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