; | |
2293 | sect += s; | |
2294 | } | |
2295 | } | |
2296 | ||
2297 | static int narrow_write_error(struct r1bio *r1_bio, int i) | |
2298 | { | |
2299 | struct mddev *mddev = r1_bio->mddev; | |
2300 | struct r1conf *conf = mddev->private; | |
2301 | struct md_rdev *rdev = conf->mirrors[i].rdev; | |
2302 | ||
2303 | /* bio has the data to be written to device 'i' where | |
2304 | * we just recently had a write error. | |
2305 | * We repeatedly clone the bio and trim down to one block, | |
2306 | * then try the write. Where the write fails we record | |
2307 | * a bad block. | |
2308 | * It is conceivable that the bio doesn't exactly align with | |
2309 | * blocks. We must handle this somehow. | |
2310 | * | |
2311 | * We currently own a reference on the rdev. | |
2312 | */ | |
2313 | ||
2314 | int block_sectors; | |
2315 | sector_t sector; | |
2316 | int sectors; | |
2317 | int sect_to_write = r1_bio->sectors; | |
2318 | int ok = 1; | |
2319 | ||
2320 | if (rdev->badblocks.shift < 0) | |
2321 | return 0; | |
2322 | ||
2323 | block_sectors = roundup(1 << rdev->badblocks.shift, | |
2324 | bdev_logical_block_size(rdev->bdev) >> 9); | |
2325 | sector = r1_bio->sector; | |
2326 | sectors = ((sector + block_sectors) | |
2327 | & ~(sector_t)(block_sectors - 1)) | |
2328 | - sector; | |
2329 | ||
2330 | while (sect_to_write) { | |
2331 | struct bio *wbio; | |
2332 | if (sectors > sect_to_write) | |
2333 | sectors = sect_to_write; | |
2334 | /* Write at 'sector' for 'sectors'*/ | |
2335 | ||
2336 | if (test_bit(R1BIO_BehindIO, &r1_bio->state)) { | |
2337 | wbio = bio_clone_fast(r1_bio->behind_master_bio, | |
2338 | GFP_NOIO, | |
2339 | mddev->bio_set); | |
2340 | } else { | |
2341 | wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO, | |
2342 | mddev->bio_set); | |
2343 | } | |
2344 | ||
2345 | bio_set_op_attrs(wbio, REQ_OP_WRITE, 0); | |
2346 | wbio->bi_iter.bi_sector = r1_bio->sector; | |
2347 | wbio->bi_iter.bi_size = r1_bio->sectors << 9; | |
2348 | ||
2349 | bio_trim(wbio, sector - r1_bio->sector, sectors); | |
2350 | wbio->bi_iter.bi_sector += rdev->data_offset; | |
2351 | bio_set_dev(wbio, rdev->bdev); | |
2352 | ||
2353 | if (submit_bio_wait(wbio) < 0) | |
2354 | /* failure! */ | |
2355 | ok = rdev_set_badblocks(rdev, sector, | |
2356 | sectors, 0) | |
2357 | && ok; | |
2358 | ||
2359 | bio_put(wbio); | |
2360 | sect_to_write -= sectors; | |
2361 | sector += sectors; | |
2362 | sectors = block_sectors; | |
2363 | } | |
2364 | return ok; | |
2365 | } | |
2366 | ||
2367 | static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio) | |
2368 | { | |
2369 | int m; | |
2370 | int s = r1_bio->sectors; | |
2371 | for (m = 0; m < conf->raid_disks * 2 ; m++) { | |
2372 | struct md_rdev *rdev = conf->mirrors[m].rdev; | |
2373 | struct bio *bio = r1_bio->bios[m]; | |
2374 | if (bio->bi_end_io == NULL) | |
2375 | continue; | |
2376 | if (!bio->bi_status && | |
2377 | test_bit(R1BIO_MadeGood, &r1_bio->state)) { | |
2378 | rdev_clear_badblocks(rdev, r1_bio->sector, s, 0); | |
2379 | } | |
2380 | if (bio->bi_status && | |
2381 | test_bit(R1BIO_WriteError, &r1_bio->state)) { | |
2382 | if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0)) | |
2383 | md_error(conf->mddev, rdev); | |
2384 | } | |
2385 | } | |
2386 | put_buf(r1_bio); | |
2387 | md_done_sync(conf->mddev, s, 1); | |
2388 | } | |
2389 | ||
2390 | static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio) | |
2391 | { | |
2392 | int m, idx; | |
2393 | bool fail = false; | |
2394 | ||
2395 | for (m = 0; m < conf->raid_disks * 2 ; m++) | |
2396 | if (r1_bio->bios[m] == IO_MADE_GOOD) { | |
2397 | struct md_rdev *rdev = conf->mirrors[m].rdev; | |
2398 | rdev_clear_badblocks(rdev, | |
2399 | r1_bio->sector, | |
2400 | r1_bio->sectors, 0); | |
2401 | rdev_dec_pending(rdev, conf->mddev); | |
2402 | } else if (r1_bio->bios[m] != NULL) { | |
2403 | /* This drive got a write error. We need to | |
2404 | * narrow down and record precise write | |
2405 | * errors. | |
2406 | */ | |
2407 | fail = true; | |
2408 | if (!narrow_write_error(r1_bio, m)) { | |
2409 | md_error(conf->mddev, | |
2410 | conf->mirrors[m].rdev); | |
2411 | /* an I/O failed, we can't clear the bitmap */ | |
2412 | set_bit(R1BIO_Degraded, &r1_bio->state); | |
2413 | } | |
2414 | rdev_dec_pending(conf->mirrors[m].rdev, | |
2415 | conf->mddev); | |
2416 | } | |
2417 | if (fail) { | |
2418 | spin_lock_irq(&conf->device_lock); | |
2419 | list_add(&r1_bio->retry_list, &conf->bio_end_io_list); | |
2420 | idx = sector_to_idx(r1_bio->sector); | |
2421 | atomic_inc(&conf->nr_queued[idx]); | |
2422 | spin_unlock_irq(&conf->device_lock); | |
2423 | /* | |
2424 | * In case freeze_array() is waiting for condition | |
2425 | * get_unqueued_pending() == extra to be true. | |
2426 | */ | |
2427 | wake_up(&conf->wait_barrier); | |
2428 | md_wakeup_thread(conf->mddev->thread); | |
2429 | } else { | |
2430 | if (test_bit(R1BIO_WriteError, &r1_bio->state)) | |
2431 | close_write(r1_bio); | |
2432 | raid_end_bio_io(r1_bio); | |
2433 | } | |
2434 | } | |
2435 | ||
2436 | static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio) | |
2437 | { | |
2438 | struct mddev *mddev = conf->mddev; | |
2439 | struct bio *bio; | |
2440 | struct md_rdev *rdev; | |
2441 | sector_t bio_sector; | |
2442 | ||
2443 | clear_bit(R1BIO_ReadError, &r1_bio->state); | |
2444 | /* we got a read error. Maybe the drive is bad. Maybe just | |
2445 | * the block and we can fix it. | |
2446 | * We freeze all other IO, and try reading the block from | |
2447 | * other devices. When we find one, we re-write | |
2448 | * and check it that fixes the read error. | |
2449 | * This is all done synchronously while the array is | |
2450 | * frozen | |
2451 | */ | |
2452 | ||
2453 | bio = r1_bio->bios[r1_bio->read_disk]; | |
2454 | bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector; | |
2455 | bio_put(bio); | |
2456 | r1_bio->bios[r1_bio->read_disk] = NULL; | |
2457 | ||
2458 | rdev = conf->mirrors[r1_bio->read_disk].rdev; | |
2459 | if (mddev->ro == 0 | |
2460 | && !test_bit(FailFast, &rdev->flags)) { | |
2461 | freeze_array(conf, 1); | |
2462 | fix_read_error(conf, r1_bio->read_disk, | |
2463 | r1_bio->sector, r1_bio->sectors); | |
2464 | unfreeze_array(conf); | |
2465 | } else { | |
2466 | r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED; | |
2467 | } | |
2468 | ||
2469 | rdev_dec_pending(rdev, conf->mddev); | |
2470 | allow_barrier(conf, r1_bio->sector); | |
2471 | bio = r1_bio->master_bio; | |
2472 | ||
2473 | /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */ | |
2474 | r1_bio->state = 0; | |
2475 | raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio); | |
2476 | } | |
2477 | ||
2478 | static void raid1d(struct md_thread *thread) | |
2479 | { | |
2480 | struct mddev *mddev = thread->mddev; | |
2481 | struct r1bio *r1_bio; | |
2482 | unsigned long flags; | |
2483 | struct r1conf *conf = mddev->private; | |
2484 | struct list_head *head = &conf->retry_list; | |
2485 | struct blk_plug plug; | |
2486 | int idx; | |
2487 | ||
2488 | md_check_recovery(mddev); | |
2489 | ||
2490 | if (!list_empty_careful(&conf->bio_end_io_list) && | |
2491 | !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) { | |
2492 | LIST_HEAD(tmp); | |
2493 | spin_lock_irqsave(&conf->device_lock, flags); | |
2494 | if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) | |
2495 | list_splice_init(&conf->bio_end_io_list, &tmp); | |
2496 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
2497 | while (!list_empty(&tmp)) { | |
2498 | r1_bio = list_first_entry(&tmp, struct r1bio, | |
2499 | retry_list); | |
2500 | list_del(&r1_bio->retry_list); | |
2501 | idx = sector_to_idx(r1_bio->sector); | |
2502 | atomic_dec(&conf->nr_queued[idx]); | |
2503 | if (mddev->degraded) | |
2504 | set_bit(R1BIO_Degraded, &r1_bio->state); | |
2505 | if (test_bit(R1BIO_WriteError, &r1_bio->state)) | |
2506 | close_write(r1_bio); | |
2507 | raid_end_bio_io(r1_bio); | |
2508 | } | |
2509 | } | |
2510 | ||
2511 | blk_start_plug(&plug); | |
2512 | for (;;) { | |
2513 | ||
2514 | flush_pending_writes(conf); | |
2515 | ||
2516 | spin_lock_irqsave(&conf->device_lock, flags); | |
2517 | if (list_empty(head)) { | |
2518 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
2519 | break; | |
2520 | } | |
2521 | r1_bio = list_entry(head->prev, struct r1bio, retry_list); | |
2522 | list_del(head->prev); | |
2523 | idx = sector_to_idx(r1_bio->sector); | |
2524 | atomic_dec(&conf->nr_queued[idx]); | |
2525 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
2526 | ||
2527 | mddev = r1_bio->mddev; | |
2528 | conf = mddev->private; | |
2529 | if (test_bit(R1BIO_IsSync, &r1_bio->state)) { | |
2530 | if (test_bit(R1BIO_MadeGood, &r1_bio->state) || | |
2531 | test_bit(R1BIO_WriteError, &r1_bio->state)) | |
2532 | handle_sync_write_finished(conf, r1_bio); | |
2533 | else | |
2534 | sync_request_write(mddev, r1_bio); | |
2535 | } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) || | |
2536 | test_bit(R1BIO_WriteError, &r1_bio->state)) | |
2537 | handle_write_finished(conf, r1_bio); | |
2538 | else if (test_bit(R1BIO_ReadError, &r1_bio->state)) | |
2539 | handle_read_error(conf, r1_bio); | |
2540 | else | |
2541 | WARN_ON_ONCE(1); | |
2542 | ||
2543 | cond_resched(); | |
2544 | if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING)) | |
2545 | md_check_recovery(mddev); | |
2546 | } | |
2547 | blk_finish_plug(&plug); | |
2548 | } | |
2549 | ||
2550 | static int init_resync(struct r1conf *conf) | |
2551 | { | |
2552 | int buffs; | |
2553 | ||
2554 | buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE; | |
2555 | BUG_ON(conf->r1buf_pool); | |
2556 | conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free, | |
2557 | conf->poolinfo); | |
2558 | if (!conf->r1buf_pool) | |
2559 | return -ENOMEM; | |
2560 | return 0; | |
2561 | } | |
2562 | ||
2563 | /* | |
2564 | * perform a "sync" on one "block" | |
2565 | * | |
2566 | * We need to make sure that no normal I/O request - particularly write | |
2567 | * requests - conflict with active sync requests. | |
2568 | * | |
2569 | * This is achieved by tracking pending requests and a 'barrier' concept | |
2570 | * that can be installed to exclude normal IO requests. | |
2571 | */ | |
2572 | ||
2573 | static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr, | |
2574 | int *skipped) | |
2575 | { | |
2576 | struct r1conf *conf = mddev->private; | |
2577 | struct r1bio *r1_bio; | |
2578 | struct bio *bio; | |
2579 | sector_t max_sector, nr_sectors; | |
2580 | int disk = -1; | |
2581 | int i; | |
2582 | int wonly = -1; | |
2583 | int write_targets = 0, read_targets = 0; | |
2584 | sector_t sync_blocks; | |
2585 | int still_degraded = 0; | |
2586 | int good_sectors = RESYNC_SECTORS; | |
2587 | int min_bad = 0; /* number of sectors that are bad in all devices */ | |
2588 | int idx = sector_to_idx(sector_nr); | |
2589 | int page_idx = 0; | |
2590 | ||
2591 | if (!conf->r1buf_pool) | |
2592 | if (init_resync(conf)) | |
2593 | return 0; | |
2594 | ||
2595 | max_sector = mddev->dev_sectors; | |
2596 | if (sector_nr >= max_sector) { | |
2597 | /* If we aborted, we need to abort the | |
2598 | * sync on the 'current' bitmap chunk (there will | |
2599 | * only be one in raid1 resync. | |
2600 | * We can find the current addess in mddev->curr_resync | |
2601 | */ | |
2602 | if (mddev->curr_resync < max_sector) /* aborted */ | |
2603 | bitmap_end_sync(mddev->bitmap, mddev->curr_resync, | |
2604 | &sync_blocks, 1); | |
2605 | else /* completed sync */ | |
2606 | conf->fullsync = 0; | |
2607 | ||
2608 | bitmap_close_sync(mddev->bitmap); | |
2609 | close_sync(conf); | |
2610 | ||
2611 | if (mddev_is_clustered(mddev)) { | |
2612 | conf->cluster_sync_low = 0; | |
2613 | conf->cluster_sync_high = 0; | |
2614 | } | |
2615 | return 0; | |
2616 | } | |
2617 | ||
2618 | if (mddev->bitmap == NULL && | |
2619 | mddev->recovery_cp == MaxSector && | |
2620 | !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) && | |
2621 | conf->fullsync == 0) { | |
2622 | *skipped = 1; | |
2623 | return max_sector - sector_nr; | |
2624 | } | |
2625 | /* before building a request, check if we can skip these blocks.. | |
2626 | * This call the bitmap_start_sync doesn't actually record anything | |
2627 | */ | |
2628 | if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) && | |
2629 | !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { | |
2630 | /* We can skip this block, and probably several more */ | |
2631 | *skipped = 1; | |
2632 | return sync_blocks; | |
2633 | } | |
2634 | ||
2635 | /* | |
2636 | * If there is non-resync activity waiting for a turn, then let it | |
2637 | * though before starting on this new sync request. | |
2638 | */ | |
2639 | if (atomic_read(&conf->nr_waiting[idx])) | |
2640 | schedule_timeout_uninterruptible(1); | |
2641 | ||
2642 | /* we are incrementing sector_nr below. To be safe, we check against | |
2643 | * sector_nr + two times RESYNC_SECTORS | |
2644 | */ | |
2645 | ||
2646 | bitmap_cond_end_sync(mddev->bitmap, sector_nr, | |
2647 | mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high)); | |
2648 | r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO); | |
2649 | ||
2650 | raise_barrier(conf, sector_nr); | |
2651 | ||
2652 | rcu_read_lock(); | |
2653 | /* | |
2654 | * If we get a correctably read error during resync or recovery, | |
2655 | * we might want to read from a different device. So we | |
2656 | * flag all drives that could conceivably be read from for READ, | |
2657 | * and any others (which will be non-In_sync devices) for WRITE. | |
2658 | * If a read fails, we try reading from something else for which READ | |
2659 | * is OK. | |
2660 | */ | |
2661 | ||
2662 | r1_bio->mddev = mddev; | |
2663 | r1_bio->sector = sector_nr; | |
2664 | r1_bio->state = 0; | |
2665 | set_bit(R1BIO_IsSync, &r1_bio->state); | |
2666 | /* make sure good_sectors won't go across barrier unit boundary */ | |
2667 | good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors); | |
2668 | ||
2669 | for (i = 0; i < conf->raid_disks * 2; i++) { | |
2670 | struct md_rdev *rdev; | |
2671 | bio = r1_bio->bios[i]; | |
2672 | ||
2673 | rdev = rcu_dereference(conf->mirrors[i].rdev); | |
2674 | if (rdev == NULL || | |
2675 | test_bit(Faulty, &rdev->flags)) { | |
2676 | if (i < conf->raid_disks) | |
2677 | still_degraded = 1; | |
2678 | } else if (!test_bit(In_sync, &rdev->flags)) { | |
2679 | bio_set_op_attrs(bio, REQ_OP_WRITE, 0); | |
2680 | bio->bi_end_io = end_sync_write; | |
2681 | write_targets ++; | |
2682 | } else { | |
2683 | /* may need to read from here */ | |
2684 | sector_t first_bad = MaxSector; | |
2685 | int bad_sectors; | |
2686 | ||
2687 | if (is_badblock(rdev, sector_nr, good_sectors, | |
2688 | &first_bad, &bad_sectors)) { | |
2689 | if (first_bad > sector_nr) | |
2690 | good_sectors = first_bad - sector_nr; | |
2691 | else { | |
2692 | bad_sectors -= (sector_nr - first_bad); | |
2693 | if (min_bad == 0 || | |
2694 | min_bad > bad_sectors) | |
2695 | min_bad = bad_sectors; | |
2696 | } | |
2697 | } | |
2698 | if (sector_nr < first_bad) { | |
2699 | if (test_bit(WriteMostly, &rdev->flags)) { | |
2700 | if (wonly < 0) | |
2701 | wonly = i; | |
2702 | } else { | |
2703 | if (disk < 0) | |
2704 | disk = i; | |
2705 | } | |
2706 | bio_set_op_attrs(bio, REQ_OP_READ, 0); | |
2707 | bio->bi_end_io = end_sync_read; | |
2708 | read_targets++; | |
2709 | } else if (!test_bit(WriteErrorSeen, &rdev->flags) && | |
2710 | test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && | |
2711 | !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) { | |
2712 | /* | |
2713 | * The device is suitable for reading (InSync), | |
2714 | * but has bad block(s) here. Let's try to correct them, | |
2715 | * if we are doing resync or repair. Otherwise, leave | |
2716 | * this device alone for this sync request. | |
2717 | */ | |
2718 | bio_set_op_attrs(bio, REQ_OP_WRITE, 0); | |
2719 | bio->bi_end_io = end_sync_write; | |
2720 | write_targets++; | |
2721 | } | |
2722 | } | |
2723 | if (bio->bi_end_io) { | |
2724 | atomic_inc(&rdev->nr_pending); | |
2725 | bio->bi_iter.bi_sector = sector_nr + rdev->data_offset; | |
2726 | bio_set_dev(bio, rdev->bdev); | |
2727 | if (test_bit(FailFast, &rdev->flags)) | |
2728 | bio->bi_opf |= MD_FAILFAST; | |
2729 | } | |
2730 | } | |
2731 | rcu_read_unlock(); | |
2732 | if (disk < 0) | |
2733 | disk = wonly; | |
2734 | r1_bio->read_disk = disk; | |
2735 | ||
2736 | if (read_targets == 0 && min_bad > 0) { | |
2737 | /* These sectors are bad on all InSync devices, so we | |
2738 | * need to mark them bad on all write targets | |
2739 | */ | |
2740 | int ok = 1; | |
2741 | for (i = 0 ; i < conf->raid_disks * 2 ; i++) | |
2742 | if (r1_bio->bios[i]->bi_end_io == end_sync_write) { | |
2743 | struct md_rdev *rdev = conf->mirrors[i].rdev; | |
2744 | ok = rdev_set_badblocks(rdev, sector_nr, | |
2745 | min_bad, 0 | |
2746 | ) && ok; | |
2747 | } | |
2748 | set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags); | |
2749 | *skipped = 1; | |
2750 | put_buf(r1_bio); | |
2751 | ||
2752 | if (!ok) { | |
2753 | /* Cannot record the badblocks, so need to | |
2754 | * abort the resync. | |
2755 | * If there are multiple read targets, could just | |
2756 | * fail the really bad ones ??? | |
2757 | */ | |
2758 | conf->recovery_disabled = mddev->recovery_disabled; | |
2759 | set_bit(MD_RECOVERY_INTR, &mddev->recovery); | |
2760 | return 0; | |
2761 | } else | |
2762 | return min_bad; | |
2763 | ||
2764 | } | |
2765 | if (min_bad > 0 && min_bad < good_sectors) { | |
2766 | /* only resync enough to reach the next bad->good | |
2767 | * transition */ | |
2768 | good_sectors = min_bad; | |
2769 | } | |
2770 | ||
2771 | if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0) | |
2772 | /* extra read targets are also write targets */ | |
2773 | write_targets += read_targets-1; | |
2774 | ||
2775 | if (write_targets == 0 || read_targets == 0) { | |
2776 | /* There is nowhere to write, so all non-sync | |
2777 | * drives must be failed - so we are finished | |
2778 | */ | |
2779 | sector_t rv; | |
2780 | if (min_bad > 0) | |
2781 | max_sector = sector_nr + min_bad; | |
2782 | rv = max_sector - sector_nr; | |
2783 | *skipped = 1; | |
2784 | put_buf(r1_bio); | |
2785 | return rv; | |
2786 | } | |
2787 | ||
2788 | if (max_sector > mddev->resync_max) | |
2789 | max_sector = mddev->resync_max; /* Don't do IO beyond here */ | |
2790 | if (max_sector > sector_nr + good_sectors) | |
2791 | max_sector = sector_nr + good_sectors; | |
2792 | nr_sectors = 0; | |
2793 | sync_blocks = 0; | |
2794 | do { | |
2795 | struct page *page; | |
2796 | int len = PAGE_SIZE; | |
2797 | if (sector_nr + (len>>9) > max_sector) | |
2798 | len = (max_sector - sector_nr) << 9; | |
2799 | if (len == 0) | |
2800 | break; | |
2801 | if (sync_blocks == 0) { | |
2802 | if (!bitmap_start_sync(mddev->bitmap, sector_nr, | |
2803 | &sync_blocks, still_degraded) && | |
2804 | !conf->fullsync && | |
2805 | !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) | |
2806 | break; | |
2807 | if ((len >> 9) > sync_blocks) | |
2808 | len = sync_blocks<<9; | |
2809 | } | |
2810 | ||
2811 | for (i = 0 ; i < conf->raid_disks * 2; i++) { | |
2812 | struct resync_pages *rp; | |
2813 | ||
2814 | bio = r1_bio->bios[i]; | |
2815 | rp = get_resync_pages(bio); | |
2816 | if (bio->bi_end_io) { | |
2817 | page = resync_fetch_page(rp, page_idx); | |
2818 | ||
2819 | /* | |
2820 | * won't fail because the vec table is big | |
2821 | * enough to hold all these pages | |
2822 | */ | |
2823 | bio_add_page(bio, page, len, 0); | |
2824 | } | |
2825 | } | |
2826 | nr_sectors += len>>9; | |
2827 | sector_nr += len>>9; | |
2828 | sync_blocks -= (len>>9); | |
2829 | } while (++page_idx < RESYNC_PAGES); | |
2830 | ||
2831 | r1_bio->sectors = nr_sectors; | |
2832 | ||
2833 | if (mddev_is_clustered(mddev) && | |
2834 | conf->cluster_sync_high < sector_nr + nr_sectors) { | |
2835 | conf->cluster_sync_low = mddev->curr_resync_completed; | |
2836 | conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS; | |
2837 | /* Send resync message */ | |
2838 | md_cluster_ops->resync_info_update(mddev, | |
2839 | conf->cluster_sync_low, | |
2840 | conf->cluster_sync_high); | |
2841 | } | |
2842 | ||
2843 | /* For a user-requested sync, we read all readable devices and do a | |
2844 | * compare | |
2845 | */ | |
2846 | if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) { | |
2847 | atomic_set(&r1_bio->remaining, read_targets); | |
2848 | for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) { | |
2849 | bio = r1_bio->bios[i]; | |
2850 | if (bio->bi_end_io == end_sync_read) { | |
2851 | read_targets--; | |
2852 | md_sync_acct_bio(bio, nr_sectors); | |
2853 | if (read_targets == 1) | |
2854 | bio->bi_opf &= ~MD_FAILFAST; | |
2855 | generic_make_request(bio); | |
2856 | } | |
2857 | } | |
2858 | } else { | |
2859 | atomic_set(&r1_bio->remaining, 1); | |
2860 | bio = r1_bio->bios[r1_bio->read_disk]; | |
2861 | md_sync_acct_bio(bio, nr_sectors); | |
2862 | if (read_targets == 1) | |
2863 | bio->bi_opf &= ~MD_FAILFAST; | |
2864 | generic_make_request(bio); | |
2865 | ||
2866 | } | |
2867 | return nr_sectors; | |
2868 | } | |
2869 | ||
2870 | static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks) | |
2871 | { | |
2872 | if (sectors) | |
2873 | return sectors; | |
2874 | ||
2875 | return mddev->dev_sectors; | |
2876 | } | |
2877 | ||
2878 | static struct r1conf *setup_conf(struct mddev *mddev) | |
2879 | { | |
2880 | struct r1conf *conf; | |
2881 | int i; | |
2882 | struct raid1_info *disk; | |
2883 | struct md_rdev *rdev; | |
2884 | int err = -ENOMEM; | |
2885 | ||
2886 | conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL); | |
2887 | if (!conf) | |
2888 | goto abort; | |
2889 | ||
2890 | conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR, | |
2891 | sizeof(atomic_t), GFP_KERNEL); | |
2892 | if (!conf->nr_pending) | |
2893 | goto abort; | |
2894 | ||
2895 | conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR, | |
2896 | sizeof(atomic_t), GFP_KERNEL); | |
2897 | if (!conf->nr_waiting) | |
2898 | goto abort; | |
2899 | ||
2900 | conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR, | |
2901 | sizeof(atomic_t), GFP_KERNEL); | |
2902 | if (!conf->nr_queued) | |
2903 | goto abort; | |
2904 | ||
2905 | conf->barrier = kcalloc(BARRIER_BUCKETS_NR, | |
2906 | sizeof(atomic_t), GFP_KERNEL); | |
2907 | if (!conf->barrier) | |
2908 | goto abort; | |
2909 | ||
2910 | conf->mirrors = kzalloc(sizeof(struct raid1_info) | |
2911 | * mddev->raid_disks * 2, | |
2912 | GFP_KERNEL); | |
2913 | if (!conf->mirrors) | |
2914 | goto abort; | |
2915 | ||
2916 | conf->tmppage = alloc_page(GFP_KERNEL); | |
2917 | if (!conf->tmppage) | |
2918 | goto abort; | |
2919 | ||
2920 | conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL); | |
2921 | if (!conf->poolinfo) | |
2922 | goto abort; | |
2923 | conf->poolinfo->raid_disks = mddev->raid_disks * 2; | |
2924 | conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, | |
2925 | r1bio_pool_free, | |
2926 | conf->poolinfo); | |
2927 | if (!conf->r1bio_pool) | |
2928 | goto abort; | |
2929 | ||
2930 | conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0); | |
2931 | if (!conf->bio_split) | |
2932 | goto abort; | |
2933 | ||
2934 | conf->poolinfo->mddev = mddev; | |
2935 | ||
2936 | err = -EINVAL; | |
2937 | spin_lock_init(&conf->device_lock); | |
2938 | rdev_for_each(rdev, mddev) { | |
2939 | int disk_idx = rdev->raid_disk; | |
2940 | if (disk_idx >= mddev->raid_disks | |
2941 | || disk_idx < 0) | |
2942 | continue; | |
2943 | if (test_bit(Replacement, &rdev->flags)) | |
2944 | disk = conf->mirrors + mddev->raid_disks + disk_idx; | |
2945 | else | |
2946 | disk = conf->mirrors + disk_idx; | |
2947 | ||
2948 | if (disk->rdev) | |
2949 | goto abort; | |
2950 | disk->rdev = rdev; | |
2951 | disk->head_position = 0; | |
2952 | disk->seq_start = MaxSector; | |
2953 | } | |
2954 | conf->raid_disks = mddev->raid_disks; | |
2955 | conf->mddev = mddev; | |
2956 | INIT_LIST_HEAD(&conf->retry_list); | |
2957 | INIT_LIST_HEAD(&conf->bio_end_io_list); | |
2958 | ||
2959 | spin_lock_init(&conf->resync_lock); | |
2960 | init_waitqueue_head(&conf->wait_barrier); | |
2961 | ||
2962 | bio_list_init(&conf->pending_bio_list); | |
2963 | conf->pending_count = 0; | |
2964 | conf->recovery_disabled = mddev->recovery_disabled - 1; | |
2965 | ||
2966 | err = -EIO; | |
2967 | for (i = 0; i < conf->raid_disks * 2; i++) { | |
2968 | ||
2969 | disk = conf->mirrors + i; | |
2970 | ||
2971 | if (i < conf->raid_disks && | |
2972 | disk[conf->raid_disks].rdev) { | |
2973 | /* This slot has a replacement. */ | |
2974 | if (!disk->rdev) { | |
2975 | /* No original, just make the replacement | |
2976 | * a recovering spare | |
2977 | */ | |
2978 | disk->rdev = | |
2979 | disk[conf->raid_disks].rdev; | |
2980 | disk[conf->raid_disks].rdev = NULL; | |
2981 | } else if (!test_bit(In_sync, &disk->rdev->flags)) | |
2982 | /* Original is not in_sync - bad */ | |
2983 | goto abort; | |
2984 | } | |
2985 | ||
2986 | if (!disk->rdev || | |
2987 | !test_bit(In_sync, &disk->rdev->flags)) { | |
2988 | disk->head_position = 0; | |
2989 | if (disk->rdev && | |
2990 | (disk->rdev->saved_raid_disk < 0)) | |
2991 | conf->fullsync = 1; | |
2992 | } | |
2993 | } | |
2994 | ||
2995 | err = -ENOMEM; | |
2996 | conf->thread = md_register_thread(raid1d, mddev, "raid1"); | |
2997 | if (!conf->thread) | |
2998 | goto abort; | |
2999 | ||
3000 | return conf; | |
3001 | ||
3002 | abort: | |
3003 | if (conf) { | |
3004 | mempool_destroy(conf->r1bio_pool); | |
3005 | kfree(conf->mirrors); | |
3006 | safe_put_page(conf->tmppage); | |
3007 | kfree(conf->poolinfo); | |
3008 | kfree(conf->nr_pending); | |
3009 | kfree(conf->nr_waiting); | |
3010 | kfree(conf->nr_queued); | |
3011 | kfree(conf->barrier); | |
3012 | if (conf->bio_split) | |
3013 | bioset_free(conf->bio_split); | |
3014 | kfree(conf); | |
3015 | } | |
3016 | return ERR_PTR(err); | |
3017 | } | |
3018 | ||
3019 | static void raid1_free(struct mddev *mddev, void *priv); | |
3020 | static int raid1_run(struct mddev *mddev) | |
3021 | { | |
3022 | struct r1conf *conf; | |
3023 | int i; | |
3024 | struct md_rdev *rdev; | |
3025 | int ret; | |
3026 | bool discard_supported = false; | |
3027 | ||
3028 | if (mddev->level != 1) { | |
3029 | pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n", | |
3030 | mdname(mddev), mddev->level); | |
3031 | return -EIO; | |
3032 | } | |
3033 | if (mddev->reshape_position != MaxSector) { | |
3034 | pr_warn("md/raid1:%s: reshape_position set but not supported\n", | |
3035 | mdname(mddev)); | |
3036 | return -EIO; | |
3037 | } | |
3038 | if (mddev_init_writes_pending(mddev) < 0) | |
3039 | return -ENOMEM; | |
3040 | /* | |
3041 | * copy the already verified devices into our private RAID1 | |
3042 | * bookkeeping area. [whatever we allocate in run(), | |
3043 | * should be freed in raid1_free()] | |
3044 | */ | |
3045 | if (mddev->private == NULL) | |
3046 | conf = setup_conf(mddev); | |
3047 | else | |
3048 | conf = mddev->private; | |
3049 | ||
3050 | if (IS_ERR(conf)) | |
3051 | return PTR_ERR(conf); | |
3052 | ||
3053 | if (mddev->queue) { | |
3054 | blk_queue_max_write_same_sectors(mddev->queue, 0); | |
3055 | blk_queue_max_write_zeroes_sectors(mddev->queue, 0); | |
3056 | } | |
3057 | ||
3058 | rdev_for_each(rdev, mddev) { | |
3059 | if (!mddev->gendisk) | |
3060 | continue; | |
3061 | disk_stack_limits(mddev->gendisk, rdev->bdev, | |
3062 | rdev->data_offset << 9); | |
3063 | if (blk_queue_discard(bdev_get_queue(rdev->bdev))) | |
3064 | discard_supported = true; | |
3065 | } | |
3066 | ||
3067 | mddev->degraded = 0; | |
3068 | for (i=0; i < conf->raid_disks; i++) | |
3069 | if (conf->mirrors[i].rdev == NULL || | |
3070 | !test_bit(In_sync, &conf->mirrors[i].rdev->flags) || | |
3071 | test_bit(Faulty, &conf->mirrors[i].rdev->flags)) | |
3072 | mddev->degraded++; | |
3073 | ||
3074 | if (conf->raid_disks - mddev->degraded == 1) | |
3075 | mddev->recovery_cp = MaxSector; | |
3076 | ||
3077 | if (mddev->recovery_cp != MaxSector) | |
3078 | pr_info("md/raid1:%s: not clean -- starting background reconstruction\n", | |
3079 | mdname(mddev)); | |
3080 | pr_info("md/raid1:%s: active with %d out of %d mirrors\n", | |
3081 | mdname(mddev), mddev->raid_disks - mddev->degraded, | |
3082 | mddev->raid_disks); | |
3083 | ||
3084 | /* | |
3085 | * Ok, everything is just fine now | |
3086 | */ | |
3087 | mddev->thread = conf->thread; | |
3088 | conf->thread = NULL; | |
3089 | mddev->private = conf; | |
3090 | set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags); | |
3091 | ||
3092 | md_set_array_sectors(mddev, raid1_size(mddev, 0, 0)); | |
3093 | ||
3094 | if (mddev->queue) { | |
3095 | if (discard_supported) | |
3096 | queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, | |
3097 | mddev->queue); | |
3098 | else | |
3099 | queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, | |
3100 | mddev->queue); | |
3101 | } | |
3102 | ||
3103 | ret = md_integrity_register(mddev); | |
3104 | if (ret) { | |
3105 | md_unregister_thread(&mddev->thread); | |
3106 | raid1_free(mddev, conf); | |
3107 | } | |
3108 | return ret; | |
3109 | } | |
3110 | ||
3111 | static void raid1_free(struct mddev *mddev, void *priv) | |
3112 | { | |
3113 | struct r1conf *conf = priv; | |
3114 | ||
3115 | mempool_destroy(conf->r1bio_pool); | |
3116 | kfree(conf->mirrors); | |
3117 | safe_put_page(conf->tmppage); | |
3118 | kfree(conf->poolinfo); | |
3119 | kfree(conf->nr_pending); | |
3120 | kfree(conf->nr_waiting); | |
3121 | kfree(conf->nr_queued); | |
3122 | kfree(conf->barrier); | |
3123 | if (conf->bio_split) | |
3124 | bioset_free(conf->bio_split); | |
3125 | kfree(conf); | |
3126 | } | |
3127 | ||
3128 | static int raid1_resize(struct mddev *mddev, sector_t sectors) | |
3129 | { | |
3130 | /* no resync is happening, and there is enough space | |
3131 | * on all devices, so we can resize. | |
3132 | * We need to make sure resync covers any new space. | |
3133 | * If the array is shrinking we should possibly wait until | |
3134 | * any io in the removed space completes, but it hardly seems | |
3135 | * worth it. | |
3136 | */ | |
3137 | sector_t newsize = raid1_size(mddev, sectors, 0); | |
3138 | if (mddev->external_size && | |
3139 | mddev->array_sectors > newsize) | |
3140 | return -EINVAL; | |
3141 | if (mddev->bitmap) { | |
3142 | int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0); | |
3143 | if (ret) | |
3144 | return ret; | |
3145 | } | |
3146 | md_set_array_sectors(mddev, newsize); | |
3147 | if (sectors > mddev->dev_sectors && | |
3148 | mddev->recovery_cp > mddev->dev_sectors) { | |
3149 | mddev->recovery_cp = mddev->dev_sectors; | |
3150 | set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); | |
3151 | } | |
3152 | mddev->dev_sectors = sectors; | |
3153 | mddev->resync_max_sectors = sectors; | |
3154 | return 0; | |
3155 | } | |
3156 | ||
3157 | static int raid1_reshape(struct mddev *mddev) | |
3158 | { | |
3159 | /* We need to: | |
3160 | * 1/ resize the r1bio_pool | |
3161 | * 2/ resize conf->mirrors | |
3162 | * | |
3163 | * We allocate a new r1bio_pool if we can. | |
3164 | * Then raise a device barrier and wait until all IO stops. | |
3165 | * Then resize conf->mirrors and swap in the new r1bio pool. | |
3166 | * | |
3167 | * At the same time, we "pack" the devices so that all the missing | |
3168 | * devices have the higher raid_disk numbers. | |
3169 | */ | |
3170 | mempool_t *newpool, *oldpool; | |
3171 | struct pool_info *newpoolinfo; | |
3172 | struct raid1_info *newmirrors; | |
3173 | struct r1conf *conf = mddev->private; | |
3174 | int cnt, raid_disks; | |
3175 | unsigned long flags; | |
3176 | int d, d2; | |
3177 | ||
3178 | /* Cannot change chunk_size, layout, or level */ | |
3179 | if (mddev->chunk_sectors != mddev->new_chunk_sectors || | |
3180 | mddev->layout != mddev->new_layout || | |
3181 | mddev->level != mddev->new_level) { | |
3182 | mddev->new_chunk_sectors = mddev->chunk_sectors; | |
3183 | mddev->new_layout = mddev->layout; | |
3184 | mddev->new_level = mddev->level; | |
3185 | return -EINVAL; | |
3186 | } | |
3187 | ||
3188 | if (!mddev_is_clustered(mddev)) | |
3189 | md_allow_write(mddev); | |
3190 | ||
3191 | raid_disks = mddev->raid_disks + mddev->delta_disks; | |
3192 | ||
3193 | if (raid_disks < conf->raid_disks) { | |
3194 | cnt=0; | |
3195 | for (d= 0; d < conf->raid_disks; d++) | |
3196 | if (conf->mirrors[d].rdev) | |
3197 | cnt++; | |
3198 | if (cnt > raid_disks) | |
3199 | return -EBUSY; | |
3200 | } | |
3201 | ||
3202 | newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL); | |
3203 | if (!newpoolinfo) | |
3204 | return -ENOMEM; | |
3205 | newpoolinfo->mddev = mddev; | |
3206 | newpoolinfo->raid_disks = raid_disks * 2; | |
3207 | ||
3208 | newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc, | |
3209 | r1bio_pool_free, newpoolinfo); | |
3210 | if (!newpool) { | |
3211 | kfree(newpoolinfo); | |
3212 | return -ENOMEM; | |
3213 | } | |
3214 | newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2, | |
3215 | GFP_KERNEL); | |
3216 | if (!newmirrors) { | |
3217 | kfree(newpoolinfo); | |
3218 | mempool_destroy(newpool); | |
3219 | return -ENOMEM; | |
3220 | } | |
3221 | ||
3222 | freeze_array(conf, 0); | |
3223 | ||
3224 | /* ok, everything is stopped */ | |
3225 | oldpool = conf->r1bio_pool; | |
3226 | conf->r1bio_pool = newpool; | |
3227 | ||
3228 | for (d = d2 = 0; d < conf->raid_disks; d++) { | |
3229 | struct md_rdev *rdev = conf->mirrors[d].rdev; | |
3230 | if (rdev && rdev->raid_disk != d2) { | |
3231 | sysfs_unlink_rdev(mddev, rdev); | |
3232 | rdev->raid_disk = d2; | |
3233 | sysfs_unlink_rdev(mddev, rdev); | |
3234 | if (sysfs_link_rdev(mddev, rdev)) | |
3235 | pr_warn("md/raid1:%s: cannot register rd%d\n", | |
3236 | mdname(mddev), rdev->raid_disk); | |
3237 | } | |
3238 | if (rdev) | |
3239 | newmirrors[d2++].rdev = rdev; | |
3240 | } | |
3241 | kfree(conf->mirrors); | |
3242 | conf->mirrors = newmirrors; | |
3243 | kfree(conf->poolinfo); | |
3244 | conf->poolinfo = newpoolinfo; | |
3245 | ||
3246 | spin_lock_irqsave(&conf->device_lock, flags); | |
3247 | mddev->degraded += (raid_disks - conf->raid_disks); | |
3248 | spin_unlock_irqrestore(&conf->device_lock, flags); | |
3249 | conf->raid_disks = mddev->raid_disks = raid_disks; | |
3250 | mddev->delta_disks = 0; | |
3251 | ||
3252 | unfreeze_array(conf); | |
3253 | ||
3254 | set_bit(MD_RECOVERY_RECOVER, &mddev->recovery); | |
3255 | set_bit(MD_RECOVERY_NEEDED, &mddev->recovery); | |
3256 | md_wakeup_thread(mddev->thread); | |
3257 | ||
3258 | mempool_destroy(oldpool); | |
3259 | return 0; | |
3260 | } | |
3261 | ||
3262 | static void raid1_quiesce(struct mddev *mddev, int state) | |
3263 | { | |
3264 | struct r1conf *conf = mddev->private; | |
3265 | ||
3266 | switch(state) { | |
3267 | case 2: /* wake for suspend */ | |
3268 | wake_up(&conf->wait_barrier); | |
3269 | break; | |
3270 | case 1: | |
3271 | freeze_array(conf, 0); | |
3272 | break; | |
3273 | case 0: | |
3274 | unfreeze_array(conf); | |
3275 | break; | |
3276 | } | |
3277 | } | |
3278 | ||
3279 | static void *raid1_takeover(struct mddev *mddev) | |
3280 | { | |
3281 | /* raid1 can take over: | |
3282 | * raid5 with 2 devices, any layout or chunk size | |
3283 | */ | |
3284 | if (mddev->level == 5 && mddev->raid_disks == 2) { | |
3285 | struct r1conf *conf; | |
3286 | mddev->new_level = 1; | |
3287 | mddev->new_layout = 0; | |
3288 | mddev->new_chunk_sectors = 0; | |
3289 | conf = setup_conf(mddev); | |
3290 | if (!IS_ERR(conf)) { | |
3291 | /* Array must appear to be quiesced */ | |
3292 | conf->array_frozen = 1; | |
3293 | mddev_clear_unsupported_flags(mddev, | |
3294 | UNSUPPORTED_MDDEV_FLAGS); | |
3295 | } | |
3296 | return conf; | |
3297 | } | |
3298 | return ERR_PTR(-EINVAL); | |
3299 | } | |
3300 | ||
3301 | static struct md_personality raid1_personality = | |
3302 | { | |
3303 | .name = "raid1", | |
3304 | .level = 1, | |
3305 | .owner = THIS_MODULE, | |
3306 | .make_request = raid1_make_request, | |
3307 | .run = raid1_run, | |
3308 | .free = raid1_free, | |
3309 | .status = raid1_status, | |
3310 | .error_handler = raid1_error, | |
3311 | .hot_add_disk = raid1_add_disk, | |
3312 | .hot_remove_disk= raid1_remove_disk, | |
3313 | .spare_active = raid1_spare_active, | |
3314 | .sync_request = raid1_sync_request, | |
3315 | .resize = raid1_resize, | |
3316 | .size = raid1_size, | |
3317 | .check_reshape = raid1_reshape, | |
3318 | .quiesce = raid1_quiesce, | |
3319 | .takeover = raid1_takeover, | |
3320 | .congested = raid1_congested, | |
3321 | }; | |
3322 | ||
3323 | static int __init raid_init(void) | |
3324 | { | |
3325 | return register_md_personality(&raid1_personality); | |
3326 | } | |
3327 | ||
3328 | static void raid_exit(void) | |
3329 | { | |
3330 | unregister_md_personality(&raid1_personality); | |
3331 | } | |
3332 | ||
3333 | module_init(raid_init); | |
3334 | module_exit(raid_exit); | |
3335 | MODULE_LICENSE("GPL"); | |
3336 | MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD"); | |
3337 | MODULE_ALIAS("md-personality-3"); /* RAID1 */ | |
3338 | MODULE_ALIAS("md-raid1"); | |
3339 | MODULE_ALIAS("md-level-1"); | |
3340 | ||
3341 | module_param(max_queued_requests, int, S_IRUGO|S_IWUSR); |