Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid1.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 *
7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 *
9 * RAID-1 management functions.
10 *
11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 *
13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 *
16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
17 * bitmapped intelligence in resync:
18 *
19 * - bitmap marked during normal i/o
20 * - bitmap used to skip nondirty blocks during sync
21 *
22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
23 * - persistent bitmap code
24 */
25
26#include <linux/slab.h>
27#include <linux/delay.h>
28#include <linux/blkdev.h>
29#include <linux/module.h>
30#include <linux/seq_file.h>
31#include <linux/ratelimit.h>
32#include <linux/interval_tree_generic.h>
33
34#include <trace/events/block.h>
35
36#include "md.h"
37#include "raid1.h"
38#include "md-bitmap.h"
39
40#define UNSUPPORTED_MDDEV_FLAGS \
41 ((1L << MD_HAS_JOURNAL) | \
42 (1L << MD_JOURNAL_CLEAN) | \
43 (1L << MD_HAS_PPL) | \
44 (1L << MD_HAS_MULTIPLE_PPLS))
45
46static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
47static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
48
49#define RAID_1_10_NAME "raid1"
50#include "raid1-10.c"
51
52#define START(node) ((node)->start)
53#define LAST(node) ((node)->last)
54INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
55 START, LAST, static inline, raid1_rb);
56
57static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
58 struct serial_info *si, int idx)
59{
60 unsigned long flags;
61 int ret = 0;
62 sector_t lo = r1_bio->sector;
63 sector_t hi = lo + r1_bio->sectors;
64 struct serial_in_rdev *serial = &rdev->serial[idx];
65
66 spin_lock_irqsave(&serial->serial_lock, flags);
67 /* collision happened */
68 if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
69 ret = -EBUSY;
70 else {
71 si->start = lo;
72 si->last = hi;
73 raid1_rb_insert(si, &serial->serial_rb);
74 }
75 spin_unlock_irqrestore(&serial->serial_lock, flags);
76
77 return ret;
78}
79
80static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
81{
82 struct mddev *mddev = rdev->mddev;
83 struct serial_info *si;
84 int idx = sector_to_idx(r1_bio->sector);
85 struct serial_in_rdev *serial = &rdev->serial[idx];
86
87 if (WARN_ON(!mddev->serial_info_pool))
88 return;
89 si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
90 wait_event(serial->serial_io_wait,
91 check_and_add_serial(rdev, r1_bio, si, idx) == 0);
92}
93
94static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
95{
96 struct serial_info *si;
97 unsigned long flags;
98 int found = 0;
99 struct mddev *mddev = rdev->mddev;
100 int idx = sector_to_idx(lo);
101 struct serial_in_rdev *serial = &rdev->serial[idx];
102
103 spin_lock_irqsave(&serial->serial_lock, flags);
104 for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
105 si; si = raid1_rb_iter_next(si, lo, hi)) {
106 if (si->start == lo && si->last == hi) {
107 raid1_rb_remove(si, &serial->serial_rb);
108 mempool_free(si, mddev->serial_info_pool);
109 found = 1;
110 break;
111 }
112 }
113 if (!found)
114 WARN(1, "The write IO is not recorded for serialization\n");
115 spin_unlock_irqrestore(&serial->serial_lock, flags);
116 wake_up(&serial->serial_io_wait);
117}
118
119/*
120 * for resync bio, r1bio pointer can be retrieved from the per-bio
121 * 'struct resync_pages'.
122 */
123static inline struct r1bio *get_resync_r1bio(struct bio *bio)
124{
125 return get_resync_pages(bio)->raid_bio;
126}
127
128static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
129{
130 struct pool_info *pi = data;
131 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
132
133 /* allocate a r1bio with room for raid_disks entries in the bios array */
134 return kzalloc(size, gfp_flags);
135}
136
137#define RESYNC_DEPTH 32
138#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
139#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
140#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
141#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
142#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
143
144static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
145{
146 struct pool_info *pi = data;
147 struct r1bio *r1_bio;
148 struct bio *bio;
149 int need_pages;
150 int j;
151 struct resync_pages *rps;
152
153 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
154 if (!r1_bio)
155 return NULL;
156
157 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
158 gfp_flags);
159 if (!rps)
160 goto out_free_r1bio;
161
162 /*
163 * Allocate bios : 1 for reading, n-1 for writing
164 */
165 for (j = pi->raid_disks ; j-- ; ) {
166 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
167 if (!bio)
168 goto out_free_bio;
169 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
170 r1_bio->bios[j] = bio;
171 }
172 /*
173 * Allocate RESYNC_PAGES data pages and attach them to
174 * the first bio.
175 * If this is a user-requested check/repair, allocate
176 * RESYNC_PAGES for each bio.
177 */
178 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
179 need_pages = pi->raid_disks;
180 else
181 need_pages = 1;
182 for (j = 0; j < pi->raid_disks; j++) {
183 struct resync_pages *rp = &rps[j];
184
185 bio = r1_bio->bios[j];
186
187 if (j < need_pages) {
188 if (resync_alloc_pages(rp, gfp_flags))
189 goto out_free_pages;
190 } else {
191 memcpy(rp, &rps[0], sizeof(*rp));
192 resync_get_all_pages(rp);
193 }
194
195 rp->raid_bio = r1_bio;
196 bio->bi_private = rp;
197 }
198
199 r1_bio->master_bio = NULL;
200
201 return r1_bio;
202
203out_free_pages:
204 while (--j >= 0)
205 resync_free_pages(&rps[j]);
206
207out_free_bio:
208 while (++j < pi->raid_disks) {
209 bio_uninit(r1_bio->bios[j]);
210 kfree(r1_bio->bios[j]);
211 }
212 kfree(rps);
213
214out_free_r1bio:
215 rbio_pool_free(r1_bio, data);
216 return NULL;
217}
218
219static void r1buf_pool_free(void *__r1_bio, void *data)
220{
221 struct pool_info *pi = data;
222 int i;
223 struct r1bio *r1bio = __r1_bio;
224 struct resync_pages *rp = NULL;
225
226 for (i = pi->raid_disks; i--; ) {
227 rp = get_resync_pages(r1bio->bios[i]);
228 resync_free_pages(rp);
229 bio_uninit(r1bio->bios[i]);
230 kfree(r1bio->bios[i]);
231 }
232
233 /* resync pages array stored in the 1st bio's .bi_private */
234 kfree(rp);
235
236 rbio_pool_free(r1bio, data);
237}
238
239static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
240{
241 int i;
242
243 for (i = 0; i < conf->raid_disks * 2; i++) {
244 struct bio **bio = r1_bio->bios + i;
245 if (!BIO_SPECIAL(*bio))
246 bio_put(*bio);
247 *bio = NULL;
248 }
249}
250
251static void free_r1bio(struct r1bio *r1_bio)
252{
253 struct r1conf *conf = r1_bio->mddev->private;
254
255 put_all_bios(conf, r1_bio);
256 mempool_free(r1_bio, &conf->r1bio_pool);
257}
258
259static void put_buf(struct r1bio *r1_bio)
260{
261 struct r1conf *conf = r1_bio->mddev->private;
262 sector_t sect = r1_bio->sector;
263 int i;
264
265 for (i = 0; i < conf->raid_disks * 2; i++) {
266 struct bio *bio = r1_bio->bios[i];
267 if (bio->bi_end_io)
268 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
269 }
270
271 mempool_free(r1_bio, &conf->r1buf_pool);
272
273 lower_barrier(conf, sect);
274}
275
276static void reschedule_retry(struct r1bio *r1_bio)
277{
278 unsigned long flags;
279 struct mddev *mddev = r1_bio->mddev;
280 struct r1conf *conf = mddev->private;
281 int idx;
282
283 idx = sector_to_idx(r1_bio->sector);
284 spin_lock_irqsave(&conf->device_lock, flags);
285 list_add(&r1_bio->retry_list, &conf->retry_list);
286 atomic_inc(&conf->nr_queued[idx]);
287 spin_unlock_irqrestore(&conf->device_lock, flags);
288
289 wake_up(&conf->wait_barrier);
290 md_wakeup_thread(mddev->thread);
291}
292
293/*
294 * raid_end_bio_io() is called when we have finished servicing a mirrored
295 * operation and are ready to return a success/failure code to the buffer
296 * cache layer.
297 */
298static void call_bio_endio(struct r1bio *r1_bio)
299{
300 struct bio *bio = r1_bio->master_bio;
301
302 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
303 bio->bi_status = BLK_STS_IOERR;
304
305 bio_endio(bio);
306}
307
308static void raid_end_bio_io(struct r1bio *r1_bio)
309{
310 struct bio *bio = r1_bio->master_bio;
311 struct r1conf *conf = r1_bio->mddev->private;
312 sector_t sector = r1_bio->sector;
313
314 /* if nobody has done the final endio yet, do it now */
315 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
316 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
317 (bio_data_dir(bio) == WRITE) ? "write" : "read",
318 (unsigned long long) bio->bi_iter.bi_sector,
319 (unsigned long long) bio_end_sector(bio) - 1);
320
321 call_bio_endio(r1_bio);
322 }
323
324 free_r1bio(r1_bio);
325 /*
326 * Wake up any possible resync thread that waits for the device
327 * to go idle. All I/Os, even write-behind writes, are done.
328 */
329 allow_barrier(conf, sector);
330}
331
332/*
333 * Update disk head position estimator based on IRQ completion info.
334 */
335static inline void update_head_pos(int disk, struct r1bio *r1_bio)
336{
337 struct r1conf *conf = r1_bio->mddev->private;
338
339 conf->mirrors[disk].head_position =
340 r1_bio->sector + (r1_bio->sectors);
341}
342
343/*
344 * Find the disk number which triggered given bio
345 */
346static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
347{
348 int mirror;
349 struct r1conf *conf = r1_bio->mddev->private;
350 int raid_disks = conf->raid_disks;
351
352 for (mirror = 0; mirror < raid_disks * 2; mirror++)
353 if (r1_bio->bios[mirror] == bio)
354 break;
355
356 BUG_ON(mirror == raid_disks * 2);
357 update_head_pos(mirror, r1_bio);
358
359 return mirror;
360}
361
362static void raid1_end_read_request(struct bio *bio)
363{
364 int uptodate = !bio->bi_status;
365 struct r1bio *r1_bio = bio->bi_private;
366 struct r1conf *conf = r1_bio->mddev->private;
367 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
368
369 /*
370 * this branch is our 'one mirror IO has finished' event handler:
371 */
372 update_head_pos(r1_bio->read_disk, r1_bio);
373
374 if (uptodate)
375 set_bit(R1BIO_Uptodate, &r1_bio->state);
376 else if (test_bit(FailFast, &rdev->flags) &&
377 test_bit(R1BIO_FailFast, &r1_bio->state))
378 /* This was a fail-fast read so we definitely
379 * want to retry */
380 ;
381 else {
382 /* If all other devices have failed, we want to return
383 * the error upwards rather than fail the last device.
384 * Here we redefine "uptodate" to mean "Don't want to retry"
385 */
386 unsigned long flags;
387 spin_lock_irqsave(&conf->device_lock, flags);
388 if (r1_bio->mddev->degraded == conf->raid_disks ||
389 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
390 test_bit(In_sync, &rdev->flags)))
391 uptodate = 1;
392 spin_unlock_irqrestore(&conf->device_lock, flags);
393 }
394
395 if (uptodate) {
396 raid_end_bio_io(r1_bio);
397 rdev_dec_pending(rdev, conf->mddev);
398 } else {
399 /*
400 * oops, read error:
401 */
402 pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
403 mdname(conf->mddev),
404 rdev->bdev,
405 (unsigned long long)r1_bio->sector);
406 set_bit(R1BIO_ReadError, &r1_bio->state);
407 reschedule_retry(r1_bio);
408 /* don't drop the reference on read_disk yet */
409 }
410}
411
412static void close_write(struct r1bio *r1_bio)
413{
414 struct mddev *mddev = r1_bio->mddev;
415
416 /* it really is the end of this request */
417 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
418 bio_free_pages(r1_bio->behind_master_bio);
419 bio_put(r1_bio->behind_master_bio);
420 r1_bio->behind_master_bio = NULL;
421 }
422
423 if (test_bit(R1BIO_BehindIO, &r1_bio->state))
424 mddev->bitmap_ops->end_behind_write(mddev);
425 md_write_end(mddev);
426}
427
428static void r1_bio_write_done(struct r1bio *r1_bio)
429{
430 if (!atomic_dec_and_test(&r1_bio->remaining))
431 return;
432
433 if (test_bit(R1BIO_WriteError, &r1_bio->state))
434 reschedule_retry(r1_bio);
435 else {
436 close_write(r1_bio);
437 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
438 reschedule_retry(r1_bio);
439 else
440 raid_end_bio_io(r1_bio);
441 }
442}
443
444static void raid1_end_write_request(struct bio *bio)
445{
446 struct r1bio *r1_bio = bio->bi_private;
447 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
448 struct r1conf *conf = r1_bio->mddev->private;
449 struct bio *to_put = NULL;
450 int mirror = find_bio_disk(r1_bio, bio);
451 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
452 bool discard_error;
453 sector_t lo = r1_bio->sector;
454 sector_t hi = r1_bio->sector + r1_bio->sectors;
455
456 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
457
458 /*
459 * 'one mirror IO has finished' event handler:
460 */
461 if (bio->bi_status && !discard_error) {
462 set_bit(WriteErrorSeen, &rdev->flags);
463 if (!test_and_set_bit(WantReplacement, &rdev->flags))
464 set_bit(MD_RECOVERY_NEEDED, &
465 conf->mddev->recovery);
466
467 if (test_bit(FailFast, &rdev->flags) &&
468 (bio->bi_opf & MD_FAILFAST) &&
469 /* We never try FailFast to WriteMostly devices */
470 !test_bit(WriteMostly, &rdev->flags)) {
471 md_error(r1_bio->mddev, rdev);
472 }
473
474 /*
475 * When the device is faulty, it is not necessary to
476 * handle write error.
477 */
478 if (!test_bit(Faulty, &rdev->flags))
479 set_bit(R1BIO_WriteError, &r1_bio->state);
480 else {
481 /* Finished with this branch */
482 r1_bio->bios[mirror] = NULL;
483 to_put = bio;
484 }
485 } else {
486 /*
487 * Set R1BIO_Uptodate in our master bio, so that we
488 * will return a good error code for to the higher
489 * levels even if IO on some other mirrored buffer
490 * fails.
491 *
492 * The 'master' represents the composite IO operation
493 * to user-side. So if something waits for IO, then it
494 * will wait for the 'master' bio.
495 */
496 r1_bio->bios[mirror] = NULL;
497 to_put = bio;
498 /*
499 * Do not set R1BIO_Uptodate if the current device is
500 * rebuilding or Faulty. This is because we cannot use
501 * such device for properly reading the data back (we could
502 * potentially use it, if the current write would have felt
503 * before rdev->recovery_offset, but for simplicity we don't
504 * check this here.
505 */
506 if (test_bit(In_sync, &rdev->flags) &&
507 !test_bit(Faulty, &rdev->flags))
508 set_bit(R1BIO_Uptodate, &r1_bio->state);
509
510 /* Maybe we can clear some bad blocks. */
511 if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
512 !discard_error) {
513 r1_bio->bios[mirror] = IO_MADE_GOOD;
514 set_bit(R1BIO_MadeGood, &r1_bio->state);
515 }
516 }
517
518 if (behind) {
519 if (test_bit(CollisionCheck, &rdev->flags))
520 remove_serial(rdev, lo, hi);
521 if (test_bit(WriteMostly, &rdev->flags))
522 atomic_dec(&r1_bio->behind_remaining);
523
524 /*
525 * In behind mode, we ACK the master bio once the I/O
526 * has safely reached all non-writemostly
527 * disks. Setting the Returned bit ensures that this
528 * gets done only once -- we don't ever want to return
529 * -EIO here, instead we'll wait
530 */
531 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
532 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
533 /* Maybe we can return now */
534 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
535 struct bio *mbio = r1_bio->master_bio;
536 pr_debug("raid1: behind end write sectors"
537 " %llu-%llu\n",
538 (unsigned long long) mbio->bi_iter.bi_sector,
539 (unsigned long long) bio_end_sector(mbio) - 1);
540 call_bio_endio(r1_bio);
541 }
542 }
543 } else if (rdev->mddev->serialize_policy)
544 remove_serial(rdev, lo, hi);
545 if (r1_bio->bios[mirror] == NULL)
546 rdev_dec_pending(rdev, conf->mddev);
547
548 /*
549 * Let's see if all mirrored write operations have finished
550 * already.
551 */
552 r1_bio_write_done(r1_bio);
553
554 if (to_put)
555 bio_put(to_put);
556}
557
558static sector_t align_to_barrier_unit_end(sector_t start_sector,
559 sector_t sectors)
560{
561 sector_t len;
562
563 WARN_ON(sectors == 0);
564 /*
565 * len is the number of sectors from start_sector to end of the
566 * barrier unit which start_sector belongs to.
567 */
568 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
569 start_sector;
570
571 if (len > sectors)
572 len = sectors;
573
574 return len;
575}
576
577static void update_read_sectors(struct r1conf *conf, int disk,
578 sector_t this_sector, int len)
579{
580 struct raid1_info *info = &conf->mirrors[disk];
581
582 atomic_inc(&info->rdev->nr_pending);
583 if (info->next_seq_sect != this_sector)
584 info->seq_start = this_sector;
585 info->next_seq_sect = this_sector + len;
586}
587
588static int choose_first_rdev(struct r1conf *conf, struct r1bio *r1_bio,
589 int *max_sectors)
590{
591 sector_t this_sector = r1_bio->sector;
592 int len = r1_bio->sectors;
593 int disk;
594
595 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
596 struct md_rdev *rdev;
597 int read_len;
598
599 if (r1_bio->bios[disk] == IO_BLOCKED)
600 continue;
601
602 rdev = conf->mirrors[disk].rdev;
603 if (!rdev || test_bit(Faulty, &rdev->flags))
604 continue;
605
606 /* choose the first disk even if it has some bad blocks. */
607 read_len = raid1_check_read_range(rdev, this_sector, &len);
608 if (read_len > 0) {
609 update_read_sectors(conf, disk, this_sector, read_len);
610 *max_sectors = read_len;
611 return disk;
612 }
613 }
614
615 return -1;
616}
617
618static bool rdev_in_recovery(struct md_rdev *rdev, struct r1bio *r1_bio)
619{
620 return !test_bit(In_sync, &rdev->flags) &&
621 rdev->recovery_offset < r1_bio->sector + r1_bio->sectors;
622}
623
624static int choose_bb_rdev(struct r1conf *conf, struct r1bio *r1_bio,
625 int *max_sectors)
626{
627 sector_t this_sector = r1_bio->sector;
628 int best_disk = -1;
629 int best_len = 0;
630 int disk;
631
632 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
633 struct md_rdev *rdev;
634 int len;
635 int read_len;
636
637 if (r1_bio->bios[disk] == IO_BLOCKED)
638 continue;
639
640 rdev = conf->mirrors[disk].rdev;
641 if (!rdev || test_bit(Faulty, &rdev->flags) ||
642 rdev_in_recovery(rdev, r1_bio) ||
643 test_bit(WriteMostly, &rdev->flags))
644 continue;
645
646 /* keep track of the disk with the most readable sectors. */
647 len = r1_bio->sectors;
648 read_len = raid1_check_read_range(rdev, this_sector, &len);
649 if (read_len > best_len) {
650 best_disk = disk;
651 best_len = read_len;
652 }
653 }
654
655 if (best_disk != -1) {
656 *max_sectors = best_len;
657 update_read_sectors(conf, best_disk, this_sector, best_len);
658 }
659
660 return best_disk;
661}
662
663static int choose_slow_rdev(struct r1conf *conf, struct r1bio *r1_bio,
664 int *max_sectors)
665{
666 sector_t this_sector = r1_bio->sector;
667 int bb_disk = -1;
668 int bb_read_len = 0;
669 int disk;
670
671 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
672 struct md_rdev *rdev;
673 int len;
674 int read_len;
675
676 if (r1_bio->bios[disk] == IO_BLOCKED)
677 continue;
678
679 rdev = conf->mirrors[disk].rdev;
680 if (!rdev || test_bit(Faulty, &rdev->flags) ||
681 !test_bit(WriteMostly, &rdev->flags) ||
682 rdev_in_recovery(rdev, r1_bio))
683 continue;
684
685 /* there are no bad blocks, we can use this disk */
686 len = r1_bio->sectors;
687 read_len = raid1_check_read_range(rdev, this_sector, &len);
688 if (read_len == r1_bio->sectors) {
689 *max_sectors = read_len;
690 update_read_sectors(conf, disk, this_sector, read_len);
691 return disk;
692 }
693
694 /*
695 * there are partial bad blocks, choose the rdev with largest
696 * read length.
697 */
698 if (read_len > bb_read_len) {
699 bb_disk = disk;
700 bb_read_len = read_len;
701 }
702 }
703
704 if (bb_disk != -1) {
705 *max_sectors = bb_read_len;
706 update_read_sectors(conf, bb_disk, this_sector, bb_read_len);
707 }
708
709 return bb_disk;
710}
711
712static bool is_sequential(struct r1conf *conf, int disk, struct r1bio *r1_bio)
713{
714 /* TODO: address issues with this check and concurrency. */
715 return conf->mirrors[disk].next_seq_sect == r1_bio->sector ||
716 conf->mirrors[disk].head_position == r1_bio->sector;
717}
718
719/*
720 * If buffered sequential IO size exceeds optimal iosize, check if there is idle
721 * disk. If yes, choose the idle disk.
722 */
723static bool should_choose_next(struct r1conf *conf, int disk)
724{
725 struct raid1_info *mirror = &conf->mirrors[disk];
726 int opt_iosize;
727
728 if (!test_bit(Nonrot, &mirror->rdev->flags))
729 return false;
730
731 opt_iosize = bdev_io_opt(mirror->rdev->bdev) >> 9;
732 return opt_iosize > 0 && mirror->seq_start != MaxSector &&
733 mirror->next_seq_sect > opt_iosize &&
734 mirror->next_seq_sect - opt_iosize >= mirror->seq_start;
735}
736
737static bool rdev_readable(struct md_rdev *rdev, struct r1bio *r1_bio)
738{
739 if (!rdev || test_bit(Faulty, &rdev->flags))
740 return false;
741
742 if (rdev_in_recovery(rdev, r1_bio))
743 return false;
744
745 /* don't read from slow disk unless have to */
746 if (test_bit(WriteMostly, &rdev->flags))
747 return false;
748
749 /* don't split IO for bad blocks unless have to */
750 if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors))
751 return false;
752
753 return true;
754}
755
756struct read_balance_ctl {
757 sector_t closest_dist;
758 int closest_dist_disk;
759 int min_pending;
760 int min_pending_disk;
761 int sequential_disk;
762 int readable_disks;
763};
764
765static int choose_best_rdev(struct r1conf *conf, struct r1bio *r1_bio)
766{
767 int disk;
768 struct read_balance_ctl ctl = {
769 .closest_dist_disk = -1,
770 .closest_dist = MaxSector,
771 .min_pending_disk = -1,
772 .min_pending = UINT_MAX,
773 .sequential_disk = -1,
774 };
775
776 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
777 struct md_rdev *rdev;
778 sector_t dist;
779 unsigned int pending;
780
781 if (r1_bio->bios[disk] == IO_BLOCKED)
782 continue;
783
784 rdev = conf->mirrors[disk].rdev;
785 if (!rdev_readable(rdev, r1_bio))
786 continue;
787
788 /* At least two disks to choose from so failfast is OK */
789 if (ctl.readable_disks++ == 1)
790 set_bit(R1BIO_FailFast, &r1_bio->state);
791
792 pending = atomic_read(&rdev->nr_pending);
793 dist = abs(r1_bio->sector - conf->mirrors[disk].head_position);
794
795 /* Don't change to another disk for sequential reads */
796 if (is_sequential(conf, disk, r1_bio)) {
797 if (!should_choose_next(conf, disk))
798 return disk;
799
800 /*
801 * Add 'pending' to avoid choosing this disk if
802 * there is other idle disk.
803 */
804 pending++;
805 /*
806 * If there is no other idle disk, this disk
807 * will be chosen.
808 */
809 ctl.sequential_disk = disk;
810 }
811
812 if (ctl.min_pending > pending) {
813 ctl.min_pending = pending;
814 ctl.min_pending_disk = disk;
815 }
816
817 if (ctl.closest_dist > dist) {
818 ctl.closest_dist = dist;
819 ctl.closest_dist_disk = disk;
820 }
821 }
822
823 /*
824 * sequential IO size exceeds optimal iosize, however, there is no other
825 * idle disk, so choose the sequential disk.
826 */
827 if (ctl.sequential_disk != -1 && ctl.min_pending != 0)
828 return ctl.sequential_disk;
829
830 /*
831 * If all disks are rotational, choose the closest disk. If any disk is
832 * non-rotational, choose the disk with less pending request even the
833 * disk is rotational, which might/might not be optimal for raids with
834 * mixed ratation/non-rotational disks depending on workload.
835 */
836 if (ctl.min_pending_disk != -1 &&
837 (READ_ONCE(conf->nonrot_disks) || ctl.min_pending == 0))
838 return ctl.min_pending_disk;
839 else
840 return ctl.closest_dist_disk;
841}
842
843/*
844 * This routine returns the disk from which the requested read should be done.
845 *
846 * 1) If resync is in progress, find the first usable disk and use it even if it
847 * has some bad blocks.
848 *
849 * 2) Now that there is no resync, loop through all disks and skipping slow
850 * disks and disks with bad blocks for now. Only pay attention to key disk
851 * choice.
852 *
853 * 3) If we've made it this far, now look for disks with bad blocks and choose
854 * the one with most number of sectors.
855 *
856 * 4) If we are all the way at the end, we have no choice but to use a disk even
857 * if it is write mostly.
858 *
859 * The rdev for the device selected will have nr_pending incremented.
860 */
861static int read_balance(struct r1conf *conf, struct r1bio *r1_bio,
862 int *max_sectors)
863{
864 int disk;
865
866 clear_bit(R1BIO_FailFast, &r1_bio->state);
867
868 if (raid1_should_read_first(conf->mddev, r1_bio->sector,
869 r1_bio->sectors))
870 return choose_first_rdev(conf, r1_bio, max_sectors);
871
872 disk = choose_best_rdev(conf, r1_bio);
873 if (disk >= 0) {
874 *max_sectors = r1_bio->sectors;
875 update_read_sectors(conf, disk, r1_bio->sector,
876 r1_bio->sectors);
877 return disk;
878 }
879
880 /*
881 * If we are here it means we didn't find a perfectly good disk so
882 * now spend a bit more time trying to find one with the most good
883 * sectors.
884 */
885 disk = choose_bb_rdev(conf, r1_bio, max_sectors);
886 if (disk >= 0)
887 return disk;
888
889 return choose_slow_rdev(conf, r1_bio, max_sectors);
890}
891
892static void wake_up_barrier(struct r1conf *conf)
893{
894 if (wq_has_sleeper(&conf->wait_barrier))
895 wake_up(&conf->wait_barrier);
896}
897
898static void flush_bio_list(struct r1conf *conf, struct bio *bio)
899{
900 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
901 raid1_prepare_flush_writes(conf->mddev);
902 wake_up_barrier(conf);
903
904 while (bio) { /* submit pending writes */
905 struct bio *next = bio->bi_next;
906
907 raid1_submit_write(bio);
908 bio = next;
909 cond_resched();
910 }
911}
912
913static void flush_pending_writes(struct r1conf *conf)
914{
915 /* Any writes that have been queued but are awaiting
916 * bitmap updates get flushed here.
917 */
918 spin_lock_irq(&conf->device_lock);
919
920 if (conf->pending_bio_list.head) {
921 struct blk_plug plug;
922 struct bio *bio;
923
924 bio = bio_list_get(&conf->pending_bio_list);
925 spin_unlock_irq(&conf->device_lock);
926
927 /*
928 * As this is called in a wait_event() loop (see freeze_array),
929 * current->state might be TASK_UNINTERRUPTIBLE which will
930 * cause a warning when we prepare to wait again. As it is
931 * rare that this path is taken, it is perfectly safe to force
932 * us to go around the wait_event() loop again, so the warning
933 * is a false-positive. Silence the warning by resetting
934 * thread state
935 */
936 __set_current_state(TASK_RUNNING);
937 blk_start_plug(&plug);
938 flush_bio_list(conf, bio);
939 blk_finish_plug(&plug);
940 } else
941 spin_unlock_irq(&conf->device_lock);
942}
943
944/* Barriers....
945 * Sometimes we need to suspend IO while we do something else,
946 * either some resync/recovery, or reconfigure the array.
947 * To do this we raise a 'barrier'.
948 * The 'barrier' is a counter that can be raised multiple times
949 * to count how many activities are happening which preclude
950 * normal IO.
951 * We can only raise the barrier if there is no pending IO.
952 * i.e. if nr_pending == 0.
953 * We choose only to raise the barrier if no-one is waiting for the
954 * barrier to go down. This means that as soon as an IO request
955 * is ready, no other operations which require a barrier will start
956 * until the IO request has had a chance.
957 *
958 * So: regular IO calls 'wait_barrier'. When that returns there
959 * is no backgroup IO happening, It must arrange to call
960 * allow_barrier when it has finished its IO.
961 * backgroup IO calls must call raise_barrier. Once that returns
962 * there is no normal IO happeing. It must arrange to call
963 * lower_barrier when the particular background IO completes.
964 *
965 * If resync/recovery is interrupted, returns -EINTR;
966 * Otherwise, returns 0.
967 */
968static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
969{
970 int idx = sector_to_idx(sector_nr);
971
972 spin_lock_irq(&conf->resync_lock);
973
974 /* Wait until no block IO is waiting */
975 wait_event_lock_irq(conf->wait_barrier,
976 !atomic_read(&conf->nr_waiting[idx]),
977 conf->resync_lock);
978
979 /* block any new IO from starting */
980 atomic_inc(&conf->barrier[idx]);
981 /*
982 * In raise_barrier() we firstly increase conf->barrier[idx] then
983 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
984 * increase conf->nr_pending[idx] then check conf->barrier[idx].
985 * A memory barrier here to make sure conf->nr_pending[idx] won't
986 * be fetched before conf->barrier[idx] is increased. Otherwise
987 * there will be a race between raise_barrier() and _wait_barrier().
988 */
989 smp_mb__after_atomic();
990
991 /* For these conditions we must wait:
992 * A: while the array is in frozen state
993 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
994 * existing in corresponding I/O barrier bucket.
995 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
996 * max resync count which allowed on current I/O barrier bucket.
997 */
998 wait_event_lock_irq(conf->wait_barrier,
999 (!conf->array_frozen &&
1000 !atomic_read(&conf->nr_pending[idx]) &&
1001 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
1002 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
1003 conf->resync_lock);
1004
1005 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
1006 atomic_dec(&conf->barrier[idx]);
1007 spin_unlock_irq(&conf->resync_lock);
1008 wake_up(&conf->wait_barrier);
1009 return -EINTR;
1010 }
1011
1012 atomic_inc(&conf->nr_sync_pending);
1013 spin_unlock_irq(&conf->resync_lock);
1014
1015 return 0;
1016}
1017
1018static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
1019{
1020 int idx = sector_to_idx(sector_nr);
1021
1022 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
1023
1024 atomic_dec(&conf->barrier[idx]);
1025 atomic_dec(&conf->nr_sync_pending);
1026 wake_up(&conf->wait_barrier);
1027}
1028
1029static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
1030{
1031 bool ret = true;
1032
1033 /*
1034 * We need to increase conf->nr_pending[idx] very early here,
1035 * then raise_barrier() can be blocked when it waits for
1036 * conf->nr_pending[idx] to be 0. Then we can avoid holding
1037 * conf->resync_lock when there is no barrier raised in same
1038 * barrier unit bucket. Also if the array is frozen, I/O
1039 * should be blocked until array is unfrozen.
1040 */
1041 atomic_inc(&conf->nr_pending[idx]);
1042 /*
1043 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
1044 * check conf->barrier[idx]. In raise_barrier() we firstly increase
1045 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
1046 * barrier is necessary here to make sure conf->barrier[idx] won't be
1047 * fetched before conf->nr_pending[idx] is increased. Otherwise there
1048 * will be a race between _wait_barrier() and raise_barrier().
1049 */
1050 smp_mb__after_atomic();
1051
1052 /*
1053 * Don't worry about checking two atomic_t variables at same time
1054 * here. If during we check conf->barrier[idx], the array is
1055 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
1056 * 0, it is safe to return and make the I/O continue. Because the
1057 * array is frozen, all I/O returned here will eventually complete
1058 * or be queued, no race will happen. See code comment in
1059 * frozen_array().
1060 */
1061 if (!READ_ONCE(conf->array_frozen) &&
1062 !atomic_read(&conf->barrier[idx]))
1063 return ret;
1064
1065 /*
1066 * After holding conf->resync_lock, conf->nr_pending[idx]
1067 * should be decreased before waiting for barrier to drop.
1068 * Otherwise, we may encounter a race condition because
1069 * raise_barrer() might be waiting for conf->nr_pending[idx]
1070 * to be 0 at same time.
1071 */
1072 spin_lock_irq(&conf->resync_lock);
1073 atomic_inc(&conf->nr_waiting[idx]);
1074 atomic_dec(&conf->nr_pending[idx]);
1075 /*
1076 * In case freeze_array() is waiting for
1077 * get_unqueued_pending() == extra
1078 */
1079 wake_up_barrier(conf);
1080 /* Wait for the barrier in same barrier unit bucket to drop. */
1081
1082 /* Return false when nowait flag is set */
1083 if (nowait) {
1084 ret = false;
1085 } else {
1086 wait_event_lock_irq(conf->wait_barrier,
1087 !conf->array_frozen &&
1088 !atomic_read(&conf->barrier[idx]),
1089 conf->resync_lock);
1090 atomic_inc(&conf->nr_pending[idx]);
1091 }
1092
1093 atomic_dec(&conf->nr_waiting[idx]);
1094 spin_unlock_irq(&conf->resync_lock);
1095 return ret;
1096}
1097
1098static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1099{
1100 int idx = sector_to_idx(sector_nr);
1101 bool ret = true;
1102
1103 /*
1104 * Very similar to _wait_barrier(). The difference is, for read
1105 * I/O we don't need wait for sync I/O, but if the whole array
1106 * is frozen, the read I/O still has to wait until the array is
1107 * unfrozen. Since there is no ordering requirement with
1108 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1109 */
1110 atomic_inc(&conf->nr_pending[idx]);
1111
1112 if (!READ_ONCE(conf->array_frozen))
1113 return ret;
1114
1115 spin_lock_irq(&conf->resync_lock);
1116 atomic_inc(&conf->nr_waiting[idx]);
1117 atomic_dec(&conf->nr_pending[idx]);
1118 /*
1119 * In case freeze_array() is waiting for
1120 * get_unqueued_pending() == extra
1121 */
1122 wake_up_barrier(conf);
1123 /* Wait for array to be unfrozen */
1124
1125 /* Return false when nowait flag is set */
1126 if (nowait) {
1127 /* Return false when nowait flag is set */
1128 ret = false;
1129 } else {
1130 wait_event_lock_irq(conf->wait_barrier,
1131 !conf->array_frozen,
1132 conf->resync_lock);
1133 atomic_inc(&conf->nr_pending[idx]);
1134 }
1135
1136 atomic_dec(&conf->nr_waiting[idx]);
1137 spin_unlock_irq(&conf->resync_lock);
1138 return ret;
1139}
1140
1141static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1142{
1143 int idx = sector_to_idx(sector_nr);
1144
1145 return _wait_barrier(conf, idx, nowait);
1146}
1147
1148static void _allow_barrier(struct r1conf *conf, int idx)
1149{
1150 atomic_dec(&conf->nr_pending[idx]);
1151 wake_up_barrier(conf);
1152}
1153
1154static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1155{
1156 int idx = sector_to_idx(sector_nr);
1157
1158 _allow_barrier(conf, idx);
1159}
1160
1161/* conf->resync_lock should be held */
1162static int get_unqueued_pending(struct r1conf *conf)
1163{
1164 int idx, ret;
1165
1166 ret = atomic_read(&conf->nr_sync_pending);
1167 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1168 ret += atomic_read(&conf->nr_pending[idx]) -
1169 atomic_read(&conf->nr_queued[idx]);
1170
1171 return ret;
1172}
1173
1174static void freeze_array(struct r1conf *conf, int extra)
1175{
1176 /* Stop sync I/O and normal I/O and wait for everything to
1177 * go quiet.
1178 * This is called in two situations:
1179 * 1) management command handlers (reshape, remove disk, quiesce).
1180 * 2) one normal I/O request failed.
1181
1182 * After array_frozen is set to 1, new sync IO will be blocked at
1183 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1184 * or wait_read_barrier(). The flying I/Os will either complete or be
1185 * queued. When everything goes quite, there are only queued I/Os left.
1186
1187 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1188 * barrier bucket index which this I/O request hits. When all sync and
1189 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1190 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1191 * in handle_read_error(), we may call freeze_array() before trying to
1192 * fix the read error. In this case, the error read I/O is not queued,
1193 * so get_unqueued_pending() == 1.
1194 *
1195 * Therefore before this function returns, we need to wait until
1196 * get_unqueued_pendings(conf) gets equal to extra. For
1197 * normal I/O context, extra is 1, in rested situations extra is 0.
1198 */
1199 spin_lock_irq(&conf->resync_lock);
1200 conf->array_frozen = 1;
1201 mddev_add_trace_msg(conf->mddev, "raid1 wait freeze");
1202 wait_event_lock_irq_cmd(
1203 conf->wait_barrier,
1204 get_unqueued_pending(conf) == extra,
1205 conf->resync_lock,
1206 flush_pending_writes(conf));
1207 spin_unlock_irq(&conf->resync_lock);
1208}
1209static void unfreeze_array(struct r1conf *conf)
1210{
1211 /* reverse the effect of the freeze */
1212 spin_lock_irq(&conf->resync_lock);
1213 conf->array_frozen = 0;
1214 spin_unlock_irq(&conf->resync_lock);
1215 wake_up(&conf->wait_barrier);
1216}
1217
1218static void alloc_behind_master_bio(struct r1bio *r1_bio,
1219 struct bio *bio)
1220{
1221 int size = bio->bi_iter.bi_size;
1222 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1223 int i = 0;
1224 struct bio *behind_bio = NULL;
1225
1226 behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1227 &r1_bio->mddev->bio_set);
1228
1229 /* discard op, we don't support writezero/writesame yet */
1230 if (!bio_has_data(bio)) {
1231 behind_bio->bi_iter.bi_size = size;
1232 goto skip_copy;
1233 }
1234
1235 while (i < vcnt && size) {
1236 struct page *page;
1237 int len = min_t(int, PAGE_SIZE, size);
1238
1239 page = alloc_page(GFP_NOIO);
1240 if (unlikely(!page))
1241 goto free_pages;
1242
1243 if (!bio_add_page(behind_bio, page, len, 0)) {
1244 put_page(page);
1245 goto free_pages;
1246 }
1247
1248 size -= len;
1249 i++;
1250 }
1251
1252 bio_copy_data(behind_bio, bio);
1253skip_copy:
1254 r1_bio->behind_master_bio = behind_bio;
1255 set_bit(R1BIO_BehindIO, &r1_bio->state);
1256
1257 return;
1258
1259free_pages:
1260 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1261 bio->bi_iter.bi_size);
1262 bio_free_pages(behind_bio);
1263 bio_put(behind_bio);
1264}
1265
1266static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1267{
1268 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1269 cb);
1270 struct mddev *mddev = plug->cb.data;
1271 struct r1conf *conf = mddev->private;
1272 struct bio *bio;
1273
1274 if (from_schedule) {
1275 spin_lock_irq(&conf->device_lock);
1276 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1277 spin_unlock_irq(&conf->device_lock);
1278 wake_up_barrier(conf);
1279 md_wakeup_thread(mddev->thread);
1280 kfree(plug);
1281 return;
1282 }
1283
1284 /* we aren't scheduling, so we can do the write-out directly. */
1285 bio = bio_list_get(&plug->pending);
1286 flush_bio_list(conf, bio);
1287 kfree(plug);
1288}
1289
1290static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1291{
1292 r1_bio->master_bio = bio;
1293 r1_bio->sectors = bio_sectors(bio);
1294 r1_bio->state = 0;
1295 r1_bio->mddev = mddev;
1296 r1_bio->sector = bio->bi_iter.bi_sector;
1297}
1298
1299static inline struct r1bio *
1300alloc_r1bio(struct mddev *mddev, struct bio *bio)
1301{
1302 struct r1conf *conf = mddev->private;
1303 struct r1bio *r1_bio;
1304
1305 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1306 /* Ensure no bio records IO_BLOCKED */
1307 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1308 init_r1bio(r1_bio, mddev, bio);
1309 return r1_bio;
1310}
1311
1312static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1313 int max_read_sectors, struct r1bio *r1_bio)
1314{
1315 struct r1conf *conf = mddev->private;
1316 struct raid1_info *mirror;
1317 struct bio *read_bio;
1318 const enum req_op op = bio_op(bio);
1319 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1320 int max_sectors;
1321 int rdisk, error;
1322 bool r1bio_existed = !!r1_bio;
1323
1324 /*
1325 * If r1_bio is set, we are blocking the raid1d thread
1326 * so there is a tiny risk of deadlock. So ask for
1327 * emergency memory if needed.
1328 */
1329 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1330
1331 /*
1332 * Still need barrier for READ in case that whole
1333 * array is frozen.
1334 */
1335 if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1336 bio->bi_opf & REQ_NOWAIT)) {
1337 bio_wouldblock_error(bio);
1338 return;
1339 }
1340
1341 if (!r1_bio)
1342 r1_bio = alloc_r1bio(mddev, bio);
1343 else
1344 init_r1bio(r1_bio, mddev, bio);
1345 r1_bio->sectors = max_read_sectors;
1346
1347 /*
1348 * make_request() can abort the operation when read-ahead is being
1349 * used and no empty request is available.
1350 */
1351 rdisk = read_balance(conf, r1_bio, &max_sectors);
1352 if (rdisk < 0) {
1353 /* couldn't find anywhere to read from */
1354 if (r1bio_existed)
1355 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
1356 mdname(mddev),
1357 conf->mirrors[r1_bio->read_disk].rdev->bdev,
1358 r1_bio->sector);
1359 raid_end_bio_io(r1_bio);
1360 return;
1361 }
1362 mirror = conf->mirrors + rdisk;
1363
1364 if (r1bio_existed)
1365 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1366 mdname(mddev),
1367 (unsigned long long)r1_bio->sector,
1368 mirror->rdev->bdev);
1369
1370 if (test_bit(WriteMostly, &mirror->rdev->flags)) {
1371 /*
1372 * Reading from a write-mostly device must take care not to
1373 * over-take any writes that are 'behind'
1374 */
1375 mddev_add_trace_msg(mddev, "raid1 wait behind writes");
1376 mddev->bitmap_ops->wait_behind_writes(mddev);
1377 }
1378
1379 if (max_sectors < bio_sectors(bio)) {
1380 struct bio *split = bio_split(bio, max_sectors,
1381 gfp, &conf->bio_split);
1382
1383 if (IS_ERR(split)) {
1384 error = PTR_ERR(split);
1385 goto err_handle;
1386 }
1387 bio_chain(split, bio);
1388 submit_bio_noacct(bio);
1389 bio = split;
1390 r1_bio->master_bio = bio;
1391 r1_bio->sectors = max_sectors;
1392 }
1393
1394 r1_bio->read_disk = rdisk;
1395 if (!r1bio_existed) {
1396 md_account_bio(mddev, &bio);
1397 r1_bio->master_bio = bio;
1398 }
1399 read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1400 &mddev->bio_set);
1401
1402 r1_bio->bios[rdisk] = read_bio;
1403
1404 read_bio->bi_iter.bi_sector = r1_bio->sector +
1405 mirror->rdev->data_offset;
1406 read_bio->bi_end_io = raid1_end_read_request;
1407 read_bio->bi_opf = op | do_sync;
1408 if (test_bit(FailFast, &mirror->rdev->flags) &&
1409 test_bit(R1BIO_FailFast, &r1_bio->state))
1410 read_bio->bi_opf |= MD_FAILFAST;
1411 read_bio->bi_private = r1_bio;
1412 mddev_trace_remap(mddev, read_bio, r1_bio->sector);
1413 submit_bio_noacct(read_bio);
1414 return;
1415
1416err_handle:
1417 atomic_dec(&mirror->rdev->nr_pending);
1418 bio->bi_status = errno_to_blk_status(error);
1419 set_bit(R1BIO_Uptodate, &r1_bio->state);
1420 raid_end_bio_io(r1_bio);
1421}
1422
1423static bool wait_blocked_rdev(struct mddev *mddev, struct bio *bio)
1424{
1425 struct r1conf *conf = mddev->private;
1426 int disks = conf->raid_disks * 2;
1427 int i;
1428
1429retry:
1430 for (i = 0; i < disks; i++) {
1431 struct md_rdev *rdev = conf->mirrors[i].rdev;
1432
1433 if (!rdev)
1434 continue;
1435
1436 /* don't write here until the bad block is acknowledged */
1437 if (test_bit(WriteErrorSeen, &rdev->flags) &&
1438 rdev_has_badblock(rdev, bio->bi_iter.bi_sector,
1439 bio_sectors(bio)) < 0)
1440 set_bit(BlockedBadBlocks, &rdev->flags);
1441
1442 if (rdev_blocked(rdev)) {
1443 if (bio->bi_opf & REQ_NOWAIT)
1444 return false;
1445
1446 mddev_add_trace_msg(rdev->mddev, "raid1 wait rdev %d blocked",
1447 rdev->raid_disk);
1448 atomic_inc(&rdev->nr_pending);
1449 md_wait_for_blocked_rdev(rdev, rdev->mddev);
1450 goto retry;
1451 }
1452 }
1453
1454 return true;
1455}
1456
1457static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1458 int max_write_sectors)
1459{
1460 struct r1conf *conf = mddev->private;
1461 struct r1bio *r1_bio;
1462 int i, disks, k, error;
1463 unsigned long flags;
1464 int first_clone;
1465 int max_sectors;
1466 bool write_behind = false;
1467 bool is_discard = (bio_op(bio) == REQ_OP_DISCARD);
1468
1469 if (mddev_is_clustered(mddev) &&
1470 md_cluster_ops->area_resyncing(mddev, WRITE,
1471 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1472
1473 DEFINE_WAIT(w);
1474 if (bio->bi_opf & REQ_NOWAIT) {
1475 bio_wouldblock_error(bio);
1476 return;
1477 }
1478 for (;;) {
1479 prepare_to_wait(&conf->wait_barrier,
1480 &w, TASK_IDLE);
1481 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1482 bio->bi_iter.bi_sector,
1483 bio_end_sector(bio)))
1484 break;
1485 schedule();
1486 }
1487 finish_wait(&conf->wait_barrier, &w);
1488 }
1489
1490 /*
1491 * Register the new request and wait if the reconstruction
1492 * thread has put up a bar for new requests.
1493 * Continue immediately if no resync is active currently.
1494 */
1495 if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1496 bio->bi_opf & REQ_NOWAIT)) {
1497 bio_wouldblock_error(bio);
1498 return;
1499 }
1500
1501 if (!wait_blocked_rdev(mddev, bio)) {
1502 bio_wouldblock_error(bio);
1503 return;
1504 }
1505
1506 r1_bio = alloc_r1bio(mddev, bio);
1507 r1_bio->sectors = max_write_sectors;
1508
1509 /* first select target devices under rcu_lock and
1510 * inc refcount on their rdev. Record them by setting
1511 * bios[x] to bio
1512 * If there are known/acknowledged bad blocks on any device on
1513 * which we have seen a write error, we want to avoid writing those
1514 * blocks.
1515 * This potentially requires several writes to write around
1516 * the bad blocks. Each set of writes gets it's own r1bio
1517 * with a set of bios attached.
1518 */
1519
1520 disks = conf->raid_disks * 2;
1521 max_sectors = r1_bio->sectors;
1522 for (i = 0; i < disks; i++) {
1523 struct md_rdev *rdev = conf->mirrors[i].rdev;
1524
1525 /*
1526 * The write-behind io is only attempted on drives marked as
1527 * write-mostly, which means we could allocate write behind
1528 * bio later.
1529 */
1530 if (!is_discard && rdev && test_bit(WriteMostly, &rdev->flags))
1531 write_behind = true;
1532
1533 r1_bio->bios[i] = NULL;
1534 if (!rdev || test_bit(Faulty, &rdev->flags))
1535 continue;
1536
1537 atomic_inc(&rdev->nr_pending);
1538 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1539 sector_t first_bad;
1540 int bad_sectors;
1541 int is_bad;
1542
1543 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1544 &first_bad, &bad_sectors);
1545 if (is_bad && first_bad <= r1_bio->sector) {
1546 /* Cannot write here at all */
1547 bad_sectors -= (r1_bio->sector - first_bad);
1548 if (bad_sectors < max_sectors)
1549 /* mustn't write more than bad_sectors
1550 * to other devices yet
1551 */
1552 max_sectors = bad_sectors;
1553 rdev_dec_pending(rdev, mddev);
1554 continue;
1555 }
1556 if (is_bad) {
1557 int good_sectors;
1558
1559 /*
1560 * We cannot atomically write this, so just
1561 * error in that case. It could be possible to
1562 * atomically write other mirrors, but the
1563 * complexity of supporting that is not worth
1564 * the benefit.
1565 */
1566 if (bio->bi_opf & REQ_ATOMIC) {
1567 error = -EIO;
1568 goto err_handle;
1569 }
1570
1571 good_sectors = first_bad - r1_bio->sector;
1572 if (good_sectors < max_sectors)
1573 max_sectors = good_sectors;
1574 }
1575 }
1576 r1_bio->bios[i] = bio;
1577 }
1578
1579 /*
1580 * When using a bitmap, we may call alloc_behind_master_bio below.
1581 * alloc_behind_master_bio allocates a copy of the data payload a page
1582 * at a time and thus needs a new bio that can fit the whole payload
1583 * this bio in page sized chunks.
1584 */
1585 if (write_behind && mddev->bitmap)
1586 max_sectors = min_t(int, max_sectors,
1587 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1588 if (max_sectors < bio_sectors(bio)) {
1589 struct bio *split = bio_split(bio, max_sectors,
1590 GFP_NOIO, &conf->bio_split);
1591
1592 if (IS_ERR(split)) {
1593 error = PTR_ERR(split);
1594 goto err_handle;
1595 }
1596 bio_chain(split, bio);
1597 submit_bio_noacct(bio);
1598 bio = split;
1599 r1_bio->master_bio = bio;
1600 r1_bio->sectors = max_sectors;
1601 }
1602
1603 md_account_bio(mddev, &bio);
1604 r1_bio->master_bio = bio;
1605 atomic_set(&r1_bio->remaining, 1);
1606 atomic_set(&r1_bio->behind_remaining, 0);
1607
1608 first_clone = 1;
1609
1610 for (i = 0; i < disks; i++) {
1611 struct bio *mbio = NULL;
1612 struct md_rdev *rdev = conf->mirrors[i].rdev;
1613 if (!r1_bio->bios[i])
1614 continue;
1615
1616 if (first_clone) {
1617 unsigned long max_write_behind =
1618 mddev->bitmap_info.max_write_behind;
1619 struct md_bitmap_stats stats;
1620 int err;
1621
1622 /* do behind I/O ?
1623 * Not if there are too many, or cannot
1624 * allocate memory, or a reader on WriteMostly
1625 * is waiting for behind writes to flush */
1626 err = mddev->bitmap_ops->get_stats(mddev->bitmap, &stats);
1627 if (!err && write_behind && !stats.behind_wait &&
1628 stats.behind_writes < max_write_behind)
1629 alloc_behind_master_bio(r1_bio, bio);
1630
1631 if (test_bit(R1BIO_BehindIO, &r1_bio->state))
1632 mddev->bitmap_ops->start_behind_write(mddev);
1633 first_clone = 0;
1634 }
1635
1636 if (r1_bio->behind_master_bio) {
1637 mbio = bio_alloc_clone(rdev->bdev,
1638 r1_bio->behind_master_bio,
1639 GFP_NOIO, &mddev->bio_set);
1640 if (test_bit(CollisionCheck, &rdev->flags))
1641 wait_for_serialization(rdev, r1_bio);
1642 if (test_bit(WriteMostly, &rdev->flags))
1643 atomic_inc(&r1_bio->behind_remaining);
1644 } else {
1645 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1646 &mddev->bio_set);
1647
1648 if (mddev->serialize_policy)
1649 wait_for_serialization(rdev, r1_bio);
1650 }
1651
1652 r1_bio->bios[i] = mbio;
1653
1654 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1655 mbio->bi_end_io = raid1_end_write_request;
1656 mbio->bi_opf = bio_op(bio) |
1657 (bio->bi_opf & (REQ_SYNC | REQ_FUA | REQ_ATOMIC));
1658 if (test_bit(FailFast, &rdev->flags) &&
1659 !test_bit(WriteMostly, &rdev->flags) &&
1660 conf->raid_disks - mddev->degraded > 1)
1661 mbio->bi_opf |= MD_FAILFAST;
1662 mbio->bi_private = r1_bio;
1663
1664 atomic_inc(&r1_bio->remaining);
1665 mddev_trace_remap(mddev, mbio, r1_bio->sector);
1666 /* flush_pending_writes() needs access to the rdev so...*/
1667 mbio->bi_bdev = (void *)rdev;
1668 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1669 spin_lock_irqsave(&conf->device_lock, flags);
1670 bio_list_add(&conf->pending_bio_list, mbio);
1671 spin_unlock_irqrestore(&conf->device_lock, flags);
1672 md_wakeup_thread(mddev->thread);
1673 }
1674 }
1675
1676 r1_bio_write_done(r1_bio);
1677
1678 /* In case raid1d snuck in to freeze_array */
1679 wake_up_barrier(conf);
1680 return;
1681err_handle:
1682 for (k = 0; k < i; k++) {
1683 if (r1_bio->bios[k]) {
1684 rdev_dec_pending(conf->mirrors[k].rdev, mddev);
1685 r1_bio->bios[k] = NULL;
1686 }
1687 }
1688
1689 bio->bi_status = errno_to_blk_status(error);
1690 set_bit(R1BIO_Uptodate, &r1_bio->state);
1691 raid_end_bio_io(r1_bio);
1692}
1693
1694static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1695{
1696 sector_t sectors;
1697
1698 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1699 && md_flush_request(mddev, bio))
1700 return true;
1701
1702 /*
1703 * There is a limit to the maximum size, but
1704 * the read/write handler might find a lower limit
1705 * due to bad blocks. To avoid multiple splits,
1706 * we pass the maximum number of sectors down
1707 * and let the lower level perform the split.
1708 */
1709 sectors = align_to_barrier_unit_end(
1710 bio->bi_iter.bi_sector, bio_sectors(bio));
1711
1712 if (bio_data_dir(bio) == READ)
1713 raid1_read_request(mddev, bio, sectors, NULL);
1714 else {
1715 md_write_start(mddev,bio);
1716 raid1_write_request(mddev, bio, sectors);
1717 }
1718 return true;
1719}
1720
1721static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1722{
1723 struct r1conf *conf = mddev->private;
1724 int i;
1725
1726 lockdep_assert_held(&mddev->lock);
1727
1728 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1729 conf->raid_disks - mddev->degraded);
1730 for (i = 0; i < conf->raid_disks; i++) {
1731 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1732
1733 seq_printf(seq, "%s",
1734 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1735 }
1736 seq_printf(seq, "]");
1737}
1738
1739/**
1740 * raid1_error() - RAID1 error handler.
1741 * @mddev: affected md device.
1742 * @rdev: member device to fail.
1743 *
1744 * The routine acknowledges &rdev failure and determines new @mddev state.
1745 * If it failed, then:
1746 * - &MD_BROKEN flag is set in &mddev->flags.
1747 * - recovery is disabled.
1748 * Otherwise, it must be degraded:
1749 * - recovery is interrupted.
1750 * - &mddev->degraded is bumped.
1751 *
1752 * @rdev is marked as &Faulty excluding case when array is failed and
1753 * &mddev->fail_last_dev is off.
1754 */
1755static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1756{
1757 struct r1conf *conf = mddev->private;
1758 unsigned long flags;
1759
1760 spin_lock_irqsave(&conf->device_lock, flags);
1761
1762 if (test_bit(In_sync, &rdev->flags) &&
1763 (conf->raid_disks - mddev->degraded) == 1) {
1764 set_bit(MD_BROKEN, &mddev->flags);
1765
1766 if (!mddev->fail_last_dev) {
1767 conf->recovery_disabled = mddev->recovery_disabled;
1768 spin_unlock_irqrestore(&conf->device_lock, flags);
1769 return;
1770 }
1771 }
1772 set_bit(Blocked, &rdev->flags);
1773 if (test_and_clear_bit(In_sync, &rdev->flags))
1774 mddev->degraded++;
1775 set_bit(Faulty, &rdev->flags);
1776 spin_unlock_irqrestore(&conf->device_lock, flags);
1777 /*
1778 * if recovery is running, make sure it aborts.
1779 */
1780 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1781 set_mask_bits(&mddev->sb_flags, 0,
1782 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1783 pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1784 "md/raid1:%s: Operation continuing on %d devices.\n",
1785 mdname(mddev), rdev->bdev,
1786 mdname(mddev), conf->raid_disks - mddev->degraded);
1787}
1788
1789static void print_conf(struct r1conf *conf)
1790{
1791 int i;
1792
1793 pr_debug("RAID1 conf printout:\n");
1794 if (!conf) {
1795 pr_debug("(!conf)\n");
1796 return;
1797 }
1798 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1799 conf->raid_disks);
1800
1801 lockdep_assert_held(&conf->mddev->reconfig_mutex);
1802 for (i = 0; i < conf->raid_disks; i++) {
1803 struct md_rdev *rdev = conf->mirrors[i].rdev;
1804 if (rdev)
1805 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1806 i, !test_bit(In_sync, &rdev->flags),
1807 !test_bit(Faulty, &rdev->flags),
1808 rdev->bdev);
1809 }
1810}
1811
1812static void close_sync(struct r1conf *conf)
1813{
1814 int idx;
1815
1816 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1817 _wait_barrier(conf, idx, false);
1818 _allow_barrier(conf, idx);
1819 }
1820
1821 mempool_exit(&conf->r1buf_pool);
1822}
1823
1824static int raid1_spare_active(struct mddev *mddev)
1825{
1826 int i;
1827 struct r1conf *conf = mddev->private;
1828 int count = 0;
1829 unsigned long flags;
1830
1831 /*
1832 * Find all failed disks within the RAID1 configuration
1833 * and mark them readable.
1834 * Called under mddev lock, so rcu protection not needed.
1835 * device_lock used to avoid races with raid1_end_read_request
1836 * which expects 'In_sync' flags and ->degraded to be consistent.
1837 */
1838 spin_lock_irqsave(&conf->device_lock, flags);
1839 for (i = 0; i < conf->raid_disks; i++) {
1840 struct md_rdev *rdev = conf->mirrors[i].rdev;
1841 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1842 if (repl
1843 && !test_bit(Candidate, &repl->flags)
1844 && repl->recovery_offset == MaxSector
1845 && !test_bit(Faulty, &repl->flags)
1846 && !test_and_set_bit(In_sync, &repl->flags)) {
1847 /* replacement has just become active */
1848 if (!rdev ||
1849 !test_and_clear_bit(In_sync, &rdev->flags))
1850 count++;
1851 if (rdev) {
1852 /* Replaced device not technically
1853 * faulty, but we need to be sure
1854 * it gets removed and never re-added
1855 */
1856 set_bit(Faulty, &rdev->flags);
1857 sysfs_notify_dirent_safe(
1858 rdev->sysfs_state);
1859 }
1860 }
1861 if (rdev
1862 && rdev->recovery_offset == MaxSector
1863 && !test_bit(Faulty, &rdev->flags)
1864 && !test_and_set_bit(In_sync, &rdev->flags)) {
1865 count++;
1866 sysfs_notify_dirent_safe(rdev->sysfs_state);
1867 }
1868 }
1869 mddev->degraded -= count;
1870 spin_unlock_irqrestore(&conf->device_lock, flags);
1871
1872 print_conf(conf);
1873 return count;
1874}
1875
1876static bool raid1_add_conf(struct r1conf *conf, struct md_rdev *rdev, int disk,
1877 bool replacement)
1878{
1879 struct raid1_info *info = conf->mirrors + disk;
1880
1881 if (replacement)
1882 info += conf->raid_disks;
1883
1884 if (info->rdev)
1885 return false;
1886
1887 if (bdev_nonrot(rdev->bdev)) {
1888 set_bit(Nonrot, &rdev->flags);
1889 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks + 1);
1890 }
1891
1892 rdev->raid_disk = disk;
1893 info->head_position = 0;
1894 info->seq_start = MaxSector;
1895 WRITE_ONCE(info->rdev, rdev);
1896
1897 return true;
1898}
1899
1900static bool raid1_remove_conf(struct r1conf *conf, int disk)
1901{
1902 struct raid1_info *info = conf->mirrors + disk;
1903 struct md_rdev *rdev = info->rdev;
1904
1905 if (!rdev || test_bit(In_sync, &rdev->flags) ||
1906 atomic_read(&rdev->nr_pending))
1907 return false;
1908
1909 /* Only remove non-faulty devices if recovery is not possible. */
1910 if (!test_bit(Faulty, &rdev->flags) &&
1911 rdev->mddev->recovery_disabled != conf->recovery_disabled &&
1912 rdev->mddev->degraded < conf->raid_disks)
1913 return false;
1914
1915 if (test_and_clear_bit(Nonrot, &rdev->flags))
1916 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks - 1);
1917
1918 WRITE_ONCE(info->rdev, NULL);
1919 return true;
1920}
1921
1922static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1923{
1924 struct r1conf *conf = mddev->private;
1925 int err = -EEXIST;
1926 int mirror = 0, repl_slot = -1;
1927 struct raid1_info *p;
1928 int first = 0;
1929 int last = conf->raid_disks - 1;
1930
1931 if (mddev->recovery_disabled == conf->recovery_disabled)
1932 return -EBUSY;
1933
1934 if (rdev->raid_disk >= 0)
1935 first = last = rdev->raid_disk;
1936
1937 /*
1938 * find the disk ... but prefer rdev->saved_raid_disk
1939 * if possible.
1940 */
1941 if (rdev->saved_raid_disk >= 0 &&
1942 rdev->saved_raid_disk >= first &&
1943 rdev->saved_raid_disk < conf->raid_disks &&
1944 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1945 first = last = rdev->saved_raid_disk;
1946
1947 for (mirror = first; mirror <= last; mirror++) {
1948 p = conf->mirrors + mirror;
1949 if (!p->rdev) {
1950 err = mddev_stack_new_rdev(mddev, rdev);
1951 if (err)
1952 return err;
1953
1954 raid1_add_conf(conf, rdev, mirror, false);
1955 /* As all devices are equivalent, we don't need a full recovery
1956 * if this was recently any drive of the array
1957 */
1958 if (rdev->saved_raid_disk < 0)
1959 conf->fullsync = 1;
1960 break;
1961 }
1962 if (test_bit(WantReplacement, &p->rdev->flags) &&
1963 p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1964 repl_slot = mirror;
1965 }
1966
1967 if (err && repl_slot >= 0) {
1968 /* Add this device as a replacement */
1969 clear_bit(In_sync, &rdev->flags);
1970 set_bit(Replacement, &rdev->flags);
1971 raid1_add_conf(conf, rdev, repl_slot, true);
1972 err = 0;
1973 conf->fullsync = 1;
1974 }
1975
1976 print_conf(conf);
1977 return err;
1978}
1979
1980static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1981{
1982 struct r1conf *conf = mddev->private;
1983 int err = 0;
1984 int number = rdev->raid_disk;
1985 struct raid1_info *p = conf->mirrors + number;
1986
1987 if (unlikely(number >= conf->raid_disks))
1988 goto abort;
1989
1990 if (rdev != p->rdev) {
1991 number += conf->raid_disks;
1992 p = conf->mirrors + number;
1993 }
1994
1995 print_conf(conf);
1996 if (rdev == p->rdev) {
1997 if (!raid1_remove_conf(conf, number)) {
1998 err = -EBUSY;
1999 goto abort;
2000 }
2001
2002 if (number < conf->raid_disks &&
2003 conf->mirrors[conf->raid_disks + number].rdev) {
2004 /* We just removed a device that is being replaced.
2005 * Move down the replacement. We drain all IO before
2006 * doing this to avoid confusion.
2007 */
2008 struct md_rdev *repl =
2009 conf->mirrors[conf->raid_disks + number].rdev;
2010 freeze_array(conf, 0);
2011 if (atomic_read(&repl->nr_pending)) {
2012 /* It means that some queued IO of retry_list
2013 * hold repl. Thus, we cannot set replacement
2014 * as NULL, avoiding rdev NULL pointer
2015 * dereference in sync_request_write and
2016 * handle_write_finished.
2017 */
2018 err = -EBUSY;
2019 unfreeze_array(conf);
2020 goto abort;
2021 }
2022 clear_bit(Replacement, &repl->flags);
2023 WRITE_ONCE(p->rdev, repl);
2024 conf->mirrors[conf->raid_disks + number].rdev = NULL;
2025 unfreeze_array(conf);
2026 }
2027
2028 clear_bit(WantReplacement, &rdev->flags);
2029 err = md_integrity_register(mddev);
2030 }
2031abort:
2032
2033 print_conf(conf);
2034 return err;
2035}
2036
2037static void end_sync_read(struct bio *bio)
2038{
2039 struct r1bio *r1_bio = get_resync_r1bio(bio);
2040
2041 update_head_pos(r1_bio->read_disk, r1_bio);
2042
2043 /*
2044 * we have read a block, now it needs to be re-written,
2045 * or re-read if the read failed.
2046 * We don't do much here, just schedule handling by raid1d
2047 */
2048 if (!bio->bi_status)
2049 set_bit(R1BIO_Uptodate, &r1_bio->state);
2050
2051 if (atomic_dec_and_test(&r1_bio->remaining))
2052 reschedule_retry(r1_bio);
2053}
2054
2055static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
2056{
2057 sector_t sync_blocks = 0;
2058 sector_t s = r1_bio->sector;
2059 long sectors_to_go = r1_bio->sectors;
2060
2061 /* make sure these bits don't get cleared. */
2062 do {
2063 mddev->bitmap_ops->end_sync(mddev, s, &sync_blocks);
2064 s += sync_blocks;
2065 sectors_to_go -= sync_blocks;
2066 } while (sectors_to_go > 0);
2067}
2068
2069static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
2070{
2071 if (atomic_dec_and_test(&r1_bio->remaining)) {
2072 struct mddev *mddev = r1_bio->mddev;
2073 int s = r1_bio->sectors;
2074
2075 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2076 test_bit(R1BIO_WriteError, &r1_bio->state))
2077 reschedule_retry(r1_bio);
2078 else {
2079 put_buf(r1_bio);
2080 md_done_sync(mddev, s, uptodate);
2081 }
2082 }
2083}
2084
2085static void end_sync_write(struct bio *bio)
2086{
2087 int uptodate = !bio->bi_status;
2088 struct r1bio *r1_bio = get_resync_r1bio(bio);
2089 struct mddev *mddev = r1_bio->mddev;
2090 struct r1conf *conf = mddev->private;
2091 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
2092
2093 if (!uptodate) {
2094 abort_sync_write(mddev, r1_bio);
2095 set_bit(WriteErrorSeen, &rdev->flags);
2096 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2097 set_bit(MD_RECOVERY_NEEDED, &
2098 mddev->recovery);
2099 set_bit(R1BIO_WriteError, &r1_bio->state);
2100 } else if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
2101 !rdev_has_badblock(conf->mirrors[r1_bio->read_disk].rdev,
2102 r1_bio->sector, r1_bio->sectors)) {
2103 set_bit(R1BIO_MadeGood, &r1_bio->state);
2104 }
2105
2106 put_sync_write_buf(r1_bio, uptodate);
2107}
2108
2109static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
2110 int sectors, struct page *page, blk_opf_t rw)
2111{
2112 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2113 /* success */
2114 return 1;
2115 if (rw == REQ_OP_WRITE) {
2116 set_bit(WriteErrorSeen, &rdev->flags);
2117 if (!test_and_set_bit(WantReplacement,
2118 &rdev->flags))
2119 set_bit(MD_RECOVERY_NEEDED, &
2120 rdev->mddev->recovery);
2121 }
2122 /* need to record an error - either for the block or the device */
2123 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2124 md_error(rdev->mddev, rdev);
2125 return 0;
2126}
2127
2128static int fix_sync_read_error(struct r1bio *r1_bio)
2129{
2130 /* Try some synchronous reads of other devices to get
2131 * good data, much like with normal read errors. Only
2132 * read into the pages we already have so we don't
2133 * need to re-issue the read request.
2134 * We don't need to freeze the array, because being in an
2135 * active sync request, there is no normal IO, and
2136 * no overlapping syncs.
2137 * We don't need to check is_badblock() again as we
2138 * made sure that anything with a bad block in range
2139 * will have bi_end_io clear.
2140 */
2141 struct mddev *mddev = r1_bio->mddev;
2142 struct r1conf *conf = mddev->private;
2143 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2144 struct page **pages = get_resync_pages(bio)->pages;
2145 sector_t sect = r1_bio->sector;
2146 int sectors = r1_bio->sectors;
2147 int idx = 0;
2148 struct md_rdev *rdev;
2149
2150 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2151 if (test_bit(FailFast, &rdev->flags)) {
2152 /* Don't try recovering from here - just fail it
2153 * ... unless it is the last working device of course */
2154 md_error(mddev, rdev);
2155 if (test_bit(Faulty, &rdev->flags))
2156 /* Don't try to read from here, but make sure
2157 * put_buf does it's thing
2158 */
2159 bio->bi_end_io = end_sync_write;
2160 }
2161
2162 while(sectors) {
2163 int s = sectors;
2164 int d = r1_bio->read_disk;
2165 int success = 0;
2166 int start;
2167
2168 if (s > (PAGE_SIZE>>9))
2169 s = PAGE_SIZE >> 9;
2170 do {
2171 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2172 /* No rcu protection needed here devices
2173 * can only be removed when no resync is
2174 * active, and resync is currently active
2175 */
2176 rdev = conf->mirrors[d].rdev;
2177 if (sync_page_io(rdev, sect, s<<9,
2178 pages[idx],
2179 REQ_OP_READ, false)) {
2180 success = 1;
2181 break;
2182 }
2183 }
2184 d++;
2185 if (d == conf->raid_disks * 2)
2186 d = 0;
2187 } while (!success && d != r1_bio->read_disk);
2188
2189 if (!success) {
2190 int abort = 0;
2191 /* Cannot read from anywhere, this block is lost.
2192 * Record a bad block on each device. If that doesn't
2193 * work just disable and interrupt the recovery.
2194 * Don't fail devices as that won't really help.
2195 */
2196 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2197 mdname(mddev), bio->bi_bdev,
2198 (unsigned long long)r1_bio->sector);
2199 for (d = 0; d < conf->raid_disks * 2; d++) {
2200 rdev = conf->mirrors[d].rdev;
2201 if (!rdev || test_bit(Faulty, &rdev->flags))
2202 continue;
2203 if (!rdev_set_badblocks(rdev, sect, s, 0))
2204 abort = 1;
2205 }
2206 if (abort) {
2207 conf->recovery_disabled =
2208 mddev->recovery_disabled;
2209 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2210 md_done_sync(mddev, r1_bio->sectors, 0);
2211 put_buf(r1_bio);
2212 return 0;
2213 }
2214 /* Try next page */
2215 sectors -= s;
2216 sect += s;
2217 idx++;
2218 continue;
2219 }
2220
2221 start = d;
2222 /* write it back and re-read */
2223 while (d != r1_bio->read_disk) {
2224 if (d == 0)
2225 d = conf->raid_disks * 2;
2226 d--;
2227 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2228 continue;
2229 rdev = conf->mirrors[d].rdev;
2230 if (r1_sync_page_io(rdev, sect, s,
2231 pages[idx],
2232 REQ_OP_WRITE) == 0) {
2233 r1_bio->bios[d]->bi_end_io = NULL;
2234 rdev_dec_pending(rdev, mddev);
2235 }
2236 }
2237 d = start;
2238 while (d != r1_bio->read_disk) {
2239 if (d == 0)
2240 d = conf->raid_disks * 2;
2241 d--;
2242 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2243 continue;
2244 rdev = conf->mirrors[d].rdev;
2245 if (r1_sync_page_io(rdev, sect, s,
2246 pages[idx],
2247 REQ_OP_READ) != 0)
2248 atomic_add(s, &rdev->corrected_errors);
2249 }
2250 sectors -= s;
2251 sect += s;
2252 idx ++;
2253 }
2254 set_bit(R1BIO_Uptodate, &r1_bio->state);
2255 bio->bi_status = 0;
2256 return 1;
2257}
2258
2259static void process_checks(struct r1bio *r1_bio)
2260{
2261 /* We have read all readable devices. If we haven't
2262 * got the block, then there is no hope left.
2263 * If we have, then we want to do a comparison
2264 * and skip the write if everything is the same.
2265 * If any blocks failed to read, then we need to
2266 * attempt an over-write
2267 */
2268 struct mddev *mddev = r1_bio->mddev;
2269 struct r1conf *conf = mddev->private;
2270 int primary;
2271 int i;
2272 int vcnt;
2273
2274 /* Fix variable parts of all bios */
2275 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2276 for (i = 0; i < conf->raid_disks * 2; i++) {
2277 blk_status_t status;
2278 struct bio *b = r1_bio->bios[i];
2279 struct resync_pages *rp = get_resync_pages(b);
2280 if (b->bi_end_io != end_sync_read)
2281 continue;
2282 /* fixup the bio for reuse, but preserve errno */
2283 status = b->bi_status;
2284 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2285 b->bi_status = status;
2286 b->bi_iter.bi_sector = r1_bio->sector +
2287 conf->mirrors[i].rdev->data_offset;
2288 b->bi_end_io = end_sync_read;
2289 rp->raid_bio = r1_bio;
2290 b->bi_private = rp;
2291
2292 /* initialize bvec table again */
2293 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2294 }
2295 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2296 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2297 !r1_bio->bios[primary]->bi_status) {
2298 r1_bio->bios[primary]->bi_end_io = NULL;
2299 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2300 break;
2301 }
2302 r1_bio->read_disk = primary;
2303 for (i = 0; i < conf->raid_disks * 2; i++) {
2304 int j = 0;
2305 struct bio *pbio = r1_bio->bios[primary];
2306 struct bio *sbio = r1_bio->bios[i];
2307 blk_status_t status = sbio->bi_status;
2308 struct page **ppages = get_resync_pages(pbio)->pages;
2309 struct page **spages = get_resync_pages(sbio)->pages;
2310 struct bio_vec *bi;
2311 int page_len[RESYNC_PAGES] = { 0 };
2312 struct bvec_iter_all iter_all;
2313
2314 if (sbio->bi_end_io != end_sync_read)
2315 continue;
2316 /* Now we can 'fixup' the error value */
2317 sbio->bi_status = 0;
2318
2319 bio_for_each_segment_all(bi, sbio, iter_all)
2320 page_len[j++] = bi->bv_len;
2321
2322 if (!status) {
2323 for (j = vcnt; j-- ; ) {
2324 if (memcmp(page_address(ppages[j]),
2325 page_address(spages[j]),
2326 page_len[j]))
2327 break;
2328 }
2329 } else
2330 j = 0;
2331 if (j >= 0)
2332 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2333 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2334 && !status)) {
2335 /* No need to write to this device. */
2336 sbio->bi_end_io = NULL;
2337 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2338 continue;
2339 }
2340
2341 bio_copy_data(sbio, pbio);
2342 }
2343}
2344
2345static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2346{
2347 struct r1conf *conf = mddev->private;
2348 int i;
2349 int disks = conf->raid_disks * 2;
2350 struct bio *wbio;
2351
2352 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2353 /* ouch - failed to read all of that. */
2354 if (!fix_sync_read_error(r1_bio))
2355 return;
2356
2357 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2358 process_checks(r1_bio);
2359
2360 /*
2361 * schedule writes
2362 */
2363 atomic_set(&r1_bio->remaining, 1);
2364 for (i = 0; i < disks ; i++) {
2365 wbio = r1_bio->bios[i];
2366 if (wbio->bi_end_io == NULL ||
2367 (wbio->bi_end_io == end_sync_read &&
2368 (i == r1_bio->read_disk ||
2369 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2370 continue;
2371 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2372 abort_sync_write(mddev, r1_bio);
2373 continue;
2374 }
2375
2376 wbio->bi_opf = REQ_OP_WRITE;
2377 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2378 wbio->bi_opf |= MD_FAILFAST;
2379
2380 wbio->bi_end_io = end_sync_write;
2381 atomic_inc(&r1_bio->remaining);
2382 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2383
2384 submit_bio_noacct(wbio);
2385 }
2386
2387 put_sync_write_buf(r1_bio, 1);
2388}
2389
2390/*
2391 * This is a kernel thread which:
2392 *
2393 * 1. Retries failed read operations on working mirrors.
2394 * 2. Updates the raid superblock when problems encounter.
2395 * 3. Performs writes following reads for array synchronising.
2396 */
2397
2398static void fix_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2399{
2400 sector_t sect = r1_bio->sector;
2401 int sectors = r1_bio->sectors;
2402 int read_disk = r1_bio->read_disk;
2403 struct mddev *mddev = conf->mddev;
2404 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2405
2406 if (exceed_read_errors(mddev, rdev)) {
2407 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2408 return;
2409 }
2410
2411 while(sectors) {
2412 int s = sectors;
2413 int d = read_disk;
2414 int success = 0;
2415 int start;
2416
2417 if (s > (PAGE_SIZE>>9))
2418 s = PAGE_SIZE >> 9;
2419
2420 do {
2421 rdev = conf->mirrors[d].rdev;
2422 if (rdev &&
2423 (test_bit(In_sync, &rdev->flags) ||
2424 (!test_bit(Faulty, &rdev->flags) &&
2425 rdev->recovery_offset >= sect + s)) &&
2426 rdev_has_badblock(rdev, sect, s) == 0) {
2427 atomic_inc(&rdev->nr_pending);
2428 if (sync_page_io(rdev, sect, s<<9,
2429 conf->tmppage, REQ_OP_READ, false))
2430 success = 1;
2431 rdev_dec_pending(rdev, mddev);
2432 if (success)
2433 break;
2434 }
2435
2436 d++;
2437 if (d == conf->raid_disks * 2)
2438 d = 0;
2439 } while (d != read_disk);
2440
2441 if (!success) {
2442 /* Cannot read from anywhere - mark it bad */
2443 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2444 if (!rdev_set_badblocks(rdev, sect, s, 0))
2445 md_error(mddev, rdev);
2446 break;
2447 }
2448 /* write it back and re-read */
2449 start = d;
2450 while (d != read_disk) {
2451 if (d==0)
2452 d = conf->raid_disks * 2;
2453 d--;
2454 rdev = conf->mirrors[d].rdev;
2455 if (rdev &&
2456 !test_bit(Faulty, &rdev->flags)) {
2457 atomic_inc(&rdev->nr_pending);
2458 r1_sync_page_io(rdev, sect, s,
2459 conf->tmppage, REQ_OP_WRITE);
2460 rdev_dec_pending(rdev, mddev);
2461 }
2462 }
2463 d = start;
2464 while (d != read_disk) {
2465 if (d==0)
2466 d = conf->raid_disks * 2;
2467 d--;
2468 rdev = conf->mirrors[d].rdev;
2469 if (rdev &&
2470 !test_bit(Faulty, &rdev->flags)) {
2471 atomic_inc(&rdev->nr_pending);
2472 if (r1_sync_page_io(rdev, sect, s,
2473 conf->tmppage, REQ_OP_READ)) {
2474 atomic_add(s, &rdev->corrected_errors);
2475 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2476 mdname(mddev), s,
2477 (unsigned long long)(sect +
2478 rdev->data_offset),
2479 rdev->bdev);
2480 }
2481 rdev_dec_pending(rdev, mddev);
2482 }
2483 }
2484 sectors -= s;
2485 sect += s;
2486 }
2487}
2488
2489static int narrow_write_error(struct r1bio *r1_bio, int i)
2490{
2491 struct mddev *mddev = r1_bio->mddev;
2492 struct r1conf *conf = mddev->private;
2493 struct md_rdev *rdev = conf->mirrors[i].rdev;
2494
2495 /* bio has the data to be written to device 'i' where
2496 * we just recently had a write error.
2497 * We repeatedly clone the bio and trim down to one block,
2498 * then try the write. Where the write fails we record
2499 * a bad block.
2500 * It is conceivable that the bio doesn't exactly align with
2501 * blocks. We must handle this somehow.
2502 *
2503 * We currently own a reference on the rdev.
2504 */
2505
2506 int block_sectors;
2507 sector_t sector;
2508 int sectors;
2509 int sect_to_write = r1_bio->sectors;
2510 int ok = 1;
2511
2512 if (rdev->badblocks.shift < 0)
2513 return 0;
2514
2515 block_sectors = roundup(1 << rdev->badblocks.shift,
2516 bdev_logical_block_size(rdev->bdev) >> 9);
2517 sector = r1_bio->sector;
2518 sectors = ((sector + block_sectors)
2519 & ~(sector_t)(block_sectors - 1))
2520 - sector;
2521
2522 while (sect_to_write) {
2523 struct bio *wbio;
2524 if (sectors > sect_to_write)
2525 sectors = sect_to_write;
2526 /* Write at 'sector' for 'sectors'*/
2527
2528 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2529 wbio = bio_alloc_clone(rdev->bdev,
2530 r1_bio->behind_master_bio,
2531 GFP_NOIO, &mddev->bio_set);
2532 } else {
2533 wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2534 GFP_NOIO, &mddev->bio_set);
2535 }
2536
2537 wbio->bi_opf = REQ_OP_WRITE;
2538 wbio->bi_iter.bi_sector = r1_bio->sector;
2539 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2540
2541 bio_trim(wbio, sector - r1_bio->sector, sectors);
2542 wbio->bi_iter.bi_sector += rdev->data_offset;
2543
2544 if (submit_bio_wait(wbio) < 0)
2545 /* failure! */
2546 ok = rdev_set_badblocks(rdev, sector,
2547 sectors, 0)
2548 && ok;
2549
2550 bio_put(wbio);
2551 sect_to_write -= sectors;
2552 sector += sectors;
2553 sectors = block_sectors;
2554 }
2555 return ok;
2556}
2557
2558static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2559{
2560 int m;
2561 int s = r1_bio->sectors;
2562 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2563 struct md_rdev *rdev = conf->mirrors[m].rdev;
2564 struct bio *bio = r1_bio->bios[m];
2565 if (bio->bi_end_io == NULL)
2566 continue;
2567 if (!bio->bi_status &&
2568 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2569 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2570 }
2571 if (bio->bi_status &&
2572 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2573 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2574 md_error(conf->mddev, rdev);
2575 }
2576 }
2577 put_buf(r1_bio);
2578 md_done_sync(conf->mddev, s, 1);
2579}
2580
2581static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2582{
2583 int m, idx;
2584 bool fail = false;
2585
2586 for (m = 0; m < conf->raid_disks * 2 ; m++)
2587 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2588 struct md_rdev *rdev = conf->mirrors[m].rdev;
2589 rdev_clear_badblocks(rdev,
2590 r1_bio->sector,
2591 r1_bio->sectors, 0);
2592 rdev_dec_pending(rdev, conf->mddev);
2593 } else if (r1_bio->bios[m] != NULL) {
2594 /* This drive got a write error. We need to
2595 * narrow down and record precise write
2596 * errors.
2597 */
2598 fail = true;
2599 if (!narrow_write_error(r1_bio, m))
2600 md_error(conf->mddev,
2601 conf->mirrors[m].rdev);
2602 /* an I/O failed, we can't clear the bitmap */
2603 rdev_dec_pending(conf->mirrors[m].rdev,
2604 conf->mddev);
2605 }
2606 if (fail) {
2607 spin_lock_irq(&conf->device_lock);
2608 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2609 idx = sector_to_idx(r1_bio->sector);
2610 atomic_inc(&conf->nr_queued[idx]);
2611 spin_unlock_irq(&conf->device_lock);
2612 /*
2613 * In case freeze_array() is waiting for condition
2614 * get_unqueued_pending() == extra to be true.
2615 */
2616 wake_up(&conf->wait_barrier);
2617 md_wakeup_thread(conf->mddev->thread);
2618 } else {
2619 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2620 close_write(r1_bio);
2621 raid_end_bio_io(r1_bio);
2622 }
2623}
2624
2625static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2626{
2627 struct mddev *mddev = conf->mddev;
2628 struct bio *bio;
2629 struct md_rdev *rdev;
2630 sector_t sector;
2631
2632 clear_bit(R1BIO_ReadError, &r1_bio->state);
2633 /* we got a read error. Maybe the drive is bad. Maybe just
2634 * the block and we can fix it.
2635 * We freeze all other IO, and try reading the block from
2636 * other devices. When we find one, we re-write
2637 * and check it that fixes the read error.
2638 * This is all done synchronously while the array is
2639 * frozen
2640 */
2641
2642 bio = r1_bio->bios[r1_bio->read_disk];
2643 bio_put(bio);
2644 r1_bio->bios[r1_bio->read_disk] = NULL;
2645
2646 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2647 if (mddev->ro == 0
2648 && !test_bit(FailFast, &rdev->flags)) {
2649 freeze_array(conf, 1);
2650 fix_read_error(conf, r1_bio);
2651 unfreeze_array(conf);
2652 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2653 md_error(mddev, rdev);
2654 } else {
2655 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2656 }
2657
2658 rdev_dec_pending(rdev, conf->mddev);
2659 sector = r1_bio->sector;
2660 bio = r1_bio->master_bio;
2661
2662 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2663 r1_bio->state = 0;
2664 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2665 allow_barrier(conf, sector);
2666}
2667
2668static void raid1d(struct md_thread *thread)
2669{
2670 struct mddev *mddev = thread->mddev;
2671 struct r1bio *r1_bio;
2672 unsigned long flags;
2673 struct r1conf *conf = mddev->private;
2674 struct list_head *head = &conf->retry_list;
2675 struct blk_plug plug;
2676 int idx;
2677
2678 md_check_recovery(mddev);
2679
2680 if (!list_empty_careful(&conf->bio_end_io_list) &&
2681 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2682 LIST_HEAD(tmp);
2683 spin_lock_irqsave(&conf->device_lock, flags);
2684 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2685 list_splice_init(&conf->bio_end_io_list, &tmp);
2686 spin_unlock_irqrestore(&conf->device_lock, flags);
2687 while (!list_empty(&tmp)) {
2688 r1_bio = list_first_entry(&tmp, struct r1bio,
2689 retry_list);
2690 list_del(&r1_bio->retry_list);
2691 idx = sector_to_idx(r1_bio->sector);
2692 atomic_dec(&conf->nr_queued[idx]);
2693 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2694 close_write(r1_bio);
2695 raid_end_bio_io(r1_bio);
2696 }
2697 }
2698
2699 blk_start_plug(&plug);
2700 for (;;) {
2701
2702 flush_pending_writes(conf);
2703
2704 spin_lock_irqsave(&conf->device_lock, flags);
2705 if (list_empty(head)) {
2706 spin_unlock_irqrestore(&conf->device_lock, flags);
2707 break;
2708 }
2709 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2710 list_del(head->prev);
2711 idx = sector_to_idx(r1_bio->sector);
2712 atomic_dec(&conf->nr_queued[idx]);
2713 spin_unlock_irqrestore(&conf->device_lock, flags);
2714
2715 mddev = r1_bio->mddev;
2716 conf = mddev->private;
2717 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2718 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2719 test_bit(R1BIO_WriteError, &r1_bio->state))
2720 handle_sync_write_finished(conf, r1_bio);
2721 else
2722 sync_request_write(mddev, r1_bio);
2723 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2724 test_bit(R1BIO_WriteError, &r1_bio->state))
2725 handle_write_finished(conf, r1_bio);
2726 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2727 handle_read_error(conf, r1_bio);
2728 else
2729 WARN_ON_ONCE(1);
2730
2731 cond_resched();
2732 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2733 md_check_recovery(mddev);
2734 }
2735 blk_finish_plug(&plug);
2736}
2737
2738static int init_resync(struct r1conf *conf)
2739{
2740 int buffs;
2741
2742 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2743 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2744
2745 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2746 r1buf_pool_free, conf->poolinfo);
2747}
2748
2749static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2750{
2751 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2752 struct resync_pages *rps;
2753 struct bio *bio;
2754 int i;
2755
2756 for (i = conf->poolinfo->raid_disks; i--; ) {
2757 bio = r1bio->bios[i];
2758 rps = bio->bi_private;
2759 bio_reset(bio, NULL, 0);
2760 bio->bi_private = rps;
2761 }
2762 r1bio->master_bio = NULL;
2763 return r1bio;
2764}
2765
2766/*
2767 * perform a "sync" on one "block"
2768 *
2769 * We need to make sure that no normal I/O request - particularly write
2770 * requests - conflict with active sync requests.
2771 *
2772 * This is achieved by tracking pending requests and a 'barrier' concept
2773 * that can be installed to exclude normal IO requests.
2774 */
2775
2776static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2777 sector_t max_sector, int *skipped)
2778{
2779 struct r1conf *conf = mddev->private;
2780 struct r1bio *r1_bio;
2781 struct bio *bio;
2782 sector_t nr_sectors;
2783 int disk = -1;
2784 int i;
2785 int wonly = -1;
2786 int write_targets = 0, read_targets = 0;
2787 sector_t sync_blocks;
2788 bool still_degraded = false;
2789 int good_sectors = RESYNC_SECTORS;
2790 int min_bad = 0; /* number of sectors that are bad in all devices */
2791 int idx = sector_to_idx(sector_nr);
2792 int page_idx = 0;
2793
2794 if (!mempool_initialized(&conf->r1buf_pool))
2795 if (init_resync(conf))
2796 return 0;
2797
2798 if (sector_nr >= max_sector) {
2799 /* If we aborted, we need to abort the
2800 * sync on the 'current' bitmap chunk (there will
2801 * only be one in raid1 resync.
2802 * We can find the current addess in mddev->curr_resync
2803 */
2804 if (mddev->curr_resync < max_sector) /* aborted */
2805 mddev->bitmap_ops->end_sync(mddev, mddev->curr_resync,
2806 &sync_blocks);
2807 else /* completed sync */
2808 conf->fullsync = 0;
2809
2810 mddev->bitmap_ops->close_sync(mddev);
2811 close_sync(conf);
2812
2813 if (mddev_is_clustered(mddev)) {
2814 conf->cluster_sync_low = 0;
2815 conf->cluster_sync_high = 0;
2816 }
2817 return 0;
2818 }
2819
2820 if (mddev->bitmap == NULL &&
2821 mddev->recovery_cp == MaxSector &&
2822 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2823 conf->fullsync == 0) {
2824 *skipped = 1;
2825 return max_sector - sector_nr;
2826 }
2827 /* before building a request, check if we can skip these blocks..
2828 * This call the bitmap_start_sync doesn't actually record anything
2829 */
2830 if (!mddev->bitmap_ops->start_sync(mddev, sector_nr, &sync_blocks, true) &&
2831 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2832 /* We can skip this block, and probably several more */
2833 *skipped = 1;
2834 return sync_blocks;
2835 }
2836
2837 /*
2838 * If there is non-resync activity waiting for a turn, then let it
2839 * though before starting on this new sync request.
2840 */
2841 if (atomic_read(&conf->nr_waiting[idx]))
2842 schedule_timeout_uninterruptible(1);
2843
2844 /* we are incrementing sector_nr below. To be safe, we check against
2845 * sector_nr + two times RESYNC_SECTORS
2846 */
2847
2848 mddev->bitmap_ops->cond_end_sync(mddev, sector_nr,
2849 mddev_is_clustered(mddev) &&
2850 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2851
2852 if (raise_barrier(conf, sector_nr))
2853 return 0;
2854
2855 r1_bio = raid1_alloc_init_r1buf(conf);
2856
2857 /*
2858 * If we get a correctably read error during resync or recovery,
2859 * we might want to read from a different device. So we
2860 * flag all drives that could conceivably be read from for READ,
2861 * and any others (which will be non-In_sync devices) for WRITE.
2862 * If a read fails, we try reading from something else for which READ
2863 * is OK.
2864 */
2865
2866 r1_bio->mddev = mddev;
2867 r1_bio->sector = sector_nr;
2868 r1_bio->state = 0;
2869 set_bit(R1BIO_IsSync, &r1_bio->state);
2870 /* make sure good_sectors won't go across barrier unit boundary */
2871 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2872
2873 for (i = 0; i < conf->raid_disks * 2; i++) {
2874 struct md_rdev *rdev;
2875 bio = r1_bio->bios[i];
2876
2877 rdev = conf->mirrors[i].rdev;
2878 if (rdev == NULL ||
2879 test_bit(Faulty, &rdev->flags)) {
2880 if (i < conf->raid_disks)
2881 still_degraded = true;
2882 } else if (!test_bit(In_sync, &rdev->flags)) {
2883 bio->bi_opf = REQ_OP_WRITE;
2884 bio->bi_end_io = end_sync_write;
2885 write_targets ++;
2886 } else {
2887 /* may need to read from here */
2888 sector_t first_bad = MaxSector;
2889 int bad_sectors;
2890
2891 if (is_badblock(rdev, sector_nr, good_sectors,
2892 &first_bad, &bad_sectors)) {
2893 if (first_bad > sector_nr)
2894 good_sectors = first_bad - sector_nr;
2895 else {
2896 bad_sectors -= (sector_nr - first_bad);
2897 if (min_bad == 0 ||
2898 min_bad > bad_sectors)
2899 min_bad = bad_sectors;
2900 }
2901 }
2902 if (sector_nr < first_bad) {
2903 if (test_bit(WriteMostly, &rdev->flags)) {
2904 if (wonly < 0)
2905 wonly = i;
2906 } else {
2907 if (disk < 0)
2908 disk = i;
2909 }
2910 bio->bi_opf = REQ_OP_READ;
2911 bio->bi_end_io = end_sync_read;
2912 read_targets++;
2913 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2914 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2915 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2916 /*
2917 * The device is suitable for reading (InSync),
2918 * but has bad block(s) here. Let's try to correct them,
2919 * if we are doing resync or repair. Otherwise, leave
2920 * this device alone for this sync request.
2921 */
2922 bio->bi_opf = REQ_OP_WRITE;
2923 bio->bi_end_io = end_sync_write;
2924 write_targets++;
2925 }
2926 }
2927 if (rdev && bio->bi_end_io) {
2928 atomic_inc(&rdev->nr_pending);
2929 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2930 bio_set_dev(bio, rdev->bdev);
2931 if (test_bit(FailFast, &rdev->flags))
2932 bio->bi_opf |= MD_FAILFAST;
2933 }
2934 }
2935 if (disk < 0)
2936 disk = wonly;
2937 r1_bio->read_disk = disk;
2938
2939 if (read_targets == 0 && min_bad > 0) {
2940 /* These sectors are bad on all InSync devices, so we
2941 * need to mark them bad on all write targets
2942 */
2943 int ok = 1;
2944 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2945 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2946 struct md_rdev *rdev = conf->mirrors[i].rdev;
2947 ok = rdev_set_badblocks(rdev, sector_nr,
2948 min_bad, 0
2949 ) && ok;
2950 }
2951 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2952 *skipped = 1;
2953 put_buf(r1_bio);
2954
2955 if (!ok) {
2956 /* Cannot record the badblocks, so need to
2957 * abort the resync.
2958 * If there are multiple read targets, could just
2959 * fail the really bad ones ???
2960 */
2961 conf->recovery_disabled = mddev->recovery_disabled;
2962 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2963 return 0;
2964 } else
2965 return min_bad;
2966
2967 }
2968 if (min_bad > 0 && min_bad < good_sectors) {
2969 /* only resync enough to reach the next bad->good
2970 * transition */
2971 good_sectors = min_bad;
2972 }
2973
2974 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2975 /* extra read targets are also write targets */
2976 write_targets += read_targets-1;
2977
2978 if (write_targets == 0 || read_targets == 0) {
2979 /* There is nowhere to write, so all non-sync
2980 * drives must be failed - so we are finished
2981 */
2982 sector_t rv;
2983 if (min_bad > 0)
2984 max_sector = sector_nr + min_bad;
2985 rv = max_sector - sector_nr;
2986 *skipped = 1;
2987 put_buf(r1_bio);
2988 return rv;
2989 }
2990
2991 if (max_sector > mddev->resync_max)
2992 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2993 if (max_sector > sector_nr + good_sectors)
2994 max_sector = sector_nr + good_sectors;
2995 nr_sectors = 0;
2996 sync_blocks = 0;
2997 do {
2998 struct page *page;
2999 int len = PAGE_SIZE;
3000 if (sector_nr + (len>>9) > max_sector)
3001 len = (max_sector - sector_nr) << 9;
3002 if (len == 0)
3003 break;
3004 if (sync_blocks == 0) {
3005 if (!mddev->bitmap_ops->start_sync(mddev, sector_nr,
3006 &sync_blocks, still_degraded) &&
3007 !conf->fullsync &&
3008 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
3009 break;
3010 if ((len >> 9) > sync_blocks)
3011 len = sync_blocks<<9;
3012 }
3013
3014 for (i = 0 ; i < conf->raid_disks * 2; i++) {
3015 struct resync_pages *rp;
3016
3017 bio = r1_bio->bios[i];
3018 rp = get_resync_pages(bio);
3019 if (bio->bi_end_io) {
3020 page = resync_fetch_page(rp, page_idx);
3021
3022 /*
3023 * won't fail because the vec table is big
3024 * enough to hold all these pages
3025 */
3026 __bio_add_page(bio, page, len, 0);
3027 }
3028 }
3029 nr_sectors += len>>9;
3030 sector_nr += len>>9;
3031 sync_blocks -= (len>>9);
3032 } while (++page_idx < RESYNC_PAGES);
3033
3034 r1_bio->sectors = nr_sectors;
3035
3036 if (mddev_is_clustered(mddev) &&
3037 conf->cluster_sync_high < sector_nr + nr_sectors) {
3038 conf->cluster_sync_low = mddev->curr_resync_completed;
3039 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
3040 /* Send resync message */
3041 md_cluster_ops->resync_info_update(mddev,
3042 conf->cluster_sync_low,
3043 conf->cluster_sync_high);
3044 }
3045
3046 /* For a user-requested sync, we read all readable devices and do a
3047 * compare
3048 */
3049 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
3050 atomic_set(&r1_bio->remaining, read_targets);
3051 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
3052 bio = r1_bio->bios[i];
3053 if (bio->bi_end_io == end_sync_read) {
3054 read_targets--;
3055 md_sync_acct_bio(bio, nr_sectors);
3056 if (read_targets == 1)
3057 bio->bi_opf &= ~MD_FAILFAST;
3058 submit_bio_noacct(bio);
3059 }
3060 }
3061 } else {
3062 atomic_set(&r1_bio->remaining, 1);
3063 bio = r1_bio->bios[r1_bio->read_disk];
3064 md_sync_acct_bio(bio, nr_sectors);
3065 if (read_targets == 1)
3066 bio->bi_opf &= ~MD_FAILFAST;
3067 submit_bio_noacct(bio);
3068 }
3069 return nr_sectors;
3070}
3071
3072static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3073{
3074 if (sectors)
3075 return sectors;
3076
3077 return mddev->dev_sectors;
3078}
3079
3080static struct r1conf *setup_conf(struct mddev *mddev)
3081{
3082 struct r1conf *conf;
3083 int i;
3084 struct raid1_info *disk;
3085 struct md_rdev *rdev;
3086 int err = -ENOMEM;
3087
3088 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
3089 if (!conf)
3090 goto abort;
3091
3092 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
3093 sizeof(atomic_t), GFP_KERNEL);
3094 if (!conf->nr_pending)
3095 goto abort;
3096
3097 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
3098 sizeof(atomic_t), GFP_KERNEL);
3099 if (!conf->nr_waiting)
3100 goto abort;
3101
3102 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
3103 sizeof(atomic_t), GFP_KERNEL);
3104 if (!conf->nr_queued)
3105 goto abort;
3106
3107 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
3108 sizeof(atomic_t), GFP_KERNEL);
3109 if (!conf->barrier)
3110 goto abort;
3111
3112 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3113 mddev->raid_disks, 2),
3114 GFP_KERNEL);
3115 if (!conf->mirrors)
3116 goto abort;
3117
3118 conf->tmppage = alloc_page(GFP_KERNEL);
3119 if (!conf->tmppage)
3120 goto abort;
3121
3122 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3123 if (!conf->poolinfo)
3124 goto abort;
3125 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3126 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3127 rbio_pool_free, conf->poolinfo);
3128 if (err)
3129 goto abort;
3130
3131 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3132 if (err)
3133 goto abort;
3134
3135 conf->poolinfo->mddev = mddev;
3136
3137 err = -EINVAL;
3138 spin_lock_init(&conf->device_lock);
3139 conf->raid_disks = mddev->raid_disks;
3140 rdev_for_each(rdev, mddev) {
3141 int disk_idx = rdev->raid_disk;
3142
3143 if (disk_idx >= conf->raid_disks || disk_idx < 0)
3144 continue;
3145
3146 if (!raid1_add_conf(conf, rdev, disk_idx,
3147 test_bit(Replacement, &rdev->flags)))
3148 goto abort;
3149 }
3150 conf->mddev = mddev;
3151 INIT_LIST_HEAD(&conf->retry_list);
3152 INIT_LIST_HEAD(&conf->bio_end_io_list);
3153
3154 spin_lock_init(&conf->resync_lock);
3155 init_waitqueue_head(&conf->wait_barrier);
3156
3157 bio_list_init(&conf->pending_bio_list);
3158 conf->recovery_disabled = mddev->recovery_disabled - 1;
3159
3160 err = -EIO;
3161 for (i = 0; i < conf->raid_disks * 2; i++) {
3162
3163 disk = conf->mirrors + i;
3164
3165 if (i < conf->raid_disks &&
3166 disk[conf->raid_disks].rdev) {
3167 /* This slot has a replacement. */
3168 if (!disk->rdev) {
3169 /* No original, just make the replacement
3170 * a recovering spare
3171 */
3172 disk->rdev =
3173 disk[conf->raid_disks].rdev;
3174 disk[conf->raid_disks].rdev = NULL;
3175 } else if (!test_bit(In_sync, &disk->rdev->flags))
3176 /* Original is not in_sync - bad */
3177 goto abort;
3178 }
3179
3180 if (!disk->rdev ||
3181 !test_bit(In_sync, &disk->rdev->flags)) {
3182 disk->head_position = 0;
3183 if (disk->rdev &&
3184 (disk->rdev->saved_raid_disk < 0))
3185 conf->fullsync = 1;
3186 }
3187 }
3188
3189 err = -ENOMEM;
3190 rcu_assign_pointer(conf->thread,
3191 md_register_thread(raid1d, mddev, "raid1"));
3192 if (!conf->thread)
3193 goto abort;
3194
3195 return conf;
3196
3197 abort:
3198 if (conf) {
3199 mempool_exit(&conf->r1bio_pool);
3200 kfree(conf->mirrors);
3201 safe_put_page(conf->tmppage);
3202 kfree(conf->poolinfo);
3203 kfree(conf->nr_pending);
3204 kfree(conf->nr_waiting);
3205 kfree(conf->nr_queued);
3206 kfree(conf->barrier);
3207 bioset_exit(&conf->bio_split);
3208 kfree(conf);
3209 }
3210 return ERR_PTR(err);
3211}
3212
3213static int raid1_set_limits(struct mddev *mddev)
3214{
3215 struct queue_limits lim;
3216 int err;
3217
3218 md_init_stacking_limits(&lim);
3219 lim.max_write_zeroes_sectors = 0;
3220 lim.features |= BLK_FEAT_ATOMIC_WRITES_STACKED;
3221 err = mddev_stack_rdev_limits(mddev, &lim, MDDEV_STACK_INTEGRITY);
3222 if (err)
3223 return err;
3224 return queue_limits_set(mddev->gendisk->queue, &lim);
3225}
3226
3227static int raid1_run(struct mddev *mddev)
3228{
3229 struct r1conf *conf;
3230 int i;
3231 int ret;
3232
3233 if (mddev->level != 1) {
3234 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3235 mdname(mddev), mddev->level);
3236 return -EIO;
3237 }
3238 if (mddev->reshape_position != MaxSector) {
3239 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3240 mdname(mddev));
3241 return -EIO;
3242 }
3243
3244 /*
3245 * copy the already verified devices into our private RAID1
3246 * bookkeeping area. [whatever we allocate in run(),
3247 * should be freed in raid1_free()]
3248 */
3249 if (mddev->private == NULL)
3250 conf = setup_conf(mddev);
3251 else
3252 conf = mddev->private;
3253
3254 if (IS_ERR(conf))
3255 return PTR_ERR(conf);
3256
3257 if (!mddev_is_dm(mddev)) {
3258 ret = raid1_set_limits(mddev);
3259 if (ret)
3260 return ret;
3261 }
3262
3263 mddev->degraded = 0;
3264 for (i = 0; i < conf->raid_disks; i++)
3265 if (conf->mirrors[i].rdev == NULL ||
3266 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3267 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3268 mddev->degraded++;
3269 /*
3270 * RAID1 needs at least one disk in active
3271 */
3272 if (conf->raid_disks - mddev->degraded < 1) {
3273 md_unregister_thread(mddev, &conf->thread);
3274 return -EINVAL;
3275 }
3276
3277 if (conf->raid_disks - mddev->degraded == 1)
3278 mddev->recovery_cp = MaxSector;
3279
3280 if (mddev->recovery_cp != MaxSector)
3281 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3282 mdname(mddev));
3283 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3284 mdname(mddev), mddev->raid_disks - mddev->degraded,
3285 mddev->raid_disks);
3286
3287 /*
3288 * Ok, everything is just fine now
3289 */
3290 rcu_assign_pointer(mddev->thread, conf->thread);
3291 rcu_assign_pointer(conf->thread, NULL);
3292 mddev->private = conf;
3293 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3294
3295 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3296
3297 ret = md_integrity_register(mddev);
3298 if (ret)
3299 md_unregister_thread(mddev, &mddev->thread);
3300 return ret;
3301}
3302
3303static void raid1_free(struct mddev *mddev, void *priv)
3304{
3305 struct r1conf *conf = priv;
3306
3307 mempool_exit(&conf->r1bio_pool);
3308 kfree(conf->mirrors);
3309 safe_put_page(conf->tmppage);
3310 kfree(conf->poolinfo);
3311 kfree(conf->nr_pending);
3312 kfree(conf->nr_waiting);
3313 kfree(conf->nr_queued);
3314 kfree(conf->barrier);
3315 bioset_exit(&conf->bio_split);
3316 kfree(conf);
3317}
3318
3319static int raid1_resize(struct mddev *mddev, sector_t sectors)
3320{
3321 /* no resync is happening, and there is enough space
3322 * on all devices, so we can resize.
3323 * We need to make sure resync covers any new space.
3324 * If the array is shrinking we should possibly wait until
3325 * any io in the removed space completes, but it hardly seems
3326 * worth it.
3327 */
3328 sector_t newsize = raid1_size(mddev, sectors, 0);
3329 int ret;
3330
3331 if (mddev->external_size &&
3332 mddev->array_sectors > newsize)
3333 return -EINVAL;
3334
3335 ret = mddev->bitmap_ops->resize(mddev, newsize, 0, false);
3336 if (ret)
3337 return ret;
3338
3339 md_set_array_sectors(mddev, newsize);
3340 if (sectors > mddev->dev_sectors &&
3341 mddev->recovery_cp > mddev->dev_sectors) {
3342 mddev->recovery_cp = mddev->dev_sectors;
3343 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3344 }
3345 mddev->dev_sectors = sectors;
3346 mddev->resync_max_sectors = sectors;
3347 return 0;
3348}
3349
3350static int raid1_reshape(struct mddev *mddev)
3351{
3352 /* We need to:
3353 * 1/ resize the r1bio_pool
3354 * 2/ resize conf->mirrors
3355 *
3356 * We allocate a new r1bio_pool if we can.
3357 * Then raise a device barrier and wait until all IO stops.
3358 * Then resize conf->mirrors and swap in the new r1bio pool.
3359 *
3360 * At the same time, we "pack" the devices so that all the missing
3361 * devices have the higher raid_disk numbers.
3362 */
3363 mempool_t newpool, oldpool;
3364 struct pool_info *newpoolinfo;
3365 struct raid1_info *newmirrors;
3366 struct r1conf *conf = mddev->private;
3367 int cnt, raid_disks;
3368 unsigned long flags;
3369 int d, d2;
3370 int ret;
3371
3372 memset(&newpool, 0, sizeof(newpool));
3373 memset(&oldpool, 0, sizeof(oldpool));
3374
3375 /* Cannot change chunk_size, layout, or level */
3376 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3377 mddev->layout != mddev->new_layout ||
3378 mddev->level != mddev->new_level) {
3379 mddev->new_chunk_sectors = mddev->chunk_sectors;
3380 mddev->new_layout = mddev->layout;
3381 mddev->new_level = mddev->level;
3382 return -EINVAL;
3383 }
3384
3385 if (!mddev_is_clustered(mddev))
3386 md_allow_write(mddev);
3387
3388 raid_disks = mddev->raid_disks + mddev->delta_disks;
3389
3390 if (raid_disks < conf->raid_disks) {
3391 cnt=0;
3392 for (d= 0; d < conf->raid_disks; d++)
3393 if (conf->mirrors[d].rdev)
3394 cnt++;
3395 if (cnt > raid_disks)
3396 return -EBUSY;
3397 }
3398
3399 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3400 if (!newpoolinfo)
3401 return -ENOMEM;
3402 newpoolinfo->mddev = mddev;
3403 newpoolinfo->raid_disks = raid_disks * 2;
3404
3405 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3406 rbio_pool_free, newpoolinfo);
3407 if (ret) {
3408 kfree(newpoolinfo);
3409 return ret;
3410 }
3411 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3412 raid_disks, 2),
3413 GFP_KERNEL);
3414 if (!newmirrors) {
3415 kfree(newpoolinfo);
3416 mempool_exit(&newpool);
3417 return -ENOMEM;
3418 }
3419
3420 freeze_array(conf, 0);
3421
3422 /* ok, everything is stopped */
3423 oldpool = conf->r1bio_pool;
3424 conf->r1bio_pool = newpool;
3425
3426 for (d = d2 = 0; d < conf->raid_disks; d++) {
3427 struct md_rdev *rdev = conf->mirrors[d].rdev;
3428 if (rdev && rdev->raid_disk != d2) {
3429 sysfs_unlink_rdev(mddev, rdev);
3430 rdev->raid_disk = d2;
3431 sysfs_unlink_rdev(mddev, rdev);
3432 if (sysfs_link_rdev(mddev, rdev))
3433 pr_warn("md/raid1:%s: cannot register rd%d\n",
3434 mdname(mddev), rdev->raid_disk);
3435 }
3436 if (rdev)
3437 newmirrors[d2++].rdev = rdev;
3438 }
3439 kfree(conf->mirrors);
3440 conf->mirrors = newmirrors;
3441 kfree(conf->poolinfo);
3442 conf->poolinfo = newpoolinfo;
3443
3444 spin_lock_irqsave(&conf->device_lock, flags);
3445 mddev->degraded += (raid_disks - conf->raid_disks);
3446 spin_unlock_irqrestore(&conf->device_lock, flags);
3447 conf->raid_disks = mddev->raid_disks = raid_disks;
3448 mddev->delta_disks = 0;
3449
3450 unfreeze_array(conf);
3451
3452 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3453 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3454 md_wakeup_thread(mddev->thread);
3455
3456 mempool_exit(&oldpool);
3457 return 0;
3458}
3459
3460static void raid1_quiesce(struct mddev *mddev, int quiesce)
3461{
3462 struct r1conf *conf = mddev->private;
3463
3464 if (quiesce)
3465 freeze_array(conf, 0);
3466 else
3467 unfreeze_array(conf);
3468}
3469
3470static void *raid1_takeover(struct mddev *mddev)
3471{
3472 /* raid1 can take over:
3473 * raid5 with 2 devices, any layout or chunk size
3474 */
3475 if (mddev->level == 5 && mddev->raid_disks == 2) {
3476 struct r1conf *conf;
3477 mddev->new_level = 1;
3478 mddev->new_layout = 0;
3479 mddev->new_chunk_sectors = 0;
3480 conf = setup_conf(mddev);
3481 if (!IS_ERR(conf)) {
3482 /* Array must appear to be quiesced */
3483 conf->array_frozen = 1;
3484 mddev_clear_unsupported_flags(mddev,
3485 UNSUPPORTED_MDDEV_FLAGS);
3486 }
3487 return conf;
3488 }
3489 return ERR_PTR(-EINVAL);
3490}
3491
3492static struct md_personality raid1_personality =
3493{
3494 .name = "raid1",
3495 .level = 1,
3496 .owner = THIS_MODULE,
3497 .make_request = raid1_make_request,
3498 .run = raid1_run,
3499 .free = raid1_free,
3500 .status = raid1_status,
3501 .error_handler = raid1_error,
3502 .hot_add_disk = raid1_add_disk,
3503 .hot_remove_disk= raid1_remove_disk,
3504 .spare_active = raid1_spare_active,
3505 .sync_request = raid1_sync_request,
3506 .resize = raid1_resize,
3507 .size = raid1_size,
3508 .check_reshape = raid1_reshape,
3509 .quiesce = raid1_quiesce,
3510 .takeover = raid1_takeover,
3511};
3512
3513static int __init raid_init(void)
3514{
3515 return register_md_personality(&raid1_personality);
3516}
3517
3518static void raid_exit(void)
3519{
3520 unregister_md_personality(&raid1_personality);
3521}
3522
3523module_init(raid_init);
3524module_exit(raid_exit);
3525MODULE_LICENSE("GPL");
3526MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3527MODULE_ALIAS("md-personality-3"); /* RAID1 */
3528MODULE_ALIAS("md-raid1");
3529MODULE_ALIAS("md-level-1");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid1.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 *
7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 *
9 * RAID-1 management functions.
10 *
11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 *
13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 *
16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
17 * bitmapped intelligence in resync:
18 *
19 * - bitmap marked during normal i/o
20 * - bitmap used to skip nondirty blocks during sync
21 *
22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
23 * - persistent bitmap code
24 */
25
26#include <linux/slab.h>
27#include <linux/delay.h>
28#include <linux/blkdev.h>
29#include <linux/module.h>
30#include <linux/seq_file.h>
31#include <linux/ratelimit.h>
32#include <linux/interval_tree_generic.h>
33
34#include <trace/events/block.h>
35
36#include "md.h"
37#include "raid1.h"
38#include "md-bitmap.h"
39
40#define UNSUPPORTED_MDDEV_FLAGS \
41 ((1L << MD_HAS_JOURNAL) | \
42 (1L << MD_JOURNAL_CLEAN) | \
43 (1L << MD_HAS_PPL) | \
44 (1L << MD_HAS_MULTIPLE_PPLS))
45
46static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
47static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
48
49#define raid1_log(md, fmt, args...) \
50 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
51
52#include "raid1-10.c"
53
54#define START(node) ((node)->start)
55#define LAST(node) ((node)->last)
56INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
57 START, LAST, static inline, raid1_rb);
58
59static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
60 struct serial_info *si, int idx)
61{
62 unsigned long flags;
63 int ret = 0;
64 sector_t lo = r1_bio->sector;
65 sector_t hi = lo + r1_bio->sectors;
66 struct serial_in_rdev *serial = &rdev->serial[idx];
67
68 spin_lock_irqsave(&serial->serial_lock, flags);
69 /* collision happened */
70 if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
71 ret = -EBUSY;
72 else {
73 si->start = lo;
74 si->last = hi;
75 raid1_rb_insert(si, &serial->serial_rb);
76 }
77 spin_unlock_irqrestore(&serial->serial_lock, flags);
78
79 return ret;
80}
81
82static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
83{
84 struct mddev *mddev = rdev->mddev;
85 struct serial_info *si;
86 int idx = sector_to_idx(r1_bio->sector);
87 struct serial_in_rdev *serial = &rdev->serial[idx];
88
89 if (WARN_ON(!mddev->serial_info_pool))
90 return;
91 si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
92 wait_event(serial->serial_io_wait,
93 check_and_add_serial(rdev, r1_bio, si, idx) == 0);
94}
95
96static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
97{
98 struct serial_info *si;
99 unsigned long flags;
100 int found = 0;
101 struct mddev *mddev = rdev->mddev;
102 int idx = sector_to_idx(lo);
103 struct serial_in_rdev *serial = &rdev->serial[idx];
104
105 spin_lock_irqsave(&serial->serial_lock, flags);
106 for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
107 si; si = raid1_rb_iter_next(si, lo, hi)) {
108 if (si->start == lo && si->last == hi) {
109 raid1_rb_remove(si, &serial->serial_rb);
110 mempool_free(si, mddev->serial_info_pool);
111 found = 1;
112 break;
113 }
114 }
115 if (!found)
116 WARN(1, "The write IO is not recorded for serialization\n");
117 spin_unlock_irqrestore(&serial->serial_lock, flags);
118 wake_up(&serial->serial_io_wait);
119}
120
121/*
122 * for resync bio, r1bio pointer can be retrieved from the per-bio
123 * 'struct resync_pages'.
124 */
125static inline struct r1bio *get_resync_r1bio(struct bio *bio)
126{
127 return get_resync_pages(bio)->raid_bio;
128}
129
130static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
131{
132 struct pool_info *pi = data;
133 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
134
135 /* allocate a r1bio with room for raid_disks entries in the bios array */
136 return kzalloc(size, gfp_flags);
137}
138
139#define RESYNC_DEPTH 32
140#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
141#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
142#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
143#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
144#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
145
146static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
147{
148 struct pool_info *pi = data;
149 struct r1bio *r1_bio;
150 struct bio *bio;
151 int need_pages;
152 int j;
153 struct resync_pages *rps;
154
155 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
156 if (!r1_bio)
157 return NULL;
158
159 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
160 gfp_flags);
161 if (!rps)
162 goto out_free_r1bio;
163
164 /*
165 * Allocate bios : 1 for reading, n-1 for writing
166 */
167 for (j = pi->raid_disks ; j-- ; ) {
168 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
169 if (!bio)
170 goto out_free_bio;
171 r1_bio->bios[j] = bio;
172 }
173 /*
174 * Allocate RESYNC_PAGES data pages and attach them to
175 * the first bio.
176 * If this is a user-requested check/repair, allocate
177 * RESYNC_PAGES for each bio.
178 */
179 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
180 need_pages = pi->raid_disks;
181 else
182 need_pages = 1;
183 for (j = 0; j < pi->raid_disks; j++) {
184 struct resync_pages *rp = &rps[j];
185
186 bio = r1_bio->bios[j];
187
188 if (j < need_pages) {
189 if (resync_alloc_pages(rp, gfp_flags))
190 goto out_free_pages;
191 } else {
192 memcpy(rp, &rps[0], sizeof(*rp));
193 resync_get_all_pages(rp);
194 }
195
196 rp->raid_bio = r1_bio;
197 bio->bi_private = rp;
198 }
199
200 r1_bio->master_bio = NULL;
201
202 return r1_bio;
203
204out_free_pages:
205 while (--j >= 0)
206 resync_free_pages(&rps[j]);
207
208out_free_bio:
209 while (++j < pi->raid_disks)
210 bio_put(r1_bio->bios[j]);
211 kfree(rps);
212
213out_free_r1bio:
214 rbio_pool_free(r1_bio, data);
215 return NULL;
216}
217
218static void r1buf_pool_free(void *__r1_bio, void *data)
219{
220 struct pool_info *pi = data;
221 int i;
222 struct r1bio *r1bio = __r1_bio;
223 struct resync_pages *rp = NULL;
224
225 for (i = pi->raid_disks; i--; ) {
226 rp = get_resync_pages(r1bio->bios[i]);
227 resync_free_pages(rp);
228 bio_put(r1bio->bios[i]);
229 }
230
231 /* resync pages array stored in the 1st bio's .bi_private */
232 kfree(rp);
233
234 rbio_pool_free(r1bio, data);
235}
236
237static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
238{
239 int i;
240
241 for (i = 0; i < conf->raid_disks * 2; i++) {
242 struct bio **bio = r1_bio->bios + i;
243 if (!BIO_SPECIAL(*bio))
244 bio_put(*bio);
245 *bio = NULL;
246 }
247}
248
249static void free_r1bio(struct r1bio *r1_bio)
250{
251 struct r1conf *conf = r1_bio->mddev->private;
252
253 put_all_bios(conf, r1_bio);
254 mempool_free(r1_bio, &conf->r1bio_pool);
255}
256
257static void put_buf(struct r1bio *r1_bio)
258{
259 struct r1conf *conf = r1_bio->mddev->private;
260 sector_t sect = r1_bio->sector;
261 int i;
262
263 for (i = 0; i < conf->raid_disks * 2; i++) {
264 struct bio *bio = r1_bio->bios[i];
265 if (bio->bi_end_io)
266 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
267 }
268
269 mempool_free(r1_bio, &conf->r1buf_pool);
270
271 lower_barrier(conf, sect);
272}
273
274static void reschedule_retry(struct r1bio *r1_bio)
275{
276 unsigned long flags;
277 struct mddev *mddev = r1_bio->mddev;
278 struct r1conf *conf = mddev->private;
279 int idx;
280
281 idx = sector_to_idx(r1_bio->sector);
282 spin_lock_irqsave(&conf->device_lock, flags);
283 list_add(&r1_bio->retry_list, &conf->retry_list);
284 atomic_inc(&conf->nr_queued[idx]);
285 spin_unlock_irqrestore(&conf->device_lock, flags);
286
287 wake_up(&conf->wait_barrier);
288 md_wakeup_thread(mddev->thread);
289}
290
291/*
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
294 * cache layer.
295 */
296static void call_bio_endio(struct r1bio *r1_bio)
297{
298 struct bio *bio = r1_bio->master_bio;
299
300 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
301 bio->bi_status = BLK_STS_IOERR;
302
303 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
304 bio_end_io_acct(bio, r1_bio->start_time);
305 bio_endio(bio);
306}
307
308static void raid_end_bio_io(struct r1bio *r1_bio)
309{
310 struct bio *bio = r1_bio->master_bio;
311 struct r1conf *conf = r1_bio->mddev->private;
312
313 /* if nobody has done the final endio yet, do it now */
314 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
315 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
316 (bio_data_dir(bio) == WRITE) ? "write" : "read",
317 (unsigned long long) bio->bi_iter.bi_sector,
318 (unsigned long long) bio_end_sector(bio) - 1);
319
320 call_bio_endio(r1_bio);
321 }
322 /*
323 * Wake up any possible resync thread that waits for the device
324 * to go idle. All I/Os, even write-behind writes, are done.
325 */
326 allow_barrier(conf, r1_bio->sector);
327
328 free_r1bio(r1_bio);
329}
330
331/*
332 * Update disk head position estimator based on IRQ completion info.
333 */
334static inline void update_head_pos(int disk, struct r1bio *r1_bio)
335{
336 struct r1conf *conf = r1_bio->mddev->private;
337
338 conf->mirrors[disk].head_position =
339 r1_bio->sector + (r1_bio->sectors);
340}
341
342/*
343 * Find the disk number which triggered given bio
344 */
345static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
346{
347 int mirror;
348 struct r1conf *conf = r1_bio->mddev->private;
349 int raid_disks = conf->raid_disks;
350
351 for (mirror = 0; mirror < raid_disks * 2; mirror++)
352 if (r1_bio->bios[mirror] == bio)
353 break;
354
355 BUG_ON(mirror == raid_disks * 2);
356 update_head_pos(mirror, r1_bio);
357
358 return mirror;
359}
360
361static void raid1_end_read_request(struct bio *bio)
362{
363 int uptodate = !bio->bi_status;
364 struct r1bio *r1_bio = bio->bi_private;
365 struct r1conf *conf = r1_bio->mddev->private;
366 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
367
368 /*
369 * this branch is our 'one mirror IO has finished' event handler:
370 */
371 update_head_pos(r1_bio->read_disk, r1_bio);
372
373 if (uptodate)
374 set_bit(R1BIO_Uptodate, &r1_bio->state);
375 else if (test_bit(FailFast, &rdev->flags) &&
376 test_bit(R1BIO_FailFast, &r1_bio->state))
377 /* This was a fail-fast read so we definitely
378 * want to retry */
379 ;
380 else {
381 /* If all other devices have failed, we want to return
382 * the error upwards rather than fail the last device.
383 * Here we redefine "uptodate" to mean "Don't want to retry"
384 */
385 unsigned long flags;
386 spin_lock_irqsave(&conf->device_lock, flags);
387 if (r1_bio->mddev->degraded == conf->raid_disks ||
388 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
389 test_bit(In_sync, &rdev->flags)))
390 uptodate = 1;
391 spin_unlock_irqrestore(&conf->device_lock, flags);
392 }
393
394 if (uptodate) {
395 raid_end_bio_io(r1_bio);
396 rdev_dec_pending(rdev, conf->mddev);
397 } else {
398 /*
399 * oops, read error:
400 */
401 char b[BDEVNAME_SIZE];
402 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
403 mdname(conf->mddev),
404 bdevname(rdev->bdev, b),
405 (unsigned long long)r1_bio->sector);
406 set_bit(R1BIO_ReadError, &r1_bio->state);
407 reschedule_retry(r1_bio);
408 /* don't drop the reference on read_disk yet */
409 }
410}
411
412static void close_write(struct r1bio *r1_bio)
413{
414 /* it really is the end of this request */
415 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
416 bio_free_pages(r1_bio->behind_master_bio);
417 bio_put(r1_bio->behind_master_bio);
418 r1_bio->behind_master_bio = NULL;
419 }
420 /* clear the bitmap if all writes complete successfully */
421 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
422 r1_bio->sectors,
423 !test_bit(R1BIO_Degraded, &r1_bio->state),
424 test_bit(R1BIO_BehindIO, &r1_bio->state));
425 md_write_end(r1_bio->mddev);
426}
427
428static void r1_bio_write_done(struct r1bio *r1_bio)
429{
430 if (!atomic_dec_and_test(&r1_bio->remaining))
431 return;
432
433 if (test_bit(R1BIO_WriteError, &r1_bio->state))
434 reschedule_retry(r1_bio);
435 else {
436 close_write(r1_bio);
437 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
438 reschedule_retry(r1_bio);
439 else
440 raid_end_bio_io(r1_bio);
441 }
442}
443
444static void raid1_end_write_request(struct bio *bio)
445{
446 struct r1bio *r1_bio = bio->bi_private;
447 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
448 struct r1conf *conf = r1_bio->mddev->private;
449 struct bio *to_put = NULL;
450 int mirror = find_bio_disk(r1_bio, bio);
451 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
452 bool discard_error;
453 sector_t lo = r1_bio->sector;
454 sector_t hi = r1_bio->sector + r1_bio->sectors;
455
456 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
457
458 /*
459 * 'one mirror IO has finished' event handler:
460 */
461 if (bio->bi_status && !discard_error) {
462 set_bit(WriteErrorSeen, &rdev->flags);
463 if (!test_and_set_bit(WantReplacement, &rdev->flags))
464 set_bit(MD_RECOVERY_NEEDED, &
465 conf->mddev->recovery);
466
467 if (test_bit(FailFast, &rdev->flags) &&
468 (bio->bi_opf & MD_FAILFAST) &&
469 /* We never try FailFast to WriteMostly devices */
470 !test_bit(WriteMostly, &rdev->flags)) {
471 md_error(r1_bio->mddev, rdev);
472 }
473
474 /*
475 * When the device is faulty, it is not necessary to
476 * handle write error.
477 */
478 if (!test_bit(Faulty, &rdev->flags))
479 set_bit(R1BIO_WriteError, &r1_bio->state);
480 else {
481 /* Fail the request */
482 set_bit(R1BIO_Degraded, &r1_bio->state);
483 /* Finished with this branch */
484 r1_bio->bios[mirror] = NULL;
485 to_put = bio;
486 }
487 } else {
488 /*
489 * Set R1BIO_Uptodate in our master bio, so that we
490 * will return a good error code for to the higher
491 * levels even if IO on some other mirrored buffer
492 * fails.
493 *
494 * The 'master' represents the composite IO operation
495 * to user-side. So if something waits for IO, then it
496 * will wait for the 'master' bio.
497 */
498 sector_t first_bad;
499 int bad_sectors;
500
501 r1_bio->bios[mirror] = NULL;
502 to_put = bio;
503 /*
504 * Do not set R1BIO_Uptodate if the current device is
505 * rebuilding or Faulty. This is because we cannot use
506 * such device for properly reading the data back (we could
507 * potentially use it, if the current write would have felt
508 * before rdev->recovery_offset, but for simplicity we don't
509 * check this here.
510 */
511 if (test_bit(In_sync, &rdev->flags) &&
512 !test_bit(Faulty, &rdev->flags))
513 set_bit(R1BIO_Uptodate, &r1_bio->state);
514
515 /* Maybe we can clear some bad blocks. */
516 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
517 &first_bad, &bad_sectors) && !discard_error) {
518 r1_bio->bios[mirror] = IO_MADE_GOOD;
519 set_bit(R1BIO_MadeGood, &r1_bio->state);
520 }
521 }
522
523 if (behind) {
524 if (test_bit(CollisionCheck, &rdev->flags))
525 remove_serial(rdev, lo, hi);
526 if (test_bit(WriteMostly, &rdev->flags))
527 atomic_dec(&r1_bio->behind_remaining);
528
529 /*
530 * In behind mode, we ACK the master bio once the I/O
531 * has safely reached all non-writemostly
532 * disks. Setting the Returned bit ensures that this
533 * gets done only once -- we don't ever want to return
534 * -EIO here, instead we'll wait
535 */
536 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
537 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
538 /* Maybe we can return now */
539 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
540 struct bio *mbio = r1_bio->master_bio;
541 pr_debug("raid1: behind end write sectors"
542 " %llu-%llu\n",
543 (unsigned long long) mbio->bi_iter.bi_sector,
544 (unsigned long long) bio_end_sector(mbio) - 1);
545 call_bio_endio(r1_bio);
546 }
547 }
548 } else if (rdev->mddev->serialize_policy)
549 remove_serial(rdev, lo, hi);
550 if (r1_bio->bios[mirror] == NULL)
551 rdev_dec_pending(rdev, conf->mddev);
552
553 /*
554 * Let's see if all mirrored write operations have finished
555 * already.
556 */
557 r1_bio_write_done(r1_bio);
558
559 if (to_put)
560 bio_put(to_put);
561}
562
563static sector_t align_to_barrier_unit_end(sector_t start_sector,
564 sector_t sectors)
565{
566 sector_t len;
567
568 WARN_ON(sectors == 0);
569 /*
570 * len is the number of sectors from start_sector to end of the
571 * barrier unit which start_sector belongs to.
572 */
573 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
574 start_sector;
575
576 if (len > sectors)
577 len = sectors;
578
579 return len;
580}
581
582/*
583 * This routine returns the disk from which the requested read should
584 * be done. There is a per-array 'next expected sequential IO' sector
585 * number - if this matches on the next IO then we use the last disk.
586 * There is also a per-disk 'last know head position' sector that is
587 * maintained from IRQ contexts, both the normal and the resync IO
588 * completion handlers update this position correctly. If there is no
589 * perfect sequential match then we pick the disk whose head is closest.
590 *
591 * If there are 2 mirrors in the same 2 devices, performance degrades
592 * because position is mirror, not device based.
593 *
594 * The rdev for the device selected will have nr_pending incremented.
595 */
596static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
597{
598 const sector_t this_sector = r1_bio->sector;
599 int sectors;
600 int best_good_sectors;
601 int best_disk, best_dist_disk, best_pending_disk;
602 int has_nonrot_disk;
603 int disk;
604 sector_t best_dist;
605 unsigned int min_pending;
606 struct md_rdev *rdev;
607 int choose_first;
608 int choose_next_idle;
609
610 rcu_read_lock();
611 /*
612 * Check if we can balance. We can balance on the whole
613 * device if no resync is going on, or below the resync window.
614 * We take the first readable disk when above the resync window.
615 */
616 retry:
617 sectors = r1_bio->sectors;
618 best_disk = -1;
619 best_dist_disk = -1;
620 best_dist = MaxSector;
621 best_pending_disk = -1;
622 min_pending = UINT_MAX;
623 best_good_sectors = 0;
624 has_nonrot_disk = 0;
625 choose_next_idle = 0;
626 clear_bit(R1BIO_FailFast, &r1_bio->state);
627
628 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
629 (mddev_is_clustered(conf->mddev) &&
630 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
631 this_sector + sectors)))
632 choose_first = 1;
633 else
634 choose_first = 0;
635
636 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
637 sector_t dist;
638 sector_t first_bad;
639 int bad_sectors;
640 unsigned int pending;
641 bool nonrot;
642
643 rdev = rcu_dereference(conf->mirrors[disk].rdev);
644 if (r1_bio->bios[disk] == IO_BLOCKED
645 || rdev == NULL
646 || test_bit(Faulty, &rdev->flags))
647 continue;
648 if (!test_bit(In_sync, &rdev->flags) &&
649 rdev->recovery_offset < this_sector + sectors)
650 continue;
651 if (test_bit(WriteMostly, &rdev->flags)) {
652 /* Don't balance among write-mostly, just
653 * use the first as a last resort */
654 if (best_dist_disk < 0) {
655 if (is_badblock(rdev, this_sector, sectors,
656 &first_bad, &bad_sectors)) {
657 if (first_bad <= this_sector)
658 /* Cannot use this */
659 continue;
660 best_good_sectors = first_bad - this_sector;
661 } else
662 best_good_sectors = sectors;
663 best_dist_disk = disk;
664 best_pending_disk = disk;
665 }
666 continue;
667 }
668 /* This is a reasonable device to use. It might
669 * even be best.
670 */
671 if (is_badblock(rdev, this_sector, sectors,
672 &first_bad, &bad_sectors)) {
673 if (best_dist < MaxSector)
674 /* already have a better device */
675 continue;
676 if (first_bad <= this_sector) {
677 /* cannot read here. If this is the 'primary'
678 * device, then we must not read beyond
679 * bad_sectors from another device..
680 */
681 bad_sectors -= (this_sector - first_bad);
682 if (choose_first && sectors > bad_sectors)
683 sectors = bad_sectors;
684 if (best_good_sectors > sectors)
685 best_good_sectors = sectors;
686
687 } else {
688 sector_t good_sectors = first_bad - this_sector;
689 if (good_sectors > best_good_sectors) {
690 best_good_sectors = good_sectors;
691 best_disk = disk;
692 }
693 if (choose_first)
694 break;
695 }
696 continue;
697 } else {
698 if ((sectors > best_good_sectors) && (best_disk >= 0))
699 best_disk = -1;
700 best_good_sectors = sectors;
701 }
702
703 if (best_disk >= 0)
704 /* At least two disks to choose from so failfast is OK */
705 set_bit(R1BIO_FailFast, &r1_bio->state);
706
707 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
708 has_nonrot_disk |= nonrot;
709 pending = atomic_read(&rdev->nr_pending);
710 dist = abs(this_sector - conf->mirrors[disk].head_position);
711 if (choose_first) {
712 best_disk = disk;
713 break;
714 }
715 /* Don't change to another disk for sequential reads */
716 if (conf->mirrors[disk].next_seq_sect == this_sector
717 || dist == 0) {
718 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
719 struct raid1_info *mirror = &conf->mirrors[disk];
720
721 best_disk = disk;
722 /*
723 * If buffered sequential IO size exceeds optimal
724 * iosize, check if there is idle disk. If yes, choose
725 * the idle disk. read_balance could already choose an
726 * idle disk before noticing it's a sequential IO in
727 * this disk. This doesn't matter because this disk
728 * will idle, next time it will be utilized after the
729 * first disk has IO size exceeds optimal iosize. In
730 * this way, iosize of the first disk will be optimal
731 * iosize at least. iosize of the second disk might be
732 * small, but not a big deal since when the second disk
733 * starts IO, the first disk is likely still busy.
734 */
735 if (nonrot && opt_iosize > 0 &&
736 mirror->seq_start != MaxSector &&
737 mirror->next_seq_sect > opt_iosize &&
738 mirror->next_seq_sect - opt_iosize >=
739 mirror->seq_start) {
740 choose_next_idle = 1;
741 continue;
742 }
743 break;
744 }
745
746 if (choose_next_idle)
747 continue;
748
749 if (min_pending > pending) {
750 min_pending = pending;
751 best_pending_disk = disk;
752 }
753
754 if (dist < best_dist) {
755 best_dist = dist;
756 best_dist_disk = disk;
757 }
758 }
759
760 /*
761 * If all disks are rotational, choose the closest disk. If any disk is
762 * non-rotational, choose the disk with less pending request even the
763 * disk is rotational, which might/might not be optimal for raids with
764 * mixed ratation/non-rotational disks depending on workload.
765 */
766 if (best_disk == -1) {
767 if (has_nonrot_disk || min_pending == 0)
768 best_disk = best_pending_disk;
769 else
770 best_disk = best_dist_disk;
771 }
772
773 if (best_disk >= 0) {
774 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
775 if (!rdev)
776 goto retry;
777 atomic_inc(&rdev->nr_pending);
778 sectors = best_good_sectors;
779
780 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
781 conf->mirrors[best_disk].seq_start = this_sector;
782
783 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
784 }
785 rcu_read_unlock();
786 *max_sectors = sectors;
787
788 return best_disk;
789}
790
791static void flush_bio_list(struct r1conf *conf, struct bio *bio)
792{
793 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
794 md_bitmap_unplug(conf->mddev->bitmap);
795 wake_up(&conf->wait_barrier);
796
797 while (bio) { /* submit pending writes */
798 struct bio *next = bio->bi_next;
799 struct md_rdev *rdev = (void *)bio->bi_bdev;
800 bio->bi_next = NULL;
801 bio_set_dev(bio, rdev->bdev);
802 if (test_bit(Faulty, &rdev->flags)) {
803 bio_io_error(bio);
804 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
805 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
806 /* Just ignore it */
807 bio_endio(bio);
808 else
809 submit_bio_noacct(bio);
810 bio = next;
811 cond_resched();
812 }
813}
814
815static void flush_pending_writes(struct r1conf *conf)
816{
817 /* Any writes that have been queued but are awaiting
818 * bitmap updates get flushed here.
819 */
820 spin_lock_irq(&conf->device_lock);
821
822 if (conf->pending_bio_list.head) {
823 struct blk_plug plug;
824 struct bio *bio;
825
826 bio = bio_list_get(&conf->pending_bio_list);
827 conf->pending_count = 0;
828 spin_unlock_irq(&conf->device_lock);
829
830 /*
831 * As this is called in a wait_event() loop (see freeze_array),
832 * current->state might be TASK_UNINTERRUPTIBLE which will
833 * cause a warning when we prepare to wait again. As it is
834 * rare that this path is taken, it is perfectly safe to force
835 * us to go around the wait_event() loop again, so the warning
836 * is a false-positive. Silence the warning by resetting
837 * thread state
838 */
839 __set_current_state(TASK_RUNNING);
840 blk_start_plug(&plug);
841 flush_bio_list(conf, bio);
842 blk_finish_plug(&plug);
843 } else
844 spin_unlock_irq(&conf->device_lock);
845}
846
847/* Barriers....
848 * Sometimes we need to suspend IO while we do something else,
849 * either some resync/recovery, or reconfigure the array.
850 * To do this we raise a 'barrier'.
851 * The 'barrier' is a counter that can be raised multiple times
852 * to count how many activities are happening which preclude
853 * normal IO.
854 * We can only raise the barrier if there is no pending IO.
855 * i.e. if nr_pending == 0.
856 * We choose only to raise the barrier if no-one is waiting for the
857 * barrier to go down. This means that as soon as an IO request
858 * is ready, no other operations which require a barrier will start
859 * until the IO request has had a chance.
860 *
861 * So: regular IO calls 'wait_barrier'. When that returns there
862 * is no backgroup IO happening, It must arrange to call
863 * allow_barrier when it has finished its IO.
864 * backgroup IO calls must call raise_barrier. Once that returns
865 * there is no normal IO happeing. It must arrange to call
866 * lower_barrier when the particular background IO completes.
867 *
868 * If resync/recovery is interrupted, returns -EINTR;
869 * Otherwise, returns 0.
870 */
871static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
872{
873 int idx = sector_to_idx(sector_nr);
874
875 spin_lock_irq(&conf->resync_lock);
876
877 /* Wait until no block IO is waiting */
878 wait_event_lock_irq(conf->wait_barrier,
879 !atomic_read(&conf->nr_waiting[idx]),
880 conf->resync_lock);
881
882 /* block any new IO from starting */
883 atomic_inc(&conf->barrier[idx]);
884 /*
885 * In raise_barrier() we firstly increase conf->barrier[idx] then
886 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
887 * increase conf->nr_pending[idx] then check conf->barrier[idx].
888 * A memory barrier here to make sure conf->nr_pending[idx] won't
889 * be fetched before conf->barrier[idx] is increased. Otherwise
890 * there will be a race between raise_barrier() and _wait_barrier().
891 */
892 smp_mb__after_atomic();
893
894 /* For these conditions we must wait:
895 * A: while the array is in frozen state
896 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
897 * existing in corresponding I/O barrier bucket.
898 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
899 * max resync count which allowed on current I/O barrier bucket.
900 */
901 wait_event_lock_irq(conf->wait_barrier,
902 (!conf->array_frozen &&
903 !atomic_read(&conf->nr_pending[idx]) &&
904 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
905 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
906 conf->resync_lock);
907
908 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
909 atomic_dec(&conf->barrier[idx]);
910 spin_unlock_irq(&conf->resync_lock);
911 wake_up(&conf->wait_barrier);
912 return -EINTR;
913 }
914
915 atomic_inc(&conf->nr_sync_pending);
916 spin_unlock_irq(&conf->resync_lock);
917
918 return 0;
919}
920
921static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
922{
923 int idx = sector_to_idx(sector_nr);
924
925 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
926
927 atomic_dec(&conf->barrier[idx]);
928 atomic_dec(&conf->nr_sync_pending);
929 wake_up(&conf->wait_barrier);
930}
931
932static void _wait_barrier(struct r1conf *conf, int idx)
933{
934 /*
935 * We need to increase conf->nr_pending[idx] very early here,
936 * then raise_barrier() can be blocked when it waits for
937 * conf->nr_pending[idx] to be 0. Then we can avoid holding
938 * conf->resync_lock when there is no barrier raised in same
939 * barrier unit bucket. Also if the array is frozen, I/O
940 * should be blocked until array is unfrozen.
941 */
942 atomic_inc(&conf->nr_pending[idx]);
943 /*
944 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
945 * check conf->barrier[idx]. In raise_barrier() we firstly increase
946 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
947 * barrier is necessary here to make sure conf->barrier[idx] won't be
948 * fetched before conf->nr_pending[idx] is increased. Otherwise there
949 * will be a race between _wait_barrier() and raise_barrier().
950 */
951 smp_mb__after_atomic();
952
953 /*
954 * Don't worry about checking two atomic_t variables at same time
955 * here. If during we check conf->barrier[idx], the array is
956 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
957 * 0, it is safe to return and make the I/O continue. Because the
958 * array is frozen, all I/O returned here will eventually complete
959 * or be queued, no race will happen. See code comment in
960 * frozen_array().
961 */
962 if (!READ_ONCE(conf->array_frozen) &&
963 !atomic_read(&conf->barrier[idx]))
964 return;
965
966 /*
967 * After holding conf->resync_lock, conf->nr_pending[idx]
968 * should be decreased before waiting for barrier to drop.
969 * Otherwise, we may encounter a race condition because
970 * raise_barrer() might be waiting for conf->nr_pending[idx]
971 * to be 0 at same time.
972 */
973 spin_lock_irq(&conf->resync_lock);
974 atomic_inc(&conf->nr_waiting[idx]);
975 atomic_dec(&conf->nr_pending[idx]);
976 /*
977 * In case freeze_array() is waiting for
978 * get_unqueued_pending() == extra
979 */
980 wake_up(&conf->wait_barrier);
981 /* Wait for the barrier in same barrier unit bucket to drop. */
982 wait_event_lock_irq(conf->wait_barrier,
983 !conf->array_frozen &&
984 !atomic_read(&conf->barrier[idx]),
985 conf->resync_lock);
986 atomic_inc(&conf->nr_pending[idx]);
987 atomic_dec(&conf->nr_waiting[idx]);
988 spin_unlock_irq(&conf->resync_lock);
989}
990
991static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
992{
993 int idx = sector_to_idx(sector_nr);
994
995 /*
996 * Very similar to _wait_barrier(). The difference is, for read
997 * I/O we don't need wait for sync I/O, but if the whole array
998 * is frozen, the read I/O still has to wait until the array is
999 * unfrozen. Since there is no ordering requirement with
1000 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1001 */
1002 atomic_inc(&conf->nr_pending[idx]);
1003
1004 if (!READ_ONCE(conf->array_frozen))
1005 return;
1006
1007 spin_lock_irq(&conf->resync_lock);
1008 atomic_inc(&conf->nr_waiting[idx]);
1009 atomic_dec(&conf->nr_pending[idx]);
1010 /*
1011 * In case freeze_array() is waiting for
1012 * get_unqueued_pending() == extra
1013 */
1014 wake_up(&conf->wait_barrier);
1015 /* Wait for array to be unfrozen */
1016 wait_event_lock_irq(conf->wait_barrier,
1017 !conf->array_frozen,
1018 conf->resync_lock);
1019 atomic_inc(&conf->nr_pending[idx]);
1020 atomic_dec(&conf->nr_waiting[idx]);
1021 spin_unlock_irq(&conf->resync_lock);
1022}
1023
1024static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1025{
1026 int idx = sector_to_idx(sector_nr);
1027
1028 _wait_barrier(conf, idx);
1029}
1030
1031static void _allow_barrier(struct r1conf *conf, int idx)
1032{
1033 atomic_dec(&conf->nr_pending[idx]);
1034 wake_up(&conf->wait_barrier);
1035}
1036
1037static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1038{
1039 int idx = sector_to_idx(sector_nr);
1040
1041 _allow_barrier(conf, idx);
1042}
1043
1044/* conf->resync_lock should be held */
1045static int get_unqueued_pending(struct r1conf *conf)
1046{
1047 int idx, ret;
1048
1049 ret = atomic_read(&conf->nr_sync_pending);
1050 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1051 ret += atomic_read(&conf->nr_pending[idx]) -
1052 atomic_read(&conf->nr_queued[idx]);
1053
1054 return ret;
1055}
1056
1057static void freeze_array(struct r1conf *conf, int extra)
1058{
1059 /* Stop sync I/O and normal I/O and wait for everything to
1060 * go quiet.
1061 * This is called in two situations:
1062 * 1) management command handlers (reshape, remove disk, quiesce).
1063 * 2) one normal I/O request failed.
1064
1065 * After array_frozen is set to 1, new sync IO will be blocked at
1066 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1067 * or wait_read_barrier(). The flying I/Os will either complete or be
1068 * queued. When everything goes quite, there are only queued I/Os left.
1069
1070 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1071 * barrier bucket index which this I/O request hits. When all sync and
1072 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1073 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1074 * in handle_read_error(), we may call freeze_array() before trying to
1075 * fix the read error. In this case, the error read I/O is not queued,
1076 * so get_unqueued_pending() == 1.
1077 *
1078 * Therefore before this function returns, we need to wait until
1079 * get_unqueued_pendings(conf) gets equal to extra. For
1080 * normal I/O context, extra is 1, in rested situations extra is 0.
1081 */
1082 spin_lock_irq(&conf->resync_lock);
1083 conf->array_frozen = 1;
1084 raid1_log(conf->mddev, "wait freeze");
1085 wait_event_lock_irq_cmd(
1086 conf->wait_barrier,
1087 get_unqueued_pending(conf) == extra,
1088 conf->resync_lock,
1089 flush_pending_writes(conf));
1090 spin_unlock_irq(&conf->resync_lock);
1091}
1092static void unfreeze_array(struct r1conf *conf)
1093{
1094 /* reverse the effect of the freeze */
1095 spin_lock_irq(&conf->resync_lock);
1096 conf->array_frozen = 0;
1097 spin_unlock_irq(&conf->resync_lock);
1098 wake_up(&conf->wait_barrier);
1099}
1100
1101static void alloc_behind_master_bio(struct r1bio *r1_bio,
1102 struct bio *bio)
1103{
1104 int size = bio->bi_iter.bi_size;
1105 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1106 int i = 0;
1107 struct bio *behind_bio = NULL;
1108
1109 behind_bio = bio_alloc_bioset(GFP_NOIO, vcnt, &r1_bio->mddev->bio_set);
1110 if (!behind_bio)
1111 return;
1112
1113 /* discard op, we don't support writezero/writesame yet */
1114 if (!bio_has_data(bio)) {
1115 behind_bio->bi_iter.bi_size = size;
1116 goto skip_copy;
1117 }
1118
1119 behind_bio->bi_write_hint = bio->bi_write_hint;
1120
1121 while (i < vcnt && size) {
1122 struct page *page;
1123 int len = min_t(int, PAGE_SIZE, size);
1124
1125 page = alloc_page(GFP_NOIO);
1126 if (unlikely(!page))
1127 goto free_pages;
1128
1129 bio_add_page(behind_bio, page, len, 0);
1130
1131 size -= len;
1132 i++;
1133 }
1134
1135 bio_copy_data(behind_bio, bio);
1136skip_copy:
1137 r1_bio->behind_master_bio = behind_bio;
1138 set_bit(R1BIO_BehindIO, &r1_bio->state);
1139
1140 return;
1141
1142free_pages:
1143 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1144 bio->bi_iter.bi_size);
1145 bio_free_pages(behind_bio);
1146 bio_put(behind_bio);
1147}
1148
1149struct raid1_plug_cb {
1150 struct blk_plug_cb cb;
1151 struct bio_list pending;
1152 int pending_cnt;
1153};
1154
1155static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1156{
1157 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1158 cb);
1159 struct mddev *mddev = plug->cb.data;
1160 struct r1conf *conf = mddev->private;
1161 struct bio *bio;
1162
1163 if (from_schedule || current->bio_list) {
1164 spin_lock_irq(&conf->device_lock);
1165 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1166 conf->pending_count += plug->pending_cnt;
1167 spin_unlock_irq(&conf->device_lock);
1168 wake_up(&conf->wait_barrier);
1169 md_wakeup_thread(mddev->thread);
1170 kfree(plug);
1171 return;
1172 }
1173
1174 /* we aren't scheduling, so we can do the write-out directly. */
1175 bio = bio_list_get(&plug->pending);
1176 flush_bio_list(conf, bio);
1177 kfree(plug);
1178}
1179
1180static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1181{
1182 r1_bio->master_bio = bio;
1183 r1_bio->sectors = bio_sectors(bio);
1184 r1_bio->state = 0;
1185 r1_bio->mddev = mddev;
1186 r1_bio->sector = bio->bi_iter.bi_sector;
1187}
1188
1189static inline struct r1bio *
1190alloc_r1bio(struct mddev *mddev, struct bio *bio)
1191{
1192 struct r1conf *conf = mddev->private;
1193 struct r1bio *r1_bio;
1194
1195 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1196 /* Ensure no bio records IO_BLOCKED */
1197 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1198 init_r1bio(r1_bio, mddev, bio);
1199 return r1_bio;
1200}
1201
1202static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1203 int max_read_sectors, struct r1bio *r1_bio)
1204{
1205 struct r1conf *conf = mddev->private;
1206 struct raid1_info *mirror;
1207 struct bio *read_bio;
1208 struct bitmap *bitmap = mddev->bitmap;
1209 const int op = bio_op(bio);
1210 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1211 int max_sectors;
1212 int rdisk;
1213 bool r1bio_existed = !!r1_bio;
1214 char b[BDEVNAME_SIZE];
1215
1216 /*
1217 * If r1_bio is set, we are blocking the raid1d thread
1218 * so there is a tiny risk of deadlock. So ask for
1219 * emergency memory if needed.
1220 */
1221 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1222
1223 if (r1bio_existed) {
1224 /* Need to get the block device name carefully */
1225 struct md_rdev *rdev;
1226 rcu_read_lock();
1227 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1228 if (rdev)
1229 bdevname(rdev->bdev, b);
1230 else
1231 strcpy(b, "???");
1232 rcu_read_unlock();
1233 }
1234
1235 /*
1236 * Still need barrier for READ in case that whole
1237 * array is frozen.
1238 */
1239 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1240
1241 if (!r1_bio)
1242 r1_bio = alloc_r1bio(mddev, bio);
1243 else
1244 init_r1bio(r1_bio, mddev, bio);
1245 r1_bio->sectors = max_read_sectors;
1246
1247 /*
1248 * make_request() can abort the operation when read-ahead is being
1249 * used and no empty request is available.
1250 */
1251 rdisk = read_balance(conf, r1_bio, &max_sectors);
1252
1253 if (rdisk < 0) {
1254 /* couldn't find anywhere to read from */
1255 if (r1bio_existed) {
1256 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1257 mdname(mddev),
1258 b,
1259 (unsigned long long)r1_bio->sector);
1260 }
1261 raid_end_bio_io(r1_bio);
1262 return;
1263 }
1264 mirror = conf->mirrors + rdisk;
1265
1266 if (r1bio_existed)
1267 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1268 mdname(mddev),
1269 (unsigned long long)r1_bio->sector,
1270 bdevname(mirror->rdev->bdev, b));
1271
1272 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1273 bitmap) {
1274 /*
1275 * Reading from a write-mostly device must take care not to
1276 * over-take any writes that are 'behind'
1277 */
1278 raid1_log(mddev, "wait behind writes");
1279 wait_event(bitmap->behind_wait,
1280 atomic_read(&bitmap->behind_writes) == 0);
1281 }
1282
1283 if (max_sectors < bio_sectors(bio)) {
1284 struct bio *split = bio_split(bio, max_sectors,
1285 gfp, &conf->bio_split);
1286 bio_chain(split, bio);
1287 submit_bio_noacct(bio);
1288 bio = split;
1289 r1_bio->master_bio = bio;
1290 r1_bio->sectors = max_sectors;
1291 }
1292
1293 r1_bio->read_disk = rdisk;
1294
1295 if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1296 r1_bio->start_time = bio_start_io_acct(bio);
1297
1298 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1299
1300 r1_bio->bios[rdisk] = read_bio;
1301
1302 read_bio->bi_iter.bi_sector = r1_bio->sector +
1303 mirror->rdev->data_offset;
1304 bio_set_dev(read_bio, mirror->rdev->bdev);
1305 read_bio->bi_end_io = raid1_end_read_request;
1306 bio_set_op_attrs(read_bio, op, do_sync);
1307 if (test_bit(FailFast, &mirror->rdev->flags) &&
1308 test_bit(R1BIO_FailFast, &r1_bio->state))
1309 read_bio->bi_opf |= MD_FAILFAST;
1310 read_bio->bi_private = r1_bio;
1311
1312 if (mddev->gendisk)
1313 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1314 r1_bio->sector);
1315
1316 submit_bio_noacct(read_bio);
1317}
1318
1319static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1320 int max_write_sectors)
1321{
1322 struct r1conf *conf = mddev->private;
1323 struct r1bio *r1_bio;
1324 int i, disks;
1325 struct bitmap *bitmap = mddev->bitmap;
1326 unsigned long flags;
1327 struct md_rdev *blocked_rdev;
1328 struct blk_plug_cb *cb;
1329 struct raid1_plug_cb *plug = NULL;
1330 int first_clone;
1331 int max_sectors;
1332 bool write_behind = false;
1333
1334 if (mddev_is_clustered(mddev) &&
1335 md_cluster_ops->area_resyncing(mddev, WRITE,
1336 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1337
1338 DEFINE_WAIT(w);
1339 for (;;) {
1340 prepare_to_wait(&conf->wait_barrier,
1341 &w, TASK_IDLE);
1342 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1343 bio->bi_iter.bi_sector,
1344 bio_end_sector(bio)))
1345 break;
1346 schedule();
1347 }
1348 finish_wait(&conf->wait_barrier, &w);
1349 }
1350
1351 /*
1352 * Register the new request and wait if the reconstruction
1353 * thread has put up a bar for new requests.
1354 * Continue immediately if no resync is active currently.
1355 */
1356 wait_barrier(conf, bio->bi_iter.bi_sector);
1357
1358 r1_bio = alloc_r1bio(mddev, bio);
1359 r1_bio->sectors = max_write_sectors;
1360
1361 if (conf->pending_count >= max_queued_requests) {
1362 md_wakeup_thread(mddev->thread);
1363 raid1_log(mddev, "wait queued");
1364 wait_event(conf->wait_barrier,
1365 conf->pending_count < max_queued_requests);
1366 }
1367 /* first select target devices under rcu_lock and
1368 * inc refcount on their rdev. Record them by setting
1369 * bios[x] to bio
1370 * If there are known/acknowledged bad blocks on any device on
1371 * which we have seen a write error, we want to avoid writing those
1372 * blocks.
1373 * This potentially requires several writes to write around
1374 * the bad blocks. Each set of writes gets it's own r1bio
1375 * with a set of bios attached.
1376 */
1377
1378 disks = conf->raid_disks * 2;
1379 retry_write:
1380 blocked_rdev = NULL;
1381 rcu_read_lock();
1382 max_sectors = r1_bio->sectors;
1383 for (i = 0; i < disks; i++) {
1384 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1385
1386 /*
1387 * The write-behind io is only attempted on drives marked as
1388 * write-mostly, which means we could allocate write behind
1389 * bio later.
1390 */
1391 if (rdev && test_bit(WriteMostly, &rdev->flags))
1392 write_behind = true;
1393
1394 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1395 atomic_inc(&rdev->nr_pending);
1396 blocked_rdev = rdev;
1397 break;
1398 }
1399 r1_bio->bios[i] = NULL;
1400 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1401 if (i < conf->raid_disks)
1402 set_bit(R1BIO_Degraded, &r1_bio->state);
1403 continue;
1404 }
1405
1406 atomic_inc(&rdev->nr_pending);
1407 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1408 sector_t first_bad;
1409 int bad_sectors;
1410 int is_bad;
1411
1412 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1413 &first_bad, &bad_sectors);
1414 if (is_bad < 0) {
1415 /* mustn't write here until the bad block is
1416 * acknowledged*/
1417 set_bit(BlockedBadBlocks, &rdev->flags);
1418 blocked_rdev = rdev;
1419 break;
1420 }
1421 if (is_bad && first_bad <= r1_bio->sector) {
1422 /* Cannot write here at all */
1423 bad_sectors -= (r1_bio->sector - first_bad);
1424 if (bad_sectors < max_sectors)
1425 /* mustn't write more than bad_sectors
1426 * to other devices yet
1427 */
1428 max_sectors = bad_sectors;
1429 rdev_dec_pending(rdev, mddev);
1430 /* We don't set R1BIO_Degraded as that
1431 * only applies if the disk is
1432 * missing, so it might be re-added,
1433 * and we want to know to recover this
1434 * chunk.
1435 * In this case the device is here,
1436 * and the fact that this chunk is not
1437 * in-sync is recorded in the bad
1438 * block log
1439 */
1440 continue;
1441 }
1442 if (is_bad) {
1443 int good_sectors = first_bad - r1_bio->sector;
1444 if (good_sectors < max_sectors)
1445 max_sectors = good_sectors;
1446 }
1447 }
1448 r1_bio->bios[i] = bio;
1449 }
1450 rcu_read_unlock();
1451
1452 if (unlikely(blocked_rdev)) {
1453 /* Wait for this device to become unblocked */
1454 int j;
1455
1456 for (j = 0; j < i; j++)
1457 if (r1_bio->bios[j])
1458 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1459 r1_bio->state = 0;
1460 allow_barrier(conf, bio->bi_iter.bi_sector);
1461 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1462 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1463 wait_barrier(conf, bio->bi_iter.bi_sector);
1464 goto retry_write;
1465 }
1466
1467 /*
1468 * When using a bitmap, we may call alloc_behind_master_bio below.
1469 * alloc_behind_master_bio allocates a copy of the data payload a page
1470 * at a time and thus needs a new bio that can fit the whole payload
1471 * this bio in page sized chunks.
1472 */
1473 if (write_behind && bitmap)
1474 max_sectors = min_t(int, max_sectors,
1475 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1476 if (max_sectors < bio_sectors(bio)) {
1477 struct bio *split = bio_split(bio, max_sectors,
1478 GFP_NOIO, &conf->bio_split);
1479 bio_chain(split, bio);
1480 submit_bio_noacct(bio);
1481 bio = split;
1482 r1_bio->master_bio = bio;
1483 r1_bio->sectors = max_sectors;
1484 }
1485
1486 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1487 r1_bio->start_time = bio_start_io_acct(bio);
1488 atomic_set(&r1_bio->remaining, 1);
1489 atomic_set(&r1_bio->behind_remaining, 0);
1490
1491 first_clone = 1;
1492
1493 for (i = 0; i < disks; i++) {
1494 struct bio *mbio = NULL;
1495 struct md_rdev *rdev = conf->mirrors[i].rdev;
1496 if (!r1_bio->bios[i])
1497 continue;
1498
1499 if (first_clone) {
1500 /* do behind I/O ?
1501 * Not if there are too many, or cannot
1502 * allocate memory, or a reader on WriteMostly
1503 * is waiting for behind writes to flush */
1504 if (bitmap &&
1505 (atomic_read(&bitmap->behind_writes)
1506 < mddev->bitmap_info.max_write_behind) &&
1507 !waitqueue_active(&bitmap->behind_wait)) {
1508 alloc_behind_master_bio(r1_bio, bio);
1509 }
1510
1511 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1512 test_bit(R1BIO_BehindIO, &r1_bio->state));
1513 first_clone = 0;
1514 }
1515
1516 if (r1_bio->behind_master_bio)
1517 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1518 GFP_NOIO, &mddev->bio_set);
1519 else
1520 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1521
1522 if (r1_bio->behind_master_bio) {
1523 if (test_bit(CollisionCheck, &rdev->flags))
1524 wait_for_serialization(rdev, r1_bio);
1525 if (test_bit(WriteMostly, &rdev->flags))
1526 atomic_inc(&r1_bio->behind_remaining);
1527 } else if (mddev->serialize_policy)
1528 wait_for_serialization(rdev, r1_bio);
1529
1530 r1_bio->bios[i] = mbio;
1531
1532 mbio->bi_iter.bi_sector = (r1_bio->sector +
1533 conf->mirrors[i].rdev->data_offset);
1534 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1535 mbio->bi_end_io = raid1_end_write_request;
1536 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1537 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1538 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1539 conf->raid_disks - mddev->degraded > 1)
1540 mbio->bi_opf |= MD_FAILFAST;
1541 mbio->bi_private = r1_bio;
1542
1543 atomic_inc(&r1_bio->remaining);
1544
1545 if (mddev->gendisk)
1546 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1547 r1_bio->sector);
1548 /* flush_pending_writes() needs access to the rdev so...*/
1549 mbio->bi_bdev = (void *)conf->mirrors[i].rdev;
1550
1551 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1552 if (cb)
1553 plug = container_of(cb, struct raid1_plug_cb, cb);
1554 else
1555 plug = NULL;
1556 if (plug) {
1557 bio_list_add(&plug->pending, mbio);
1558 plug->pending_cnt++;
1559 } else {
1560 spin_lock_irqsave(&conf->device_lock, flags);
1561 bio_list_add(&conf->pending_bio_list, mbio);
1562 conf->pending_count++;
1563 spin_unlock_irqrestore(&conf->device_lock, flags);
1564 md_wakeup_thread(mddev->thread);
1565 }
1566 }
1567
1568 r1_bio_write_done(r1_bio);
1569
1570 /* In case raid1d snuck in to freeze_array */
1571 wake_up(&conf->wait_barrier);
1572}
1573
1574static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1575{
1576 sector_t sectors;
1577
1578 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1579 && md_flush_request(mddev, bio))
1580 return true;
1581
1582 /*
1583 * There is a limit to the maximum size, but
1584 * the read/write handler might find a lower limit
1585 * due to bad blocks. To avoid multiple splits,
1586 * we pass the maximum number of sectors down
1587 * and let the lower level perform the split.
1588 */
1589 sectors = align_to_barrier_unit_end(
1590 bio->bi_iter.bi_sector, bio_sectors(bio));
1591
1592 if (bio_data_dir(bio) == READ)
1593 raid1_read_request(mddev, bio, sectors, NULL);
1594 else {
1595 if (!md_write_start(mddev,bio))
1596 return false;
1597 raid1_write_request(mddev, bio, sectors);
1598 }
1599 return true;
1600}
1601
1602static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1603{
1604 struct r1conf *conf = mddev->private;
1605 int i;
1606
1607 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1608 conf->raid_disks - mddev->degraded);
1609 rcu_read_lock();
1610 for (i = 0; i < conf->raid_disks; i++) {
1611 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1612 seq_printf(seq, "%s",
1613 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1614 }
1615 rcu_read_unlock();
1616 seq_printf(seq, "]");
1617}
1618
1619static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1620{
1621 char b[BDEVNAME_SIZE];
1622 struct r1conf *conf = mddev->private;
1623 unsigned long flags;
1624
1625 /*
1626 * If it is not operational, then we have already marked it as dead
1627 * else if it is the last working disks with "fail_last_dev == false",
1628 * ignore the error, let the next level up know.
1629 * else mark the drive as failed
1630 */
1631 spin_lock_irqsave(&conf->device_lock, flags);
1632 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1633 && (conf->raid_disks - mddev->degraded) == 1) {
1634 /*
1635 * Don't fail the drive, act as though we were just a
1636 * normal single drive.
1637 * However don't try a recovery from this drive as
1638 * it is very likely to fail.
1639 */
1640 conf->recovery_disabled = mddev->recovery_disabled;
1641 spin_unlock_irqrestore(&conf->device_lock, flags);
1642 return;
1643 }
1644 set_bit(Blocked, &rdev->flags);
1645 if (test_and_clear_bit(In_sync, &rdev->flags))
1646 mddev->degraded++;
1647 set_bit(Faulty, &rdev->flags);
1648 spin_unlock_irqrestore(&conf->device_lock, flags);
1649 /*
1650 * if recovery is running, make sure it aborts.
1651 */
1652 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1653 set_mask_bits(&mddev->sb_flags, 0,
1654 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1655 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1656 "md/raid1:%s: Operation continuing on %d devices.\n",
1657 mdname(mddev), bdevname(rdev->bdev, b),
1658 mdname(mddev), conf->raid_disks - mddev->degraded);
1659}
1660
1661static void print_conf(struct r1conf *conf)
1662{
1663 int i;
1664
1665 pr_debug("RAID1 conf printout:\n");
1666 if (!conf) {
1667 pr_debug("(!conf)\n");
1668 return;
1669 }
1670 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1671 conf->raid_disks);
1672
1673 rcu_read_lock();
1674 for (i = 0; i < conf->raid_disks; i++) {
1675 char b[BDEVNAME_SIZE];
1676 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1677 if (rdev)
1678 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1679 i, !test_bit(In_sync, &rdev->flags),
1680 !test_bit(Faulty, &rdev->flags),
1681 bdevname(rdev->bdev,b));
1682 }
1683 rcu_read_unlock();
1684}
1685
1686static void close_sync(struct r1conf *conf)
1687{
1688 int idx;
1689
1690 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1691 _wait_barrier(conf, idx);
1692 _allow_barrier(conf, idx);
1693 }
1694
1695 mempool_exit(&conf->r1buf_pool);
1696}
1697
1698static int raid1_spare_active(struct mddev *mddev)
1699{
1700 int i;
1701 struct r1conf *conf = mddev->private;
1702 int count = 0;
1703 unsigned long flags;
1704
1705 /*
1706 * Find all failed disks within the RAID1 configuration
1707 * and mark them readable.
1708 * Called under mddev lock, so rcu protection not needed.
1709 * device_lock used to avoid races with raid1_end_read_request
1710 * which expects 'In_sync' flags and ->degraded to be consistent.
1711 */
1712 spin_lock_irqsave(&conf->device_lock, flags);
1713 for (i = 0; i < conf->raid_disks; i++) {
1714 struct md_rdev *rdev = conf->mirrors[i].rdev;
1715 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1716 if (repl
1717 && !test_bit(Candidate, &repl->flags)
1718 && repl->recovery_offset == MaxSector
1719 && !test_bit(Faulty, &repl->flags)
1720 && !test_and_set_bit(In_sync, &repl->flags)) {
1721 /* replacement has just become active */
1722 if (!rdev ||
1723 !test_and_clear_bit(In_sync, &rdev->flags))
1724 count++;
1725 if (rdev) {
1726 /* Replaced device not technically
1727 * faulty, but we need to be sure
1728 * it gets removed and never re-added
1729 */
1730 set_bit(Faulty, &rdev->flags);
1731 sysfs_notify_dirent_safe(
1732 rdev->sysfs_state);
1733 }
1734 }
1735 if (rdev
1736 && rdev->recovery_offset == MaxSector
1737 && !test_bit(Faulty, &rdev->flags)
1738 && !test_and_set_bit(In_sync, &rdev->flags)) {
1739 count++;
1740 sysfs_notify_dirent_safe(rdev->sysfs_state);
1741 }
1742 }
1743 mddev->degraded -= count;
1744 spin_unlock_irqrestore(&conf->device_lock, flags);
1745
1746 print_conf(conf);
1747 return count;
1748}
1749
1750static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1751{
1752 struct r1conf *conf = mddev->private;
1753 int err = -EEXIST;
1754 int mirror = 0;
1755 struct raid1_info *p;
1756 int first = 0;
1757 int last = conf->raid_disks - 1;
1758
1759 if (mddev->recovery_disabled == conf->recovery_disabled)
1760 return -EBUSY;
1761
1762 if (md_integrity_add_rdev(rdev, mddev))
1763 return -ENXIO;
1764
1765 if (rdev->raid_disk >= 0)
1766 first = last = rdev->raid_disk;
1767
1768 /*
1769 * find the disk ... but prefer rdev->saved_raid_disk
1770 * if possible.
1771 */
1772 if (rdev->saved_raid_disk >= 0 &&
1773 rdev->saved_raid_disk >= first &&
1774 rdev->saved_raid_disk < conf->raid_disks &&
1775 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1776 first = last = rdev->saved_raid_disk;
1777
1778 for (mirror = first; mirror <= last; mirror++) {
1779 p = conf->mirrors + mirror;
1780 if (!p->rdev) {
1781 if (mddev->gendisk)
1782 disk_stack_limits(mddev->gendisk, rdev->bdev,
1783 rdev->data_offset << 9);
1784
1785 p->head_position = 0;
1786 rdev->raid_disk = mirror;
1787 err = 0;
1788 /* As all devices are equivalent, we don't need a full recovery
1789 * if this was recently any drive of the array
1790 */
1791 if (rdev->saved_raid_disk < 0)
1792 conf->fullsync = 1;
1793 rcu_assign_pointer(p->rdev, rdev);
1794 break;
1795 }
1796 if (test_bit(WantReplacement, &p->rdev->flags) &&
1797 p[conf->raid_disks].rdev == NULL) {
1798 /* Add this device as a replacement */
1799 clear_bit(In_sync, &rdev->flags);
1800 set_bit(Replacement, &rdev->flags);
1801 rdev->raid_disk = mirror;
1802 err = 0;
1803 conf->fullsync = 1;
1804 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1805 break;
1806 }
1807 }
1808 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1809 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1810 print_conf(conf);
1811 return err;
1812}
1813
1814static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1815{
1816 struct r1conf *conf = mddev->private;
1817 int err = 0;
1818 int number = rdev->raid_disk;
1819 struct raid1_info *p = conf->mirrors + number;
1820
1821 if (rdev != p->rdev)
1822 p = conf->mirrors + conf->raid_disks + number;
1823
1824 print_conf(conf);
1825 if (rdev == p->rdev) {
1826 if (test_bit(In_sync, &rdev->flags) ||
1827 atomic_read(&rdev->nr_pending)) {
1828 err = -EBUSY;
1829 goto abort;
1830 }
1831 /* Only remove non-faulty devices if recovery
1832 * is not possible.
1833 */
1834 if (!test_bit(Faulty, &rdev->flags) &&
1835 mddev->recovery_disabled != conf->recovery_disabled &&
1836 mddev->degraded < conf->raid_disks) {
1837 err = -EBUSY;
1838 goto abort;
1839 }
1840 p->rdev = NULL;
1841 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1842 synchronize_rcu();
1843 if (atomic_read(&rdev->nr_pending)) {
1844 /* lost the race, try later */
1845 err = -EBUSY;
1846 p->rdev = rdev;
1847 goto abort;
1848 }
1849 }
1850 if (conf->mirrors[conf->raid_disks + number].rdev) {
1851 /* We just removed a device that is being replaced.
1852 * Move down the replacement. We drain all IO before
1853 * doing this to avoid confusion.
1854 */
1855 struct md_rdev *repl =
1856 conf->mirrors[conf->raid_disks + number].rdev;
1857 freeze_array(conf, 0);
1858 if (atomic_read(&repl->nr_pending)) {
1859 /* It means that some queued IO of retry_list
1860 * hold repl. Thus, we cannot set replacement
1861 * as NULL, avoiding rdev NULL pointer
1862 * dereference in sync_request_write and
1863 * handle_write_finished.
1864 */
1865 err = -EBUSY;
1866 unfreeze_array(conf);
1867 goto abort;
1868 }
1869 clear_bit(Replacement, &repl->flags);
1870 p->rdev = repl;
1871 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1872 unfreeze_array(conf);
1873 }
1874
1875 clear_bit(WantReplacement, &rdev->flags);
1876 err = md_integrity_register(mddev);
1877 }
1878abort:
1879
1880 print_conf(conf);
1881 return err;
1882}
1883
1884static void end_sync_read(struct bio *bio)
1885{
1886 struct r1bio *r1_bio = get_resync_r1bio(bio);
1887
1888 update_head_pos(r1_bio->read_disk, r1_bio);
1889
1890 /*
1891 * we have read a block, now it needs to be re-written,
1892 * or re-read if the read failed.
1893 * We don't do much here, just schedule handling by raid1d
1894 */
1895 if (!bio->bi_status)
1896 set_bit(R1BIO_Uptodate, &r1_bio->state);
1897
1898 if (atomic_dec_and_test(&r1_bio->remaining))
1899 reschedule_retry(r1_bio);
1900}
1901
1902static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1903{
1904 sector_t sync_blocks = 0;
1905 sector_t s = r1_bio->sector;
1906 long sectors_to_go = r1_bio->sectors;
1907
1908 /* make sure these bits don't get cleared. */
1909 do {
1910 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1911 s += sync_blocks;
1912 sectors_to_go -= sync_blocks;
1913 } while (sectors_to_go > 0);
1914}
1915
1916static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1917{
1918 if (atomic_dec_and_test(&r1_bio->remaining)) {
1919 struct mddev *mddev = r1_bio->mddev;
1920 int s = r1_bio->sectors;
1921
1922 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1923 test_bit(R1BIO_WriteError, &r1_bio->state))
1924 reschedule_retry(r1_bio);
1925 else {
1926 put_buf(r1_bio);
1927 md_done_sync(mddev, s, uptodate);
1928 }
1929 }
1930}
1931
1932static void end_sync_write(struct bio *bio)
1933{
1934 int uptodate = !bio->bi_status;
1935 struct r1bio *r1_bio = get_resync_r1bio(bio);
1936 struct mddev *mddev = r1_bio->mddev;
1937 struct r1conf *conf = mddev->private;
1938 sector_t first_bad;
1939 int bad_sectors;
1940 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1941
1942 if (!uptodate) {
1943 abort_sync_write(mddev, r1_bio);
1944 set_bit(WriteErrorSeen, &rdev->flags);
1945 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1946 set_bit(MD_RECOVERY_NEEDED, &
1947 mddev->recovery);
1948 set_bit(R1BIO_WriteError, &r1_bio->state);
1949 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1950 &first_bad, &bad_sectors) &&
1951 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1952 r1_bio->sector,
1953 r1_bio->sectors,
1954 &first_bad, &bad_sectors)
1955 )
1956 set_bit(R1BIO_MadeGood, &r1_bio->state);
1957
1958 put_sync_write_buf(r1_bio, uptodate);
1959}
1960
1961static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1962 int sectors, struct page *page, int rw)
1963{
1964 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1965 /* success */
1966 return 1;
1967 if (rw == WRITE) {
1968 set_bit(WriteErrorSeen, &rdev->flags);
1969 if (!test_and_set_bit(WantReplacement,
1970 &rdev->flags))
1971 set_bit(MD_RECOVERY_NEEDED, &
1972 rdev->mddev->recovery);
1973 }
1974 /* need to record an error - either for the block or the device */
1975 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1976 md_error(rdev->mddev, rdev);
1977 return 0;
1978}
1979
1980static int fix_sync_read_error(struct r1bio *r1_bio)
1981{
1982 /* Try some synchronous reads of other devices to get
1983 * good data, much like with normal read errors. Only
1984 * read into the pages we already have so we don't
1985 * need to re-issue the read request.
1986 * We don't need to freeze the array, because being in an
1987 * active sync request, there is no normal IO, and
1988 * no overlapping syncs.
1989 * We don't need to check is_badblock() again as we
1990 * made sure that anything with a bad block in range
1991 * will have bi_end_io clear.
1992 */
1993 struct mddev *mddev = r1_bio->mddev;
1994 struct r1conf *conf = mddev->private;
1995 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1996 struct page **pages = get_resync_pages(bio)->pages;
1997 sector_t sect = r1_bio->sector;
1998 int sectors = r1_bio->sectors;
1999 int idx = 0;
2000 struct md_rdev *rdev;
2001
2002 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2003 if (test_bit(FailFast, &rdev->flags)) {
2004 /* Don't try recovering from here - just fail it
2005 * ... unless it is the last working device of course */
2006 md_error(mddev, rdev);
2007 if (test_bit(Faulty, &rdev->flags))
2008 /* Don't try to read from here, but make sure
2009 * put_buf does it's thing
2010 */
2011 bio->bi_end_io = end_sync_write;
2012 }
2013
2014 while(sectors) {
2015 int s = sectors;
2016 int d = r1_bio->read_disk;
2017 int success = 0;
2018 int start;
2019
2020 if (s > (PAGE_SIZE>>9))
2021 s = PAGE_SIZE >> 9;
2022 do {
2023 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2024 /* No rcu protection needed here devices
2025 * can only be removed when no resync is
2026 * active, and resync is currently active
2027 */
2028 rdev = conf->mirrors[d].rdev;
2029 if (sync_page_io(rdev, sect, s<<9,
2030 pages[idx],
2031 REQ_OP_READ, 0, false)) {
2032 success = 1;
2033 break;
2034 }
2035 }
2036 d++;
2037 if (d == conf->raid_disks * 2)
2038 d = 0;
2039 } while (!success && d != r1_bio->read_disk);
2040
2041 if (!success) {
2042 char b[BDEVNAME_SIZE];
2043 int abort = 0;
2044 /* Cannot read from anywhere, this block is lost.
2045 * Record a bad block on each device. If that doesn't
2046 * work just disable and interrupt the recovery.
2047 * Don't fail devices as that won't really help.
2048 */
2049 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2050 mdname(mddev), bio_devname(bio, b),
2051 (unsigned long long)r1_bio->sector);
2052 for (d = 0; d < conf->raid_disks * 2; d++) {
2053 rdev = conf->mirrors[d].rdev;
2054 if (!rdev || test_bit(Faulty, &rdev->flags))
2055 continue;
2056 if (!rdev_set_badblocks(rdev, sect, s, 0))
2057 abort = 1;
2058 }
2059 if (abort) {
2060 conf->recovery_disabled =
2061 mddev->recovery_disabled;
2062 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2063 md_done_sync(mddev, r1_bio->sectors, 0);
2064 put_buf(r1_bio);
2065 return 0;
2066 }
2067 /* Try next page */
2068 sectors -= s;
2069 sect += s;
2070 idx++;
2071 continue;
2072 }
2073
2074 start = d;
2075 /* write it back and re-read */
2076 while (d != r1_bio->read_disk) {
2077 if (d == 0)
2078 d = conf->raid_disks * 2;
2079 d--;
2080 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2081 continue;
2082 rdev = conf->mirrors[d].rdev;
2083 if (r1_sync_page_io(rdev, sect, s,
2084 pages[idx],
2085 WRITE) == 0) {
2086 r1_bio->bios[d]->bi_end_io = NULL;
2087 rdev_dec_pending(rdev, mddev);
2088 }
2089 }
2090 d = start;
2091 while (d != r1_bio->read_disk) {
2092 if (d == 0)
2093 d = conf->raid_disks * 2;
2094 d--;
2095 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2096 continue;
2097 rdev = conf->mirrors[d].rdev;
2098 if (r1_sync_page_io(rdev, sect, s,
2099 pages[idx],
2100 READ) != 0)
2101 atomic_add(s, &rdev->corrected_errors);
2102 }
2103 sectors -= s;
2104 sect += s;
2105 idx ++;
2106 }
2107 set_bit(R1BIO_Uptodate, &r1_bio->state);
2108 bio->bi_status = 0;
2109 return 1;
2110}
2111
2112static void process_checks(struct r1bio *r1_bio)
2113{
2114 /* We have read all readable devices. If we haven't
2115 * got the block, then there is no hope left.
2116 * If we have, then we want to do a comparison
2117 * and skip the write if everything is the same.
2118 * If any blocks failed to read, then we need to
2119 * attempt an over-write
2120 */
2121 struct mddev *mddev = r1_bio->mddev;
2122 struct r1conf *conf = mddev->private;
2123 int primary;
2124 int i;
2125 int vcnt;
2126
2127 /* Fix variable parts of all bios */
2128 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2129 for (i = 0; i < conf->raid_disks * 2; i++) {
2130 blk_status_t status;
2131 struct bio *b = r1_bio->bios[i];
2132 struct resync_pages *rp = get_resync_pages(b);
2133 if (b->bi_end_io != end_sync_read)
2134 continue;
2135 /* fixup the bio for reuse, but preserve errno */
2136 status = b->bi_status;
2137 bio_reset(b);
2138 b->bi_status = status;
2139 b->bi_iter.bi_sector = r1_bio->sector +
2140 conf->mirrors[i].rdev->data_offset;
2141 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2142 b->bi_end_io = end_sync_read;
2143 rp->raid_bio = r1_bio;
2144 b->bi_private = rp;
2145
2146 /* initialize bvec table again */
2147 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2148 }
2149 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2150 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2151 !r1_bio->bios[primary]->bi_status) {
2152 r1_bio->bios[primary]->bi_end_io = NULL;
2153 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2154 break;
2155 }
2156 r1_bio->read_disk = primary;
2157 for (i = 0; i < conf->raid_disks * 2; i++) {
2158 int j = 0;
2159 struct bio *pbio = r1_bio->bios[primary];
2160 struct bio *sbio = r1_bio->bios[i];
2161 blk_status_t status = sbio->bi_status;
2162 struct page **ppages = get_resync_pages(pbio)->pages;
2163 struct page **spages = get_resync_pages(sbio)->pages;
2164 struct bio_vec *bi;
2165 int page_len[RESYNC_PAGES] = { 0 };
2166 struct bvec_iter_all iter_all;
2167
2168 if (sbio->bi_end_io != end_sync_read)
2169 continue;
2170 /* Now we can 'fixup' the error value */
2171 sbio->bi_status = 0;
2172
2173 bio_for_each_segment_all(bi, sbio, iter_all)
2174 page_len[j++] = bi->bv_len;
2175
2176 if (!status) {
2177 for (j = vcnt; j-- ; ) {
2178 if (memcmp(page_address(ppages[j]),
2179 page_address(spages[j]),
2180 page_len[j]))
2181 break;
2182 }
2183 } else
2184 j = 0;
2185 if (j >= 0)
2186 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2187 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2188 && !status)) {
2189 /* No need to write to this device. */
2190 sbio->bi_end_io = NULL;
2191 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2192 continue;
2193 }
2194
2195 bio_copy_data(sbio, pbio);
2196 }
2197}
2198
2199static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2200{
2201 struct r1conf *conf = mddev->private;
2202 int i;
2203 int disks = conf->raid_disks * 2;
2204 struct bio *wbio;
2205
2206 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2207 /* ouch - failed to read all of that. */
2208 if (!fix_sync_read_error(r1_bio))
2209 return;
2210
2211 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2212 process_checks(r1_bio);
2213
2214 /*
2215 * schedule writes
2216 */
2217 atomic_set(&r1_bio->remaining, 1);
2218 for (i = 0; i < disks ; i++) {
2219 wbio = r1_bio->bios[i];
2220 if (wbio->bi_end_io == NULL ||
2221 (wbio->bi_end_io == end_sync_read &&
2222 (i == r1_bio->read_disk ||
2223 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2224 continue;
2225 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2226 abort_sync_write(mddev, r1_bio);
2227 continue;
2228 }
2229
2230 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2231 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2232 wbio->bi_opf |= MD_FAILFAST;
2233
2234 wbio->bi_end_io = end_sync_write;
2235 atomic_inc(&r1_bio->remaining);
2236 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2237
2238 submit_bio_noacct(wbio);
2239 }
2240
2241 put_sync_write_buf(r1_bio, 1);
2242}
2243
2244/*
2245 * This is a kernel thread which:
2246 *
2247 * 1. Retries failed read operations on working mirrors.
2248 * 2. Updates the raid superblock when problems encounter.
2249 * 3. Performs writes following reads for array synchronising.
2250 */
2251
2252static void fix_read_error(struct r1conf *conf, int read_disk,
2253 sector_t sect, int sectors)
2254{
2255 struct mddev *mddev = conf->mddev;
2256 while(sectors) {
2257 int s = sectors;
2258 int d = read_disk;
2259 int success = 0;
2260 int start;
2261 struct md_rdev *rdev;
2262
2263 if (s > (PAGE_SIZE>>9))
2264 s = PAGE_SIZE >> 9;
2265
2266 do {
2267 sector_t first_bad;
2268 int bad_sectors;
2269
2270 rcu_read_lock();
2271 rdev = rcu_dereference(conf->mirrors[d].rdev);
2272 if (rdev &&
2273 (test_bit(In_sync, &rdev->flags) ||
2274 (!test_bit(Faulty, &rdev->flags) &&
2275 rdev->recovery_offset >= sect + s)) &&
2276 is_badblock(rdev, sect, s,
2277 &first_bad, &bad_sectors) == 0) {
2278 atomic_inc(&rdev->nr_pending);
2279 rcu_read_unlock();
2280 if (sync_page_io(rdev, sect, s<<9,
2281 conf->tmppage, REQ_OP_READ, 0, false))
2282 success = 1;
2283 rdev_dec_pending(rdev, mddev);
2284 if (success)
2285 break;
2286 } else
2287 rcu_read_unlock();
2288 d++;
2289 if (d == conf->raid_disks * 2)
2290 d = 0;
2291 } while (!success && d != read_disk);
2292
2293 if (!success) {
2294 /* Cannot read from anywhere - mark it bad */
2295 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2296 if (!rdev_set_badblocks(rdev, sect, s, 0))
2297 md_error(mddev, rdev);
2298 break;
2299 }
2300 /* write it back and re-read */
2301 start = d;
2302 while (d != read_disk) {
2303 if (d==0)
2304 d = conf->raid_disks * 2;
2305 d--;
2306 rcu_read_lock();
2307 rdev = rcu_dereference(conf->mirrors[d].rdev);
2308 if (rdev &&
2309 !test_bit(Faulty, &rdev->flags)) {
2310 atomic_inc(&rdev->nr_pending);
2311 rcu_read_unlock();
2312 r1_sync_page_io(rdev, sect, s,
2313 conf->tmppage, WRITE);
2314 rdev_dec_pending(rdev, mddev);
2315 } else
2316 rcu_read_unlock();
2317 }
2318 d = start;
2319 while (d != read_disk) {
2320 char b[BDEVNAME_SIZE];
2321 if (d==0)
2322 d = conf->raid_disks * 2;
2323 d--;
2324 rcu_read_lock();
2325 rdev = rcu_dereference(conf->mirrors[d].rdev);
2326 if (rdev &&
2327 !test_bit(Faulty, &rdev->flags)) {
2328 atomic_inc(&rdev->nr_pending);
2329 rcu_read_unlock();
2330 if (r1_sync_page_io(rdev, sect, s,
2331 conf->tmppage, READ)) {
2332 atomic_add(s, &rdev->corrected_errors);
2333 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2334 mdname(mddev), s,
2335 (unsigned long long)(sect +
2336 rdev->data_offset),
2337 bdevname(rdev->bdev, b));
2338 }
2339 rdev_dec_pending(rdev, mddev);
2340 } else
2341 rcu_read_unlock();
2342 }
2343 sectors -= s;
2344 sect += s;
2345 }
2346}
2347
2348static int narrow_write_error(struct r1bio *r1_bio, int i)
2349{
2350 struct mddev *mddev = r1_bio->mddev;
2351 struct r1conf *conf = mddev->private;
2352 struct md_rdev *rdev = conf->mirrors[i].rdev;
2353
2354 /* bio has the data to be written to device 'i' where
2355 * we just recently had a write error.
2356 * We repeatedly clone the bio and trim down to one block,
2357 * then try the write. Where the write fails we record
2358 * a bad block.
2359 * It is conceivable that the bio doesn't exactly align with
2360 * blocks. We must handle this somehow.
2361 *
2362 * We currently own a reference on the rdev.
2363 */
2364
2365 int block_sectors;
2366 sector_t sector;
2367 int sectors;
2368 int sect_to_write = r1_bio->sectors;
2369 int ok = 1;
2370
2371 if (rdev->badblocks.shift < 0)
2372 return 0;
2373
2374 block_sectors = roundup(1 << rdev->badblocks.shift,
2375 bdev_logical_block_size(rdev->bdev) >> 9);
2376 sector = r1_bio->sector;
2377 sectors = ((sector + block_sectors)
2378 & ~(sector_t)(block_sectors - 1))
2379 - sector;
2380
2381 while (sect_to_write) {
2382 struct bio *wbio;
2383 if (sectors > sect_to_write)
2384 sectors = sect_to_write;
2385 /* Write at 'sector' for 'sectors'*/
2386
2387 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2388 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2389 GFP_NOIO,
2390 &mddev->bio_set);
2391 } else {
2392 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2393 &mddev->bio_set);
2394 }
2395
2396 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2397 wbio->bi_iter.bi_sector = r1_bio->sector;
2398 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2399
2400 bio_trim(wbio, sector - r1_bio->sector, sectors);
2401 wbio->bi_iter.bi_sector += rdev->data_offset;
2402 bio_set_dev(wbio, rdev->bdev);
2403
2404 if (submit_bio_wait(wbio) < 0)
2405 /* failure! */
2406 ok = rdev_set_badblocks(rdev, sector,
2407 sectors, 0)
2408 && ok;
2409
2410 bio_put(wbio);
2411 sect_to_write -= sectors;
2412 sector += sectors;
2413 sectors = block_sectors;
2414 }
2415 return ok;
2416}
2417
2418static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2419{
2420 int m;
2421 int s = r1_bio->sectors;
2422 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2423 struct md_rdev *rdev = conf->mirrors[m].rdev;
2424 struct bio *bio = r1_bio->bios[m];
2425 if (bio->bi_end_io == NULL)
2426 continue;
2427 if (!bio->bi_status &&
2428 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2429 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2430 }
2431 if (bio->bi_status &&
2432 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2433 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2434 md_error(conf->mddev, rdev);
2435 }
2436 }
2437 put_buf(r1_bio);
2438 md_done_sync(conf->mddev, s, 1);
2439}
2440
2441static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2442{
2443 int m, idx;
2444 bool fail = false;
2445
2446 for (m = 0; m < conf->raid_disks * 2 ; m++)
2447 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2448 struct md_rdev *rdev = conf->mirrors[m].rdev;
2449 rdev_clear_badblocks(rdev,
2450 r1_bio->sector,
2451 r1_bio->sectors, 0);
2452 rdev_dec_pending(rdev, conf->mddev);
2453 } else if (r1_bio->bios[m] != NULL) {
2454 /* This drive got a write error. We need to
2455 * narrow down and record precise write
2456 * errors.
2457 */
2458 fail = true;
2459 if (!narrow_write_error(r1_bio, m)) {
2460 md_error(conf->mddev,
2461 conf->mirrors[m].rdev);
2462 /* an I/O failed, we can't clear the bitmap */
2463 set_bit(R1BIO_Degraded, &r1_bio->state);
2464 }
2465 rdev_dec_pending(conf->mirrors[m].rdev,
2466 conf->mddev);
2467 }
2468 if (fail) {
2469 spin_lock_irq(&conf->device_lock);
2470 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2471 idx = sector_to_idx(r1_bio->sector);
2472 atomic_inc(&conf->nr_queued[idx]);
2473 spin_unlock_irq(&conf->device_lock);
2474 /*
2475 * In case freeze_array() is waiting for condition
2476 * get_unqueued_pending() == extra to be true.
2477 */
2478 wake_up(&conf->wait_barrier);
2479 md_wakeup_thread(conf->mddev->thread);
2480 } else {
2481 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2482 close_write(r1_bio);
2483 raid_end_bio_io(r1_bio);
2484 }
2485}
2486
2487static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2488{
2489 struct mddev *mddev = conf->mddev;
2490 struct bio *bio;
2491 struct md_rdev *rdev;
2492
2493 clear_bit(R1BIO_ReadError, &r1_bio->state);
2494 /* we got a read error. Maybe the drive is bad. Maybe just
2495 * the block and we can fix it.
2496 * We freeze all other IO, and try reading the block from
2497 * other devices. When we find one, we re-write
2498 * and check it that fixes the read error.
2499 * This is all done synchronously while the array is
2500 * frozen
2501 */
2502
2503 bio = r1_bio->bios[r1_bio->read_disk];
2504 bio_put(bio);
2505 r1_bio->bios[r1_bio->read_disk] = NULL;
2506
2507 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2508 if (mddev->ro == 0
2509 && !test_bit(FailFast, &rdev->flags)) {
2510 freeze_array(conf, 1);
2511 fix_read_error(conf, r1_bio->read_disk,
2512 r1_bio->sector, r1_bio->sectors);
2513 unfreeze_array(conf);
2514 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2515 md_error(mddev, rdev);
2516 } else {
2517 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2518 }
2519
2520 rdev_dec_pending(rdev, conf->mddev);
2521 allow_barrier(conf, r1_bio->sector);
2522 bio = r1_bio->master_bio;
2523
2524 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2525 r1_bio->state = 0;
2526 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2527}
2528
2529static void raid1d(struct md_thread *thread)
2530{
2531 struct mddev *mddev = thread->mddev;
2532 struct r1bio *r1_bio;
2533 unsigned long flags;
2534 struct r1conf *conf = mddev->private;
2535 struct list_head *head = &conf->retry_list;
2536 struct blk_plug plug;
2537 int idx;
2538
2539 md_check_recovery(mddev);
2540
2541 if (!list_empty_careful(&conf->bio_end_io_list) &&
2542 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2543 LIST_HEAD(tmp);
2544 spin_lock_irqsave(&conf->device_lock, flags);
2545 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2546 list_splice_init(&conf->bio_end_io_list, &tmp);
2547 spin_unlock_irqrestore(&conf->device_lock, flags);
2548 while (!list_empty(&tmp)) {
2549 r1_bio = list_first_entry(&tmp, struct r1bio,
2550 retry_list);
2551 list_del(&r1_bio->retry_list);
2552 idx = sector_to_idx(r1_bio->sector);
2553 atomic_dec(&conf->nr_queued[idx]);
2554 if (mddev->degraded)
2555 set_bit(R1BIO_Degraded, &r1_bio->state);
2556 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2557 close_write(r1_bio);
2558 raid_end_bio_io(r1_bio);
2559 }
2560 }
2561
2562 blk_start_plug(&plug);
2563 for (;;) {
2564
2565 flush_pending_writes(conf);
2566
2567 spin_lock_irqsave(&conf->device_lock, flags);
2568 if (list_empty(head)) {
2569 spin_unlock_irqrestore(&conf->device_lock, flags);
2570 break;
2571 }
2572 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2573 list_del(head->prev);
2574 idx = sector_to_idx(r1_bio->sector);
2575 atomic_dec(&conf->nr_queued[idx]);
2576 spin_unlock_irqrestore(&conf->device_lock, flags);
2577
2578 mddev = r1_bio->mddev;
2579 conf = mddev->private;
2580 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2581 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2582 test_bit(R1BIO_WriteError, &r1_bio->state))
2583 handle_sync_write_finished(conf, r1_bio);
2584 else
2585 sync_request_write(mddev, r1_bio);
2586 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2587 test_bit(R1BIO_WriteError, &r1_bio->state))
2588 handle_write_finished(conf, r1_bio);
2589 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2590 handle_read_error(conf, r1_bio);
2591 else
2592 WARN_ON_ONCE(1);
2593
2594 cond_resched();
2595 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2596 md_check_recovery(mddev);
2597 }
2598 blk_finish_plug(&plug);
2599}
2600
2601static int init_resync(struct r1conf *conf)
2602{
2603 int buffs;
2604
2605 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2606 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2607
2608 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2609 r1buf_pool_free, conf->poolinfo);
2610}
2611
2612static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2613{
2614 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2615 struct resync_pages *rps;
2616 struct bio *bio;
2617 int i;
2618
2619 for (i = conf->poolinfo->raid_disks; i--; ) {
2620 bio = r1bio->bios[i];
2621 rps = bio->bi_private;
2622 bio_reset(bio);
2623 bio->bi_private = rps;
2624 }
2625 r1bio->master_bio = NULL;
2626 return r1bio;
2627}
2628
2629/*
2630 * perform a "sync" on one "block"
2631 *
2632 * We need to make sure that no normal I/O request - particularly write
2633 * requests - conflict with active sync requests.
2634 *
2635 * This is achieved by tracking pending requests and a 'barrier' concept
2636 * that can be installed to exclude normal IO requests.
2637 */
2638
2639static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2640 int *skipped)
2641{
2642 struct r1conf *conf = mddev->private;
2643 struct r1bio *r1_bio;
2644 struct bio *bio;
2645 sector_t max_sector, nr_sectors;
2646 int disk = -1;
2647 int i;
2648 int wonly = -1;
2649 int write_targets = 0, read_targets = 0;
2650 sector_t sync_blocks;
2651 int still_degraded = 0;
2652 int good_sectors = RESYNC_SECTORS;
2653 int min_bad = 0; /* number of sectors that are bad in all devices */
2654 int idx = sector_to_idx(sector_nr);
2655 int page_idx = 0;
2656
2657 if (!mempool_initialized(&conf->r1buf_pool))
2658 if (init_resync(conf))
2659 return 0;
2660
2661 max_sector = mddev->dev_sectors;
2662 if (sector_nr >= max_sector) {
2663 /* If we aborted, we need to abort the
2664 * sync on the 'current' bitmap chunk (there will
2665 * only be one in raid1 resync.
2666 * We can find the current addess in mddev->curr_resync
2667 */
2668 if (mddev->curr_resync < max_sector) /* aborted */
2669 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2670 &sync_blocks, 1);
2671 else /* completed sync */
2672 conf->fullsync = 0;
2673
2674 md_bitmap_close_sync(mddev->bitmap);
2675 close_sync(conf);
2676
2677 if (mddev_is_clustered(mddev)) {
2678 conf->cluster_sync_low = 0;
2679 conf->cluster_sync_high = 0;
2680 }
2681 return 0;
2682 }
2683
2684 if (mddev->bitmap == NULL &&
2685 mddev->recovery_cp == MaxSector &&
2686 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2687 conf->fullsync == 0) {
2688 *skipped = 1;
2689 return max_sector - sector_nr;
2690 }
2691 /* before building a request, check if we can skip these blocks..
2692 * This call the bitmap_start_sync doesn't actually record anything
2693 */
2694 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2695 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2696 /* We can skip this block, and probably several more */
2697 *skipped = 1;
2698 return sync_blocks;
2699 }
2700
2701 /*
2702 * If there is non-resync activity waiting for a turn, then let it
2703 * though before starting on this new sync request.
2704 */
2705 if (atomic_read(&conf->nr_waiting[idx]))
2706 schedule_timeout_uninterruptible(1);
2707
2708 /* we are incrementing sector_nr below. To be safe, we check against
2709 * sector_nr + two times RESYNC_SECTORS
2710 */
2711
2712 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2713 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2714
2715
2716 if (raise_barrier(conf, sector_nr))
2717 return 0;
2718
2719 r1_bio = raid1_alloc_init_r1buf(conf);
2720
2721 rcu_read_lock();
2722 /*
2723 * If we get a correctably read error during resync or recovery,
2724 * we might want to read from a different device. So we
2725 * flag all drives that could conceivably be read from for READ,
2726 * and any others (which will be non-In_sync devices) for WRITE.
2727 * If a read fails, we try reading from something else for which READ
2728 * is OK.
2729 */
2730
2731 r1_bio->mddev = mddev;
2732 r1_bio->sector = sector_nr;
2733 r1_bio->state = 0;
2734 set_bit(R1BIO_IsSync, &r1_bio->state);
2735 /* make sure good_sectors won't go across barrier unit boundary */
2736 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2737
2738 for (i = 0; i < conf->raid_disks * 2; i++) {
2739 struct md_rdev *rdev;
2740 bio = r1_bio->bios[i];
2741
2742 rdev = rcu_dereference(conf->mirrors[i].rdev);
2743 if (rdev == NULL ||
2744 test_bit(Faulty, &rdev->flags)) {
2745 if (i < conf->raid_disks)
2746 still_degraded = 1;
2747 } else if (!test_bit(In_sync, &rdev->flags)) {
2748 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2749 bio->bi_end_io = end_sync_write;
2750 write_targets ++;
2751 } else {
2752 /* may need to read from here */
2753 sector_t first_bad = MaxSector;
2754 int bad_sectors;
2755
2756 if (is_badblock(rdev, sector_nr, good_sectors,
2757 &first_bad, &bad_sectors)) {
2758 if (first_bad > sector_nr)
2759 good_sectors = first_bad - sector_nr;
2760 else {
2761 bad_sectors -= (sector_nr - first_bad);
2762 if (min_bad == 0 ||
2763 min_bad > bad_sectors)
2764 min_bad = bad_sectors;
2765 }
2766 }
2767 if (sector_nr < first_bad) {
2768 if (test_bit(WriteMostly, &rdev->flags)) {
2769 if (wonly < 0)
2770 wonly = i;
2771 } else {
2772 if (disk < 0)
2773 disk = i;
2774 }
2775 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2776 bio->bi_end_io = end_sync_read;
2777 read_targets++;
2778 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2779 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2780 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2781 /*
2782 * The device is suitable for reading (InSync),
2783 * but has bad block(s) here. Let's try to correct them,
2784 * if we are doing resync or repair. Otherwise, leave
2785 * this device alone for this sync request.
2786 */
2787 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2788 bio->bi_end_io = end_sync_write;
2789 write_targets++;
2790 }
2791 }
2792 if (rdev && bio->bi_end_io) {
2793 atomic_inc(&rdev->nr_pending);
2794 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2795 bio_set_dev(bio, rdev->bdev);
2796 if (test_bit(FailFast, &rdev->flags))
2797 bio->bi_opf |= MD_FAILFAST;
2798 }
2799 }
2800 rcu_read_unlock();
2801 if (disk < 0)
2802 disk = wonly;
2803 r1_bio->read_disk = disk;
2804
2805 if (read_targets == 0 && min_bad > 0) {
2806 /* These sectors are bad on all InSync devices, so we
2807 * need to mark them bad on all write targets
2808 */
2809 int ok = 1;
2810 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2811 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2812 struct md_rdev *rdev = conf->mirrors[i].rdev;
2813 ok = rdev_set_badblocks(rdev, sector_nr,
2814 min_bad, 0
2815 ) && ok;
2816 }
2817 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2818 *skipped = 1;
2819 put_buf(r1_bio);
2820
2821 if (!ok) {
2822 /* Cannot record the badblocks, so need to
2823 * abort the resync.
2824 * If there are multiple read targets, could just
2825 * fail the really bad ones ???
2826 */
2827 conf->recovery_disabled = mddev->recovery_disabled;
2828 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2829 return 0;
2830 } else
2831 return min_bad;
2832
2833 }
2834 if (min_bad > 0 && min_bad < good_sectors) {
2835 /* only resync enough to reach the next bad->good
2836 * transition */
2837 good_sectors = min_bad;
2838 }
2839
2840 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2841 /* extra read targets are also write targets */
2842 write_targets += read_targets-1;
2843
2844 if (write_targets == 0 || read_targets == 0) {
2845 /* There is nowhere to write, so all non-sync
2846 * drives must be failed - so we are finished
2847 */
2848 sector_t rv;
2849 if (min_bad > 0)
2850 max_sector = sector_nr + min_bad;
2851 rv = max_sector - sector_nr;
2852 *skipped = 1;
2853 put_buf(r1_bio);
2854 return rv;
2855 }
2856
2857 if (max_sector > mddev->resync_max)
2858 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2859 if (max_sector > sector_nr + good_sectors)
2860 max_sector = sector_nr + good_sectors;
2861 nr_sectors = 0;
2862 sync_blocks = 0;
2863 do {
2864 struct page *page;
2865 int len = PAGE_SIZE;
2866 if (sector_nr + (len>>9) > max_sector)
2867 len = (max_sector - sector_nr) << 9;
2868 if (len == 0)
2869 break;
2870 if (sync_blocks == 0) {
2871 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2872 &sync_blocks, still_degraded) &&
2873 !conf->fullsync &&
2874 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2875 break;
2876 if ((len >> 9) > sync_blocks)
2877 len = sync_blocks<<9;
2878 }
2879
2880 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2881 struct resync_pages *rp;
2882
2883 bio = r1_bio->bios[i];
2884 rp = get_resync_pages(bio);
2885 if (bio->bi_end_io) {
2886 page = resync_fetch_page(rp, page_idx);
2887
2888 /*
2889 * won't fail because the vec table is big
2890 * enough to hold all these pages
2891 */
2892 bio_add_page(bio, page, len, 0);
2893 }
2894 }
2895 nr_sectors += len>>9;
2896 sector_nr += len>>9;
2897 sync_blocks -= (len>>9);
2898 } while (++page_idx < RESYNC_PAGES);
2899
2900 r1_bio->sectors = nr_sectors;
2901
2902 if (mddev_is_clustered(mddev) &&
2903 conf->cluster_sync_high < sector_nr + nr_sectors) {
2904 conf->cluster_sync_low = mddev->curr_resync_completed;
2905 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2906 /* Send resync message */
2907 md_cluster_ops->resync_info_update(mddev,
2908 conf->cluster_sync_low,
2909 conf->cluster_sync_high);
2910 }
2911
2912 /* For a user-requested sync, we read all readable devices and do a
2913 * compare
2914 */
2915 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2916 atomic_set(&r1_bio->remaining, read_targets);
2917 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2918 bio = r1_bio->bios[i];
2919 if (bio->bi_end_io == end_sync_read) {
2920 read_targets--;
2921 md_sync_acct_bio(bio, nr_sectors);
2922 if (read_targets == 1)
2923 bio->bi_opf &= ~MD_FAILFAST;
2924 submit_bio_noacct(bio);
2925 }
2926 }
2927 } else {
2928 atomic_set(&r1_bio->remaining, 1);
2929 bio = r1_bio->bios[r1_bio->read_disk];
2930 md_sync_acct_bio(bio, nr_sectors);
2931 if (read_targets == 1)
2932 bio->bi_opf &= ~MD_FAILFAST;
2933 submit_bio_noacct(bio);
2934 }
2935 return nr_sectors;
2936}
2937
2938static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2939{
2940 if (sectors)
2941 return sectors;
2942
2943 return mddev->dev_sectors;
2944}
2945
2946static struct r1conf *setup_conf(struct mddev *mddev)
2947{
2948 struct r1conf *conf;
2949 int i;
2950 struct raid1_info *disk;
2951 struct md_rdev *rdev;
2952 int err = -ENOMEM;
2953
2954 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2955 if (!conf)
2956 goto abort;
2957
2958 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2959 sizeof(atomic_t), GFP_KERNEL);
2960 if (!conf->nr_pending)
2961 goto abort;
2962
2963 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2964 sizeof(atomic_t), GFP_KERNEL);
2965 if (!conf->nr_waiting)
2966 goto abort;
2967
2968 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2969 sizeof(atomic_t), GFP_KERNEL);
2970 if (!conf->nr_queued)
2971 goto abort;
2972
2973 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2974 sizeof(atomic_t), GFP_KERNEL);
2975 if (!conf->barrier)
2976 goto abort;
2977
2978 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2979 mddev->raid_disks, 2),
2980 GFP_KERNEL);
2981 if (!conf->mirrors)
2982 goto abort;
2983
2984 conf->tmppage = alloc_page(GFP_KERNEL);
2985 if (!conf->tmppage)
2986 goto abort;
2987
2988 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2989 if (!conf->poolinfo)
2990 goto abort;
2991 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2992 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2993 rbio_pool_free, conf->poolinfo);
2994 if (err)
2995 goto abort;
2996
2997 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2998 if (err)
2999 goto abort;
3000
3001 conf->poolinfo->mddev = mddev;
3002
3003 err = -EINVAL;
3004 spin_lock_init(&conf->device_lock);
3005 rdev_for_each(rdev, mddev) {
3006 int disk_idx = rdev->raid_disk;
3007 if (disk_idx >= mddev->raid_disks
3008 || disk_idx < 0)
3009 continue;
3010 if (test_bit(Replacement, &rdev->flags))
3011 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3012 else
3013 disk = conf->mirrors + disk_idx;
3014
3015 if (disk->rdev)
3016 goto abort;
3017 disk->rdev = rdev;
3018 disk->head_position = 0;
3019 disk->seq_start = MaxSector;
3020 }
3021 conf->raid_disks = mddev->raid_disks;
3022 conf->mddev = mddev;
3023 INIT_LIST_HEAD(&conf->retry_list);
3024 INIT_LIST_HEAD(&conf->bio_end_io_list);
3025
3026 spin_lock_init(&conf->resync_lock);
3027 init_waitqueue_head(&conf->wait_barrier);
3028
3029 bio_list_init(&conf->pending_bio_list);
3030 conf->pending_count = 0;
3031 conf->recovery_disabled = mddev->recovery_disabled - 1;
3032
3033 err = -EIO;
3034 for (i = 0; i < conf->raid_disks * 2; i++) {
3035
3036 disk = conf->mirrors + i;
3037
3038 if (i < conf->raid_disks &&
3039 disk[conf->raid_disks].rdev) {
3040 /* This slot has a replacement. */
3041 if (!disk->rdev) {
3042 /* No original, just make the replacement
3043 * a recovering spare
3044 */
3045 disk->rdev =
3046 disk[conf->raid_disks].rdev;
3047 disk[conf->raid_disks].rdev = NULL;
3048 } else if (!test_bit(In_sync, &disk->rdev->flags))
3049 /* Original is not in_sync - bad */
3050 goto abort;
3051 }
3052
3053 if (!disk->rdev ||
3054 !test_bit(In_sync, &disk->rdev->flags)) {
3055 disk->head_position = 0;
3056 if (disk->rdev &&
3057 (disk->rdev->saved_raid_disk < 0))
3058 conf->fullsync = 1;
3059 }
3060 }
3061
3062 err = -ENOMEM;
3063 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3064 if (!conf->thread)
3065 goto abort;
3066
3067 return conf;
3068
3069 abort:
3070 if (conf) {
3071 mempool_exit(&conf->r1bio_pool);
3072 kfree(conf->mirrors);
3073 safe_put_page(conf->tmppage);
3074 kfree(conf->poolinfo);
3075 kfree(conf->nr_pending);
3076 kfree(conf->nr_waiting);
3077 kfree(conf->nr_queued);
3078 kfree(conf->barrier);
3079 bioset_exit(&conf->bio_split);
3080 kfree(conf);
3081 }
3082 return ERR_PTR(err);
3083}
3084
3085static void raid1_free(struct mddev *mddev, void *priv);
3086static int raid1_run(struct mddev *mddev)
3087{
3088 struct r1conf *conf;
3089 int i;
3090 struct md_rdev *rdev;
3091 int ret;
3092 bool discard_supported = false;
3093
3094 if (mddev->level != 1) {
3095 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3096 mdname(mddev), mddev->level);
3097 return -EIO;
3098 }
3099 if (mddev->reshape_position != MaxSector) {
3100 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3101 mdname(mddev));
3102 return -EIO;
3103 }
3104 if (mddev_init_writes_pending(mddev) < 0)
3105 return -ENOMEM;
3106 /*
3107 * copy the already verified devices into our private RAID1
3108 * bookkeeping area. [whatever we allocate in run(),
3109 * should be freed in raid1_free()]
3110 */
3111 if (mddev->private == NULL)
3112 conf = setup_conf(mddev);
3113 else
3114 conf = mddev->private;
3115
3116 if (IS_ERR(conf))
3117 return PTR_ERR(conf);
3118
3119 if (mddev->queue) {
3120 blk_queue_max_write_same_sectors(mddev->queue, 0);
3121 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3122 }
3123
3124 rdev_for_each(rdev, mddev) {
3125 if (!mddev->gendisk)
3126 continue;
3127 disk_stack_limits(mddev->gendisk, rdev->bdev,
3128 rdev->data_offset << 9);
3129 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3130 discard_supported = true;
3131 }
3132
3133 mddev->degraded = 0;
3134 for (i = 0; i < conf->raid_disks; i++)
3135 if (conf->mirrors[i].rdev == NULL ||
3136 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3137 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3138 mddev->degraded++;
3139 /*
3140 * RAID1 needs at least one disk in active
3141 */
3142 if (conf->raid_disks - mddev->degraded < 1) {
3143 ret = -EINVAL;
3144 goto abort;
3145 }
3146
3147 if (conf->raid_disks - mddev->degraded == 1)
3148 mddev->recovery_cp = MaxSector;
3149
3150 if (mddev->recovery_cp != MaxSector)
3151 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3152 mdname(mddev));
3153 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3154 mdname(mddev), mddev->raid_disks - mddev->degraded,
3155 mddev->raid_disks);
3156
3157 /*
3158 * Ok, everything is just fine now
3159 */
3160 mddev->thread = conf->thread;
3161 conf->thread = NULL;
3162 mddev->private = conf;
3163 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3164
3165 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3166
3167 if (mddev->queue) {
3168 if (discard_supported)
3169 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3170 mddev->queue);
3171 else
3172 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3173 mddev->queue);
3174 }
3175
3176 ret = md_integrity_register(mddev);
3177 if (ret) {
3178 md_unregister_thread(&mddev->thread);
3179 goto abort;
3180 }
3181 return 0;
3182
3183abort:
3184 raid1_free(mddev, conf);
3185 return ret;
3186}
3187
3188static void raid1_free(struct mddev *mddev, void *priv)
3189{
3190 struct r1conf *conf = priv;
3191
3192 mempool_exit(&conf->r1bio_pool);
3193 kfree(conf->mirrors);
3194 safe_put_page(conf->tmppage);
3195 kfree(conf->poolinfo);
3196 kfree(conf->nr_pending);
3197 kfree(conf->nr_waiting);
3198 kfree(conf->nr_queued);
3199 kfree(conf->barrier);
3200 bioset_exit(&conf->bio_split);
3201 kfree(conf);
3202}
3203
3204static int raid1_resize(struct mddev *mddev, sector_t sectors)
3205{
3206 /* no resync is happening, and there is enough space
3207 * on all devices, so we can resize.
3208 * We need to make sure resync covers any new space.
3209 * If the array is shrinking we should possibly wait until
3210 * any io in the removed space completes, but it hardly seems
3211 * worth it.
3212 */
3213 sector_t newsize = raid1_size(mddev, sectors, 0);
3214 if (mddev->external_size &&
3215 mddev->array_sectors > newsize)
3216 return -EINVAL;
3217 if (mddev->bitmap) {
3218 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3219 if (ret)
3220 return ret;
3221 }
3222 md_set_array_sectors(mddev, newsize);
3223 if (sectors > mddev->dev_sectors &&
3224 mddev->recovery_cp > mddev->dev_sectors) {
3225 mddev->recovery_cp = mddev->dev_sectors;
3226 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3227 }
3228 mddev->dev_sectors = sectors;
3229 mddev->resync_max_sectors = sectors;
3230 return 0;
3231}
3232
3233static int raid1_reshape(struct mddev *mddev)
3234{
3235 /* We need to:
3236 * 1/ resize the r1bio_pool
3237 * 2/ resize conf->mirrors
3238 *
3239 * We allocate a new r1bio_pool if we can.
3240 * Then raise a device barrier and wait until all IO stops.
3241 * Then resize conf->mirrors and swap in the new r1bio pool.
3242 *
3243 * At the same time, we "pack" the devices so that all the missing
3244 * devices have the higher raid_disk numbers.
3245 */
3246 mempool_t newpool, oldpool;
3247 struct pool_info *newpoolinfo;
3248 struct raid1_info *newmirrors;
3249 struct r1conf *conf = mddev->private;
3250 int cnt, raid_disks;
3251 unsigned long flags;
3252 int d, d2;
3253 int ret;
3254
3255 memset(&newpool, 0, sizeof(newpool));
3256 memset(&oldpool, 0, sizeof(oldpool));
3257
3258 /* Cannot change chunk_size, layout, or level */
3259 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3260 mddev->layout != mddev->new_layout ||
3261 mddev->level != mddev->new_level) {
3262 mddev->new_chunk_sectors = mddev->chunk_sectors;
3263 mddev->new_layout = mddev->layout;
3264 mddev->new_level = mddev->level;
3265 return -EINVAL;
3266 }
3267
3268 if (!mddev_is_clustered(mddev))
3269 md_allow_write(mddev);
3270
3271 raid_disks = mddev->raid_disks + mddev->delta_disks;
3272
3273 if (raid_disks < conf->raid_disks) {
3274 cnt=0;
3275 for (d= 0; d < conf->raid_disks; d++)
3276 if (conf->mirrors[d].rdev)
3277 cnt++;
3278 if (cnt > raid_disks)
3279 return -EBUSY;
3280 }
3281
3282 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3283 if (!newpoolinfo)
3284 return -ENOMEM;
3285 newpoolinfo->mddev = mddev;
3286 newpoolinfo->raid_disks = raid_disks * 2;
3287
3288 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3289 rbio_pool_free, newpoolinfo);
3290 if (ret) {
3291 kfree(newpoolinfo);
3292 return ret;
3293 }
3294 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3295 raid_disks, 2),
3296 GFP_KERNEL);
3297 if (!newmirrors) {
3298 kfree(newpoolinfo);
3299 mempool_exit(&newpool);
3300 return -ENOMEM;
3301 }
3302
3303 freeze_array(conf, 0);
3304
3305 /* ok, everything is stopped */
3306 oldpool = conf->r1bio_pool;
3307 conf->r1bio_pool = newpool;
3308
3309 for (d = d2 = 0; d < conf->raid_disks; d++) {
3310 struct md_rdev *rdev = conf->mirrors[d].rdev;
3311 if (rdev && rdev->raid_disk != d2) {
3312 sysfs_unlink_rdev(mddev, rdev);
3313 rdev->raid_disk = d2;
3314 sysfs_unlink_rdev(mddev, rdev);
3315 if (sysfs_link_rdev(mddev, rdev))
3316 pr_warn("md/raid1:%s: cannot register rd%d\n",
3317 mdname(mddev), rdev->raid_disk);
3318 }
3319 if (rdev)
3320 newmirrors[d2++].rdev = rdev;
3321 }
3322 kfree(conf->mirrors);
3323 conf->mirrors = newmirrors;
3324 kfree(conf->poolinfo);
3325 conf->poolinfo = newpoolinfo;
3326
3327 spin_lock_irqsave(&conf->device_lock, flags);
3328 mddev->degraded += (raid_disks - conf->raid_disks);
3329 spin_unlock_irqrestore(&conf->device_lock, flags);
3330 conf->raid_disks = mddev->raid_disks = raid_disks;
3331 mddev->delta_disks = 0;
3332
3333 unfreeze_array(conf);
3334
3335 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3336 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3337 md_wakeup_thread(mddev->thread);
3338
3339 mempool_exit(&oldpool);
3340 return 0;
3341}
3342
3343static void raid1_quiesce(struct mddev *mddev, int quiesce)
3344{
3345 struct r1conf *conf = mddev->private;
3346
3347 if (quiesce)
3348 freeze_array(conf, 0);
3349 else
3350 unfreeze_array(conf);
3351}
3352
3353static void *raid1_takeover(struct mddev *mddev)
3354{
3355 /* raid1 can take over:
3356 * raid5 with 2 devices, any layout or chunk size
3357 */
3358 if (mddev->level == 5 && mddev->raid_disks == 2) {
3359 struct r1conf *conf;
3360 mddev->new_level = 1;
3361 mddev->new_layout = 0;
3362 mddev->new_chunk_sectors = 0;
3363 conf = setup_conf(mddev);
3364 if (!IS_ERR(conf)) {
3365 /* Array must appear to be quiesced */
3366 conf->array_frozen = 1;
3367 mddev_clear_unsupported_flags(mddev,
3368 UNSUPPORTED_MDDEV_FLAGS);
3369 }
3370 return conf;
3371 }
3372 return ERR_PTR(-EINVAL);
3373}
3374
3375static struct md_personality raid1_personality =
3376{
3377 .name = "raid1",
3378 .level = 1,
3379 .owner = THIS_MODULE,
3380 .make_request = raid1_make_request,
3381 .run = raid1_run,
3382 .free = raid1_free,
3383 .status = raid1_status,
3384 .error_handler = raid1_error,
3385 .hot_add_disk = raid1_add_disk,
3386 .hot_remove_disk= raid1_remove_disk,
3387 .spare_active = raid1_spare_active,
3388 .sync_request = raid1_sync_request,
3389 .resize = raid1_resize,
3390 .size = raid1_size,
3391 .check_reshape = raid1_reshape,
3392 .quiesce = raid1_quiesce,
3393 .takeover = raid1_takeover,
3394};
3395
3396static int __init raid_init(void)
3397{
3398 return register_md_personality(&raid1_personality);
3399}
3400
3401static void raid_exit(void)
3402{
3403 unregister_md_personality(&raid1_personality);
3404}
3405
3406module_init(raid_init);
3407module_exit(raid_exit);
3408MODULE_LICENSE("GPL");
3409MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3410MODULE_ALIAS("md-personality-3"); /* RAID1 */
3411MODULE_ALIAS("md-raid1");
3412MODULE_ALIAS("md-level-1");
3413
3414module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);