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 Ƙstergaard" <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
41#include <trace/events/block.h>
42
43#include "md.h"
44#include "raid1.h"
45#include "md-bitmap.h"
46
47#define UNSUPPORTED_MDDEV_FLAGS \
48 ((1L << MD_HAS_JOURNAL) | \
49 (1L << MD_JOURNAL_CLEAN) | \
50 (1L << MD_HAS_PPL) | \
51 (1L << MD_HAS_MULTIPLE_PPLS))
52
53/*
54 * Number of guaranteed r1bios in case of extreme VM load:
55 */
56#define NR_RAID1_BIOS 256
57
58/* when we get a read error on a read-only array, we redirect to another
59 * device without failing the first device, or trying to over-write to
60 * correct the read error. To keep track of bad blocks on a per-bio
61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
62 */
63#define IO_BLOCKED ((struct bio *)1)
64/* When we successfully write to a known bad-block, we need to remove the
65 * bad-block marking which must be done from process context. So we record
66 * the success by setting devs[n].bio to IO_MADE_GOOD
67 */
68#define IO_MADE_GOOD ((struct bio *)2)
69
70#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
71
72/* When there are this many requests queue to be written by
73 * the raid1 thread, we become 'congested' to provide back-pressure
74 * for writeback.
75 */
76static int max_queued_requests = 1024;
77
78static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
79static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
80
81#define raid1_log(md, fmt, args...) \
82 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
83
84#include "raid1-10.c"
85
86/*
87 * for resync bio, r1bio pointer can be retrieved from the per-bio
88 * 'struct resync_pages'.
89 */
90static inline struct r1bio *get_resync_r1bio(struct bio *bio)
91{
92 return get_resync_pages(bio)->raid_bio;
93}
94
95static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
96{
97 struct pool_info *pi = data;
98 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
99
100 /* allocate a r1bio with room for raid_disks entries in the bios array */
101 return kzalloc(size, gfp_flags);
102}
103
104static void r1bio_pool_free(void *r1_bio, void *data)
105{
106 kfree(r1_bio);
107}
108
109#define RESYNC_DEPTH 32
110#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
111#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
112#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
113#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
114#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
115
116static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
117{
118 struct pool_info *pi = data;
119 struct r1bio *r1_bio;
120 struct bio *bio;
121 int need_pages;
122 int j;
123 struct resync_pages *rps;
124
125 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
126 if (!r1_bio)
127 return NULL;
128
129 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
130 gfp_flags);
131 if (!rps)
132 goto out_free_r1bio;
133
134 /*
135 * Allocate bios : 1 for reading, n-1 for writing
136 */
137 for (j = pi->raid_disks ; j-- ; ) {
138 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
139 if (!bio)
140 goto out_free_bio;
141 r1_bio->bios[j] = bio;
142 }
143 /*
144 * Allocate RESYNC_PAGES data pages and attach them to
145 * the first bio.
146 * If this is a user-requested check/repair, allocate
147 * RESYNC_PAGES for each bio.
148 */
149 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
150 need_pages = pi->raid_disks;
151 else
152 need_pages = 1;
153 for (j = 0; j < pi->raid_disks; j++) {
154 struct resync_pages *rp = &rps[j];
155
156 bio = r1_bio->bios[j];
157
158 if (j < need_pages) {
159 if (resync_alloc_pages(rp, gfp_flags))
160 goto out_free_pages;
161 } else {
162 memcpy(rp, &rps[0], sizeof(*rp));
163 resync_get_all_pages(rp);
164 }
165
166 rp->raid_bio = r1_bio;
167 bio->bi_private = rp;
168 }
169
170 r1_bio->master_bio = NULL;
171
172 return r1_bio;
173
174out_free_pages:
175 while (--j >= 0)
176 resync_free_pages(&rps[j]);
177
178out_free_bio:
179 while (++j < pi->raid_disks)
180 bio_put(r1_bio->bios[j]);
181 kfree(rps);
182
183out_free_r1bio:
184 r1bio_pool_free(r1_bio, data);
185 return NULL;
186}
187
188static void r1buf_pool_free(void *__r1_bio, void *data)
189{
190 struct pool_info *pi = data;
191 int i;
192 struct r1bio *r1bio = __r1_bio;
193 struct resync_pages *rp = NULL;
194
195 for (i = pi->raid_disks; i--; ) {
196 rp = get_resync_pages(r1bio->bios[i]);
197 resync_free_pages(rp);
198 bio_put(r1bio->bios[i]);
199 }
200
201 /* resync pages array stored in the 1st bio's .bi_private */
202 kfree(rp);
203
204 r1bio_pool_free(r1bio, data);
205}
206
207static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
208{
209 int i;
210
211 for (i = 0; i < conf->raid_disks * 2; i++) {
212 struct bio **bio = r1_bio->bios + i;
213 if (!BIO_SPECIAL(*bio))
214 bio_put(*bio);
215 *bio = NULL;
216 }
217}
218
219static void free_r1bio(struct r1bio *r1_bio)
220{
221 struct r1conf *conf = r1_bio->mddev->private;
222
223 put_all_bios(conf, r1_bio);
224 mempool_free(r1_bio, &conf->r1bio_pool);
225}
226
227static void put_buf(struct r1bio *r1_bio)
228{
229 struct r1conf *conf = r1_bio->mddev->private;
230 sector_t sect = r1_bio->sector;
231 int i;
232
233 for (i = 0; i < conf->raid_disks * 2; i++) {
234 struct bio *bio = r1_bio->bios[i];
235 if (bio->bi_end_io)
236 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
237 }
238
239 mempool_free(r1_bio, &conf->r1buf_pool);
240
241 lower_barrier(conf, sect);
242}
243
244static void reschedule_retry(struct r1bio *r1_bio)
245{
246 unsigned long flags;
247 struct mddev *mddev = r1_bio->mddev;
248 struct r1conf *conf = mddev->private;
249 int idx;
250
251 idx = sector_to_idx(r1_bio->sector);
252 spin_lock_irqsave(&conf->device_lock, flags);
253 list_add(&r1_bio->retry_list, &conf->retry_list);
254 atomic_inc(&conf->nr_queued[idx]);
255 spin_unlock_irqrestore(&conf->device_lock, flags);
256
257 wake_up(&conf->wait_barrier);
258 md_wakeup_thread(mddev->thread);
259}
260
261/*
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
264 * cache layer.
265 */
266static void call_bio_endio(struct r1bio *r1_bio)
267{
268 struct bio *bio = r1_bio->master_bio;
269 struct r1conf *conf = r1_bio->mddev->private;
270
271 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
272 bio->bi_status = BLK_STS_IOERR;
273
274 bio_endio(bio);
275 /*
276 * Wake up any possible resync thread that waits for the device
277 * to go idle.
278 */
279 allow_barrier(conf, r1_bio->sector);
280}
281
282static void raid_end_bio_io(struct r1bio *r1_bio)
283{
284 struct bio *bio = r1_bio->master_bio;
285
286 /* if nobody has done the final endio yet, do it now */
287 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
288 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
289 (bio_data_dir(bio) == WRITE) ? "write" : "read",
290 (unsigned long long) bio->bi_iter.bi_sector,
291 (unsigned long long) bio_end_sector(bio) - 1);
292
293 call_bio_endio(r1_bio);
294 }
295 free_r1bio(r1_bio);
296}
297
298/*
299 * Update disk head position estimator based on IRQ completion info.
300 */
301static inline void update_head_pos(int disk, struct r1bio *r1_bio)
302{
303 struct r1conf *conf = r1_bio->mddev->private;
304
305 conf->mirrors[disk].head_position =
306 r1_bio->sector + (r1_bio->sectors);
307}
308
309/*
310 * Find the disk number which triggered given bio
311 */
312static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
313{
314 int mirror;
315 struct r1conf *conf = r1_bio->mddev->private;
316 int raid_disks = conf->raid_disks;
317
318 for (mirror = 0; mirror < raid_disks * 2; mirror++)
319 if (r1_bio->bios[mirror] == bio)
320 break;
321
322 BUG_ON(mirror == raid_disks * 2);
323 update_head_pos(mirror, r1_bio);
324
325 return mirror;
326}
327
328static void raid1_end_read_request(struct bio *bio)
329{
330 int uptodate = !bio->bi_status;
331 struct r1bio *r1_bio = bio->bi_private;
332 struct r1conf *conf = r1_bio->mddev->private;
333 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
334
335 /*
336 * this branch is our 'one mirror IO has finished' event handler:
337 */
338 update_head_pos(r1_bio->read_disk, r1_bio);
339
340 if (uptodate)
341 set_bit(R1BIO_Uptodate, &r1_bio->state);
342 else if (test_bit(FailFast, &rdev->flags) &&
343 test_bit(R1BIO_FailFast, &r1_bio->state))
344 /* This was a fail-fast read so we definitely
345 * want to retry */
346 ;
347 else {
348 /* If all other devices have failed, we want to return
349 * the error upwards rather than fail the last device.
350 * Here we redefine "uptodate" to mean "Don't want to retry"
351 */
352 unsigned long flags;
353 spin_lock_irqsave(&conf->device_lock, flags);
354 if (r1_bio->mddev->degraded == conf->raid_disks ||
355 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
356 test_bit(In_sync, &rdev->flags)))
357 uptodate = 1;
358 spin_unlock_irqrestore(&conf->device_lock, flags);
359 }
360
361 if (uptodate) {
362 raid_end_bio_io(r1_bio);
363 rdev_dec_pending(rdev, conf->mddev);
364 } else {
365 /*
366 * oops, read error:
367 */
368 char b[BDEVNAME_SIZE];
369 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
370 mdname(conf->mddev),
371 bdevname(rdev->bdev, b),
372 (unsigned long long)r1_bio->sector);
373 set_bit(R1BIO_ReadError, &r1_bio->state);
374 reschedule_retry(r1_bio);
375 /* don't drop the reference on read_disk yet */
376 }
377}
378
379static void close_write(struct r1bio *r1_bio)
380{
381 /* it really is the end of this request */
382 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383 bio_free_pages(r1_bio->behind_master_bio);
384 bio_put(r1_bio->behind_master_bio);
385 r1_bio->behind_master_bio = NULL;
386 }
387 /* clear the bitmap if all writes complete successfully */
388 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
389 r1_bio->sectors,
390 !test_bit(R1BIO_Degraded, &r1_bio->state),
391 test_bit(R1BIO_BehindIO, &r1_bio->state));
392 md_write_end(r1_bio->mddev);
393}
394
395static void r1_bio_write_done(struct r1bio *r1_bio)
396{
397 if (!atomic_dec_and_test(&r1_bio->remaining))
398 return;
399
400 if (test_bit(R1BIO_WriteError, &r1_bio->state))
401 reschedule_retry(r1_bio);
402 else {
403 close_write(r1_bio);
404 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
405 reschedule_retry(r1_bio);
406 else
407 raid_end_bio_io(r1_bio);
408 }
409}
410
411static void raid1_end_write_request(struct bio *bio)
412{
413 struct r1bio *r1_bio = bio->bi_private;
414 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
415 struct r1conf *conf = r1_bio->mddev->private;
416 struct bio *to_put = NULL;
417 int mirror = find_bio_disk(r1_bio, bio);
418 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
419 bool discard_error;
420
421 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
422
423 /*
424 * 'one mirror IO has finished' event handler:
425 */
426 if (bio->bi_status && !discard_error) {
427 set_bit(WriteErrorSeen, &rdev->flags);
428 if (!test_and_set_bit(WantReplacement, &rdev->flags))
429 set_bit(MD_RECOVERY_NEEDED, &
430 conf->mddev->recovery);
431
432 if (test_bit(FailFast, &rdev->flags) &&
433 (bio->bi_opf & MD_FAILFAST) &&
434 /* We never try FailFast to WriteMostly devices */
435 !test_bit(WriteMostly, &rdev->flags)) {
436 md_error(r1_bio->mddev, rdev);
437 if (!test_bit(Faulty, &rdev->flags))
438 /* This is the only remaining device,
439 * We need to retry the write without
440 * FailFast
441 */
442 set_bit(R1BIO_WriteError, &r1_bio->state);
443 else {
444 /* Finished with this branch */
445 r1_bio->bios[mirror] = NULL;
446 to_put = bio;
447 }
448 } else
449 set_bit(R1BIO_WriteError, &r1_bio->state);
450 } else {
451 /*
452 * Set R1BIO_Uptodate in our master bio, so that we
453 * will return a good error code for to the higher
454 * levels even if IO on some other mirrored buffer
455 * fails.
456 *
457 * The 'master' represents the composite IO operation
458 * to user-side. So if something waits for IO, then it
459 * will wait for the 'master' bio.
460 */
461 sector_t first_bad;
462 int bad_sectors;
463
464 r1_bio->bios[mirror] = NULL;
465 to_put = bio;
466 /*
467 * Do not set R1BIO_Uptodate if the current device is
468 * rebuilding or Faulty. This is because we cannot use
469 * such device for properly reading the data back (we could
470 * potentially use it, if the current write would have felt
471 * before rdev->recovery_offset, but for simplicity we don't
472 * check this here.
473 */
474 if (test_bit(In_sync, &rdev->flags) &&
475 !test_bit(Faulty, &rdev->flags))
476 set_bit(R1BIO_Uptodate, &r1_bio->state);
477
478 /* Maybe we can clear some bad blocks. */
479 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
480 &first_bad, &bad_sectors) && !discard_error) {
481 r1_bio->bios[mirror] = IO_MADE_GOOD;
482 set_bit(R1BIO_MadeGood, &r1_bio->state);
483 }
484 }
485
486 if (behind) {
487 if (test_bit(WriteMostly, &rdev->flags))
488 atomic_dec(&r1_bio->behind_remaining);
489
490 /*
491 * In behind mode, we ACK the master bio once the I/O
492 * has safely reached all non-writemostly
493 * disks. Setting the Returned bit ensures that this
494 * gets done only once -- we don't ever want to return
495 * -EIO here, instead we'll wait
496 */
497 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
498 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
499 /* Maybe we can return now */
500 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
501 struct bio *mbio = r1_bio->master_bio;
502 pr_debug("raid1: behind end write sectors"
503 " %llu-%llu\n",
504 (unsigned long long) mbio->bi_iter.bi_sector,
505 (unsigned long long) bio_end_sector(mbio) - 1);
506 call_bio_endio(r1_bio);
507 }
508 }
509 }
510 if (r1_bio->bios[mirror] == NULL)
511 rdev_dec_pending(rdev, conf->mddev);
512
513 /*
514 * Let's see if all mirrored write operations have finished
515 * already.
516 */
517 r1_bio_write_done(r1_bio);
518
519 if (to_put)
520 bio_put(to_put);
521}
522
523static sector_t align_to_barrier_unit_end(sector_t start_sector,
524 sector_t sectors)
525{
526 sector_t len;
527
528 WARN_ON(sectors == 0);
529 /*
530 * len is the number of sectors from start_sector to end of the
531 * barrier unit which start_sector belongs to.
532 */
533 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
534 start_sector;
535
536 if (len > sectors)
537 len = sectors;
538
539 return len;
540}
541
542/*
543 * This routine returns the disk from which the requested read should
544 * be done. There is a per-array 'next expected sequential IO' sector
545 * number - if this matches on the next IO then we use the last disk.
546 * There is also a per-disk 'last know head position' sector that is
547 * maintained from IRQ contexts, both the normal and the resync IO
548 * completion handlers update this position correctly. If there is no
549 * perfect sequential match then we pick the disk whose head is closest.
550 *
551 * If there are 2 mirrors in the same 2 devices, performance degrades
552 * because position is mirror, not device based.
553 *
554 * The rdev for the device selected will have nr_pending incremented.
555 */
556static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
557{
558 const sector_t this_sector = r1_bio->sector;
559 int sectors;
560 int best_good_sectors;
561 int best_disk, best_dist_disk, best_pending_disk;
562 int has_nonrot_disk;
563 int disk;
564 sector_t best_dist;
565 unsigned int min_pending;
566 struct md_rdev *rdev;
567 int choose_first;
568 int choose_next_idle;
569
570 rcu_read_lock();
571 /*
572 * Check if we can balance. We can balance on the whole
573 * device if no resync is going on, or below the resync window.
574 * We take the first readable disk when above the resync window.
575 */
576 retry:
577 sectors = r1_bio->sectors;
578 best_disk = -1;
579 best_dist_disk = -1;
580 best_dist = MaxSector;
581 best_pending_disk = -1;
582 min_pending = UINT_MAX;
583 best_good_sectors = 0;
584 has_nonrot_disk = 0;
585 choose_next_idle = 0;
586 clear_bit(R1BIO_FailFast, &r1_bio->state);
587
588 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
589 (mddev_is_clustered(conf->mddev) &&
590 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
591 this_sector + sectors)))
592 choose_first = 1;
593 else
594 choose_first = 0;
595
596 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
597 sector_t dist;
598 sector_t first_bad;
599 int bad_sectors;
600 unsigned int pending;
601 bool nonrot;
602
603 rdev = rcu_dereference(conf->mirrors[disk].rdev);
604 if (r1_bio->bios[disk] == IO_BLOCKED
605 || rdev == NULL
606 || test_bit(Faulty, &rdev->flags))
607 continue;
608 if (!test_bit(In_sync, &rdev->flags) &&
609 rdev->recovery_offset < this_sector + sectors)
610 continue;
611 if (test_bit(WriteMostly, &rdev->flags)) {
612 /* Don't balance among write-mostly, just
613 * use the first as a last resort */
614 if (best_dist_disk < 0) {
615 if (is_badblock(rdev, this_sector, sectors,
616 &first_bad, &bad_sectors)) {
617 if (first_bad <= this_sector)
618 /* Cannot use this */
619 continue;
620 best_good_sectors = first_bad - this_sector;
621 } else
622 best_good_sectors = sectors;
623 best_dist_disk = disk;
624 best_pending_disk = disk;
625 }
626 continue;
627 }
628 /* This is a reasonable device to use. It might
629 * even be best.
630 */
631 if (is_badblock(rdev, this_sector, sectors,
632 &first_bad, &bad_sectors)) {
633 if (best_dist < MaxSector)
634 /* already have a better device */
635 continue;
636 if (first_bad <= this_sector) {
637 /* cannot read here. If this is the 'primary'
638 * device, then we must not read beyond
639 * bad_sectors from another device..
640 */
641 bad_sectors -= (this_sector - first_bad);
642 if (choose_first && sectors > bad_sectors)
643 sectors = bad_sectors;
644 if (best_good_sectors > sectors)
645 best_good_sectors = sectors;
646
647 } else {
648 sector_t good_sectors = first_bad - this_sector;
649 if (good_sectors > best_good_sectors) {
650 best_good_sectors = good_sectors;
651 best_disk = disk;
652 }
653 if (choose_first)
654 break;
655 }
656 continue;
657 } else {
658 if ((sectors > best_good_sectors) && (best_disk >= 0))
659 best_disk = -1;
660 best_good_sectors = sectors;
661 }
662
663 if (best_disk >= 0)
664 /* At least two disks to choose from so failfast is OK */
665 set_bit(R1BIO_FailFast, &r1_bio->state);
666
667 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
668 has_nonrot_disk |= nonrot;
669 pending = atomic_read(&rdev->nr_pending);
670 dist = abs(this_sector - conf->mirrors[disk].head_position);
671 if (choose_first) {
672 best_disk = disk;
673 break;
674 }
675 /* Don't change to another disk for sequential reads */
676 if (conf->mirrors[disk].next_seq_sect == this_sector
677 || dist == 0) {
678 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
679 struct raid1_info *mirror = &conf->mirrors[disk];
680
681 best_disk = disk;
682 /*
683 * If buffered sequential IO size exceeds optimal
684 * iosize, check if there is idle disk. If yes, choose
685 * the idle disk. read_balance could already choose an
686 * idle disk before noticing it's a sequential IO in
687 * this disk. This doesn't matter because this disk
688 * will idle, next time it will be utilized after the
689 * first disk has IO size exceeds optimal iosize. In
690 * this way, iosize of the first disk will be optimal
691 * iosize at least. iosize of the second disk might be
692 * small, but not a big deal since when the second disk
693 * starts IO, the first disk is likely still busy.
694 */
695 if (nonrot && opt_iosize > 0 &&
696 mirror->seq_start != MaxSector &&
697 mirror->next_seq_sect > opt_iosize &&
698 mirror->next_seq_sect - opt_iosize >=
699 mirror->seq_start) {
700 choose_next_idle = 1;
701 continue;
702 }
703 break;
704 }
705
706 if (choose_next_idle)
707 continue;
708
709 if (min_pending > pending) {
710 min_pending = pending;
711 best_pending_disk = disk;
712 }
713
714 if (dist < best_dist) {
715 best_dist = dist;
716 best_dist_disk = disk;
717 }
718 }
719
720 /*
721 * If all disks are rotational, choose the closest disk. If any disk is
722 * non-rotational, choose the disk with less pending request even the
723 * disk is rotational, which might/might not be optimal for raids with
724 * mixed ratation/non-rotational disks depending on workload.
725 */
726 if (best_disk == -1) {
727 if (has_nonrot_disk || min_pending == 0)
728 best_disk = best_pending_disk;
729 else
730 best_disk = best_dist_disk;
731 }
732
733 if (best_disk >= 0) {
734 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
735 if (!rdev)
736 goto retry;
737 atomic_inc(&rdev->nr_pending);
738 sectors = best_good_sectors;
739
740 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
741 conf->mirrors[best_disk].seq_start = this_sector;
742
743 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
744 }
745 rcu_read_unlock();
746 *max_sectors = sectors;
747
748 return best_disk;
749}
750
751static int raid1_congested(struct mddev *mddev, int bits)
752{
753 struct r1conf *conf = mddev->private;
754 int i, ret = 0;
755
756 if ((bits & (1 << WB_async_congested)) &&
757 conf->pending_count >= max_queued_requests)
758 return 1;
759
760 rcu_read_lock();
761 for (i = 0; i < conf->raid_disks * 2; i++) {
762 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
763 if (rdev && !test_bit(Faulty, &rdev->flags)) {
764 struct request_queue *q = bdev_get_queue(rdev->bdev);
765
766 BUG_ON(!q);
767
768 /* Note the '|| 1' - when read_balance prefers
769 * non-congested targets, it can be removed
770 */
771 if ((bits & (1 << WB_async_congested)) || 1)
772 ret |= bdi_congested(q->backing_dev_info, bits);
773 else
774 ret &= bdi_congested(q->backing_dev_info, bits);
775 }
776 }
777 rcu_read_unlock();
778 return ret;
779}
780
781static void flush_bio_list(struct r1conf *conf, struct bio *bio)
782{
783 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
784 md_bitmap_unplug(conf->mddev->bitmap);
785 wake_up(&conf->wait_barrier);
786
787 while (bio) { /* submit pending writes */
788 struct bio *next = bio->bi_next;
789 struct md_rdev *rdev = (void *)bio->bi_disk;
790 bio->bi_next = NULL;
791 bio_set_dev(bio, rdev->bdev);
792 if (test_bit(Faulty, &rdev->flags)) {
793 bio_io_error(bio);
794 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
795 !blk_queue_discard(bio->bi_disk->queue)))
796 /* Just ignore it */
797 bio_endio(bio);
798 else
799 generic_make_request(bio);
800 bio = next;
801 }
802}
803
804static void flush_pending_writes(struct r1conf *conf)
805{
806 /* Any writes that have been queued but are awaiting
807 * bitmap updates get flushed here.
808 */
809 spin_lock_irq(&conf->device_lock);
810
811 if (conf->pending_bio_list.head) {
812 struct blk_plug plug;
813 struct bio *bio;
814
815 bio = bio_list_get(&conf->pending_bio_list);
816 conf->pending_count = 0;
817 spin_unlock_irq(&conf->device_lock);
818
819 /*
820 * As this is called in a wait_event() loop (see freeze_array),
821 * current->state might be TASK_UNINTERRUPTIBLE which will
822 * cause a warning when we prepare to wait again. As it is
823 * rare that this path is taken, it is perfectly safe to force
824 * us to go around the wait_event() loop again, so the warning
825 * is a false-positive. Silence the warning by resetting
826 * thread state
827 */
828 __set_current_state(TASK_RUNNING);
829 blk_start_plug(&plug);
830 flush_bio_list(conf, bio);
831 blk_finish_plug(&plug);
832 } else
833 spin_unlock_irq(&conf->device_lock);
834}
835
836/* Barriers....
837 * Sometimes we need to suspend IO while we do something else,
838 * either some resync/recovery, or reconfigure the array.
839 * To do this we raise a 'barrier'.
840 * The 'barrier' is a counter that can be raised multiple times
841 * to count how many activities are happening which preclude
842 * normal IO.
843 * We can only raise the barrier if there is no pending IO.
844 * i.e. if nr_pending == 0.
845 * We choose only to raise the barrier if no-one is waiting for the
846 * barrier to go down. This means that as soon as an IO request
847 * is ready, no other operations which require a barrier will start
848 * until the IO request has had a chance.
849 *
850 * So: regular IO calls 'wait_barrier'. When that returns there
851 * is no backgroup IO happening, It must arrange to call
852 * allow_barrier when it has finished its IO.
853 * backgroup IO calls must call raise_barrier. Once that returns
854 * there is no normal IO happeing. It must arrange to call
855 * lower_barrier when the particular background IO completes.
856 */
857static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
858{
859 int idx = sector_to_idx(sector_nr);
860
861 spin_lock_irq(&conf->resync_lock);
862
863 /* Wait until no block IO is waiting */
864 wait_event_lock_irq(conf->wait_barrier,
865 !atomic_read(&conf->nr_waiting[idx]),
866 conf->resync_lock);
867
868 /* block any new IO from starting */
869 atomic_inc(&conf->barrier[idx]);
870 /*
871 * In raise_barrier() we firstly increase conf->barrier[idx] then
872 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
873 * increase conf->nr_pending[idx] then check conf->barrier[idx].
874 * A memory barrier here to make sure conf->nr_pending[idx] won't
875 * be fetched before conf->barrier[idx] is increased. Otherwise
876 * there will be a race between raise_barrier() and _wait_barrier().
877 */
878 smp_mb__after_atomic();
879
880 /* For these conditions we must wait:
881 * A: while the array is in frozen state
882 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
883 * existing in corresponding I/O barrier bucket.
884 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
885 * max resync count which allowed on current I/O barrier bucket.
886 */
887 wait_event_lock_irq(conf->wait_barrier,
888 (!conf->array_frozen &&
889 !atomic_read(&conf->nr_pending[idx]) &&
890 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
891 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
892 conf->resync_lock);
893
894 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
895 atomic_dec(&conf->barrier[idx]);
896 spin_unlock_irq(&conf->resync_lock);
897 wake_up(&conf->wait_barrier);
898 return -EINTR;
899 }
900
901 atomic_inc(&conf->nr_sync_pending);
902 spin_unlock_irq(&conf->resync_lock);
903
904 return 0;
905}
906
907static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
908{
909 int idx = sector_to_idx(sector_nr);
910
911 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
912
913 atomic_dec(&conf->barrier[idx]);
914 atomic_dec(&conf->nr_sync_pending);
915 wake_up(&conf->wait_barrier);
916}
917
918static void _wait_barrier(struct r1conf *conf, int idx)
919{
920 /*
921 * We need to increase conf->nr_pending[idx] very early here,
922 * then raise_barrier() can be blocked when it waits for
923 * conf->nr_pending[idx] to be 0. Then we can avoid holding
924 * conf->resync_lock when there is no barrier raised in same
925 * barrier unit bucket. Also if the array is frozen, I/O
926 * should be blocked until array is unfrozen.
927 */
928 atomic_inc(&conf->nr_pending[idx]);
929 /*
930 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
931 * check conf->barrier[idx]. In raise_barrier() we firstly increase
932 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
933 * barrier is necessary here to make sure conf->barrier[idx] won't be
934 * fetched before conf->nr_pending[idx] is increased. Otherwise there
935 * will be a race between _wait_barrier() and raise_barrier().
936 */
937 smp_mb__after_atomic();
938
939 /*
940 * Don't worry about checking two atomic_t variables at same time
941 * here. If during we check conf->barrier[idx], the array is
942 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
943 * 0, it is safe to return and make the I/O continue. Because the
944 * array is frozen, all I/O returned here will eventually complete
945 * or be queued, no race will happen. See code comment in
946 * frozen_array().
947 */
948 if (!READ_ONCE(conf->array_frozen) &&
949 !atomic_read(&conf->barrier[idx]))
950 return;
951
952 /*
953 * After holding conf->resync_lock, conf->nr_pending[idx]
954 * should be decreased before waiting for barrier to drop.
955 * Otherwise, we may encounter a race condition because
956 * raise_barrer() might be waiting for conf->nr_pending[idx]
957 * to be 0 at same time.
958 */
959 spin_lock_irq(&conf->resync_lock);
960 atomic_inc(&conf->nr_waiting[idx]);
961 atomic_dec(&conf->nr_pending[idx]);
962 /*
963 * In case freeze_array() is waiting for
964 * get_unqueued_pending() == extra
965 */
966 wake_up(&conf->wait_barrier);
967 /* Wait for the barrier in same barrier unit bucket to drop. */
968 wait_event_lock_irq(conf->wait_barrier,
969 !conf->array_frozen &&
970 !atomic_read(&conf->barrier[idx]),
971 conf->resync_lock);
972 atomic_inc(&conf->nr_pending[idx]);
973 atomic_dec(&conf->nr_waiting[idx]);
974 spin_unlock_irq(&conf->resync_lock);
975}
976
977static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
978{
979 int idx = sector_to_idx(sector_nr);
980
981 /*
982 * Very similar to _wait_barrier(). The difference is, for read
983 * I/O we don't need wait for sync I/O, but if the whole array
984 * is frozen, the read I/O still has to wait until the array is
985 * unfrozen. Since there is no ordering requirement with
986 * conf->barrier[idx] here, memory barrier is unnecessary as well.
987 */
988 atomic_inc(&conf->nr_pending[idx]);
989
990 if (!READ_ONCE(conf->array_frozen))
991 return;
992
993 spin_lock_irq(&conf->resync_lock);
994 atomic_inc(&conf->nr_waiting[idx]);
995 atomic_dec(&conf->nr_pending[idx]);
996 /*
997 * In case freeze_array() is waiting for
998 * get_unqueued_pending() == extra
999 */
1000 wake_up(&conf->wait_barrier);
1001 /* Wait for array to be unfrozen */
1002 wait_event_lock_irq(conf->wait_barrier,
1003 !conf->array_frozen,
1004 conf->resync_lock);
1005 atomic_inc(&conf->nr_pending[idx]);
1006 atomic_dec(&conf->nr_waiting[idx]);
1007 spin_unlock_irq(&conf->resync_lock);
1008}
1009
1010static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1011{
1012 int idx = sector_to_idx(sector_nr);
1013
1014 _wait_barrier(conf, idx);
1015}
1016
1017static void _allow_barrier(struct r1conf *conf, int idx)
1018{
1019 atomic_dec(&conf->nr_pending[idx]);
1020 wake_up(&conf->wait_barrier);
1021}
1022
1023static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1024{
1025 int idx = sector_to_idx(sector_nr);
1026
1027 _allow_barrier(conf, idx);
1028}
1029
1030/* conf->resync_lock should be held */
1031static int get_unqueued_pending(struct r1conf *conf)
1032{
1033 int idx, ret;
1034
1035 ret = atomic_read(&conf->nr_sync_pending);
1036 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1037 ret += atomic_read(&conf->nr_pending[idx]) -
1038 atomic_read(&conf->nr_queued[idx]);
1039
1040 return ret;
1041}
1042
1043static void freeze_array(struct r1conf *conf, int extra)
1044{
1045 /* Stop sync I/O and normal I/O and wait for everything to
1046 * go quiet.
1047 * This is called in two situations:
1048 * 1) management command handlers (reshape, remove disk, quiesce).
1049 * 2) one normal I/O request failed.
1050
1051 * After array_frozen is set to 1, new sync IO will be blocked at
1052 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1053 * or wait_read_barrier(). The flying I/Os will either complete or be
1054 * queued. When everything goes quite, there are only queued I/Os left.
1055
1056 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1057 * barrier bucket index which this I/O request hits. When all sync and
1058 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1059 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1060 * in handle_read_error(), we may call freeze_array() before trying to
1061 * fix the read error. In this case, the error read I/O is not queued,
1062 * so get_unqueued_pending() == 1.
1063 *
1064 * Therefore before this function returns, we need to wait until
1065 * get_unqueued_pendings(conf) gets equal to extra. For
1066 * normal I/O context, extra is 1, in rested situations extra is 0.
1067 */
1068 spin_lock_irq(&conf->resync_lock);
1069 conf->array_frozen = 1;
1070 raid1_log(conf->mddev, "wait freeze");
1071 wait_event_lock_irq_cmd(
1072 conf->wait_barrier,
1073 get_unqueued_pending(conf) == extra,
1074 conf->resync_lock,
1075 flush_pending_writes(conf));
1076 spin_unlock_irq(&conf->resync_lock);
1077}
1078static void unfreeze_array(struct r1conf *conf)
1079{
1080 /* reverse the effect of the freeze */
1081 spin_lock_irq(&conf->resync_lock);
1082 conf->array_frozen = 0;
1083 spin_unlock_irq(&conf->resync_lock);
1084 wake_up(&conf->wait_barrier);
1085}
1086
1087static void alloc_behind_master_bio(struct r1bio *r1_bio,
1088 struct bio *bio)
1089{
1090 int size = bio->bi_iter.bi_size;
1091 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1092 int i = 0;
1093 struct bio *behind_bio = NULL;
1094
1095 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1096 if (!behind_bio)
1097 return;
1098
1099 /* discard op, we don't support writezero/writesame yet */
1100 if (!bio_has_data(bio)) {
1101 behind_bio->bi_iter.bi_size = size;
1102 goto skip_copy;
1103 }
1104
1105 behind_bio->bi_write_hint = bio->bi_write_hint;
1106
1107 while (i < vcnt && size) {
1108 struct page *page;
1109 int len = min_t(int, PAGE_SIZE, size);
1110
1111 page = alloc_page(GFP_NOIO);
1112 if (unlikely(!page))
1113 goto free_pages;
1114
1115 bio_add_page(behind_bio, page, len, 0);
1116
1117 size -= len;
1118 i++;
1119 }
1120
1121 bio_copy_data(behind_bio, bio);
1122skip_copy:
1123 r1_bio->behind_master_bio = behind_bio;
1124 set_bit(R1BIO_BehindIO, &r1_bio->state);
1125
1126 return;
1127
1128free_pages:
1129 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1130 bio->bi_iter.bi_size);
1131 bio_free_pages(behind_bio);
1132 bio_put(behind_bio);
1133}
1134
1135struct raid1_plug_cb {
1136 struct blk_plug_cb cb;
1137 struct bio_list pending;
1138 int pending_cnt;
1139};
1140
1141static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1142{
1143 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1144 cb);
1145 struct mddev *mddev = plug->cb.data;
1146 struct r1conf *conf = mddev->private;
1147 struct bio *bio;
1148
1149 if (from_schedule || current->bio_list) {
1150 spin_lock_irq(&conf->device_lock);
1151 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1152 conf->pending_count += plug->pending_cnt;
1153 spin_unlock_irq(&conf->device_lock);
1154 wake_up(&conf->wait_barrier);
1155 md_wakeup_thread(mddev->thread);
1156 kfree(plug);
1157 return;
1158 }
1159
1160 /* we aren't scheduling, so we can do the write-out directly. */
1161 bio = bio_list_get(&plug->pending);
1162 flush_bio_list(conf, bio);
1163 kfree(plug);
1164}
1165
1166static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1167{
1168 r1_bio->master_bio = bio;
1169 r1_bio->sectors = bio_sectors(bio);
1170 r1_bio->state = 0;
1171 r1_bio->mddev = mddev;
1172 r1_bio->sector = bio->bi_iter.bi_sector;
1173}
1174
1175static inline struct r1bio *
1176alloc_r1bio(struct mddev *mddev, struct bio *bio)
1177{
1178 struct r1conf *conf = mddev->private;
1179 struct r1bio *r1_bio;
1180
1181 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1182 /* Ensure no bio records IO_BLOCKED */
1183 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1184 init_r1bio(r1_bio, mddev, bio);
1185 return r1_bio;
1186}
1187
1188static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1189 int max_read_sectors, struct r1bio *r1_bio)
1190{
1191 struct r1conf *conf = mddev->private;
1192 struct raid1_info *mirror;
1193 struct bio *read_bio;
1194 struct bitmap *bitmap = mddev->bitmap;
1195 const int op = bio_op(bio);
1196 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1197 int max_sectors;
1198 int rdisk;
1199 bool print_msg = !!r1_bio;
1200 char b[BDEVNAME_SIZE];
1201
1202 /*
1203 * If r1_bio is set, we are blocking the raid1d thread
1204 * so there is a tiny risk of deadlock. So ask for
1205 * emergency memory if needed.
1206 */
1207 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1208
1209 if (print_msg) {
1210 /* Need to get the block device name carefully */
1211 struct md_rdev *rdev;
1212 rcu_read_lock();
1213 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1214 if (rdev)
1215 bdevname(rdev->bdev, b);
1216 else
1217 strcpy(b, "???");
1218 rcu_read_unlock();
1219 }
1220
1221 /*
1222 * Still need barrier for READ in case that whole
1223 * array is frozen.
1224 */
1225 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1226
1227 if (!r1_bio)
1228 r1_bio = alloc_r1bio(mddev, bio);
1229 else
1230 init_r1bio(r1_bio, mddev, bio);
1231 r1_bio->sectors = max_read_sectors;
1232
1233 /*
1234 * make_request() can abort the operation when read-ahead is being
1235 * used and no empty request is available.
1236 */
1237 rdisk = read_balance(conf, r1_bio, &max_sectors);
1238
1239 if (rdisk < 0) {
1240 /* couldn't find anywhere to read from */
1241 if (print_msg) {
1242 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1243 mdname(mddev),
1244 b,
1245 (unsigned long long)r1_bio->sector);
1246 }
1247 raid_end_bio_io(r1_bio);
1248 return;
1249 }
1250 mirror = conf->mirrors + rdisk;
1251
1252 if (print_msg)
1253 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1254 mdname(mddev),
1255 (unsigned long long)r1_bio->sector,
1256 bdevname(mirror->rdev->bdev, b));
1257
1258 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1259 bitmap) {
1260 /*
1261 * Reading from a write-mostly device must take care not to
1262 * over-take any writes that are 'behind'
1263 */
1264 raid1_log(mddev, "wait behind writes");
1265 wait_event(bitmap->behind_wait,
1266 atomic_read(&bitmap->behind_writes) == 0);
1267 }
1268
1269 if (max_sectors < bio_sectors(bio)) {
1270 struct bio *split = bio_split(bio, max_sectors,
1271 gfp, &conf->bio_split);
1272 bio_chain(split, bio);
1273 generic_make_request(bio);
1274 bio = split;
1275 r1_bio->master_bio = bio;
1276 r1_bio->sectors = max_sectors;
1277 }
1278
1279 r1_bio->read_disk = rdisk;
1280
1281 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1282
1283 r1_bio->bios[rdisk] = read_bio;
1284
1285 read_bio->bi_iter.bi_sector = r1_bio->sector +
1286 mirror->rdev->data_offset;
1287 bio_set_dev(read_bio, mirror->rdev->bdev);
1288 read_bio->bi_end_io = raid1_end_read_request;
1289 bio_set_op_attrs(read_bio, op, do_sync);
1290 if (test_bit(FailFast, &mirror->rdev->flags) &&
1291 test_bit(R1BIO_FailFast, &r1_bio->state))
1292 read_bio->bi_opf |= MD_FAILFAST;
1293 read_bio->bi_private = r1_bio;
1294
1295 if (mddev->gendisk)
1296 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1297 disk_devt(mddev->gendisk), r1_bio->sector);
1298
1299 generic_make_request(read_bio);
1300}
1301
1302static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1303 int max_write_sectors)
1304{
1305 struct r1conf *conf = mddev->private;
1306 struct r1bio *r1_bio;
1307 int i, disks;
1308 struct bitmap *bitmap = mddev->bitmap;
1309 unsigned long flags;
1310 struct md_rdev *blocked_rdev;
1311 struct blk_plug_cb *cb;
1312 struct raid1_plug_cb *plug = NULL;
1313 int first_clone;
1314 int max_sectors;
1315
1316 if (mddev_is_clustered(mddev) &&
1317 md_cluster_ops->area_resyncing(mddev, WRITE,
1318 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1319
1320 DEFINE_WAIT(w);
1321 for (;;) {
1322 prepare_to_wait(&conf->wait_barrier,
1323 &w, TASK_IDLE);
1324 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325 bio->bi_iter.bi_sector,
1326 bio_end_sector(bio)))
1327 break;
1328 schedule();
1329 }
1330 finish_wait(&conf->wait_barrier, &w);
1331 }
1332
1333 /*
1334 * Register the new request and wait if the reconstruction
1335 * thread has put up a bar for new requests.
1336 * Continue immediately if no resync is active currently.
1337 */
1338 wait_barrier(conf, bio->bi_iter.bi_sector);
1339
1340 r1_bio = alloc_r1bio(mddev, bio);
1341 r1_bio->sectors = max_write_sectors;
1342
1343 if (conf->pending_count >= max_queued_requests) {
1344 md_wakeup_thread(mddev->thread);
1345 raid1_log(mddev, "wait queued");
1346 wait_event(conf->wait_barrier,
1347 conf->pending_count < max_queued_requests);
1348 }
1349 /* first select target devices under rcu_lock and
1350 * inc refcount on their rdev. Record them by setting
1351 * bios[x] to bio
1352 * If there are known/acknowledged bad blocks on any device on
1353 * which we have seen a write error, we want to avoid writing those
1354 * blocks.
1355 * This potentially requires several writes to write around
1356 * the bad blocks. Each set of writes gets it's own r1bio
1357 * with a set of bios attached.
1358 */
1359
1360 disks = conf->raid_disks * 2;
1361 retry_write:
1362 blocked_rdev = NULL;
1363 rcu_read_lock();
1364 max_sectors = r1_bio->sectors;
1365 for (i = 0; i < disks; i++) {
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1367 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1368 atomic_inc(&rdev->nr_pending);
1369 blocked_rdev = rdev;
1370 break;
1371 }
1372 r1_bio->bios[i] = NULL;
1373 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1374 if (i < conf->raid_disks)
1375 set_bit(R1BIO_Degraded, &r1_bio->state);
1376 continue;
1377 }
1378
1379 atomic_inc(&rdev->nr_pending);
1380 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1381 sector_t first_bad;
1382 int bad_sectors;
1383 int is_bad;
1384
1385 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1386 &first_bad, &bad_sectors);
1387 if (is_bad < 0) {
1388 /* mustn't write here until the bad block is
1389 * acknowledged*/
1390 set_bit(BlockedBadBlocks, &rdev->flags);
1391 blocked_rdev = rdev;
1392 break;
1393 }
1394 if (is_bad && first_bad <= r1_bio->sector) {
1395 /* Cannot write here at all */
1396 bad_sectors -= (r1_bio->sector - first_bad);
1397 if (bad_sectors < max_sectors)
1398 /* mustn't write more than bad_sectors
1399 * to other devices yet
1400 */
1401 max_sectors = bad_sectors;
1402 rdev_dec_pending(rdev, mddev);
1403 /* We don't set R1BIO_Degraded as that
1404 * only applies if the disk is
1405 * missing, so it might be re-added,
1406 * and we want to know to recover this
1407 * chunk.
1408 * In this case the device is here,
1409 * and the fact that this chunk is not
1410 * in-sync is recorded in the bad
1411 * block log
1412 */
1413 continue;
1414 }
1415 if (is_bad) {
1416 int good_sectors = first_bad - r1_bio->sector;
1417 if (good_sectors < max_sectors)
1418 max_sectors = good_sectors;
1419 }
1420 }
1421 r1_bio->bios[i] = bio;
1422 }
1423 rcu_read_unlock();
1424
1425 if (unlikely(blocked_rdev)) {
1426 /* Wait for this device to become unblocked */
1427 int j;
1428
1429 for (j = 0; j < i; j++)
1430 if (r1_bio->bios[j])
1431 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1432 r1_bio->state = 0;
1433 allow_barrier(conf, bio->bi_iter.bi_sector);
1434 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1435 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1436 wait_barrier(conf, bio->bi_iter.bi_sector);
1437 goto retry_write;
1438 }
1439
1440 if (max_sectors < bio_sectors(bio)) {
1441 struct bio *split = bio_split(bio, max_sectors,
1442 GFP_NOIO, &conf->bio_split);
1443 bio_chain(split, bio);
1444 generic_make_request(bio);
1445 bio = split;
1446 r1_bio->master_bio = bio;
1447 r1_bio->sectors = max_sectors;
1448 }
1449
1450 atomic_set(&r1_bio->remaining, 1);
1451 atomic_set(&r1_bio->behind_remaining, 0);
1452
1453 first_clone = 1;
1454
1455 for (i = 0; i < disks; i++) {
1456 struct bio *mbio = NULL;
1457 if (!r1_bio->bios[i])
1458 continue;
1459
1460
1461 if (first_clone) {
1462 /* do behind I/O ?
1463 * Not if there are too many, or cannot
1464 * allocate memory, or a reader on WriteMostly
1465 * is waiting for behind writes to flush */
1466 if (bitmap &&
1467 (atomic_read(&bitmap->behind_writes)
1468 < mddev->bitmap_info.max_write_behind) &&
1469 !waitqueue_active(&bitmap->behind_wait)) {
1470 alloc_behind_master_bio(r1_bio, bio);
1471 }
1472
1473 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1474 test_bit(R1BIO_BehindIO, &r1_bio->state));
1475 first_clone = 0;
1476 }
1477
1478 if (r1_bio->behind_master_bio)
1479 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1480 GFP_NOIO, &mddev->bio_set);
1481 else
1482 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1483
1484 if (r1_bio->behind_master_bio) {
1485 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1486 atomic_inc(&r1_bio->behind_remaining);
1487 }
1488
1489 r1_bio->bios[i] = mbio;
1490
1491 mbio->bi_iter.bi_sector = (r1_bio->sector +
1492 conf->mirrors[i].rdev->data_offset);
1493 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1494 mbio->bi_end_io = raid1_end_write_request;
1495 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1496 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1497 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1498 conf->raid_disks - mddev->degraded > 1)
1499 mbio->bi_opf |= MD_FAILFAST;
1500 mbio->bi_private = r1_bio;
1501
1502 atomic_inc(&r1_bio->remaining);
1503
1504 if (mddev->gendisk)
1505 trace_block_bio_remap(mbio->bi_disk->queue,
1506 mbio, disk_devt(mddev->gendisk),
1507 r1_bio->sector);
1508 /* flush_pending_writes() needs access to the rdev so...*/
1509 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1510
1511 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1512 if (cb)
1513 plug = container_of(cb, struct raid1_plug_cb, cb);
1514 else
1515 plug = NULL;
1516 if (plug) {
1517 bio_list_add(&plug->pending, mbio);
1518 plug->pending_cnt++;
1519 } else {
1520 spin_lock_irqsave(&conf->device_lock, flags);
1521 bio_list_add(&conf->pending_bio_list, mbio);
1522 conf->pending_count++;
1523 spin_unlock_irqrestore(&conf->device_lock, flags);
1524 md_wakeup_thread(mddev->thread);
1525 }
1526 }
1527
1528 r1_bio_write_done(r1_bio);
1529
1530 /* In case raid1d snuck in to freeze_array */
1531 wake_up(&conf->wait_barrier);
1532}
1533
1534static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1535{
1536 sector_t sectors;
1537
1538 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1539 md_flush_request(mddev, bio);
1540 return true;
1541 }
1542
1543 /*
1544 * There is a limit to the maximum size, but
1545 * the read/write handler might find a lower limit
1546 * due to bad blocks. To avoid multiple splits,
1547 * we pass the maximum number of sectors down
1548 * and let the lower level perform the split.
1549 */
1550 sectors = align_to_barrier_unit_end(
1551 bio->bi_iter.bi_sector, bio_sectors(bio));
1552
1553 if (bio_data_dir(bio) == READ)
1554 raid1_read_request(mddev, bio, sectors, NULL);
1555 else {
1556 if (!md_write_start(mddev,bio))
1557 return false;
1558 raid1_write_request(mddev, bio, sectors);
1559 }
1560 return true;
1561}
1562
1563static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1564{
1565 struct r1conf *conf = mddev->private;
1566 int i;
1567
1568 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1569 conf->raid_disks - mddev->degraded);
1570 rcu_read_lock();
1571 for (i = 0; i < conf->raid_disks; i++) {
1572 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1573 seq_printf(seq, "%s",
1574 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1575 }
1576 rcu_read_unlock();
1577 seq_printf(seq, "]");
1578}
1579
1580static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1581{
1582 char b[BDEVNAME_SIZE];
1583 struct r1conf *conf = mddev->private;
1584 unsigned long flags;
1585
1586 /*
1587 * If it is not operational, then we have already marked it as dead
1588 * else if it is the last working disks, ignore the error, let the
1589 * next level up know.
1590 * else mark the drive as failed
1591 */
1592 spin_lock_irqsave(&conf->device_lock, flags);
1593 if (test_bit(In_sync, &rdev->flags)
1594 && (conf->raid_disks - mddev->degraded) == 1) {
1595 /*
1596 * Don't fail the drive, act as though we were just a
1597 * normal single drive.
1598 * However don't try a recovery from this drive as
1599 * it is very likely to fail.
1600 */
1601 conf->recovery_disabled = mddev->recovery_disabled;
1602 spin_unlock_irqrestore(&conf->device_lock, flags);
1603 return;
1604 }
1605 set_bit(Blocked, &rdev->flags);
1606 if (test_and_clear_bit(In_sync, &rdev->flags))
1607 mddev->degraded++;
1608 set_bit(Faulty, &rdev->flags);
1609 spin_unlock_irqrestore(&conf->device_lock, flags);
1610 /*
1611 * if recovery is running, make sure it aborts.
1612 */
1613 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1614 set_mask_bits(&mddev->sb_flags, 0,
1615 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1616 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1617 "md/raid1:%s: Operation continuing on %d devices.\n",
1618 mdname(mddev), bdevname(rdev->bdev, b),
1619 mdname(mddev), conf->raid_disks - mddev->degraded);
1620}
1621
1622static void print_conf(struct r1conf *conf)
1623{
1624 int i;
1625
1626 pr_debug("RAID1 conf printout:\n");
1627 if (!conf) {
1628 pr_debug("(!conf)\n");
1629 return;
1630 }
1631 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1632 conf->raid_disks);
1633
1634 rcu_read_lock();
1635 for (i = 0; i < conf->raid_disks; i++) {
1636 char b[BDEVNAME_SIZE];
1637 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1638 if (rdev)
1639 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1640 i, !test_bit(In_sync, &rdev->flags),
1641 !test_bit(Faulty, &rdev->flags),
1642 bdevname(rdev->bdev,b));
1643 }
1644 rcu_read_unlock();
1645}
1646
1647static void close_sync(struct r1conf *conf)
1648{
1649 int idx;
1650
1651 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1652 _wait_barrier(conf, idx);
1653 _allow_barrier(conf, idx);
1654 }
1655
1656 mempool_exit(&conf->r1buf_pool);
1657}
1658
1659static int raid1_spare_active(struct mddev *mddev)
1660{
1661 int i;
1662 struct r1conf *conf = mddev->private;
1663 int count = 0;
1664 unsigned long flags;
1665
1666 /*
1667 * Find all failed disks within the RAID1 configuration
1668 * and mark them readable.
1669 * Called under mddev lock, so rcu protection not needed.
1670 * device_lock used to avoid races with raid1_end_read_request
1671 * which expects 'In_sync' flags and ->degraded to be consistent.
1672 */
1673 spin_lock_irqsave(&conf->device_lock, flags);
1674 for (i = 0; i < conf->raid_disks; i++) {
1675 struct md_rdev *rdev = conf->mirrors[i].rdev;
1676 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1677 if (repl
1678 && !test_bit(Candidate, &repl->flags)
1679 && repl->recovery_offset == MaxSector
1680 && !test_bit(Faulty, &repl->flags)
1681 && !test_and_set_bit(In_sync, &repl->flags)) {
1682 /* replacement has just become active */
1683 if (!rdev ||
1684 !test_and_clear_bit(In_sync, &rdev->flags))
1685 count++;
1686 if (rdev) {
1687 /* Replaced device not technically
1688 * faulty, but we need to be sure
1689 * it gets removed and never re-added
1690 */
1691 set_bit(Faulty, &rdev->flags);
1692 sysfs_notify_dirent_safe(
1693 rdev->sysfs_state);
1694 }
1695 }
1696 if (rdev
1697 && rdev->recovery_offset == MaxSector
1698 && !test_bit(Faulty, &rdev->flags)
1699 && !test_and_set_bit(In_sync, &rdev->flags)) {
1700 count++;
1701 sysfs_notify_dirent_safe(rdev->sysfs_state);
1702 }
1703 }
1704 mddev->degraded -= count;
1705 spin_unlock_irqrestore(&conf->device_lock, flags);
1706
1707 print_conf(conf);
1708 return count;
1709}
1710
1711static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1712{
1713 struct r1conf *conf = mddev->private;
1714 int err = -EEXIST;
1715 int mirror = 0;
1716 struct raid1_info *p;
1717 int first = 0;
1718 int last = conf->raid_disks - 1;
1719
1720 if (mddev->recovery_disabled == conf->recovery_disabled)
1721 return -EBUSY;
1722
1723 if (md_integrity_add_rdev(rdev, mddev))
1724 return -ENXIO;
1725
1726 if (rdev->raid_disk >= 0)
1727 first = last = rdev->raid_disk;
1728
1729 /*
1730 * find the disk ... but prefer rdev->saved_raid_disk
1731 * if possible.
1732 */
1733 if (rdev->saved_raid_disk >= 0 &&
1734 rdev->saved_raid_disk >= first &&
1735 rdev->saved_raid_disk < conf->raid_disks &&
1736 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1737 first = last = rdev->saved_raid_disk;
1738
1739 for (mirror = first; mirror <= last; mirror++) {
1740 p = conf->mirrors+mirror;
1741 if (!p->rdev) {
1742
1743 if (mddev->gendisk)
1744 disk_stack_limits(mddev->gendisk, rdev->bdev,
1745 rdev->data_offset << 9);
1746
1747 p->head_position = 0;
1748 rdev->raid_disk = mirror;
1749 err = 0;
1750 /* As all devices are equivalent, we don't need a full recovery
1751 * if this was recently any drive of the array
1752 */
1753 if (rdev->saved_raid_disk < 0)
1754 conf->fullsync = 1;
1755 rcu_assign_pointer(p->rdev, rdev);
1756 break;
1757 }
1758 if (test_bit(WantReplacement, &p->rdev->flags) &&
1759 p[conf->raid_disks].rdev == NULL) {
1760 /* Add this device as a replacement */
1761 clear_bit(In_sync, &rdev->flags);
1762 set_bit(Replacement, &rdev->flags);
1763 rdev->raid_disk = mirror;
1764 err = 0;
1765 conf->fullsync = 1;
1766 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1767 break;
1768 }
1769 }
1770 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1771 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1772 print_conf(conf);
1773 return err;
1774}
1775
1776static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1777{
1778 struct r1conf *conf = mddev->private;
1779 int err = 0;
1780 int number = rdev->raid_disk;
1781 struct raid1_info *p = conf->mirrors + number;
1782
1783 if (rdev != p->rdev)
1784 p = conf->mirrors + conf->raid_disks + number;
1785
1786 print_conf(conf);
1787 if (rdev == p->rdev) {
1788 if (test_bit(In_sync, &rdev->flags) ||
1789 atomic_read(&rdev->nr_pending)) {
1790 err = -EBUSY;
1791 goto abort;
1792 }
1793 /* Only remove non-faulty devices if recovery
1794 * is not possible.
1795 */
1796 if (!test_bit(Faulty, &rdev->flags) &&
1797 mddev->recovery_disabled != conf->recovery_disabled &&
1798 mddev->degraded < conf->raid_disks) {
1799 err = -EBUSY;
1800 goto abort;
1801 }
1802 p->rdev = NULL;
1803 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1804 synchronize_rcu();
1805 if (atomic_read(&rdev->nr_pending)) {
1806 /* lost the race, try later */
1807 err = -EBUSY;
1808 p->rdev = rdev;
1809 goto abort;
1810 }
1811 }
1812 if (conf->mirrors[conf->raid_disks + number].rdev) {
1813 /* We just removed a device that is being replaced.
1814 * Move down the replacement. We drain all IO before
1815 * doing this to avoid confusion.
1816 */
1817 struct md_rdev *repl =
1818 conf->mirrors[conf->raid_disks + number].rdev;
1819 freeze_array(conf, 0);
1820 if (atomic_read(&repl->nr_pending)) {
1821 /* It means that some queued IO of retry_list
1822 * hold repl. Thus, we cannot set replacement
1823 * as NULL, avoiding rdev NULL pointer
1824 * dereference in sync_request_write and
1825 * handle_write_finished.
1826 */
1827 err = -EBUSY;
1828 unfreeze_array(conf);
1829 goto abort;
1830 }
1831 clear_bit(Replacement, &repl->flags);
1832 p->rdev = repl;
1833 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1834 unfreeze_array(conf);
1835 }
1836
1837 clear_bit(WantReplacement, &rdev->flags);
1838 err = md_integrity_register(mddev);
1839 }
1840abort:
1841
1842 print_conf(conf);
1843 return err;
1844}
1845
1846static void end_sync_read(struct bio *bio)
1847{
1848 struct r1bio *r1_bio = get_resync_r1bio(bio);
1849
1850 update_head_pos(r1_bio->read_disk, r1_bio);
1851
1852 /*
1853 * we have read a block, now it needs to be re-written,
1854 * or re-read if the read failed.
1855 * We don't do much here, just schedule handling by raid1d
1856 */
1857 if (!bio->bi_status)
1858 set_bit(R1BIO_Uptodate, &r1_bio->state);
1859
1860 if (atomic_dec_and_test(&r1_bio->remaining))
1861 reschedule_retry(r1_bio);
1862}
1863
1864static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1865{
1866 sector_t sync_blocks = 0;
1867 sector_t s = r1_bio->sector;
1868 long sectors_to_go = r1_bio->sectors;
1869
1870 /* make sure these bits don't get cleared. */
1871 do {
1872 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1873 s += sync_blocks;
1874 sectors_to_go -= sync_blocks;
1875 } while (sectors_to_go > 0);
1876}
1877
1878static void end_sync_write(struct bio *bio)
1879{
1880 int uptodate = !bio->bi_status;
1881 struct r1bio *r1_bio = get_resync_r1bio(bio);
1882 struct mddev *mddev = r1_bio->mddev;
1883 struct r1conf *conf = mddev->private;
1884 sector_t first_bad;
1885 int bad_sectors;
1886 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1887
1888 if (!uptodate) {
1889 abort_sync_write(mddev, r1_bio);
1890 set_bit(WriteErrorSeen, &rdev->flags);
1891 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1892 set_bit(MD_RECOVERY_NEEDED, &
1893 mddev->recovery);
1894 set_bit(R1BIO_WriteError, &r1_bio->state);
1895 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1896 &first_bad, &bad_sectors) &&
1897 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1898 r1_bio->sector,
1899 r1_bio->sectors,
1900 &first_bad, &bad_sectors)
1901 )
1902 set_bit(R1BIO_MadeGood, &r1_bio->state);
1903
1904 if (atomic_dec_and_test(&r1_bio->remaining)) {
1905 int s = r1_bio->sectors;
1906 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1907 test_bit(R1BIO_WriteError, &r1_bio->state))
1908 reschedule_retry(r1_bio);
1909 else {
1910 put_buf(r1_bio);
1911 md_done_sync(mddev, s, uptodate);
1912 }
1913 }
1914}
1915
1916static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1917 int sectors, struct page *page, int rw)
1918{
1919 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1920 /* success */
1921 return 1;
1922 if (rw == WRITE) {
1923 set_bit(WriteErrorSeen, &rdev->flags);
1924 if (!test_and_set_bit(WantReplacement,
1925 &rdev->flags))
1926 set_bit(MD_RECOVERY_NEEDED, &
1927 rdev->mddev->recovery);
1928 }
1929 /* need to record an error - either for the block or the device */
1930 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1931 md_error(rdev->mddev, rdev);
1932 return 0;
1933}
1934
1935static int fix_sync_read_error(struct r1bio *r1_bio)
1936{
1937 /* Try some synchronous reads of other devices to get
1938 * good data, much like with normal read errors. Only
1939 * read into the pages we already have so we don't
1940 * need to re-issue the read request.
1941 * We don't need to freeze the array, because being in an
1942 * active sync request, there is no normal IO, and
1943 * no overlapping syncs.
1944 * We don't need to check is_badblock() again as we
1945 * made sure that anything with a bad block in range
1946 * will have bi_end_io clear.
1947 */
1948 struct mddev *mddev = r1_bio->mddev;
1949 struct r1conf *conf = mddev->private;
1950 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1951 struct page **pages = get_resync_pages(bio)->pages;
1952 sector_t sect = r1_bio->sector;
1953 int sectors = r1_bio->sectors;
1954 int idx = 0;
1955 struct md_rdev *rdev;
1956
1957 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1958 if (test_bit(FailFast, &rdev->flags)) {
1959 /* Don't try recovering from here - just fail it
1960 * ... unless it is the last working device of course */
1961 md_error(mddev, rdev);
1962 if (test_bit(Faulty, &rdev->flags))
1963 /* Don't try to read from here, but make sure
1964 * put_buf does it's thing
1965 */
1966 bio->bi_end_io = end_sync_write;
1967 }
1968
1969 while(sectors) {
1970 int s = sectors;
1971 int d = r1_bio->read_disk;
1972 int success = 0;
1973 int start;
1974
1975 if (s > (PAGE_SIZE>>9))
1976 s = PAGE_SIZE >> 9;
1977 do {
1978 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1979 /* No rcu protection needed here devices
1980 * can only be removed when no resync is
1981 * active, and resync is currently active
1982 */
1983 rdev = conf->mirrors[d].rdev;
1984 if (sync_page_io(rdev, sect, s<<9,
1985 pages[idx],
1986 REQ_OP_READ, 0, false)) {
1987 success = 1;
1988 break;
1989 }
1990 }
1991 d++;
1992 if (d == conf->raid_disks * 2)
1993 d = 0;
1994 } while (!success && d != r1_bio->read_disk);
1995
1996 if (!success) {
1997 char b[BDEVNAME_SIZE];
1998 int abort = 0;
1999 /* Cannot read from anywhere, this block is lost.
2000 * Record a bad block on each device. If that doesn't
2001 * work just disable and interrupt the recovery.
2002 * Don't fail devices as that won't really help.
2003 */
2004 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2005 mdname(mddev), bio_devname(bio, b),
2006 (unsigned long long)r1_bio->sector);
2007 for (d = 0; d < conf->raid_disks * 2; d++) {
2008 rdev = conf->mirrors[d].rdev;
2009 if (!rdev || test_bit(Faulty, &rdev->flags))
2010 continue;
2011 if (!rdev_set_badblocks(rdev, sect, s, 0))
2012 abort = 1;
2013 }
2014 if (abort) {
2015 conf->recovery_disabled =
2016 mddev->recovery_disabled;
2017 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2018 md_done_sync(mddev, r1_bio->sectors, 0);
2019 put_buf(r1_bio);
2020 return 0;
2021 }
2022 /* Try next page */
2023 sectors -= s;
2024 sect += s;
2025 idx++;
2026 continue;
2027 }
2028
2029 start = d;
2030 /* write it back and re-read */
2031 while (d != r1_bio->read_disk) {
2032 if (d == 0)
2033 d = conf->raid_disks * 2;
2034 d--;
2035 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2036 continue;
2037 rdev = conf->mirrors[d].rdev;
2038 if (r1_sync_page_io(rdev, sect, s,
2039 pages[idx],
2040 WRITE) == 0) {
2041 r1_bio->bios[d]->bi_end_io = NULL;
2042 rdev_dec_pending(rdev, mddev);
2043 }
2044 }
2045 d = start;
2046 while (d != r1_bio->read_disk) {
2047 if (d == 0)
2048 d = conf->raid_disks * 2;
2049 d--;
2050 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2051 continue;
2052 rdev = conf->mirrors[d].rdev;
2053 if (r1_sync_page_io(rdev, sect, s,
2054 pages[idx],
2055 READ) != 0)
2056 atomic_add(s, &rdev->corrected_errors);
2057 }
2058 sectors -= s;
2059 sect += s;
2060 idx ++;
2061 }
2062 set_bit(R1BIO_Uptodate, &r1_bio->state);
2063 bio->bi_status = 0;
2064 return 1;
2065}
2066
2067static void process_checks(struct r1bio *r1_bio)
2068{
2069 /* We have read all readable devices. If we haven't
2070 * got the block, then there is no hope left.
2071 * If we have, then we want to do a comparison
2072 * and skip the write if everything is the same.
2073 * If any blocks failed to read, then we need to
2074 * attempt an over-write
2075 */
2076 struct mddev *mddev = r1_bio->mddev;
2077 struct r1conf *conf = mddev->private;
2078 int primary;
2079 int i;
2080 int vcnt;
2081
2082 /* Fix variable parts of all bios */
2083 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2084 for (i = 0; i < conf->raid_disks * 2; i++) {
2085 blk_status_t status;
2086 struct bio *b = r1_bio->bios[i];
2087 struct resync_pages *rp = get_resync_pages(b);
2088 if (b->bi_end_io != end_sync_read)
2089 continue;
2090 /* fixup the bio for reuse, but preserve errno */
2091 status = b->bi_status;
2092 bio_reset(b);
2093 b->bi_status = status;
2094 b->bi_iter.bi_sector = r1_bio->sector +
2095 conf->mirrors[i].rdev->data_offset;
2096 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2097 b->bi_end_io = end_sync_read;
2098 rp->raid_bio = r1_bio;
2099 b->bi_private = rp;
2100
2101 /* initialize bvec table again */
2102 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2103 }
2104 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2105 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2106 !r1_bio->bios[primary]->bi_status) {
2107 r1_bio->bios[primary]->bi_end_io = NULL;
2108 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2109 break;
2110 }
2111 r1_bio->read_disk = primary;
2112 for (i = 0; i < conf->raid_disks * 2; i++) {
2113 int j;
2114 struct bio *pbio = r1_bio->bios[primary];
2115 struct bio *sbio = r1_bio->bios[i];
2116 blk_status_t status = sbio->bi_status;
2117 struct page **ppages = get_resync_pages(pbio)->pages;
2118 struct page **spages = get_resync_pages(sbio)->pages;
2119 struct bio_vec *bi;
2120 int page_len[RESYNC_PAGES] = { 0 };
2121 struct bvec_iter_all iter_all;
2122
2123 if (sbio->bi_end_io != end_sync_read)
2124 continue;
2125 /* Now we can 'fixup' the error value */
2126 sbio->bi_status = 0;
2127
2128 bio_for_each_segment_all(bi, sbio, j, iter_all)
2129 page_len[j] = bi->bv_len;
2130
2131 if (!status) {
2132 for (j = vcnt; j-- ; ) {
2133 if (memcmp(page_address(ppages[j]),
2134 page_address(spages[j]),
2135 page_len[j]))
2136 break;
2137 }
2138 } else
2139 j = 0;
2140 if (j >= 0)
2141 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2142 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2143 && !status)) {
2144 /* No need to write to this device. */
2145 sbio->bi_end_io = NULL;
2146 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2147 continue;
2148 }
2149
2150 bio_copy_data(sbio, pbio);
2151 }
2152}
2153
2154static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2155{
2156 struct r1conf *conf = mddev->private;
2157 int i;
2158 int disks = conf->raid_disks * 2;
2159 struct bio *wbio;
2160
2161 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2162 /* ouch - failed to read all of that. */
2163 if (!fix_sync_read_error(r1_bio))
2164 return;
2165
2166 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2167 process_checks(r1_bio);
2168
2169 /*
2170 * schedule writes
2171 */
2172 atomic_set(&r1_bio->remaining, 1);
2173 for (i = 0; i < disks ; i++) {
2174 wbio = r1_bio->bios[i];
2175 if (wbio->bi_end_io == NULL ||
2176 (wbio->bi_end_io == end_sync_read &&
2177 (i == r1_bio->read_disk ||
2178 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2179 continue;
2180 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2181 abort_sync_write(mddev, r1_bio);
2182 continue;
2183 }
2184
2185 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2186 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2187 wbio->bi_opf |= MD_FAILFAST;
2188
2189 wbio->bi_end_io = end_sync_write;
2190 atomic_inc(&r1_bio->remaining);
2191 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2192
2193 generic_make_request(wbio);
2194 }
2195
2196 if (atomic_dec_and_test(&r1_bio->remaining)) {
2197 /* if we're here, all write(s) have completed, so clean up */
2198 int s = r1_bio->sectors;
2199 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2200 test_bit(R1BIO_WriteError, &r1_bio->state))
2201 reschedule_retry(r1_bio);
2202 else {
2203 put_buf(r1_bio);
2204 md_done_sync(mddev, s, 1);
2205 }
2206 }
2207}
2208
2209/*
2210 * This is a kernel thread which:
2211 *
2212 * 1. Retries failed read operations on working mirrors.
2213 * 2. Updates the raid superblock when problems encounter.
2214 * 3. Performs writes following reads for array synchronising.
2215 */
2216
2217static void fix_read_error(struct r1conf *conf, int read_disk,
2218 sector_t sect, int sectors)
2219{
2220 struct mddev *mddev = conf->mddev;
2221 while(sectors) {
2222 int s = sectors;
2223 int d = read_disk;
2224 int success = 0;
2225 int start;
2226 struct md_rdev *rdev;
2227
2228 if (s > (PAGE_SIZE>>9))
2229 s = PAGE_SIZE >> 9;
2230
2231 do {
2232 sector_t first_bad;
2233 int bad_sectors;
2234
2235 rcu_read_lock();
2236 rdev = rcu_dereference(conf->mirrors[d].rdev);
2237 if (rdev &&
2238 (test_bit(In_sync, &rdev->flags) ||
2239 (!test_bit(Faulty, &rdev->flags) &&
2240 rdev->recovery_offset >= sect + s)) &&
2241 is_badblock(rdev, sect, s,
2242 &first_bad, &bad_sectors) == 0) {
2243 atomic_inc(&rdev->nr_pending);
2244 rcu_read_unlock();
2245 if (sync_page_io(rdev, sect, s<<9,
2246 conf->tmppage, REQ_OP_READ, 0, false))
2247 success = 1;
2248 rdev_dec_pending(rdev, mddev);
2249 if (success)
2250 break;
2251 } else
2252 rcu_read_unlock();
2253 d++;
2254 if (d == conf->raid_disks * 2)
2255 d = 0;
2256 } while (!success && d != read_disk);
2257
2258 if (!success) {
2259 /* Cannot read from anywhere - mark it bad */
2260 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2261 if (!rdev_set_badblocks(rdev, sect, s, 0))
2262 md_error(mddev, rdev);
2263 break;
2264 }
2265 /* write it back and re-read */
2266 start = d;
2267 while (d != read_disk) {
2268 if (d==0)
2269 d = conf->raid_disks * 2;
2270 d--;
2271 rcu_read_lock();
2272 rdev = rcu_dereference(conf->mirrors[d].rdev);
2273 if (rdev &&
2274 !test_bit(Faulty, &rdev->flags)) {
2275 atomic_inc(&rdev->nr_pending);
2276 rcu_read_unlock();
2277 r1_sync_page_io(rdev, sect, s,
2278 conf->tmppage, WRITE);
2279 rdev_dec_pending(rdev, mddev);
2280 } else
2281 rcu_read_unlock();
2282 }
2283 d = start;
2284 while (d != read_disk) {
2285 char b[BDEVNAME_SIZE];
2286 if (d==0)
2287 d = conf->raid_disks * 2;
2288 d--;
2289 rcu_read_lock();
2290 rdev = rcu_dereference(conf->mirrors[d].rdev);
2291 if (rdev &&
2292 !test_bit(Faulty, &rdev->flags)) {
2293 atomic_inc(&rdev->nr_pending);
2294 rcu_read_unlock();
2295 if (r1_sync_page_io(rdev, sect, s,
2296 conf->tmppage, READ)) {
2297 atomic_add(s, &rdev->corrected_errors);
2298 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2299 mdname(mddev), s,
2300 (unsigned long long)(sect +
2301 rdev->data_offset),
2302 bdevname(rdev->bdev, b));
2303 }
2304 rdev_dec_pending(rdev, mddev);
2305 } else
2306 rcu_read_unlock();
2307 }
2308 sectors -= s;
2309 sect += s;
2310 }
2311}
2312
2313static int narrow_write_error(struct r1bio *r1_bio, int i)
2314{
2315 struct mddev *mddev = r1_bio->mddev;
2316 struct r1conf *conf = mddev->private;
2317 struct md_rdev *rdev = conf->mirrors[i].rdev;
2318
2319 /* bio has the data to be written to device 'i' where
2320 * we just recently had a write error.
2321 * We repeatedly clone the bio and trim down to one block,
2322 * then try the write. Where the write fails we record
2323 * a bad block.
2324 * It is conceivable that the bio doesn't exactly align with
2325 * blocks. We must handle this somehow.
2326 *
2327 * We currently own a reference on the rdev.
2328 */
2329
2330 int block_sectors;
2331 sector_t sector;
2332 int sectors;
2333 int sect_to_write = r1_bio->sectors;
2334 int ok = 1;
2335
2336 if (rdev->badblocks.shift < 0)
2337 return 0;
2338
2339 block_sectors = roundup(1 << rdev->badblocks.shift,
2340 bdev_logical_block_size(rdev->bdev) >> 9);
2341 sector = r1_bio->sector;
2342 sectors = ((sector + block_sectors)
2343 & ~(sector_t)(block_sectors - 1))
2344 - sector;
2345
2346 while (sect_to_write) {
2347 struct bio *wbio;
2348 if (sectors > sect_to_write)
2349 sectors = sect_to_write;
2350 /* Write at 'sector' for 'sectors'*/
2351
2352 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2353 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2354 GFP_NOIO,
2355 &mddev->bio_set);
2356 } else {
2357 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2358 &mddev->bio_set);
2359 }
2360
2361 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2362 wbio->bi_iter.bi_sector = r1_bio->sector;
2363 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2364
2365 bio_trim(wbio, sector - r1_bio->sector, sectors);
2366 wbio->bi_iter.bi_sector += rdev->data_offset;
2367 bio_set_dev(wbio, rdev->bdev);
2368
2369 if (submit_bio_wait(wbio) < 0)
2370 /* failure! */
2371 ok = rdev_set_badblocks(rdev, sector,
2372 sectors, 0)
2373 && ok;
2374
2375 bio_put(wbio);
2376 sect_to_write -= sectors;
2377 sector += sectors;
2378 sectors = block_sectors;
2379 }
2380 return ok;
2381}
2382
2383static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2384{
2385 int m;
2386 int s = r1_bio->sectors;
2387 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2388 struct md_rdev *rdev = conf->mirrors[m].rdev;
2389 struct bio *bio = r1_bio->bios[m];
2390 if (bio->bi_end_io == NULL)
2391 continue;
2392 if (!bio->bi_status &&
2393 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2394 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2395 }
2396 if (bio->bi_status &&
2397 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2398 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2399 md_error(conf->mddev, rdev);
2400 }
2401 }
2402 put_buf(r1_bio);
2403 md_done_sync(conf->mddev, s, 1);
2404}
2405
2406static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2407{
2408 int m, idx;
2409 bool fail = false;
2410
2411 for (m = 0; m < conf->raid_disks * 2 ; m++)
2412 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2413 struct md_rdev *rdev = conf->mirrors[m].rdev;
2414 rdev_clear_badblocks(rdev,
2415 r1_bio->sector,
2416 r1_bio->sectors, 0);
2417 rdev_dec_pending(rdev, conf->mddev);
2418 } else if (r1_bio->bios[m] != NULL) {
2419 /* This drive got a write error. We need to
2420 * narrow down and record precise write
2421 * errors.
2422 */
2423 fail = true;
2424 if (!narrow_write_error(r1_bio, m)) {
2425 md_error(conf->mddev,
2426 conf->mirrors[m].rdev);
2427 /* an I/O failed, we can't clear the bitmap */
2428 set_bit(R1BIO_Degraded, &r1_bio->state);
2429 }
2430 rdev_dec_pending(conf->mirrors[m].rdev,
2431 conf->mddev);
2432 }
2433 if (fail) {
2434 spin_lock_irq(&conf->device_lock);
2435 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2436 idx = sector_to_idx(r1_bio->sector);
2437 atomic_inc(&conf->nr_queued[idx]);
2438 spin_unlock_irq(&conf->device_lock);
2439 /*
2440 * In case freeze_array() is waiting for condition
2441 * get_unqueued_pending() == extra to be true.
2442 */
2443 wake_up(&conf->wait_barrier);
2444 md_wakeup_thread(conf->mddev->thread);
2445 } else {
2446 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2447 close_write(r1_bio);
2448 raid_end_bio_io(r1_bio);
2449 }
2450}
2451
2452static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2453{
2454 struct mddev *mddev = conf->mddev;
2455 struct bio *bio;
2456 struct md_rdev *rdev;
2457
2458 clear_bit(R1BIO_ReadError, &r1_bio->state);
2459 /* we got a read error. Maybe the drive is bad. Maybe just
2460 * the block and we can fix it.
2461 * We freeze all other IO, and try reading the block from
2462 * other devices. When we find one, we re-write
2463 * and check it that fixes the read error.
2464 * This is all done synchronously while the array is
2465 * frozen
2466 */
2467
2468 bio = r1_bio->bios[r1_bio->read_disk];
2469 bio_put(bio);
2470 r1_bio->bios[r1_bio->read_disk] = NULL;
2471
2472 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2473 if (mddev->ro == 0
2474 && !test_bit(FailFast, &rdev->flags)) {
2475 freeze_array(conf, 1);
2476 fix_read_error(conf, r1_bio->read_disk,
2477 r1_bio->sector, r1_bio->sectors);
2478 unfreeze_array(conf);
2479 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2480 md_error(mddev, rdev);
2481 } else {
2482 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2483 }
2484
2485 rdev_dec_pending(rdev, conf->mddev);
2486 allow_barrier(conf, r1_bio->sector);
2487 bio = r1_bio->master_bio;
2488
2489 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2490 r1_bio->state = 0;
2491 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2492}
2493
2494static void raid1d(struct md_thread *thread)
2495{
2496 struct mddev *mddev = thread->mddev;
2497 struct r1bio *r1_bio;
2498 unsigned long flags;
2499 struct r1conf *conf = mddev->private;
2500 struct list_head *head = &conf->retry_list;
2501 struct blk_plug plug;
2502 int idx;
2503
2504 md_check_recovery(mddev);
2505
2506 if (!list_empty_careful(&conf->bio_end_io_list) &&
2507 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2508 LIST_HEAD(tmp);
2509 spin_lock_irqsave(&conf->device_lock, flags);
2510 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2511 list_splice_init(&conf->bio_end_io_list, &tmp);
2512 spin_unlock_irqrestore(&conf->device_lock, flags);
2513 while (!list_empty(&tmp)) {
2514 r1_bio = list_first_entry(&tmp, struct r1bio,
2515 retry_list);
2516 list_del(&r1_bio->retry_list);
2517 idx = sector_to_idx(r1_bio->sector);
2518 atomic_dec(&conf->nr_queued[idx]);
2519 if (mddev->degraded)
2520 set_bit(R1BIO_Degraded, &r1_bio->state);
2521 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2522 close_write(r1_bio);
2523 raid_end_bio_io(r1_bio);
2524 }
2525 }
2526
2527 blk_start_plug(&plug);
2528 for (;;) {
2529
2530 flush_pending_writes(conf);
2531
2532 spin_lock_irqsave(&conf->device_lock, flags);
2533 if (list_empty(head)) {
2534 spin_unlock_irqrestore(&conf->device_lock, flags);
2535 break;
2536 }
2537 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2538 list_del(head->prev);
2539 idx = sector_to_idx(r1_bio->sector);
2540 atomic_dec(&conf->nr_queued[idx]);
2541 spin_unlock_irqrestore(&conf->device_lock, flags);
2542
2543 mddev = r1_bio->mddev;
2544 conf = mddev->private;
2545 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2546 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2547 test_bit(R1BIO_WriteError, &r1_bio->state))
2548 handle_sync_write_finished(conf, r1_bio);
2549 else
2550 sync_request_write(mddev, r1_bio);
2551 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2552 test_bit(R1BIO_WriteError, &r1_bio->state))
2553 handle_write_finished(conf, r1_bio);
2554 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2555 handle_read_error(conf, r1_bio);
2556 else
2557 WARN_ON_ONCE(1);
2558
2559 cond_resched();
2560 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2561 md_check_recovery(mddev);
2562 }
2563 blk_finish_plug(&plug);
2564}
2565
2566static int init_resync(struct r1conf *conf)
2567{
2568 int buffs;
2569
2570 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2571 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2572
2573 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2574 r1buf_pool_free, conf->poolinfo);
2575}
2576
2577static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2578{
2579 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2580 struct resync_pages *rps;
2581 struct bio *bio;
2582 int i;
2583
2584 for (i = conf->poolinfo->raid_disks; i--; ) {
2585 bio = r1bio->bios[i];
2586 rps = bio->bi_private;
2587 bio_reset(bio);
2588 bio->bi_private = rps;
2589 }
2590 r1bio->master_bio = NULL;
2591 return r1bio;
2592}
2593
2594/*
2595 * perform a "sync" on one "block"
2596 *
2597 * We need to make sure that no normal I/O request - particularly write
2598 * requests - conflict with active sync requests.
2599 *
2600 * This is achieved by tracking pending requests and a 'barrier' concept
2601 * that can be installed to exclude normal IO requests.
2602 */
2603
2604static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2605 int *skipped)
2606{
2607 struct r1conf *conf = mddev->private;
2608 struct r1bio *r1_bio;
2609 struct bio *bio;
2610 sector_t max_sector, nr_sectors;
2611 int disk = -1;
2612 int i;
2613 int wonly = -1;
2614 int write_targets = 0, read_targets = 0;
2615 sector_t sync_blocks;
2616 int still_degraded = 0;
2617 int good_sectors = RESYNC_SECTORS;
2618 int min_bad = 0; /* number of sectors that are bad in all devices */
2619 int idx = sector_to_idx(sector_nr);
2620 int page_idx = 0;
2621
2622 if (!mempool_initialized(&conf->r1buf_pool))
2623 if (init_resync(conf))
2624 return 0;
2625
2626 max_sector = mddev->dev_sectors;
2627 if (sector_nr >= max_sector) {
2628 /* If we aborted, we need to abort the
2629 * sync on the 'current' bitmap chunk (there will
2630 * only be one in raid1 resync.
2631 * We can find the current addess in mddev->curr_resync
2632 */
2633 if (mddev->curr_resync < max_sector) /* aborted */
2634 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2635 &sync_blocks, 1);
2636 else /* completed sync */
2637 conf->fullsync = 0;
2638
2639 md_bitmap_close_sync(mddev->bitmap);
2640 close_sync(conf);
2641
2642 if (mddev_is_clustered(mddev)) {
2643 conf->cluster_sync_low = 0;
2644 conf->cluster_sync_high = 0;
2645 }
2646 return 0;
2647 }
2648
2649 if (mddev->bitmap == NULL &&
2650 mddev->recovery_cp == MaxSector &&
2651 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2652 conf->fullsync == 0) {
2653 *skipped = 1;
2654 return max_sector - sector_nr;
2655 }
2656 /* before building a request, check if we can skip these blocks..
2657 * This call the bitmap_start_sync doesn't actually record anything
2658 */
2659 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2660 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2661 /* We can skip this block, and probably several more */
2662 *skipped = 1;
2663 return sync_blocks;
2664 }
2665
2666 /*
2667 * If there is non-resync activity waiting for a turn, then let it
2668 * though before starting on this new sync request.
2669 */
2670 if (atomic_read(&conf->nr_waiting[idx]))
2671 schedule_timeout_uninterruptible(1);
2672
2673 /* we are incrementing sector_nr below. To be safe, we check against
2674 * sector_nr + two times RESYNC_SECTORS
2675 */
2676
2677 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2678 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2679
2680
2681 if (raise_barrier(conf, sector_nr))
2682 return 0;
2683
2684 r1_bio = raid1_alloc_init_r1buf(conf);
2685
2686 rcu_read_lock();
2687 /*
2688 * If we get a correctably read error during resync or recovery,
2689 * we might want to read from a different device. So we
2690 * flag all drives that could conceivably be read from for READ,
2691 * and any others (which will be non-In_sync devices) for WRITE.
2692 * If a read fails, we try reading from something else for which READ
2693 * is OK.
2694 */
2695
2696 r1_bio->mddev = mddev;
2697 r1_bio->sector = sector_nr;
2698 r1_bio->state = 0;
2699 set_bit(R1BIO_IsSync, &r1_bio->state);
2700 /* make sure good_sectors won't go across barrier unit boundary */
2701 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2702
2703 for (i = 0; i < conf->raid_disks * 2; i++) {
2704 struct md_rdev *rdev;
2705 bio = r1_bio->bios[i];
2706
2707 rdev = rcu_dereference(conf->mirrors[i].rdev);
2708 if (rdev == NULL ||
2709 test_bit(Faulty, &rdev->flags)) {
2710 if (i < conf->raid_disks)
2711 still_degraded = 1;
2712 } else if (!test_bit(In_sync, &rdev->flags)) {
2713 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2714 bio->bi_end_io = end_sync_write;
2715 write_targets ++;
2716 } else {
2717 /* may need to read from here */
2718 sector_t first_bad = MaxSector;
2719 int bad_sectors;
2720
2721 if (is_badblock(rdev, sector_nr, good_sectors,
2722 &first_bad, &bad_sectors)) {
2723 if (first_bad > sector_nr)
2724 good_sectors = first_bad - sector_nr;
2725 else {
2726 bad_sectors -= (sector_nr - first_bad);
2727 if (min_bad == 0 ||
2728 min_bad > bad_sectors)
2729 min_bad = bad_sectors;
2730 }
2731 }
2732 if (sector_nr < first_bad) {
2733 if (test_bit(WriteMostly, &rdev->flags)) {
2734 if (wonly < 0)
2735 wonly = i;
2736 } else {
2737 if (disk < 0)
2738 disk = i;
2739 }
2740 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2741 bio->bi_end_io = end_sync_read;
2742 read_targets++;
2743 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2744 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2745 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2746 /*
2747 * The device is suitable for reading (InSync),
2748 * but has bad block(s) here. Let's try to correct them,
2749 * if we are doing resync or repair. Otherwise, leave
2750 * this device alone for this sync request.
2751 */
2752 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2753 bio->bi_end_io = end_sync_write;
2754 write_targets++;
2755 }
2756 }
2757 if (bio->bi_end_io) {
2758 atomic_inc(&rdev->nr_pending);
2759 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2760 bio_set_dev(bio, rdev->bdev);
2761 if (test_bit(FailFast, &rdev->flags))
2762 bio->bi_opf |= MD_FAILFAST;
2763 }
2764 }
2765 rcu_read_unlock();
2766 if (disk < 0)
2767 disk = wonly;
2768 r1_bio->read_disk = disk;
2769
2770 if (read_targets == 0 && min_bad > 0) {
2771 /* These sectors are bad on all InSync devices, so we
2772 * need to mark them bad on all write targets
2773 */
2774 int ok = 1;
2775 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2776 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2777 struct md_rdev *rdev = conf->mirrors[i].rdev;
2778 ok = rdev_set_badblocks(rdev, sector_nr,
2779 min_bad, 0
2780 ) && ok;
2781 }
2782 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2783 *skipped = 1;
2784 put_buf(r1_bio);
2785
2786 if (!ok) {
2787 /* Cannot record the badblocks, so need to
2788 * abort the resync.
2789 * If there are multiple read targets, could just
2790 * fail the really bad ones ???
2791 */
2792 conf->recovery_disabled = mddev->recovery_disabled;
2793 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2794 return 0;
2795 } else
2796 return min_bad;
2797
2798 }
2799 if (min_bad > 0 && min_bad < good_sectors) {
2800 /* only resync enough to reach the next bad->good
2801 * transition */
2802 good_sectors = min_bad;
2803 }
2804
2805 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2806 /* extra read targets are also write targets */
2807 write_targets += read_targets-1;
2808
2809 if (write_targets == 0 || read_targets == 0) {
2810 /* There is nowhere to write, so all non-sync
2811 * drives must be failed - so we are finished
2812 */
2813 sector_t rv;
2814 if (min_bad > 0)
2815 max_sector = sector_nr + min_bad;
2816 rv = max_sector - sector_nr;
2817 *skipped = 1;
2818 put_buf(r1_bio);
2819 return rv;
2820 }
2821
2822 if (max_sector > mddev->resync_max)
2823 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2824 if (max_sector > sector_nr + good_sectors)
2825 max_sector = sector_nr + good_sectors;
2826 nr_sectors = 0;
2827 sync_blocks = 0;
2828 do {
2829 struct page *page;
2830 int len = PAGE_SIZE;
2831 if (sector_nr + (len>>9) > max_sector)
2832 len = (max_sector - sector_nr) << 9;
2833 if (len == 0)
2834 break;
2835 if (sync_blocks == 0) {
2836 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2837 &sync_blocks, still_degraded) &&
2838 !conf->fullsync &&
2839 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2840 break;
2841 if ((len >> 9) > sync_blocks)
2842 len = sync_blocks<<9;
2843 }
2844
2845 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2846 struct resync_pages *rp;
2847
2848 bio = r1_bio->bios[i];
2849 rp = get_resync_pages(bio);
2850 if (bio->bi_end_io) {
2851 page = resync_fetch_page(rp, page_idx);
2852
2853 /*
2854 * won't fail because the vec table is big
2855 * enough to hold all these pages
2856 */
2857 bio_add_page(bio, page, len, 0);
2858 }
2859 }
2860 nr_sectors += len>>9;
2861 sector_nr += len>>9;
2862 sync_blocks -= (len>>9);
2863 } while (++page_idx < RESYNC_PAGES);
2864
2865 r1_bio->sectors = nr_sectors;
2866
2867 if (mddev_is_clustered(mddev) &&
2868 conf->cluster_sync_high < sector_nr + nr_sectors) {
2869 conf->cluster_sync_low = mddev->curr_resync_completed;
2870 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2871 /* Send resync message */
2872 md_cluster_ops->resync_info_update(mddev,
2873 conf->cluster_sync_low,
2874 conf->cluster_sync_high);
2875 }
2876
2877 /* For a user-requested sync, we read all readable devices and do a
2878 * compare
2879 */
2880 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2881 atomic_set(&r1_bio->remaining, read_targets);
2882 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2883 bio = r1_bio->bios[i];
2884 if (bio->bi_end_io == end_sync_read) {
2885 read_targets--;
2886 md_sync_acct_bio(bio, nr_sectors);
2887 if (read_targets == 1)
2888 bio->bi_opf &= ~MD_FAILFAST;
2889 generic_make_request(bio);
2890 }
2891 }
2892 } else {
2893 atomic_set(&r1_bio->remaining, 1);
2894 bio = r1_bio->bios[r1_bio->read_disk];
2895 md_sync_acct_bio(bio, nr_sectors);
2896 if (read_targets == 1)
2897 bio->bi_opf &= ~MD_FAILFAST;
2898 generic_make_request(bio);
2899
2900 }
2901 return nr_sectors;
2902}
2903
2904static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2905{
2906 if (sectors)
2907 return sectors;
2908
2909 return mddev->dev_sectors;
2910}
2911
2912static struct r1conf *setup_conf(struct mddev *mddev)
2913{
2914 struct r1conf *conf;
2915 int i;
2916 struct raid1_info *disk;
2917 struct md_rdev *rdev;
2918 int err = -ENOMEM;
2919
2920 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2921 if (!conf)
2922 goto abort;
2923
2924 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2925 sizeof(atomic_t), GFP_KERNEL);
2926 if (!conf->nr_pending)
2927 goto abort;
2928
2929 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2930 sizeof(atomic_t), GFP_KERNEL);
2931 if (!conf->nr_waiting)
2932 goto abort;
2933
2934 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2935 sizeof(atomic_t), GFP_KERNEL);
2936 if (!conf->nr_queued)
2937 goto abort;
2938
2939 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2940 sizeof(atomic_t), GFP_KERNEL);
2941 if (!conf->barrier)
2942 goto abort;
2943
2944 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2945 mddev->raid_disks, 2),
2946 GFP_KERNEL);
2947 if (!conf->mirrors)
2948 goto abort;
2949
2950 conf->tmppage = alloc_page(GFP_KERNEL);
2951 if (!conf->tmppage)
2952 goto abort;
2953
2954 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2955 if (!conf->poolinfo)
2956 goto abort;
2957 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2958 err = mempool_init(&conf->r1bio_pool, NR_RAID1_BIOS, r1bio_pool_alloc,
2959 r1bio_pool_free, conf->poolinfo);
2960 if (err)
2961 goto abort;
2962
2963 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2964 if (err)
2965 goto abort;
2966
2967 conf->poolinfo->mddev = mddev;
2968
2969 err = -EINVAL;
2970 spin_lock_init(&conf->device_lock);
2971 rdev_for_each(rdev, mddev) {
2972 int disk_idx = rdev->raid_disk;
2973 if (disk_idx >= mddev->raid_disks
2974 || disk_idx < 0)
2975 continue;
2976 if (test_bit(Replacement, &rdev->flags))
2977 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2978 else
2979 disk = conf->mirrors + disk_idx;
2980
2981 if (disk->rdev)
2982 goto abort;
2983 disk->rdev = rdev;
2984 disk->head_position = 0;
2985 disk->seq_start = MaxSector;
2986 }
2987 conf->raid_disks = mddev->raid_disks;
2988 conf->mddev = mddev;
2989 INIT_LIST_HEAD(&conf->retry_list);
2990 INIT_LIST_HEAD(&conf->bio_end_io_list);
2991
2992 spin_lock_init(&conf->resync_lock);
2993 init_waitqueue_head(&conf->wait_barrier);
2994
2995 bio_list_init(&conf->pending_bio_list);
2996 conf->pending_count = 0;
2997 conf->recovery_disabled = mddev->recovery_disabled - 1;
2998
2999 err = -EIO;
3000 for (i = 0; i < conf->raid_disks * 2; i++) {
3001
3002 disk = conf->mirrors + i;
3003
3004 if (i < conf->raid_disks &&
3005 disk[conf->raid_disks].rdev) {
3006 /* This slot has a replacement. */
3007 if (!disk->rdev) {
3008 /* No original, just make the replacement
3009 * a recovering spare
3010 */
3011 disk->rdev =
3012 disk[conf->raid_disks].rdev;
3013 disk[conf->raid_disks].rdev = NULL;
3014 } else if (!test_bit(In_sync, &disk->rdev->flags))
3015 /* Original is not in_sync - bad */
3016 goto abort;
3017 }
3018
3019 if (!disk->rdev ||
3020 !test_bit(In_sync, &disk->rdev->flags)) {
3021 disk->head_position = 0;
3022 if (disk->rdev &&
3023 (disk->rdev->saved_raid_disk < 0))
3024 conf->fullsync = 1;
3025 }
3026 }
3027
3028 err = -ENOMEM;
3029 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3030 if (!conf->thread)
3031 goto abort;
3032
3033 return conf;
3034
3035 abort:
3036 if (conf) {
3037 mempool_exit(&conf->r1bio_pool);
3038 kfree(conf->mirrors);
3039 safe_put_page(conf->tmppage);
3040 kfree(conf->poolinfo);
3041 kfree(conf->nr_pending);
3042 kfree(conf->nr_waiting);
3043 kfree(conf->nr_queued);
3044 kfree(conf->barrier);
3045 bioset_exit(&conf->bio_split);
3046 kfree(conf);
3047 }
3048 return ERR_PTR(err);
3049}
3050
3051static void raid1_free(struct mddev *mddev, void *priv);
3052static int raid1_run(struct mddev *mddev)
3053{
3054 struct r1conf *conf;
3055 int i;
3056 struct md_rdev *rdev;
3057 int ret;
3058 bool discard_supported = false;
3059
3060 if (mddev->level != 1) {
3061 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3062 mdname(mddev), mddev->level);
3063 return -EIO;
3064 }
3065 if (mddev->reshape_position != MaxSector) {
3066 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3067 mdname(mddev));
3068 return -EIO;
3069 }
3070 if (mddev_init_writes_pending(mddev) < 0)
3071 return -ENOMEM;
3072 /*
3073 * copy the already verified devices into our private RAID1
3074 * bookkeeping area. [whatever we allocate in run(),
3075 * should be freed in raid1_free()]
3076 */
3077 if (mddev->private == NULL)
3078 conf = setup_conf(mddev);
3079 else
3080 conf = mddev->private;
3081
3082 if (IS_ERR(conf))
3083 return PTR_ERR(conf);
3084
3085 if (mddev->queue) {
3086 blk_queue_max_write_same_sectors(mddev->queue, 0);
3087 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3088 }
3089
3090 rdev_for_each(rdev, mddev) {
3091 if (!mddev->gendisk)
3092 continue;
3093 disk_stack_limits(mddev->gendisk, rdev->bdev,
3094 rdev->data_offset << 9);
3095 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3096 discard_supported = true;
3097 }
3098
3099 mddev->degraded = 0;
3100 for (i=0; i < conf->raid_disks; i++)
3101 if (conf->mirrors[i].rdev == NULL ||
3102 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3103 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3104 mddev->degraded++;
3105
3106 if (conf->raid_disks - mddev->degraded == 1)
3107 mddev->recovery_cp = MaxSector;
3108
3109 if (mddev->recovery_cp != MaxSector)
3110 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3111 mdname(mddev));
3112 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3113 mdname(mddev), mddev->raid_disks - mddev->degraded,
3114 mddev->raid_disks);
3115
3116 /*
3117 * Ok, everything is just fine now
3118 */
3119 mddev->thread = conf->thread;
3120 conf->thread = NULL;
3121 mddev->private = conf;
3122 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3123
3124 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3125
3126 if (mddev->queue) {
3127 if (discard_supported)
3128 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3129 mddev->queue);
3130 else
3131 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3132 mddev->queue);
3133 }
3134
3135 ret = md_integrity_register(mddev);
3136 if (ret) {
3137 md_unregister_thread(&mddev->thread);
3138 raid1_free(mddev, conf);
3139 }
3140 return ret;
3141}
3142
3143static void raid1_free(struct mddev *mddev, void *priv)
3144{
3145 struct r1conf *conf = priv;
3146
3147 mempool_exit(&conf->r1bio_pool);
3148 kfree(conf->mirrors);
3149 safe_put_page(conf->tmppage);
3150 kfree(conf->poolinfo);
3151 kfree(conf->nr_pending);
3152 kfree(conf->nr_waiting);
3153 kfree(conf->nr_queued);
3154 kfree(conf->barrier);
3155 bioset_exit(&conf->bio_split);
3156 kfree(conf);
3157}
3158
3159static int raid1_resize(struct mddev *mddev, sector_t sectors)
3160{
3161 /* no resync is happening, and there is enough space
3162 * on all devices, so we can resize.
3163 * We need to make sure resync covers any new space.
3164 * If the array is shrinking we should possibly wait until
3165 * any io in the removed space completes, but it hardly seems
3166 * worth it.
3167 */
3168 sector_t newsize = raid1_size(mddev, sectors, 0);
3169 if (mddev->external_size &&
3170 mddev->array_sectors > newsize)
3171 return -EINVAL;
3172 if (mddev->bitmap) {
3173 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3174 if (ret)
3175 return ret;
3176 }
3177 md_set_array_sectors(mddev, newsize);
3178 if (sectors > mddev->dev_sectors &&
3179 mddev->recovery_cp > mddev->dev_sectors) {
3180 mddev->recovery_cp = mddev->dev_sectors;
3181 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3182 }
3183 mddev->dev_sectors = sectors;
3184 mddev->resync_max_sectors = sectors;
3185 return 0;
3186}
3187
3188static int raid1_reshape(struct mddev *mddev)
3189{
3190 /* We need to:
3191 * 1/ resize the r1bio_pool
3192 * 2/ resize conf->mirrors
3193 *
3194 * We allocate a new r1bio_pool if we can.
3195 * Then raise a device barrier and wait until all IO stops.
3196 * Then resize conf->mirrors and swap in the new r1bio pool.
3197 *
3198 * At the same time, we "pack" the devices so that all the missing
3199 * devices have the higher raid_disk numbers.
3200 */
3201 mempool_t newpool, oldpool;
3202 struct pool_info *newpoolinfo;
3203 struct raid1_info *newmirrors;
3204 struct r1conf *conf = mddev->private;
3205 int cnt, raid_disks;
3206 unsigned long flags;
3207 int d, d2;
3208 int ret;
3209
3210 memset(&newpool, 0, sizeof(newpool));
3211 memset(&oldpool, 0, sizeof(oldpool));
3212
3213 /* Cannot change chunk_size, layout, or level */
3214 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3215 mddev->layout != mddev->new_layout ||
3216 mddev->level != mddev->new_level) {
3217 mddev->new_chunk_sectors = mddev->chunk_sectors;
3218 mddev->new_layout = mddev->layout;
3219 mddev->new_level = mddev->level;
3220 return -EINVAL;
3221 }
3222
3223 if (!mddev_is_clustered(mddev))
3224 md_allow_write(mddev);
3225
3226 raid_disks = mddev->raid_disks + mddev->delta_disks;
3227
3228 if (raid_disks < conf->raid_disks) {
3229 cnt=0;
3230 for (d= 0; d < conf->raid_disks; d++)
3231 if (conf->mirrors[d].rdev)
3232 cnt++;
3233 if (cnt > raid_disks)
3234 return -EBUSY;
3235 }
3236
3237 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3238 if (!newpoolinfo)
3239 return -ENOMEM;
3240 newpoolinfo->mddev = mddev;
3241 newpoolinfo->raid_disks = raid_disks * 2;
3242
3243 ret = mempool_init(&newpool, NR_RAID1_BIOS, r1bio_pool_alloc,
3244 r1bio_pool_free, newpoolinfo);
3245 if (ret) {
3246 kfree(newpoolinfo);
3247 return ret;
3248 }
3249 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3250 raid_disks, 2),
3251 GFP_KERNEL);
3252 if (!newmirrors) {
3253 kfree(newpoolinfo);
3254 mempool_exit(&newpool);
3255 return -ENOMEM;
3256 }
3257
3258 freeze_array(conf, 0);
3259
3260 /* ok, everything is stopped */
3261 oldpool = conf->r1bio_pool;
3262 conf->r1bio_pool = newpool;
3263
3264 for (d = d2 = 0; d < conf->raid_disks; d++) {
3265 struct md_rdev *rdev = conf->mirrors[d].rdev;
3266 if (rdev && rdev->raid_disk != d2) {
3267 sysfs_unlink_rdev(mddev, rdev);
3268 rdev->raid_disk = d2;
3269 sysfs_unlink_rdev(mddev, rdev);
3270 if (sysfs_link_rdev(mddev, rdev))
3271 pr_warn("md/raid1:%s: cannot register rd%d\n",
3272 mdname(mddev), rdev->raid_disk);
3273 }
3274 if (rdev)
3275 newmirrors[d2++].rdev = rdev;
3276 }
3277 kfree(conf->mirrors);
3278 conf->mirrors = newmirrors;
3279 kfree(conf->poolinfo);
3280 conf->poolinfo = newpoolinfo;
3281
3282 spin_lock_irqsave(&conf->device_lock, flags);
3283 mddev->degraded += (raid_disks - conf->raid_disks);
3284 spin_unlock_irqrestore(&conf->device_lock, flags);
3285 conf->raid_disks = mddev->raid_disks = raid_disks;
3286 mddev->delta_disks = 0;
3287
3288 unfreeze_array(conf);
3289
3290 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3291 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3292 md_wakeup_thread(mddev->thread);
3293
3294 mempool_exit(&oldpool);
3295 return 0;
3296}
3297
3298static void raid1_quiesce(struct mddev *mddev, int quiesce)
3299{
3300 struct r1conf *conf = mddev->private;
3301
3302 if (quiesce)
3303 freeze_array(conf, 0);
3304 else
3305 unfreeze_array(conf);
3306}
3307
3308static void *raid1_takeover(struct mddev *mddev)
3309{
3310 /* raid1 can take over:
3311 * raid5 with 2 devices, any layout or chunk size
3312 */
3313 if (mddev->level == 5 && mddev->raid_disks == 2) {
3314 struct r1conf *conf;
3315 mddev->new_level = 1;
3316 mddev->new_layout = 0;
3317 mddev->new_chunk_sectors = 0;
3318 conf = setup_conf(mddev);
3319 if (!IS_ERR(conf)) {
3320 /* Array m