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/*
2 * raid1.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5 *
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7 *
8 * RAID-1 management functions.
9 *
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11 *
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
27 * any later version.
28 *
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32 */
33
34#include <linux/slab.h>
35#include <linux/delay.h>
36#include <linux/blkdev.h>
37#include <linux/module.h>
38#include <linux/seq_file.h>
39#include <linux/ratelimit.h>
40
41#include <trace/events/block.h>
42
43#include "md.h"
44#include "raid1.h"
45#include "md-bitmap.h"
46
47#define UNSUPPORTED_MDDEV_FLAGS \
48 ((1L << MD_HAS_JOURNAL) | \
49 (1L << MD_JOURNAL_CLEAN) | \
50 (1L << MD_HAS_PPL) | \
51 (1L << MD_HAS_MULTIPLE_PPLS))
52
53/*
54 * Number of guaranteed r1bios in case of extreme VM load:
55 */
56#define NR_RAID1_BIOS 256
57
58/* when we get a read error on a read-only array, we redirect to another
59 * device without failing the first device, or trying to over-write to
60 * correct the read error. To keep track of bad blocks on a per-bio
61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
62 */
63#define IO_BLOCKED ((struct bio *)1)
64/* When we successfully write to a known bad-block, we need to remove the
65 * bad-block marking which must be done from process context. So we record
66 * the success by setting devs[n].bio to IO_MADE_GOOD
67 */
68#define IO_MADE_GOOD ((struct bio *)2)
69
70#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
71
72/* When there are this many requests queue to be written by
73 * the raid1 thread, we become 'congested' to provide back-pressure
74 * for writeback.
75 */
76static int max_queued_requests = 1024;
77
78static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
79static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
80
81#define raid1_log(md, fmt, args...) \
82 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
83
84#include "raid1-10.c"
85
86/*
87 * for resync bio, r1bio pointer can be retrieved from the per-bio
88 * 'struct resync_pages'.
89 */
90static inline struct r1bio *get_resync_r1bio(struct bio *bio)
91{
92 return get_resync_pages(bio)->raid_bio;
93}
94
95static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
96{
97 struct pool_info *pi = data;
98 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
99
100 /* allocate a r1bio with room for raid_disks entries in the bios array */
101 return kzalloc(size, gfp_flags);
102}
103
104static void r1bio_pool_free(void *r1_bio, void *data)
105{
106 kfree(r1_bio);
107}
108
109#define RESYNC_DEPTH 32
110#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
111#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
112#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
113#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
114#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
115
116static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
117{
118 struct pool_info *pi = data;
119 struct r1bio *r1_bio;
120 struct bio *bio;
121 int need_pages;
122 int j;
123 struct resync_pages *rps;
124
125 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
126 if (!r1_bio)
127 return NULL;
128
129 rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
130 gfp_flags);
131 if (!rps)
132 goto out_free_r1bio;
133
134 /*
135 * Allocate bios : 1 for reading, n-1 for writing
136 */
137 for (j = pi->raid_disks ; j-- ; ) {
138 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
139 if (!bio)
140 goto out_free_bio;
141 r1_bio->bios[j] = bio;
142 }
143 /*
144 * Allocate RESYNC_PAGES data pages and attach them to
145 * the first bio.
146 * If this is a user-requested check/repair, allocate
147 * RESYNC_PAGES for each bio.
148 */
149 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
150 need_pages = pi->raid_disks;
151 else
152 need_pages = 1;
153 for (j = 0; j < pi->raid_disks; j++) {
154 struct resync_pages *rp = &rps[j];
155
156 bio = r1_bio->bios[j];
157
158 if (j < need_pages) {
159 if (resync_alloc_pages(rp, gfp_flags))
160 goto out_free_pages;
161 } else {
162 memcpy(rp, &rps[0], sizeof(*rp));
163 resync_get_all_pages(rp);
164 }
165
166 rp->raid_bio = r1_bio;
167 bio->bi_private = rp;
168 }
169
170 r1_bio->master_bio = NULL;
171
172 return r1_bio;
173
174out_free_pages:
175 while (--j >= 0)
176 resync_free_pages(&rps[j]);
177
178out_free_bio:
179 while (++j < pi->raid_disks)
180 bio_put(r1_bio->bios[j]);
181 kfree(rps);
182
183out_free_r1bio:
184 r1bio_pool_free(r1_bio, data);
185 return NULL;
186}
187
188static void r1buf_pool_free(void *__r1_bio, void *data)
189{
190 struct pool_info *pi = data;
191 int i;
192 struct r1bio *r1bio = __r1_bio;
193 struct resync_pages *rp = NULL;
194
195 for (i = pi->raid_disks; i--; ) {
196 rp = get_resync_pages(r1bio->bios[i]);
197 resync_free_pages(rp);
198 bio_put(r1bio->bios[i]);
199 }
200
201 /* resync pages array stored in the 1st bio's .bi_private */
202 kfree(rp);
203
204 r1bio_pool_free(r1bio, data);
205}
206
207static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
208{
209 int i;
210
211 for (i = 0; i < conf->raid_disks * 2; i++) {
212 struct bio **bio = r1_bio->bios + i;
213 if (!BIO_SPECIAL(*bio))
214 bio_put(*bio);
215 *bio = NULL;
216 }
217}
218
219static void free_r1bio(struct r1bio *r1_bio)
220{
221 struct r1conf *conf = r1_bio->mddev->private;
222
223 put_all_bios(conf, r1_bio);
224 mempool_free(r1_bio, conf->r1bio_pool);
225}
226
227static void put_buf(struct r1bio *r1_bio)
228{
229 struct r1conf *conf = r1_bio->mddev->private;
230 sector_t sect = r1_bio->sector;
231 int i;
232
233 for (i = 0; i < conf->raid_disks * 2; i++) {
234 struct bio *bio = r1_bio->bios[i];
235 if (bio->bi_end_io)
236 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
237 }
238
239 mempool_free(r1_bio, conf->r1buf_pool);
240
241 lower_barrier(conf, sect);
242}
243
244static void reschedule_retry(struct r1bio *r1_bio)
245{
246 unsigned long flags;
247 struct mddev *mddev = r1_bio->mddev;
248 struct r1conf *conf = mddev->private;
249 int idx;
250
251 idx = sector_to_idx(r1_bio->sector);
252 spin_lock_irqsave(&conf->device_lock, flags);
253 list_add(&r1_bio->retry_list, &conf->retry_list);
254 atomic_inc(&conf->nr_queued[idx]);
255 spin_unlock_irqrestore(&conf->device_lock, flags);
256
257 wake_up(&conf->wait_barrier);
258 md_wakeup_thread(mddev->thread);
259}
260
261/*
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
264 * cache layer.
265 */
266static void call_bio_endio(struct r1bio *r1_bio)
267{
268 struct bio *bio = r1_bio->master_bio;
269 struct r1conf *conf = r1_bio->mddev->private;
270
271 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
272 bio->bi_status = BLK_STS_IOERR;
273
274 bio_endio(bio);
275 /*
276 * Wake up any possible resync thread that waits for the device
277 * to go idle.
278 */
279 allow_barrier(conf, r1_bio->sector);
280}
281
282static void raid_end_bio_io(struct r1bio *r1_bio)
283{
284 struct bio *bio = r1_bio->master_bio;
285
286 /* if nobody has done the final endio yet, do it now */
287 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
288 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
289 (bio_data_dir(bio) == WRITE) ? "write" : "read",
290 (unsigned long long) bio->bi_iter.bi_sector,
291 (unsigned long long) bio_end_sector(bio) - 1);
292
293 call_bio_endio(r1_bio);
294 }
295 free_r1bio(r1_bio);
296}
297
298/*
299 * Update disk head position estimator based on IRQ completion info.
300 */
301static inline void update_head_pos(int disk, struct r1bio *r1_bio)
302{
303 struct r1conf *conf = r1_bio->mddev->private;
304
305 conf->mirrors[disk].head_position =
306 r1_bio->sector + (r1_bio->sectors);
307}
308
309/*
310 * Find the disk number which triggered given bio
311 */
312static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
313{
314 int mirror;
315 struct r1conf *conf = r1_bio->mddev->private;
316 int raid_disks = conf->raid_disks;
317
318 for (mirror = 0; mirror < raid_disks * 2; mirror++)
319 if (r1_bio->bios[mirror] == bio)
320 break;
321
322 BUG_ON(mirror == raid_disks * 2);
323 update_head_pos(mirror, r1_bio);
324
325 return mirror;
326}
327
328static void raid1_end_read_request(struct bio *bio)
329{
330 int uptodate = !bio->bi_status;
331 struct r1bio *r1_bio = bio->bi_private;
332 struct r1conf *conf = r1_bio->mddev->private;
333 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
334
335 /*
336 * this branch is our 'one mirror IO has finished' event handler:
337 */
338 update_head_pos(r1_bio->read_disk, r1_bio);
339
340 if (uptodate)
341 set_bit(R1BIO_Uptodate, &r1_bio->state);
342 else if (test_bit(FailFast, &rdev->flags) &&
343 test_bit(R1BIO_FailFast, &r1_bio->state))
344 /* This was a fail-fast read so we definitely
345 * want to retry */
346 ;
347 else {
348 /* If all other devices have failed, we want to return
349 * the error upwards rather than fail the last device.
350 * Here we redefine "uptodate" to mean "Don't want to retry"
351 */
352 unsigned long flags;
353 spin_lock_irqsave(&conf->device_lock, flags);
354 if (r1_bio->mddev->degraded == conf->raid_disks ||
355 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
356 test_bit(In_sync, &rdev->flags)))
357 uptodate = 1;
358 spin_unlock_irqrestore(&conf->device_lock, flags);
359 }
360
361 if (uptodate) {
362 raid_end_bio_io(r1_bio);
363 rdev_dec_pending(rdev, conf->mddev);
364 } else {
365 /*
366 * oops, read error:
367 */
368 char b[BDEVNAME_SIZE];
369 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
370 mdname(conf->mddev),
371 bdevname(rdev->bdev, b),
372 (unsigned long long)r1_bio->sector);
373 set_bit(R1BIO_ReadError, &r1_bio->state);
374 reschedule_retry(r1_bio);
375 /* don't drop the reference on read_disk yet */
376 }
377}
378
379static void close_write(struct r1bio *r1_bio)
380{
381 /* it really is the end of this request */
382 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383 bio_free_pages(r1_bio->behind_master_bio);
384 bio_put(r1_bio->behind_master_bio);
385 r1_bio->behind_master_bio = NULL;
386 }
387 /* clear the bitmap if all writes complete successfully */
388 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
389 r1_bio->sectors,
390 !test_bit(R1BIO_Degraded, &r1_bio->state),
391 test_bit(R1BIO_BehindIO, &r1_bio->state));
392 md_write_end(r1_bio->mddev);
393}
394
395static void r1_bio_write_done(struct r1bio *r1_bio)
396{
397 if (!atomic_dec_and_test(&r1_bio->remaining))
398 return;
399
400 if (test_bit(R1BIO_WriteError, &r1_bio->state))
401 reschedule_retry(r1_bio);
402 else {
403 close_write(r1_bio);
404 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
405 reschedule_retry(r1_bio);
406 else
407 raid_end_bio_io(r1_bio);
408 }
409}
410
411static void raid1_end_write_request(struct bio *bio)
412{
413 struct r1bio *r1_bio = bio->bi_private;
414 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
415 struct r1conf *conf = r1_bio->mddev->private;
416 struct bio *to_put = NULL;
417 int mirror = find_bio_disk(r1_bio, bio);
418 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
419 bool discard_error;
420
421 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
422
423 /*
424 * 'one mirror IO has finished' event handler:
425 */
426 if (bio->bi_status && !discard_error) {
427 set_bit(WriteErrorSeen, &rdev->flags);
428 if (!test_and_set_bit(WantReplacement, &rdev->flags))
429 set_bit(MD_RECOVERY_NEEDED, &
430 conf->mddev->recovery);
431
432 if (test_bit(FailFast, &rdev->flags) &&
433 (bio->bi_opf & MD_FAILFAST) &&
434 /* We never try FailFast to WriteMostly devices */
435 !test_bit(WriteMostly, &rdev->flags)) {
436 md_error(r1_bio->mddev, rdev);
437 if (!test_bit(Faulty, &rdev->flags))
438 /* This is the only remaining device,
439 * We need to retry the write without
440 * FailFast
441 */
442 set_bit(R1BIO_WriteError, &r1_bio->state);
443 else {
444 /* Finished with this branch */
445 r1_bio->bios[mirror] = NULL;
446 to_put = bio;
447 }
448 } else
449 set_bit(R1BIO_WriteError, &r1_bio->state);
450 } else {
451 /*
452 * Set R1BIO_Uptodate in our master bio, so that we
453 * will return a good error code for to the higher
454 * levels even if IO on some other mirrored buffer
455 * fails.
456 *
457 * The 'master' represents the composite IO operation
458 * to user-side. So if something waits for IO, then it
459 * will wait for the 'master' bio.
460 */
461 sector_t first_bad;
462 int bad_sectors;
463
464 r1_bio->bios[mirror] = NULL;
465 to_put = bio;
466 /*
467 * Do not set R1BIO_Uptodate if the current device is
468 * rebuilding or Faulty. This is because we cannot use
469 * such device for properly reading the data back (we could
470 * potentially use it, if the current write would have felt
471 * before rdev->recovery_offset, but for simplicity we don't
472 * check this here.
473 */
474 if (test_bit(In_sync, &rdev->flags) &&
475 !test_bit(Faulty, &rdev->flags))
476 set_bit(R1BIO_Uptodate, &r1_bio->state);
477
478 /* Maybe we can clear some bad blocks. */
479 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
480 &first_bad, &bad_sectors) && !discard_error) {
481 r1_bio->bios[mirror] = IO_MADE_GOOD;
482 set_bit(R1BIO_MadeGood, &r1_bio->state);
483 }
484 }
485
486 if (behind) {
487 if (test_bit(WriteMostly, &rdev->flags))
488 atomic_dec(&r1_bio->behind_remaining);
489
490 /*
491 * In behind mode, we ACK the master bio once the I/O
492 * has safely reached all non-writemostly
493 * disks. Setting the Returned bit ensures that this
494 * gets done only once -- we don't ever want to return
495 * -EIO here, instead we'll wait
496 */
497 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
498 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
499 /* Maybe we can return now */
500 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
501 struct bio *mbio = r1_bio->master_bio;
502 pr_debug("raid1: behind end write sectors"
503 " %llu-%llu\n",
504 (unsigned long long) mbio->bi_iter.bi_sector,
505 (unsigned long long) bio_end_sector(mbio) - 1);
506 call_bio_endio(r1_bio);
507 }
508 }
509 }
510 if (r1_bio->bios[mirror] == NULL)
511 rdev_dec_pending(rdev, conf->mddev);
512
513 /*
514 * Let's see if all mirrored write operations have finished
515 * already.
516 */
517 r1_bio_write_done(r1_bio);
518
519 if (to_put)
520 bio_put(to_put);
521}
522
523static sector_t align_to_barrier_unit_end(sector_t start_sector,
524 sector_t sectors)
525{
526 sector_t len;
527
528 WARN_ON(sectors == 0);
529 /*
530 * len is the number of sectors from start_sector to end of the
531 * barrier unit which start_sector belongs to.
532 */
533 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
534 start_sector;
535
536 if (len > sectors)
537 len = sectors;
538
539 return len;
540}
541
542/*
543 * This routine returns the disk from which the requested read should
544 * be done. There is a per-array 'next expected sequential IO' sector
545 * number - if this matches on the next IO then we use the last disk.
546 * There is also a per-disk 'last know head position' sector that is
547 * maintained from IRQ contexts, both the normal and the resync IO
548 * completion handlers update this position correctly. If there is no
549 * perfect sequential match then we pick the disk whose head is closest.
550 *
551 * If there are 2 mirrors in the same 2 devices, performance degrades
552 * because position is mirror, not device based.
553 *
554 * The rdev for the device selected will have nr_pending incremented.
555 */
556static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
557{
558 const sector_t this_sector = r1_bio->sector;
559 int sectors;
560 int best_good_sectors;
561 int best_disk, best_dist_disk, best_pending_disk;
562 int has_nonrot_disk;
563 int disk;
564 sector_t best_dist;
565 unsigned int min_pending;
566 struct md_rdev *rdev;
567 int choose_first;
568 int choose_next_idle;
569
570 rcu_read_lock();
571 /*
572 * Check if we can balance. We can balance on the whole
573 * device if no resync is going on, or below the resync window.
574 * We take the first readable disk when above the resync window.
575 */
576 retry:
577 sectors = r1_bio->sectors;
578 best_disk = -1;
579 best_dist_disk = -1;
580 best_dist = MaxSector;
581 best_pending_disk = -1;
582 min_pending = UINT_MAX;
583 best_good_sectors = 0;
584 has_nonrot_disk = 0;
585 choose_next_idle = 0;
586 clear_bit(R1BIO_FailFast, &r1_bio->state);
587
588 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
589 (mddev_is_clustered(conf->mddev) &&
590 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
591 this_sector + sectors)))
592 choose_first = 1;
593 else
594 choose_first = 0;
595
596 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
597 sector_t dist;
598 sector_t first_bad;
599 int bad_sectors;
600 unsigned int pending;
601 bool nonrot;
602
603 rdev = rcu_dereference(conf->mirrors[disk].rdev);
604 if (r1_bio->bios[disk] == IO_BLOCKED
605 || rdev == NULL
606 || test_bit(Faulty, &rdev->flags))
607 continue;
608 if (!test_bit(In_sync, &rdev->flags) &&
609 rdev->recovery_offset < this_sector + sectors)
610 continue;
611 if (test_bit(WriteMostly, &rdev->flags)) {
612 /* Don't balance among write-mostly, just
613 * use the first as a last resort */
614 if (best_dist_disk < 0) {
615 if (is_badblock(rdev, this_sector, sectors,
616 &first_bad, &bad_sectors)) {
617 if (first_bad <= this_sector)
618 /* Cannot use this */
619 continue;
620 best_good_sectors = first_bad - this_sector;
621 } else
622 best_good_sectors = sectors;
623 best_dist_disk = disk;
624 best_pending_disk = disk;
625 }
626 continue;
627 }
628 /* This is a reasonable device to use. It might
629 * even be best.
630 */
631 if (is_badblock(rdev, this_sector, sectors,
632 &first_bad, &bad_sectors)) {
633 if (best_dist < MaxSector)
634 /* already have a better device */
635 continue;
636 if (first_bad <= this_sector) {
637 /* cannot read here. If this is the 'primary'
638 * device, then we must not read beyond
639 * bad_sectors from another device..
640 */
641 bad_sectors -= (this_sector - first_bad);
642 if (choose_first && sectors > bad_sectors)
643 sectors = bad_sectors;
644 if (best_good_sectors > sectors)
645 best_good_sectors = sectors;
646
647 } else {
648 sector_t good_sectors = first_bad - this_sector;
649 if (good_sectors > best_good_sectors) {
650 best_good_sectors = good_sectors;
651 best_disk = disk;
652 }
653 if (choose_first)
654 break;
655 }
656 continue;
657 } else {
658 if ((sectors > best_good_sectors) && (best_disk >= 0))
659 best_disk = -1;
660 best_good_sectors = sectors;
661 }
662
663 if (best_disk >= 0)
664 /* At least two disks to choose from so failfast is OK */
665 set_bit(R1BIO_FailFast, &r1_bio->state);
666
667 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
668 has_nonrot_disk |= nonrot;
669 pending = atomic_read(&rdev->nr_pending);
670 dist = abs(this_sector - conf->mirrors[disk].head_position);
671 if (choose_first) {
672 best_disk = disk;
673 break;
674 }
675 /* Don't change to another disk for sequential reads */
676 if (conf->mirrors[disk].next_seq_sect == this_sector
677 || dist == 0) {
678 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
679 struct raid1_info *mirror = &conf->mirrors[disk];
680
681 best_disk = disk;
682 /*
683 * If buffered sequential IO size exceeds optimal
684 * iosize, check if there is idle disk. If yes, choose
685 * the idle disk. read_balance could already choose an
686 * idle disk before noticing it's a sequential IO in
687 * this disk. This doesn't matter because this disk
688 * will idle, next time it will be utilized after the
689 * first disk has IO size exceeds optimal iosize. In
690 * this way, iosize of the first disk will be optimal
691 * iosize at least. iosize of the second disk might be
692 * small, but not a big deal since when the second disk
693 * starts IO, the first disk is likely still busy.
694 */
695 if (nonrot && opt_iosize > 0 &&
696 mirror->seq_start != MaxSector &&
697 mirror->next_seq_sect > opt_iosize &&
698 mirror->next_seq_sect - opt_iosize >=
699 mirror->seq_start) {
700 choose_next_idle = 1;
701 continue;
702 }
703 break;
704 }
705
706 if (choose_next_idle)
707 continue;
708
709 if (min_pending > pending) {
710 min_pending = pending;
711 best_pending_disk = disk;
712 }
713
714 if (dist < best_dist) {
715 best_dist = dist;
716 best_dist_disk = disk;
717 }
718 }
719
720 /*
721 * If all disks are rotational, choose the closest disk. If any disk is
722 * non-rotational, choose the disk with less pending request even the
723 * disk is rotational, which might/might not be optimal for raids with
724 * mixed ratation/non-rotational disks depending on workload.
725 */
726 if (best_disk == -1) {
727 if (has_nonrot_disk || min_pending == 0)
728 best_disk = best_pending_disk;
729 else
730 best_disk = best_dist_disk;
731 }
732
733 if (best_disk >= 0) {
734 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
735 if (!rdev)
736 goto retry;
737 atomic_inc(&rdev->nr_pending);
738 sectors = best_good_sectors;
739
740 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
741 conf->mirrors[best_disk].seq_start = this_sector;
742
743 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
744 }
745 rcu_read_unlock();
746 *max_sectors = sectors;
747
748 return best_disk;
749}
750
751static int raid1_congested(struct mddev *mddev, int bits)
752{
753 struct r1conf *conf = mddev->private;
754 int i, ret = 0;
755
756 if ((bits & (1 << WB_async_congested)) &&
757 conf->pending_count >= max_queued_requests)
758 return 1;
759
760 rcu_read_lock();
761 for (i = 0; i < conf->raid_disks * 2; i++) {
762 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
763 if (rdev && !test_bit(Faulty, &rdev->flags)) {
764 struct request_queue *q = bdev_get_queue(rdev->bdev);
765
766 BUG_ON(!q);
767
768 /* Note the '|| 1' - when read_balance prefers
769 * non-congested targets, it can be removed
770 */
771 if ((bits & (1 << WB_async_congested)) || 1)
772 ret |= bdi_congested(q->backing_dev_info, bits);
773 else
774 ret &= bdi_congested(q->backing_dev_info, bits);
775 }
776 }
777 rcu_read_unlock();
778 return ret;
779}
780
781static void flush_bio_list(struct r1conf *conf, struct bio *bio)
782{
783 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
784 bitmap_unplug(conf->mddev->bitmap);
785 wake_up(&conf->wait_barrier);
786
787 while (bio) { /* submit pending writes */
788 struct bio *next = bio->bi_next;
789 struct md_rdev *rdev = (void *)bio->bi_disk;
790 bio->bi_next = NULL;
791 bio_set_dev(bio, rdev->bdev);
792 if (test_bit(Faulty, &rdev->flags)) {
793 bio_io_error(bio);
794 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
795 !blk_queue_discard(bio->bi_disk->queue)))
796 /* Just ignore it */
797 bio_endio(bio);
798 else
799 generic_make_request(bio);
800 bio = next;
801 }
802}
803
804static void flush_pending_writes(struct r1conf *conf)
805{
806 /* Any writes that have been queued but are awaiting
807 * bitmap updates get flushed here.
808 */
809 spin_lock_irq(&conf->device_lock);
810
811 if (conf->pending_bio_list.head) {
812 struct blk_plug plug;
813 struct bio *bio;
814
815 bio = bio_list_get(&conf->pending_bio_list);
816 conf->pending_count = 0;
817 spin_unlock_irq(&conf->device_lock);
818
819 /*
820 * As this is called in a wait_event() loop (see freeze_array),
821 * current->state might be TASK_UNINTERRUPTIBLE which will
822 * cause a warning when we prepare to wait again. As it is
823 * rare that this path is taken, it is perfectly safe to force
824 * us to go around the wait_event() loop again, so the warning
825 * is a false-positive. Silence the warning by resetting
826 * thread state
827 */
828 __set_current_state(TASK_RUNNING);
829 blk_start_plug(&plug);
830 flush_bio_list(conf, bio);
831 blk_finish_plug(&plug);
832 } else
833 spin_unlock_irq(&conf->device_lock);
834}
835
836/* Barriers....
837 * Sometimes we need to suspend IO while we do something else,
838 * either some resync/recovery, or reconfigure the array.
839 * To do this we raise a 'barrier'.
840 * The 'barrier' is a counter that can be raised multiple times
841 * to count how many activities are happening which preclude
842 * normal IO.
843 * We can only raise the barrier if there is no pending IO.
844 * i.e. if nr_pending == 0.
845 * We choose only to raise the barrier if no-one is waiting for the
846 * barrier to go down. This means that as soon as an IO request
847 * is ready, no other operations which require a barrier will start
848 * until the IO request has had a chance.
849 *
850 * So: regular IO calls 'wait_barrier'. When that returns there
851 * is no backgroup IO happening, It must arrange to call
852 * allow_barrier when it has finished its IO.
853 * backgroup IO calls must call raise_barrier. Once that returns
854 * there is no normal IO happeing. It must arrange to call
855 * lower_barrier when the particular background IO completes.
856 */
857static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
858{
859 int idx = sector_to_idx(sector_nr);
860
861 spin_lock_irq(&conf->resync_lock);
862
863 /* Wait until no block IO is waiting */
864 wait_event_lock_irq(conf->wait_barrier,
865 !atomic_read(&conf->nr_waiting[idx]),
866 conf->resync_lock);
867
868 /* block any new IO from starting */
869 atomic_inc(&conf->barrier[idx]);
870 /*
871 * In raise_barrier() we firstly increase conf->barrier[idx] then
872 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
873 * increase conf->nr_pending[idx] then check conf->barrier[idx].
874 * A memory barrier here to make sure conf->nr_pending[idx] won't
875 * be fetched before conf->barrier[idx] is increased. Otherwise
876 * there will be a race between raise_barrier() and _wait_barrier().
877 */
878 smp_mb__after_atomic();
879
880 /* For these conditions we must wait:
881 * A: while the array is in frozen state
882 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
883 * existing in corresponding I/O barrier bucket.
884 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
885 * max resync count which allowed on current I/O barrier bucket.
886 */
887 wait_event_lock_irq(conf->wait_barrier,
888 (!conf->array_frozen &&
889 !atomic_read(&conf->nr_pending[idx]) &&
890 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
891 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
892 conf->resync_lock);
893
894 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
895 atomic_dec(&conf->barrier[idx]);
896 spin_unlock_irq(&conf->resync_lock);
897 wake_up(&conf->wait_barrier);
898 return -EINTR;
899 }
900
901 atomic_inc(&conf->nr_sync_pending);
902 spin_unlock_irq(&conf->resync_lock);
903
904 return 0;
905}
906
907static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
908{
909 int idx = sector_to_idx(sector_nr);
910
911 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
912
913 atomic_dec(&conf->barrier[idx]);
914 atomic_dec(&conf->nr_sync_pending);
915 wake_up(&conf->wait_barrier);
916}
917
918static void _wait_barrier(struct r1conf *conf, int idx)
919{
920 /*
921 * We need to increase conf->nr_pending[idx] very early here,
922 * then raise_barrier() can be blocked when it waits for
923 * conf->nr_pending[idx] to be 0. Then we can avoid holding
924 * conf->resync_lock when there is no barrier raised in same
925 * barrier unit bucket. Also if the array is frozen, I/O
926 * should be blocked until array is unfrozen.
927 */
928 atomic_inc(&conf->nr_pending[idx]);
929 /*
930 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
931 * check conf->barrier[idx]. In raise_barrier() we firstly increase
932 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
933 * barrier is necessary here to make sure conf->barrier[idx] won't be
934 * fetched before conf->nr_pending[idx] is increased. Otherwise there
935 * will be a race between _wait_barrier() and raise_barrier().
936 */
937 smp_mb__after_atomic();
938
939 /*
940 * Don't worry about checking two atomic_t variables at same time
941 * here. If during we check conf->barrier[idx], the array is
942 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
943 * 0, it is safe to return and make the I/O continue. Because the
944 * array is frozen, all I/O returned here will eventually complete
945 * or be queued, no race will happen. See code comment in
946 * frozen_array().
947 */
948 if (!READ_ONCE(conf->array_frozen) &&
949 !atomic_read(&conf->barrier[idx]))
950 return;
951
952 /*
953 * After holding conf->resync_lock, conf->nr_pending[idx]
954 * should be decreased before waiting for barrier to drop.
955 * Otherwise, we may encounter a race condition because
956 * raise_barrer() might be waiting for conf->nr_pending[idx]
957 * to be 0 at same time.
958 */
959 spin_lock_irq(&conf->resync_lock);
960 atomic_inc(&conf->nr_waiting[idx]);
961 atomic_dec(&conf->nr_pending[idx]);
962 /*
963 * In case freeze_array() is waiting for
964 * get_unqueued_pending() == extra
965 */
966 wake_up(&conf->wait_barrier);
967 /* Wait for the barrier in same barrier unit bucket to drop. */
968 wait_event_lock_irq(conf->wait_barrier,
969 !conf->array_frozen &&
970 !atomic_read(&conf->barrier[idx]),
971 conf->resync_lock);
972 atomic_inc(&conf->nr_pending[idx]);
973 atomic_dec(&conf->nr_waiting[idx]);
974 spin_unlock_irq(&conf->resync_lock);
975}
976
977static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
978{
979 int idx = sector_to_idx(sector_nr);
980
981 /*
982 * Very similar to _wait_barrier(). The difference is, for read
983 * I/O we don't need wait for sync I/O, but if the whole array
984 * is frozen, the read I/O still has to wait until the array is
985 * unfrozen. Since there is no ordering requirement with
986 * conf->barrier[idx] here, memory barrier is unnecessary as well.
987 */
988 atomic_inc(&conf->nr_pending[idx]);
989
990 if (!READ_ONCE(conf->array_frozen))
991 return;
992
993 spin_lock_irq(&conf->resync_lock);
994 atomic_inc(&conf->nr_waiting[idx]);
995 atomic_dec(&conf->nr_pending[idx]);
996 /*
997 * In case freeze_array() is waiting for
998 * get_unqueued_pending() == extra
999 */
1000 wake_up(&conf->wait_barrier);
1001 /* Wait for array to be unfrozen */
1002 wait_event_lock_irq(conf->wait_barrier,
1003 !conf->array_frozen,
1004 conf->resync_lock);
1005 atomic_inc(&conf->nr_pending[idx]);
1006 atomic_dec(&conf->nr_waiting[idx]);
1007 spin_unlock_irq(&conf->resync_lock);
1008}
1009
1010static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1011{
1012 int idx = sector_to_idx(sector_nr);
1013
1014 _wait_barrier(conf, idx);
1015}
1016
1017static void _allow_barrier(struct r1conf *conf, int idx)
1018{
1019 atomic_dec(&conf->nr_pending[idx]);
1020 wake_up(&conf->wait_barrier);
1021}
1022
1023static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1024{
1025 int idx = sector_to_idx(sector_nr);
1026
1027 _allow_barrier(conf, idx);
1028}
1029
1030/* conf->resync_lock should be held */
1031static int get_unqueued_pending(struct r1conf *conf)
1032{
1033 int idx, ret;
1034
1035 ret = atomic_read(&conf->nr_sync_pending);
1036 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1037 ret += atomic_read(&conf->nr_pending[idx]) -
1038 atomic_read(&conf->nr_queued[idx]);
1039
1040 return ret;
1041}
1042
1043static void freeze_array(struct r1conf *conf, int extra)
1044{
1045 /* Stop sync I/O and normal I/O and wait for everything to
1046 * go quiet.
1047 * This is called in two situations:
1048 * 1) management command handlers (reshape, remove disk, quiesce).
1049 * 2) one normal I/O request failed.
1050
1051 * After array_frozen is set to 1, new sync IO will be blocked at
1052 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1053 * or wait_read_barrier(). The flying I/Os will either complete or be
1054 * queued. When everything goes quite, there are only queued I/Os left.
1055
1056 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1057 * barrier bucket index which this I/O request hits. When all sync and
1058 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1059 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1060 * in handle_read_error(), we may call freeze_array() before trying to
1061 * fix the read error. In this case, the error read I/O is not queued,
1062 * so get_unqueued_pending() == 1.
1063 *
1064 * Therefore before this function returns, we need to wait until
1065 * get_unqueued_pendings(conf) gets equal to extra. For
1066 * normal I/O context, extra is 1, in rested situations extra is 0.
1067 */
1068 spin_lock_irq(&conf->resync_lock);
1069 conf->array_frozen = 1;
1070 raid1_log(conf->mddev, "wait freeze");
1071 wait_event_lock_irq_cmd(
1072 conf->wait_barrier,
1073 get_unqueued_pending(conf) == extra,
1074 conf->resync_lock,
1075 flush_pending_writes(conf));
1076 spin_unlock_irq(&conf->resync_lock);
1077}
1078static void unfreeze_array(struct r1conf *conf)
1079{
1080 /* reverse the effect of the freeze */
1081 spin_lock_irq(&conf->resync_lock);
1082 conf->array_frozen = 0;
1083 spin_unlock_irq(&conf->resync_lock);
1084 wake_up(&conf->wait_barrier);
1085}
1086
1087static void alloc_behind_master_bio(struct r1bio *r1_bio,
1088 struct bio *bio)
1089{
1090 int size = bio->bi_iter.bi_size;
1091 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1092 int i = 0;
1093 struct bio *behind_bio = NULL;
1094
1095 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1096 if (!behind_bio)
1097 return;
1098
1099 /* discard op, we don't support writezero/writesame yet */
1100 if (!bio_has_data(bio)) {
1101 behind_bio->bi_iter.bi_size = size;
1102 goto skip_copy;
1103 }
1104
1105 behind_bio->bi_write_hint = bio->bi_write_hint;
1106
1107 while (i < vcnt && size) {
1108 struct page *page;
1109 int len = min_t(int, PAGE_SIZE, size);
1110
1111 page = alloc_page(GFP_NOIO);
1112 if (unlikely(!page))
1113 goto free_pages;
1114
1115 bio_add_page(behind_bio, page, len, 0);
1116
1117 size -= len;
1118 i++;
1119 }
1120
1121 bio_copy_data(behind_bio, bio);
1122skip_copy:
1123 r1_bio->behind_master_bio = behind_bio;
1124 set_bit(R1BIO_BehindIO, &r1_bio->state);
1125
1126 return;
1127
1128free_pages:
1129 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1130 bio->bi_iter.bi_size);
1131 bio_free_pages(behind_bio);
1132 bio_put(behind_bio);
1133}
1134
1135struct raid1_plug_cb {
1136 struct blk_plug_cb cb;
1137 struct bio_list pending;
1138 int pending_cnt;
1139};
1140
1141static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1142{
1143 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1144 cb);
1145 struct mddev *mddev = plug->cb.data;
1146 struct r1conf *conf = mddev->private;
1147 struct bio *bio;
1148
1149 if (from_schedule || current->bio_list) {
1150 spin_lock_irq(&conf->device_lock);
1151 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1152 conf->pending_count += plug->pending_cnt;
1153 spin_unlock_irq(&conf->device_lock);
1154 wake_up(&conf->wait_barrier);
1155 md_wakeup_thread(mddev->thread);
1156 kfree(plug);
1157 return;
1158 }
1159
1160 /* we aren't scheduling, so we can do the write-out directly. */
1161 bio = bio_list_get(&plug->pending);
1162 flush_bio_list(conf, bio);
1163 kfree(plug);
1164}
1165
1166static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1167{
1168 r1_bio->master_bio = bio;
1169 r1_bio->sectors = bio_sectors(bio);
1170 r1_bio->state = 0;
1171 r1_bio->mddev = mddev;
1172 r1_bio->sector = bio->bi_iter.bi_sector;
1173}
1174
1175static inline struct r1bio *
1176alloc_r1bio(struct mddev *mddev, struct bio *bio)
1177{
1178 struct r1conf *conf = mddev->private;
1179 struct r1bio *r1_bio;
1180
1181 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1182 /* Ensure no bio records IO_BLOCKED */
1183 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1184 init_r1bio(r1_bio, mddev, bio);
1185 return r1_bio;
1186}
1187
1188static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1189 int max_read_sectors, struct r1bio *r1_bio)
1190{
1191 struct r1conf *conf = mddev->private;
1192 struct raid1_info *mirror;
1193 struct bio *read_bio;
1194 struct bitmap *bitmap = mddev->bitmap;
1195 const int op = bio_op(bio);
1196 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1197 int max_sectors;
1198 int rdisk;
1199 bool print_msg = !!r1_bio;
1200 char b[BDEVNAME_SIZE];
1201
1202 /*
1203 * If r1_bio is set, we are blocking the raid1d thread
1204 * so there is a tiny risk of deadlock. So ask for
1205 * emergency memory if needed.
1206 */
1207 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1208
1209 if (print_msg) {
1210 /* Need to get the block device name carefully */
1211 struct md_rdev *rdev;
1212 rcu_read_lock();
1213 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1214 if (rdev)
1215 bdevname(rdev->bdev, b);
1216 else
1217 strcpy(b, "???");
1218 rcu_read_unlock();
1219 }
1220
1221 /*
1222 * Still need barrier for READ in case that whole
1223 * array is frozen.
1224 */
1225 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1226
1227 if (!r1_bio)
1228 r1_bio = alloc_r1bio(mddev, bio);
1229 else
1230 init_r1bio(r1_bio, mddev, bio);
1231 r1_bio->sectors = max_read_sectors;
1232
1233 /*
1234 * make_request() can abort the operation when read-ahead is being
1235 * used and no empty request is available.
1236 */
1237 rdisk = read_balance(conf, r1_bio, &max_sectors);
1238
1239 if (rdisk < 0) {
1240 /* couldn't find anywhere to read from */
1241 if (print_msg) {
1242 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1243 mdname(mddev),
1244 b,
1245 (unsigned long long)r1_bio->sector);
1246 }
1247 raid_end_bio_io(r1_bio);
1248 return;
1249 }
1250 mirror = conf->mirrors + rdisk;
1251
1252 if (print_msg)
1253 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1254 mdname(mddev),
1255 (unsigned long long)r1_bio->sector,
1256 bdevname(mirror->rdev->bdev, b));
1257
1258 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1259 bitmap) {
1260 /*
1261 * Reading from a write-mostly device must take care not to
1262 * over-take any writes that are 'behind'
1263 */
1264 raid1_log(mddev, "wait behind writes");
1265 wait_event(bitmap->behind_wait,
1266 atomic_read(&bitmap->behind_writes) == 0);
1267 }
1268
1269 if (max_sectors < bio_sectors(bio)) {
1270 struct bio *split = bio_split(bio, max_sectors,
1271 gfp, conf->bio_split);
1272 bio_chain(split, bio);
1273 generic_make_request(bio);
1274 bio = split;
1275 r1_bio->master_bio = bio;
1276 r1_bio->sectors = max_sectors;
1277 }
1278
1279 r1_bio->read_disk = rdisk;
1280
1281 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1282
1283 r1_bio->bios[rdisk] = read_bio;
1284
1285 read_bio->bi_iter.bi_sector = r1_bio->sector +
1286 mirror->rdev->data_offset;
1287 bio_set_dev(read_bio, mirror->rdev->bdev);
1288 read_bio->bi_end_io = raid1_end_read_request;
1289 bio_set_op_attrs(read_bio, op, do_sync);
1290 if (test_bit(FailFast, &mirror->rdev->flags) &&
1291 test_bit(R1BIO_FailFast, &r1_bio->state))
1292 read_bio->bi_opf |= MD_FAILFAST;
1293 read_bio->bi_private = r1_bio;
1294
1295 if (mddev->gendisk)
1296 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1297 disk_devt(mddev->gendisk), r1_bio->sector);
1298
1299 generic_make_request(read_bio);
1300}
1301
1302static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1303 int max_write_sectors)
1304{
1305 struct r1conf *conf = mddev->private;
1306 struct r1bio *r1_bio;
1307 int i, disks;
1308 struct bitmap *bitmap = mddev->bitmap;
1309 unsigned long flags;
1310 struct md_rdev *blocked_rdev;
1311 struct blk_plug_cb *cb;
1312 struct raid1_plug_cb *plug = NULL;
1313 int first_clone;
1314 int max_sectors;
1315
1316 if (mddev_is_clustered(mddev) &&
1317 md_cluster_ops->area_resyncing(mddev, WRITE,
1318 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1319
1320 DEFINE_WAIT(w);
1321 for (;;) {
1322 prepare_to_wait(&conf->wait_barrier,
1323 &w, TASK_IDLE);
1324 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325 bio->bi_iter.bi_sector,
1326 bio_end_sector(bio)))
1327 break;
1328 schedule();
1329 }
1330 finish_wait(&conf->wait_barrier, &w);
1331 }
1332
1333 /*
1334 * Register the new request and wait if the reconstruction
1335 * thread has put up a bar for new requests.
1336 * Continue immediately if no resync is active currently.
1337 */
1338 wait_barrier(conf, bio->bi_iter.bi_sector);
1339
1340 r1_bio = alloc_r1bio(mddev, bio);
1341 r1_bio->sectors = max_write_sectors;
1342
1343 if (conf->pending_count >= max_queued_requests) {
1344 md_wakeup_thread(mddev->thread);
1345 raid1_log(mddev, "wait queued");
1346 wait_event(conf->wait_barrier,
1347 conf->pending_count < max_queued_requests);
1348 }
1349 /* first select target devices under rcu_lock and
1350 * inc refcount on their rdev. Record them by setting
1351 * bios[x] to bio
1352 * If there are known/acknowledged bad blocks on any device on
1353 * which we have seen a write error, we want to avoid writing those
1354 * blocks.
1355 * This potentially requires several writes to write around
1356 * the bad blocks. Each set of writes gets it's own r1bio
1357 * with a set of bios attached.
1358 */
1359
1360 disks = conf->raid_disks * 2;
1361 retry_write:
1362 blocked_rdev = NULL;
1363 rcu_read_lock();
1364 max_sectors = r1_bio->sectors;
1365 for (i = 0; i < disks; i++) {
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1367 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1368 atomic_inc(&rdev->nr_pending);
1369 blocked_rdev = rdev;
1370 break;
1371 }
1372 r1_bio->bios[i] = NULL;
1373 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1374 if (i < conf->raid_disks)
1375 set_bit(R1BIO_Degraded, &r1_bio->state);
1376 continue;
1377 }
1378
1379 atomic_inc(&rdev->nr_pending);
1380 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1381 sector_t first_bad;
1382 int bad_sectors;
1383 int is_bad;
1384
1385 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1386 &first_bad, &bad_sectors);
1387 if (is_bad < 0) {
1388 /* mustn't write here until the bad block is
1389 * acknowledged*/
1390 set_bit(BlockedBadBlocks, &rdev->flags);
1391 blocked_rdev = rdev;
1392 break;
1393 }
1394 if (is_bad && first_bad <= r1_bio->sector) {
1395 /* Cannot write here at all */
1396 bad_sectors -= (r1_bio->sector - first_bad);
1397 if (bad_sectors < max_sectors)
1398 /* mustn't write more than bad_sectors
1399 * to other devices yet
1400 */
1401 max_sectors = bad_sectors;
1402 rdev_dec_pending(rdev, mddev);
1403 /* We don't set R1BIO_Degraded as that
1404 * only applies if the disk is
1405 * missing, so it might be re-added,
1406 * and we want to know to recover this
1407 * chunk.
1408 * In this case the device is here,
1409 * and the fact that this chunk is not
1410 * in-sync is recorded in the bad
1411 * block log
1412 */
1413 continue;
1414 }
1415 if (is_bad) {
1416 int good_sectors = first_bad - r1_bio->sector;
1417 if (good_sectors < max_sectors)
1418 max_sectors = good_sectors;
1419 }
1420 }
1421 r1_bio->bios[i] = bio;
1422 }
1423 rcu_read_unlock();
1424
1425 if (unlikely(blocked_rdev)) {
1426 /* Wait for this device to become unblocked */
1427 int j;
1428
1429 for (j = 0; j < i; j++)
1430 if (r1_bio->bios[j])
1431 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1432 r1_bio->state = 0;
1433 allow_barrier(conf, bio->bi_iter.bi_sector);
1434 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1435 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1436 wait_barrier(conf, bio->bi_iter.bi_sector);
1437 goto retry_write;
1438 }
1439
1440 if (max_sectors < bio_sectors(bio)) {
1441 struct bio *split = bio_split(bio, max_sectors,
1442 GFP_NOIO, conf->bio_split);
1443 bio_chain(split, bio);
1444 generic_make_request(bio);
1445 bio = split;
1446 r1_bio->master_bio = bio;
1447 r1_bio->sectors = max_sectors;
1448 }
1449
1450 atomic_set(&r1_bio->remaining, 1);
1451 atomic_set(&r1_bio->behind_remaining, 0);
1452
1453 first_clone = 1;
1454
1455 for (i = 0; i < disks; i++) {
1456 struct bio *mbio = NULL;
1457 if (!r1_bio->bios[i])
1458 continue;
1459
1460
1461 if (first_clone) {
1462 /* do behind I/O ?
1463 * Not if there are too many, or cannot
1464 * allocate memory, or a reader on WriteMostly
1465 * is waiting for behind writes to flush */
1466 if (bitmap &&
1467 (atomic_read(&bitmap->behind_writes)
1468 < mddev->bitmap_info.max_write_behind) &&
1469 !waitqueue_active(&bitmap->behind_wait)) {
1470 alloc_behind_master_bio(r1_bio, bio);
1471 }
1472
1473 bitmap_startwrite(bitmap, r1_bio->sector,
1474 r1_bio->sectors,
1475 test_bit(R1BIO_BehindIO,
1476 &r1_bio->state));
1477 first_clone = 0;
1478 }
1479
1480 if (r1_bio->behind_master_bio)
1481 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1482 GFP_NOIO, mddev->bio_set);
1483 else
1484 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1485
1486 if (r1_bio->behind_master_bio) {
1487 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1488 atomic_inc(&r1_bio->behind_remaining);
1489 }
1490
1491 r1_bio->bios[i] = mbio;
1492
1493 mbio->bi_iter.bi_sector = (r1_bio->sector +
1494 conf->mirrors[i].rdev->data_offset);
1495 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1496 mbio->bi_end_io = raid1_end_write_request;
1497 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1498 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1499 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1500 conf->raid_disks - mddev->degraded > 1)
1501 mbio->bi_opf |= MD_FAILFAST;
1502 mbio->bi_private = r1_bio;
1503
1504 atomic_inc(&r1_bio->remaining);
1505
1506 if (mddev->gendisk)
1507 trace_block_bio_remap(mbio->bi_disk->queue,
1508 mbio, disk_devt(mddev->gendisk),
1509 r1_bio->sector);
1510 /* flush_pending_writes() needs access to the rdev so...*/
1511 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1512
1513 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1514 if (cb)
1515 plug = container_of(cb, struct raid1_plug_cb, cb);
1516 else
1517 plug = NULL;
1518 if (plug) {
1519 bio_list_add(&plug->pending, mbio);
1520 plug->pending_cnt++;
1521 } else {
1522 spin_lock_irqsave(&conf->device_lock, flags);
1523 bio_list_add(&conf->pending_bio_list, mbio);
1524 conf->pending_count++;
1525 spin_unlock_irqrestore(&conf->device_lock, flags);
1526 md_wakeup_thread(mddev->thread);
1527 }
1528 }
1529
1530 r1_bio_write_done(r1_bio);
1531
1532 /* In case raid1d snuck in to freeze_array */
1533 wake_up(&conf->wait_barrier);
1534}
1535
1536static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1537{
1538 sector_t sectors;
1539
1540 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1541 md_flush_request(mddev, bio);
1542 return true;
1543 }
1544
1545 /*
1546 * There is a limit to the maximum size, but
1547 * the read/write handler might find a lower limit
1548 * due to bad blocks. To avoid multiple splits,
1549 * we pass the maximum number of sectors down
1550 * and let the lower level perform the split.
1551 */
1552 sectors = align_to_barrier_unit_end(
1553 bio->bi_iter.bi_sector, bio_sectors(bio));
1554
1555 if (bio_data_dir(bio) == READ)
1556 raid1_read_request(mddev, bio, sectors, NULL);
1557 else {
1558 if (!md_write_start(mddev,bio))
1559 return false;
1560 raid1_write_request(mddev, bio, sectors);
1561 }
1562 return true;
1563}
1564
1565static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1566{
1567 struct r1conf *conf = mddev->private;
1568 int i;
1569
1570 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1571 conf->raid_disks - mddev->degraded);
1572 rcu_read_lock();
1573 for (i = 0; i < conf->raid_disks; i++) {
1574 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1575 seq_printf(seq, "%s",
1576 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1577 }
1578 rcu_read_unlock();
1579 seq_printf(seq, "]");
1580}
1581
1582static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1583{
1584 char b[BDEVNAME_SIZE];
1585 struct r1conf *conf = mddev->private;
1586 unsigned long flags;
1587
1588 /*
1589 * If it is not operational, then we have already marked it as dead
1590 * else if it is the last working disks, ignore the error, let the
1591 * next level up know.
1592 * else mark the drive as failed
1593 */
1594 spin_lock_irqsave(&conf->device_lock, flags);
1595 if (test_bit(In_sync, &rdev->flags)
1596 && (conf->raid_disks - mddev->degraded) == 1) {
1597 /*
1598 * Don't fail the drive, act as though we were just a
1599 * normal single drive.
1600 * However don't try a recovery from this drive as
1601 * it is very likely to fail.
1602 */
1603 conf->recovery_disabled = mddev->recovery_disabled;
1604 spin_unlock_irqrestore(&conf->device_lock, flags);
1605 return;
1606 }
1607 set_bit(Blocked, &rdev->flags);
1608 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1609 mddev->degraded++;
1610 set_bit(Faulty, &rdev->flags);
1611 } else
1612 set_bit(Faulty, &rdev->flags);
1613 spin_unlock_irqrestore(&conf->device_lock, flags);
1614 /*
1615 * if recovery is running, make sure it aborts.
1616 */
1617 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1618 set_mask_bits(&mddev->sb_flags, 0,
1619 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1620 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1621 "md/raid1:%s: Operation continuing on %d devices.\n",
1622 mdname(mddev), bdevname(rdev->bdev, b),
1623 mdname(mddev), conf->raid_disks - mddev->degraded);
1624}
1625
1626static void print_conf(struct r1conf *conf)
1627{
1628 int i;
1629
1630 pr_debug("RAID1 conf printout:\n");
1631 if (!conf) {
1632 pr_debug("(!conf)\n");
1633 return;
1634 }
1635 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1636 conf->raid_disks);
1637
1638 rcu_read_lock();
1639 for (i = 0; i < conf->raid_disks; i++) {
1640 char b[BDEVNAME_SIZE];
1641 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1642 if (rdev)
1643 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1644 i, !test_bit(In_sync, &rdev->flags),
1645 !test_bit(Faulty, &rdev->flags),
1646 bdevname(rdev->bdev,b));
1647 }
1648 rcu_read_unlock();
1649}
1650
1651static void close_sync(struct r1conf *conf)
1652{
1653 int idx;
1654
1655 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1656 _wait_barrier(conf, idx);
1657 _allow_barrier(conf, idx);
1658 }
1659
1660 mempool_destroy(conf->r1buf_pool);
1661 conf->r1buf_pool = NULL;
1662}
1663
1664static int raid1_spare_active(struct mddev *mddev)
1665{
1666 int i;
1667 struct r1conf *conf = mddev->private;
1668 int count = 0;
1669 unsigned long flags;
1670
1671 /*
1672 * Find all failed disks within the RAID1 configuration
1673 * and mark them readable.
1674 * Called under mddev lock, so rcu protection not needed.
1675 * device_lock used to avoid races with raid1_end_read_request
1676 * which expects 'In_sync' flags and ->degraded to be consistent.
1677 */
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 for (i = 0; i < conf->raid_disks; i++) {
1680 struct md_rdev *rdev = conf->mirrors[i].rdev;
1681 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1682 if (repl
1683 && !test_bit(Candidate, &repl->flags)
1684 && repl->recovery_offset == MaxSector
1685 && !test_bit(Faulty, &repl->flags)
1686 && !test_and_set_bit(In_sync, &repl->flags)) {
1687 /* replacement has just become active */
1688 if (!rdev ||
1689 !test_and_clear_bit(In_sync, &rdev->flags))
1690 count++;
1691 if (rdev) {
1692 /* Replaced device not technically
1693 * faulty, but we need to be sure
1694 * it gets removed and never re-added
1695 */
1696 set_bit(Faulty, &rdev->flags);
1697 sysfs_notify_dirent_safe(
1698 rdev->sysfs_state);
1699 }
1700 }
1701 if (rdev
1702 && rdev->recovery_offset == MaxSector
1703 && !test_bit(Faulty, &rdev->flags)
1704 && !test_and_set_bit(In_sync, &rdev->flags)) {
1705 count++;
1706 sysfs_notify_dirent_safe(rdev->sysfs_state);
1707 }
1708 }
1709 mddev->degraded -= count;
1710 spin_unlock_irqrestore(&conf->device_lock, flags);
1711
1712 print_conf(conf);
1713 return count;
1714}
1715
1716static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1717{
1718 struct r1conf *conf = mddev->private;
1719 int err = -EEXIST;
1720 int mirror = 0;
1721 struct raid1_info *p;
1722 int first = 0;
1723 int last = conf->raid_disks - 1;
1724
1725 if (mddev->recovery_disabled == conf->recovery_disabled)
1726 return -EBUSY;
1727
1728 if (md_integrity_add_rdev(rdev, mddev))
1729 return -ENXIO;
1730
1731 if (rdev->raid_disk >= 0)
1732 first = last = rdev->raid_disk;
1733
1734 /*
1735 * find the disk ... but prefer rdev->saved_raid_disk
1736 * if possible.
1737 */
1738 if (rdev->saved_raid_disk >= 0 &&
1739 rdev->saved_raid_disk >= first &&
1740 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1741 first = last = rdev->saved_raid_disk;
1742
1743 for (mirror = first; mirror <= last; mirror++) {
1744 p = conf->mirrors+mirror;
1745 if (!p->rdev) {
1746
1747 if (mddev->gendisk)
1748 disk_stack_limits(mddev->gendisk, rdev->bdev,
1749 rdev->data_offset << 9);
1750
1751 p->head_position = 0;
1752 rdev->raid_disk = mirror;
1753 err = 0;
1754 /* As all devices are equivalent, we don't need a full recovery
1755 * if this was recently any drive of the array
1756 */
1757 if (rdev->saved_raid_disk < 0)
1758 conf->fullsync = 1;
1759 rcu_assign_pointer(p->rdev, rdev);
1760 break;
1761 }
1762 if (test_bit(WantReplacement, &p->rdev->flags) &&
1763 p[conf->raid_disks].rdev == NULL) {
1764 /* Add this device as a replacement */
1765 clear_bit(In_sync, &rdev->flags);
1766 set_bit(Replacement, &rdev->flags);
1767 rdev->raid_disk = mirror;
1768 err = 0;
1769 conf->fullsync = 1;
1770 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1771 break;
1772 }
1773 }
1774 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1775 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1776 print_conf(conf);
1777 return err;
1778}
1779
1780static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1781{
1782 struct r1conf *conf = mddev->private;
1783 int err = 0;
1784 int number = rdev->raid_disk;
1785 struct raid1_info *p = conf->mirrors + number;
1786
1787 if (rdev != p->rdev)
1788 p = conf->mirrors + conf->raid_disks + number;
1789
1790 print_conf(conf);
1791 if (rdev == p->rdev) {
1792 if (test_bit(In_sync, &rdev->flags) ||
1793 atomic_read(&rdev->nr_pending)) {
1794 err = -EBUSY;
1795 goto abort;
1796 }
1797 /* Only remove non-faulty devices if recovery
1798 * is not possible.
1799 */
1800 if (!test_bit(Faulty, &rdev->flags) &&
1801 mddev->recovery_disabled != conf->recovery_disabled &&
1802 mddev->degraded < conf->raid_disks) {
1803 err = -EBUSY;
1804 goto abort;
1805 }
1806 p->rdev = NULL;
1807 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1808 synchronize_rcu();
1809 if (atomic_read(&rdev->nr_pending)) {
1810 /* lost the race, try later */
1811 err = -EBUSY;
1812 p->rdev = rdev;
1813 goto abort;
1814 }
1815 }
1816 if (conf->mirrors[conf->raid_disks + number].rdev) {
1817 /* We just removed a device that is being replaced.
1818 * Move down the replacement. We drain all IO before
1819 * doing this to avoid confusion.
1820 */
1821 struct md_rdev *repl =
1822 conf->mirrors[conf->raid_disks + number].rdev;
1823 freeze_array(conf, 0);
1824 if (atomic_read(&repl->nr_pending)) {
1825 /* It means that some queued IO of retry_list
1826 * hold repl. Thus, we cannot set replacement
1827 * as NULL, avoiding rdev NULL pointer
1828 * dereference in sync_request_write and
1829 * handle_write_finished.
1830 */
1831 err = -EBUSY;
1832 unfreeze_array(conf);
1833 goto abort;
1834 }
1835 clear_bit(Replacement, &repl->flags);
1836 p->rdev = repl;
1837 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1838 unfreeze_array(conf);
1839 }
1840
1841 clear_bit(WantReplacement, &rdev->flags);
1842 err = md_integrity_register(mddev);
1843 }
1844abort:
1845
1846 print_conf(conf);
1847 return err;
1848}
1849
1850static void end_sync_read(struct bio *bio)
1851{
1852 struct r1bio *r1_bio = get_resync_r1bio(bio);
1853
1854 update_head_pos(r1_bio->read_disk, r1_bio);
1855
1856 /*
1857 * we have read a block, now it needs to be re-written,
1858 * or re-read if the read failed.
1859 * We don't do much here, just schedule handling by raid1d
1860 */
1861 if (!bio->bi_status)
1862 set_bit(R1BIO_Uptodate, &r1_bio->state);
1863
1864 if (atomic_dec_and_test(&r1_bio->remaining))
1865 reschedule_retry(r1_bio);
1866}
1867
1868static void end_sync_write(struct bio *bio)
1869{
1870 int uptodate = !bio->bi_status;
1871 struct r1bio *r1_bio = get_resync_r1bio(bio);
1872 struct mddev *mddev = r1_bio->mddev;
1873 struct r1conf *conf = mddev->private;
1874 sector_t first_bad;
1875 int bad_sectors;
1876 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1877
1878 if (!uptodate) {
1879 sector_t sync_blocks = 0;
1880 sector_t s = r1_bio->sector;
1881 long sectors_to_go = r1_bio->sectors;
1882 /* make sure these bits doesn't get cleared. */
1883 do {
1884 bitmap_end_sync(mddev->bitmap, s,
1885 &sync_blocks, 1);
1886 s += sync_blocks;
1887 sectors_to_go -= sync_blocks;
1888 } while (sectors_to_go > 0);
1889 set_bit(WriteErrorSeen, &rdev->flags);
1890 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1891 set_bit(MD_RECOVERY_NEEDED, &
1892 mddev->recovery);
1893 set_bit(R1BIO_WriteError, &r1_bio->state);
1894 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1895 &first_bad, &bad_sectors) &&
1896 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1897 r1_bio->sector,
1898 r1_bio->sectors,
1899 &first_bad, &bad_sectors)
1900 )
1901 set_bit(R1BIO_MadeGood, &r1_bio->state);
1902
1903 if (atomic_dec_and_test(&r1_bio->remaining)) {
1904 int s = r1_bio->sectors;
1905 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1906 test_bit(R1BIO_WriteError, &r1_bio->state))
1907 reschedule_retry(r1_bio);
1908 else {
1909 put_buf(r1_bio);
1910 md_done_sync(mddev, s, uptodate);
1911 }
1912 }
1913}
1914
1915static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1916 int sectors, struct page *page, int rw)
1917{
1918 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1919 /* success */
1920 return 1;
1921 if (rw == WRITE) {
1922 set_bit(WriteErrorSeen, &rdev->flags);
1923 if (!test_and_set_bit(WantReplacement,
1924 &rdev->flags))
1925 set_bit(MD_RECOVERY_NEEDED, &
1926 rdev->mddev->recovery);
1927 }
1928 /* need to record an error - either for the block or the device */
1929 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1930 md_error(rdev->mddev, rdev);
1931 return 0;
1932}
1933
1934static int fix_sync_read_error(struct r1bio *r1_bio)
1935{
1936 /* Try some synchronous reads of other devices to get
1937 * good data, much like with normal read errors. Only
1938 * read into the pages we already have so we don't
1939 * need to re-issue the read request.
1940 * We don't need to freeze the array, because being in an
1941 * active sync request, there is no normal IO, and
1942 * no overlapping syncs.
1943 * We don't need to check is_badblock() again as we
1944 * made sure that anything with a bad block in range
1945 * will have bi_end_io clear.
1946 */
1947 struct mddev *mddev = r1_bio->mddev;
1948 struct r1conf *conf = mddev->private;
1949 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1950 struct page **pages = get_resync_pages(bio)->pages;
1951 sector_t sect = r1_bio->sector;
1952 int sectors = r1_bio->sectors;
1953 int idx = 0;
1954 struct md_rdev *rdev;
1955
1956 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1957 if (test_bit(FailFast, &rdev->flags)) {
1958 /* Don't try recovering from here - just fail it
1959 * ... unless it is the last working device of course */
1960 md_error(mddev, rdev);
1961 if (test_bit(Faulty, &rdev->flags))
1962 /* Don't try to read from here, but make sure
1963 * put_buf does it's thing
1964 */
1965 bio->bi_end_io = end_sync_write;
1966 }
1967
1968 while(sectors) {
1969 int s = sectors;
1970 int d = r1_bio->read_disk;
1971 int success = 0;
1972 int start;
1973
1974 if (s > (PAGE_SIZE>>9))
1975 s = PAGE_SIZE >> 9;
1976 do {
1977 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1978 /* No rcu protection needed here devices
1979 * can only be removed when no resync is
1980 * active, and resync is currently active
1981 */
1982 rdev = conf->mirrors[d].rdev;
1983 if (sync_page_io(rdev, sect, s<<9,
1984 pages[idx],
1985 REQ_OP_READ, 0, false)) {
1986 success = 1;
1987 break;
1988 }
1989 }
1990 d++;
1991 if (d == conf->raid_disks * 2)
1992 d = 0;
1993 } while (!success && d != r1_bio->read_disk);
1994
1995 if (!success) {
1996 char b[BDEVNAME_SIZE];
1997 int abort = 0;
1998 /* Cannot read from anywhere, this block is lost.
1999 * Record a bad block on each device. If that doesn't
2000 * work just disable and interrupt the recovery.
2001 * Don't fail devices as that won't really help.
2002 */
2003 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2004 mdname(mddev), bio_devname(bio, b),
2005 (unsigned long long)r1_bio->sector);
2006 for (d = 0; d < conf->raid_disks * 2; d++) {
2007 rdev = conf->mirrors[d].rdev;
2008 if (!rdev || test_bit(Faulty, &rdev->flags))
2009 continue;
2010 if (!rdev_set_badblocks(rdev, sect, s, 0))
2011 abort = 1;
2012 }
2013 if (abort) {
2014 conf->recovery_disabled =
2015 mddev->recovery_disabled;
2016 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2017 md_done_sync(mddev, r1_bio->sectors, 0);
2018 put_buf(r1_bio);
2019 return 0;
2020 }
2021 /* Try next page */
2022 sectors -= s;
2023 sect += s;
2024 idx++;
2025 continue;
2026 }
2027
2028 start = d;
2029 /* write it back and re-read */
2030 while (d != r1_bio->read_disk) {
2031 if (d == 0)
2032 d = conf->raid_disks * 2;
2033 d--;
2034 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2035 continue;
2036 rdev = conf->mirrors[d].rdev;
2037 if (r1_sync_page_io(rdev, sect, s,
2038 pages[idx],
2039 WRITE) == 0) {
2040 r1_bio->bios[d]->bi_end_io = NULL;
2041 rdev_dec_pending(rdev, mddev);
2042 }
2043 }
2044 d = start;
2045 while (d != r1_bio->read_disk) {
2046 if (d == 0)
2047 d = conf->raid_disks * 2;
2048 d--;
2049 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2050 continue;
2051 rdev = conf->mirrors[d].rdev;
2052 if (r1_sync_page_io(rdev, sect, s,
2053 pages[idx],
2054 READ) != 0)
2055 atomic_add(s, &rdev->corrected_errors);
2056 }
2057 sectors -= s;
2058 sect += s;
2059 idx ++;
2060 }
2061 set_bit(R1BIO_Uptodate, &r1_bio->state);
2062 bio->bi_status = 0;
2063 return 1;
2064}
2065
2066static void process_checks(struct r1bio *r1_bio)
2067{
2068 /* We have read all readable devices. If we haven't
2069 * got the block, then there is no hope left.
2070 * If we have, then we want to do a comparison
2071 * and skip the write if everything is the same.
2072 * If any blocks failed to read, then we need to
2073 * attempt an over-write
2074 */
2075 struct mddev *mddev = r1_bio->mddev;
2076 struct r1conf *conf = mddev->private;
2077 int primary;
2078 int i;
2079 int vcnt;
2080
2081 /* Fix variable parts of all bios */
2082 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2083 for (i = 0; i < conf->raid_disks * 2; i++) {
2084 blk_status_t status;
2085 struct bio *b = r1_bio->bios[i];
2086 struct resync_pages *rp = get_resync_pages(b);
2087 if (b->bi_end_io != end_sync_read)
2088 continue;
2089 /* fixup the bio for reuse, but preserve errno */
2090 status = b->bi_status;
2091 bio_reset(b);
2092 b->bi_status = status;
2093 b->bi_iter.bi_sector = r1_bio->sector +
2094 conf->mirrors[i].rdev->data_offset;
2095 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2096 b->bi_end_io = end_sync_read;
2097 rp->raid_bio = r1_bio;
2098 b->bi_private = rp;
2099
2100 /* initialize bvec table again */
2101 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2102 }
2103 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2104 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2105 !r1_bio->bios[primary]->bi_status) {
2106 r1_bio->bios[primary]->bi_end_io = NULL;
2107 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2108 break;
2109 }
2110 r1_bio->read_disk = primary;
2111 for (i = 0; i < conf->raid_disks * 2; i++) {
2112 int j;
2113 struct bio *pbio = r1_bio->bios[primary];
2114 struct bio *sbio = r1_bio->bios[i];
2115 blk_status_t status = sbio->bi_status;
2116 struct page **ppages = get_resync_pages(pbio)->pages;
2117 struct page **spages = get_resync_pages(sbio)->pages;
2118 struct bio_vec *bi;
2119 int page_len[RESYNC_PAGES] = { 0 };
2120
2121 if (sbio->bi_end_io != end_sync_read)
2122 continue;
2123 /* Now we can 'fixup' the error value */
2124 sbio->bi_status = 0;
2125
2126 bio_for_each_segment_all(bi, sbio, j)
2127 page_len[j] = bi->bv_len;
2128
2129 if (!status) {
2130 for (j = vcnt; j-- ; ) {
2131 if (memcmp(page_address(ppages[j]),
2132 page_address(spages[j]),
2133 page_len[j]))
2134 break;
2135 }
2136 } else
2137 j = 0;
2138 if (j >= 0)
2139 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2140 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2141 && !status)) {
2142 /* No need to write to this device. */
2143 sbio->bi_end_io = NULL;
2144 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2145 continue;
2146 }
2147
2148 bio_copy_data(sbio, pbio);
2149 }
2150}
2151
2152static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2153{
2154 struct r1conf *conf = mddev->private;
2155 int i;
2156 int disks = conf->raid_disks * 2;
2157 struct bio *wbio;
2158
2159 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2160 /* ouch - failed to read all of that. */
2161 if (!fix_sync_read_error(r1_bio))
2162 return;
2163
2164 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2165 process_checks(r1_bio);
2166
2167 /*
2168 * schedule writes
2169 */
2170 atomic_set(&r1_bio->remaining, 1);
2171 for (i = 0; i < disks ; i++) {
2172 wbio = r1_bio->bios[i];
2173 if (wbio->bi_end_io == NULL ||
2174 (wbio->bi_end_io == end_sync_read &&
2175 (i == r1_bio->read_disk ||
2176 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2177 continue;
2178 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2179 continue;
2180
2181 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2182 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2183 wbio->bi_opf |= MD_FAILFAST;
2184
2185 wbio->bi_end_io = end_sync_write;
2186 atomic_inc(&r1_bio->remaining);
2187 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2188
2189 generic_make_request(wbio);
2190 }
2191
2192 if (atomic_dec_and_test(&r1_bio->remaining)) {
2193 /* if we're here, all write(s) have completed, so clean up */
2194 int s = r1_bio->sectors;
2195 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2196 test_bit(R1BIO_WriteError, &r1_bio->state))
2197 reschedule_retry(r1_bio);
2198 else {
2199 put_buf(r1_bio);
2200 md_done_sync(mddev, s, 1);
2201 }
2202 }
2203}
2204
2205/*
2206 * This is a kernel thread which:
2207 *
2208 * 1. Retries failed read operations on working mirrors.
2209 * 2. Updates the raid superblock when problems encounter.
2210 * 3. Performs writes following reads for array synchronising.
2211 */
2212
2213static void fix_read_error(struct r1conf *conf, int read_disk,
2214 sector_t sect, int sectors)
2215{
2216 struct mddev *mddev = conf->mddev;
2217 while(sectors) {
2218 int s = sectors;
2219 int d = read_disk;
2220 int success = 0;
2221 int start;
2222 struct md_rdev *rdev;
2223
2224 if (s > (PAGE_SIZE>>9))
2225 s = PAGE_SIZE >> 9;
2226
2227 do {
2228 sector_t first_bad;
2229 int bad_sectors;
2230
2231 rcu_read_lock();
2232 rdev = rcu_dereference(conf->mirrors[d].rdev);
2233 if (rdev &&
2234 (test_bit(In_sync, &rdev->flags) ||
2235 (!test_bit(Faulty, &rdev->flags) &&
2236 rdev->recovery_offset >= sect + s)) &&
2237 is_badblock(rdev, sect, s,
2238 &first_bad, &bad_sectors) == 0) {
2239 atomic_inc(&rdev->nr_pending);
2240 rcu_read_unlock();
2241 if (sync_page_io(rdev, sect, s<<9,
2242 conf->tmppage, REQ_OP_READ, 0, false))
2243 success = 1;
2244 rdev_dec_pending(rdev, mddev);
2245 if (success)
2246 break;
2247 } else
2248 rcu_read_unlock();
2249 d++;
2250 if (d == conf->raid_disks * 2)
2251 d = 0;
2252 } while (!success && d != read_disk);
2253
2254 if (!success) {
2255 /* Cannot read from anywhere - mark it bad */
2256 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2257 if (!rdev_set_badblocks(rdev, sect, s, 0))
2258 md_error(mddev, rdev);
2259 break;
2260 }
2261 /* write it back and re-read */
2262 start = d;
2263 while (d != read_disk) {
2264 if (d==0)
2265 d = conf->raid_disks * 2;
2266 d--;
2267 rcu_read_lock();
2268 rdev = rcu_dereference(conf->mirrors[d].rdev);
2269 if (rdev &&
2270 !test_bit(Faulty, &rdev->flags)) {
2271 atomic_inc(&rdev->nr_pending);
2272 rcu_read_unlock();
2273 r1_sync_page_io(rdev, sect, s,
2274 conf->tmppage, WRITE);
2275 rdev_dec_pending(rdev, mddev);
2276 } else
2277 rcu_read_unlock();
2278 }
2279 d = start;
2280 while (d != read_disk) {
2281 char b[BDEVNAME_SIZE];
2282 if (d==0)
2283 d = conf->raid_disks * 2;
2284 d--;
2285 rcu_read_lock();
2286 rdev = rcu_dereference(conf->mirrors[d].rdev);
2287 if (rdev &&
2288 !test_bit(Faulty, &rdev->flags)) {
2289 atomic_inc(&rdev->nr_pending);
2290 rcu_read_unlock();
2291 if (r1_sync_page_io(rdev, sect, s,
2292 conf->tmppage, READ)) {
2293 atomic_add(s, &rdev->corrected_errors);
2294 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2295 mdname(mddev), s,
2296 (unsigned long long)(sect +
2297 rdev->data_offset),
2298 bdevname(rdev->bdev, b));
2299 }
2300 rdev_dec_pending(rdev, mddev);
2301 } else
2302 rcu_read_unlock();
2303 }
2304 sectors -= s;
2305 sect += s;
2306 }
2307}
2308
2309static int narrow_write_error(struct r1bio *r1_bio, int i)
2310{
2311 struct mddev *mddev = r1_bio->mddev;
2312 struct r1conf *conf = mddev->private;
2313 struct md_rdev *rdev = conf->mirrors[i].rdev;
2314
2315 /* bio has the data to be written to device 'i' where
2316 * we just recently had a write error.
2317 * We repeatedly clone the bio and trim down to one block,
2318 * then try the write. Where the write fails we record
2319 * a bad block.
2320 * It is conceivable that the bio doesn't exactly align with
2321 * blocks. We must handle this somehow.
2322 *
2323 * We currently own a reference on the rdev.
2324 */
2325
2326 int block_sectors;
2327 sector_t sector;
2328 int sectors;
2329 int sect_to_write = r1_bio->sectors;
2330 int ok = 1;
2331
2332 if (rdev->badblocks.shift < 0)
2333 return 0;
2334
2335 block_sectors = roundup(1 << rdev->badblocks.shift,
2336 bdev_logical_block_size(rdev->bdev) >> 9);
2337 sector = r1_bio->sector;
2338 sectors = ((sector + block_sectors)
2339 & ~(sector_t)(block_sectors - 1))
2340 - sector;
2341
2342 while (sect_to_write) {
2343 struct bio *wbio;
2344 if (sectors > sect_to_write)
2345 sectors = sect_to_write;
2346 /* Write at 'sector' for 'sectors'*/
2347
2348 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2349 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2350 GFP_NOIO,
2351 mddev->bio_set);
2352 } else {
2353 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2354 mddev->bio_set);
2355 }
2356
2357 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2358 wbio->bi_iter.bi_sector = r1_bio->sector;
2359 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2360
2361 bio_trim(wbio, sector - r1_bio->sector, sectors);
2362 wbio->bi_iter.bi_sector += rdev->data_offset;
2363 bio_set_dev(wbio, rdev->bdev);
2364
2365 if (submit_bio_wait(wbio) < 0)
2366 /* failure! */
2367 ok = rdev_set_badblocks(rdev, sector,
2368 sectors, 0)
2369 && ok;
2370
2371 bio_put(wbio);
2372 sect_to_write -= sectors;
2373 sector += sectors;
2374 sectors = block_sectors;
2375 }
2376 return ok;
2377}
2378
2379static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2380{
2381 int m;
2382 int s = r1_bio->sectors;
2383 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2384 struct md_rdev *rdev = conf->mirrors[m].rdev;
2385 struct bio *bio = r1_bio->bios[m];
2386 if (bio->bi_end_io == NULL)
2387 continue;
2388 if (!bio->bi_status &&
2389 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2390 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2391 }
2392 if (bio->bi_status &&
2393 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2394 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2395 md_error(conf->mddev, rdev);
2396 }
2397 }
2398 put_buf(r1_bio);
2399 md_done_sync(conf->mddev, s, 1);
2400}
2401
2402static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2403{
2404 int m, idx;
2405 bool fail = false;
2406
2407 for (m = 0; m < conf->raid_disks * 2 ; m++)
2408 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2409 struct md_rdev *rdev = conf->mirrors[m].rdev;
2410 rdev_clear_badblocks(rdev,
2411 r1_bio->sector,
2412 r1_bio->sectors, 0);
2413 rdev_dec_pending(rdev, conf->mddev);
2414 } else if (r1_bio->bios[m] != NULL) {
2415 /* This drive got a write error. We need to
2416 * narrow down and record precise write
2417 * errors.
2418 */
2419 fail = true;
2420 if (!narrow_write_error(r1_bio, m)) {
2421 md_error(conf->mddev,
2422 conf->mirrors[m].rdev);
2423 /* an I/O failed, we can't clear the bitmap */
2424 set_bit(R1BIO_Degraded, &r1_bio->state);
2425 }
2426 rdev_dec_pending(conf->mirrors[m].rdev,
2427 conf->mddev);
2428 }
2429 if (fail) {
2430 spin_lock_irq(&conf->device_lock);
2431 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2432 idx = sector_to_idx(r1_bio->sector);
2433 atomic_inc(&conf->nr_queued[idx]);
2434 spin_unlock_irq(&conf->device_lock);
2435 /*
2436 * In case freeze_array() is waiting for condition
2437 * get_unqueued_pending() == extra to be true.
2438 */
2439 wake_up(&conf->wait_barrier);
2440 md_wakeup_thread(conf->mddev->thread);
2441 } else {
2442 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2443 close_write(r1_bio);
2444 raid_end_bio_io(r1_bio);
2445 }
2446}
2447
2448static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2449{
2450 struct mddev *mddev = conf->mddev;
2451 struct bio *bio;
2452 struct md_rdev *rdev;
2453 sector_t bio_sector;
2454
2455 clear_bit(R1BIO_ReadError, &r1_bio->state);
2456 /* we got a read error. Maybe the drive is bad. Maybe just
2457 * the block and we can fix it.
2458 * We freeze all other IO, and try reading the block from
2459 * other devices. When we find one, we re-write
2460 * and check it that fixes the read error.
2461 * This is all done synchronously while the array is
2462 * frozen
2463 */
2464
2465 bio = r1_bio->bios[r1_bio->read_disk];
2466 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2467 bio_put(bio);
2468 r1_bio->bios[r1_bio->read_disk] = NULL;
2469
2470 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2471 if (mddev->ro == 0
2472 && !test_bit(FailFast, &rdev->flags)) {
2473 freeze_array(conf, 1);
2474 fix_read_error(conf, r1_bio->read_disk,
2475 r1_bio->sector, r1_bio->sectors);
2476 unfreeze_array(conf);
2477 } else {
2478 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2479 }
2480
2481 rdev_dec_pending(rdev, conf->mddev);
2482 allow_barrier(conf, r1_bio->sector);
2483 bio = r1_bio->master_bio;
2484
2485 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2486 r1_bio->state = 0;
2487 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2488}
2489
2490static void raid1d(struct md_thread *thread)
2491{
2492 struct mddev *mddev = thread->mddev;
2493 struct r1bio *r1_bio;
2494 unsigned long flags;
2495 struct r1conf *conf = mddev->private;
2496 struct list_head *head = &conf->retry_list;
2497 struct blk_plug plug;
2498 int idx;
2499
2500 md_check_recovery(mddev);
2501
2502 if (!list_empty_careful(&conf->bio_end_io_list) &&
2503 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2504 LIST_HEAD(tmp);
2505 spin_lock_irqsave(&conf->device_lock, flags);
2506 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2507 list_splice_init(&conf->bio_end_io_list, &tmp);
2508 spin_unlock_irqrestore(&conf->device_lock, flags);
2509 while (!list_empty(&tmp)) {
2510 r1_bio = list_first_entry(&tmp, struct r1bio,
2511 retry_list);
2512 list_del(&r1_bio->retry_list);
2513 idx = sector_to_idx(r1_bio->sector);
2514 atomic_dec(&conf->nr_queued[idx]);
2515 if (mddev->degraded)
2516 set_bit(R1BIO_Degraded, &r1_bio->state);
2517 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2518 close_write(r1_bio);
2519 raid_end_bio_io(r1_bio);
2520 }
2521 }
2522
2523 blk_start_plug(&plug);
2524 for (;;) {
2525
2526 flush_pending_writes(conf);
2527
2528 spin_lock_irqsave(&conf->device_lock, flags);
2529 if (list_empty(head)) {
2530 spin_unlock_irqrestore(&conf->device_lock, flags);
2531 break;
2532 }
2533 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2534 list_del(head->prev);
2535 idx = sector_to_idx(r1_bio->sector);
2536 atomic_dec(&conf->nr_queued[idx]);
2537 spin_unlock_irqrestore(&conf->device_lock, flags);
2538
2539 mddev = r1_bio->mddev;
2540 conf = mddev->private;
2541 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2542 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2543 test_bit(R1BIO_WriteError, &r1_bio->state))
2544 handle_sync_write_finished(conf, r1_bio);
2545 else
2546 sync_request_write(mddev, r1_bio);
2547 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2548 test_bit(R1BIO_WriteError, &r1_bio->state))
2549 handle_write_finished(conf, r1_bio);
2550 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2551 handle_read_error(conf, r1_bio);
2552 else
2553 WARN_ON_ONCE(1);
2554
2555 cond_resched();
2556 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2557 md_check_recovery(mddev);
2558 }
2559 blk_finish_plug(&plug);
2560}
2561
2562static int init_resync(struct r1conf *conf)
2563{
2564 int buffs;
2565
2566 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2567 BUG_ON(conf->r1buf_pool);
2568 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2569 conf->poolinfo);
2570 if (!conf->r1buf_pool)
2571 return -ENOMEM;
2572 return 0;
2573}
2574
2575static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2576{
2577 struct r1bio *r1bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2578 struct resync_pages *rps;
2579 struct bio *bio;
2580 int i;
2581
2582 for (i = conf->poolinfo->raid_disks; i--; ) {
2583 bio = r1bio->bios[i];
2584 rps = bio->bi_private;
2585 bio_reset(bio);
2586 bio->bi_private = rps;
2587 }
2588 r1bio->master_bio = NULL;
2589 return r1bio;
2590}
2591
2592/*
2593 * perform a "sync" on one "block"
2594 *
2595 * We need to make sure that no normal I/O request - particularly write
2596 * requests - conflict with active sync requests.
2597 *
2598 * This is achieved by tracking pending requests and a 'barrier' concept
2599 * that can be installed to exclude normal IO requests.
2600 */
2601
2602static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2603 int *skipped)
2604{
2605 struct r1conf *conf = mddev->private;
2606 struct r1bio *r1_bio;
2607 struct bio *bio;
2608 sector_t max_sector, nr_sectors;
2609 int disk = -1;
2610 int i;
2611 int wonly = -1;
2612 int write_targets = 0, read_targets = 0;
2613 sector_t sync_blocks;
2614 int still_degraded = 0;
2615 int good_sectors = RESYNC_SECTORS;
2616 int min_bad = 0; /* number of sectors that are bad in all devices */
2617 int idx = sector_to_idx(sector_nr);
2618 int page_idx = 0;
2619
2620 if (!conf->r1buf_pool)
2621 if (init_resync(conf))
2622 return 0;
2623
2624 max_sector = mddev->dev_sectors;
2625 if (sector_nr >= max_sector) {
2626 /* If we aborted, we need to abort the
2627 * sync on the 'current' bitmap chunk (there will
2628 * only be one in raid1 resync.
2629 * We can find the current addess in mddev->curr_resync
2630 */
2631 if (mddev->curr_resync < max_sector) /* aborted */
2632 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2633 &sync_blocks, 1);
2634 else /* completed sync */
2635 conf->fullsync = 0;
2636
2637 bitmap_close_sync(mddev->bitmap);
2638 close_sync(conf);
2639
2640 if (mddev_is_clustered(mddev)) {
2641 conf->cluster_sync_low = 0;
2642 conf->cluster_sync_high = 0;
2643 }
2644 return 0;
2645 }
2646
2647 if (mddev->bitmap == NULL &&
2648 mddev->recovery_cp == MaxSector &&
2649 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2650 conf->fullsync == 0) {
2651 *skipped = 1;
2652 return max_sector - sector_nr;
2653 }
2654 /* before building a request, check if we can skip these blocks..
2655 * This call the bitmap_start_sync doesn't actually record anything
2656 */
2657 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2658 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2659 /* We can skip this block, and probably several more */
2660 *skipped = 1;
2661 return sync_blocks;
2662 }
2663
2664 /*
2665 * If there is non-resync activity waiting for a turn, then let it
2666 * though before starting on this new sync request.
2667 */
2668 if (atomic_read(&conf->nr_waiting[idx]))
2669 schedule_timeout_uninterruptible(1);
2670
2671 /* we are incrementing sector_nr below. To be safe, we check against
2672 * sector_nr + two times RESYNC_SECTORS
2673 */
2674
2675 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2676 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2677
2678
2679 if (raise_barrier(conf, sector_nr))
2680 return 0;
2681
2682 r1_bio = raid1_alloc_init_r1buf(conf);
2683
2684 rcu_read_lock();
2685 /*
2686 * If we get a correctably read error during resync or recovery,
2687 * we might want to read from a different device. So we
2688 * flag all drives that could conceivably be read from for READ,
2689 * and any others (which will be non-In_sync devices) for WRITE.
2690 * If a read fails, we try reading from something else for which READ
2691 * is OK.
2692 */
2693
2694 r1_bio->mddev = mddev;
2695 r1_bio->sector = sector_nr;
2696 r1_bio->state = 0;
2697 set_bit(R1BIO_IsSync, &r1_bio->state);
2698 /* make sure good_sectors won't go across barrier unit boundary */
2699 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2700
2701 for (i = 0; i < conf->raid_disks * 2; i++) {
2702 struct md_rdev *rdev;
2703 bio = r1_bio->bios[i];
2704
2705 rdev = rcu_dereference(conf->mirrors[i].rdev);
2706 if (rdev == NULL ||
2707 test_bit(Faulty, &rdev->flags)) {
2708 if (i < conf->raid_disks)
2709 still_degraded = 1;
2710 } else if (!test_bit(In_sync, &rdev->flags)) {
2711 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2712 bio->bi_end_io = end_sync_write;
2713 write_targets ++;
2714 } else {
2715 /* may need to read from here */
2716 sector_t first_bad = MaxSector;
2717 int bad_sectors;
2718
2719 if (is_badblock(rdev, sector_nr, good_sectors,
2720 &first_bad, &bad_sectors)) {
2721 if (first_bad > sector_nr)
2722 good_sectors = first_bad - sector_nr;
2723 else {
2724 bad_sectors -= (sector_nr - first_bad);
2725 if (min_bad == 0 ||
2726 min_bad > bad_sectors)
2727 min_bad = bad_sectors;
2728 }
2729 }
2730 if (sector_nr < first_bad) {
2731 if (test_bit(WriteMostly, &rdev->flags)) {
2732 if (wonly < 0)
2733 wonly = i;
2734 } else {
2735 if (disk < 0)
2736 disk = i;
2737 }
2738 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2739 bio->bi_end_io = end_sync_read;
2740 read_targets++;
2741 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2742 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2743 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2744 /*
2745 * The device is suitable for reading (InSync),
2746 * but has bad block(s) here. Let's try to correct them,
2747 * if we are doing resync or repair. Otherwise, leave
2748 * this device alone for this sync request.
2749 */
2750 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2751 bio->bi_end_io = end_sync_write;
2752 write_targets++;
2753 }
2754 }
2755 if (bio->bi_end_io) {
2756 atomic_inc(&rdev->nr_pending);
2757 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2758 bio_set_dev(bio, rdev->bdev);
2759 if (test_bit(FailFast, &rdev->flags))
2760 bio->bi_opf |= MD_FAILFAST;
2761 }
2762 }
2763 rcu_read_unlock();
2764 if (disk < 0)
2765 disk = wonly;
2766 r1_bio->read_disk = disk;
2767
2768 if (read_targets == 0 && min_bad > 0) {
2769 /* These sectors are bad on all InSync devices, so we
2770 * need to mark them bad on all write targets
2771 */
2772 int ok = 1;
2773 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2774 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2775 struct md_rdev *rdev = conf->mirrors[i].rdev;
2776 ok = rdev_set_badblocks(rdev, sector_nr,
2777 min_bad, 0
2778 ) && ok;
2779 }
2780 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2781 *skipped = 1;
2782 put_buf(r1_bio);
2783
2784 if (!ok) {
2785 /* Cannot record the badblocks, so need to
2786 * abort the resync.
2787 * If there are multiple read targets, could just
2788 * fail the really bad ones ???
2789 */
2790 conf->recovery_disabled = mddev->recovery_disabled;
2791 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2792 return 0;
2793 } else
2794 return min_bad;
2795
2796 }
2797 if (min_bad > 0 && min_bad < good_sectors) {
2798 /* only resync enough to reach the next bad->good
2799 * transition */
2800 good_sectors = min_bad;
2801 }
2802
2803 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2804 /* extra read targets are also write targets */
2805 write_targets += read_targets-1;
2806
2807 if (write_targets == 0 || read_targets == 0) {
2808 /* There is nowhere to write, so all non-sync
2809 * drives must be failed - so we are finished
2810 */
2811 sector_t rv;
2812 if (min_bad > 0)
2813 max_sector = sector_nr + min_bad;
2814 rv = max_sector - sector_nr;
2815 *skipped = 1;
2816 put_buf(r1_bio);
2817 return rv;
2818 }
2819
2820 if (max_sector > mddev->resync_max)
2821 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2822 if (max_sector > sector_nr + good_sectors)
2823 max_sector = sector_nr + good_sectors;
2824 nr_sectors = 0;
2825 sync_blocks = 0;
2826 do {
2827 struct page *page;
2828 int len = PAGE_SIZE;
2829 if (sector_nr + (len>>9) > max_sector)
2830 len = (max_sector - sector_nr) << 9;
2831 if (len == 0)
2832 break;
2833 if (sync_blocks == 0) {
2834 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2835 &sync_blocks, still_degraded) &&
2836 !conf->fullsync &&
2837 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2838 break;
2839 if ((len >> 9) > sync_blocks)
2840 len = sync_blocks<<9;
2841 }
2842
2843 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2844 struct resync_pages *rp;
2845
2846 bio = r1_bio->bios[i];
2847 rp = get_resync_pages(bio);
2848 if (bio->bi_end_io) {
2849 page = resync_fetch_page(rp, page_idx);
2850
2851 /*
2852 * won't fail because the vec table is big
2853 * enough to hold all these pages
2854 */
2855 bio_add_page(bio, page, len, 0);
2856 }
2857 }
2858 nr_sectors += len>>9;
2859 sector_nr += len>>9;
2860 sync_blocks -= (len>>9);
2861 } while (++page_idx < RESYNC_PAGES);
2862
2863 r1_bio->sectors = nr_sectors;
2864
2865 if (mddev_is_clustered(mddev) &&
2866 conf->cluster_sync_high < sector_nr + nr_sectors) {
2867 conf->cluster_sync_low = mddev->curr_resync_completed;
2868 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2869 /* Send resync message */
2870 md_cluster_ops->resync_info_update(mddev,
2871 conf->cluster_sync_low,
2872 conf->cluster_sync_high);
2873 }
2874
2875 /* For a user-requested sync, we read all readable devices and do a
2876 * compare
2877 */
2878 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2879 atomic_set(&r1_bio->remaining, read_targets);
2880 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2881 bio = r1_bio->bios[i];
2882 if (bio->bi_end_io == end_sync_read) {
2883 read_targets--;
2884 md_sync_acct_bio(bio, nr_sectors);
2885 if (read_targets == 1)
2886 bio->bi_opf &= ~MD_FAILFAST;
2887 generic_make_request(bio);
2888 }
2889 }
2890 } else {
2891 atomic_set(&r1_bio->remaining, 1);
2892 bio = r1_bio->bios[r1_bio->read_disk];
2893 md_sync_acct_bio(bio, nr_sectors);
2894 if (read_targets == 1)
2895 bio->bi_opf &= ~MD_FAILFAST;
2896 generic_make_request(bio);
2897
2898 }
2899 return nr_sectors;
2900}
2901
2902static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2903{
2904 if (sectors)
2905 return sectors;
2906
2907 return mddev->dev_sectors;
2908}
2909
2910static struct r1conf *setup_conf(struct mddev *mddev)
2911{
2912 struct r1conf *conf;
2913 int i;
2914 struct raid1_info *disk;
2915 struct md_rdev *rdev;
2916 int err = -ENOMEM;
2917
2918 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2919 if (!conf)
2920 goto abort;
2921
2922 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2923 sizeof(atomic_t), GFP_KERNEL);
2924 if (!conf->nr_pending)
2925 goto abort;
2926
2927 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2928 sizeof(atomic_t), GFP_KERNEL);
2929 if (!conf->nr_waiting)
2930 goto abort;
2931
2932 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2933 sizeof(atomic_t), GFP_KERNEL);
2934 if (!conf->nr_queued)
2935 goto abort;
2936
2937 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2938 sizeof(atomic_t), GFP_KERNEL);
2939 if (!conf->barrier)
2940 goto abort;
2941
2942 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2943 * mddev->raid_disks * 2,
2944 GFP_KERNEL);
2945 if (!conf->mirrors)
2946 goto abort;
2947
2948 conf->tmppage = alloc_page(GFP_KERNEL);
2949 if (!conf->tmppage)
2950 goto abort;
2951
2952 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2953 if (!conf->poolinfo)
2954 goto abort;
2955 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2956 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2957 r1bio_pool_free,
2958 conf->poolinfo);
2959 if (!conf->r1bio_pool)
2960 goto abort;
2961
2962 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2963 if (!conf->bio_split)
2964 goto abort;
2965
2966 conf->poolinfo->mddev = mddev;
2967
2968 err = -EINVAL;
2969 spin_lock_init(&conf->device_lock);
2970 rdev_for_each(rdev, mddev) {
2971 int disk_idx = rdev->raid_disk;
2972 if (disk_idx >= mddev->raid_disks
2973 || disk_idx < 0)
2974 continue;
2975 if (test_bit(Replacement, &rdev->flags))
2976 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2977 else
2978 disk = conf->mirrors + disk_idx;
2979
2980 if (disk->rdev)
2981 goto abort;
2982 disk->rdev = rdev;
2983 disk->head_position = 0;
2984 disk->seq_start = MaxSector;
2985 }
2986 conf->raid_disks = mddev->raid_disks;
2987 conf->mddev = mddev;
2988 INIT_LIST_HEAD(&conf->retry_list);
2989 INIT_LIST_HEAD(&conf->bio_end_io_list);
2990
2991 spin_lock_init(&conf->resync_lock);
2992 init_waitqueue_head(&conf->wait_barrier);
2993
2994 bio_list_init(&conf->pending_bio_list);
2995 conf->pending_count = 0;
2996 conf->recovery_disabled = mddev->recovery_disabled - 1;
2997
2998 err = -EIO;
2999 for (i = 0; i < conf->raid_disks * 2; i++) {
3000
3001 disk = conf->mirrors + i;
3002
3003 if (i < conf->raid_disks &&
3004 disk[conf->raid_disks].rdev) {
3005 /* This slot has a replacement. */
3006 if (!disk->rdev) {
3007 /* No original, just make the replacement
3008 * a recovering spare
3009 */
3010 disk->rdev =
3011 disk[conf->raid_disks].rdev;
3012 disk[conf->raid_disks].rdev = NULL;
3013 } else if (!test_bit(In_sync, &disk->rdev->flags))
3014 /* Original is not in_sync - bad */
3015 goto abort;
3016 }
3017
3018 if (!disk->rdev ||
3019 !test_bit(In_sync, &disk->rdev->flags)) {
3020 disk->head_position = 0;
3021 if (disk->rdev &&
3022 (disk->rdev->saved_raid_disk < 0))
3023 conf->fullsync = 1;
3024 }
3025 }
3026
3027 err = -ENOMEM;
3028 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3029 if (!conf->thread)
3030 goto abort;
3031
3032 return conf;
3033
3034 abort:
3035 if (conf) {
3036 mempool_destroy(conf->r1bio_pool);
3037 kfree(conf->mirrors);
3038 safe_put_page(conf->tmppage);
3039 kfree(conf->poolinfo);
3040 kfree(conf->nr_pending);
3041 kfree(conf->nr_waiting);
3042 kfree(conf->nr_queued);
3043 kfree(conf->barrier);
3044 if (conf->bio_split)
3045 bioset_free(conf->bio_split);
3046 kfree(conf);
3047 }
3048 return ERR_PTR(err);
3049}
3050
3051static void raid1_free(struct mddev *mddev, void *priv);
3052static int raid1_run(struct mddev *mddev)
3053{
3054 struct r1conf *conf;
3055 int i;
3056 struct md_rdev *rdev;
3057 int ret;
3058 bool discard_supported = false;
3059
3060 if (mddev->level != 1) {
3061 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3062 mdname(mddev), mddev->level);
3063 return -EIO;
3064 }
3065 if (mddev->reshape_position != MaxSector) {
3066 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3067 mdname(mddev));
3068 return -EIO;
3069 }
3070 if (mddev_init_writes_pending(mddev) < 0)
3071 return -ENOMEM;
3072 /*
3073 * copy the already verified devices into our private RAID1
3074 * bookkeeping area. [whatever we allocate in run(),
3075 * should be freed in raid1_free()]
3076 */
3077 if (mddev->private == NULL)
3078 conf = setup_conf(mddev);
3079 else
3080 conf = mddev->private;
3081
3082 if (IS_ERR(conf))
3083 return PTR_ERR(conf);
3084
3085 if (mddev->queue) {
3086 blk_queue_max_write_same_sectors(mddev->queue, 0);
3087 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3088 }
3089
3090 rdev_for_each(rdev, mddev) {
3091 if (!mddev->gendisk)
3092 continue;
3093 disk_stack_limits(mddev->gendisk, rdev->bdev,
3094 rdev->data_offset << 9);
3095 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3096 discard_supported = true;
3097 }
3098
3099 mddev->degraded = 0;
3100 for (i=0; i < conf->raid_disks; i++)
3101 if (conf->mirrors[i].rdev == NULL ||
3102 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3103 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3104 mddev->degraded++;
3105
3106 if (conf->raid_disks - mddev->degraded == 1)
3107 mddev->recovery_cp = MaxSector;
3108
3109 if (mddev->recovery_cp != MaxSector)
3110 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3111 mdname(mddev));
3112 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3113 mdname(mddev), mddev->raid_disks - mddev->degraded,
3114 mddev->raid_disks);
3115
3116 /*
3117 * Ok, everything is just fine now
3118 */
3119 mddev->thread = conf->thread;
3120 conf->thread = NULL;
3121 mddev->private = conf;
3122 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3123
3124 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3125
3126 if (mddev->queue) {
3127 if (discard_supported)
3128 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3129 mddev->queue);
3130 else
3131 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3132 mddev->queue);
3133 }
3134
3135 ret = md_integrity_register(mddev);
3136 if (ret) {
3137 md_unregister_thread(&mddev->thread);
3138 raid1_free(mddev, conf);
3139 }
3140 return ret;
3141}
3142
3143static void raid1_free(struct mddev *mddev, void *priv)
3144{
3145 struct r1conf *conf = priv;
3146
3147 mempool_destroy(conf->r1bio_pool);
3148 kfree(conf->mirrors);
3149 safe_put_page(conf->tmppage);
3150 kfree(conf->poolinfo);
3151 kfree(conf->nr_pending);
3152 kfree(conf->nr_waiting);
3153 kfree(conf->nr_queued);
3154 kfree(conf->barrier);
3155 if (conf->bio_split)
3156 bioset_free(conf->bio_split);
3157 kfree(conf);
3158}
3159
3160static int raid1_resize(struct mddev *mddev, sector_t sectors)
3161{
3162 /* no resync is happening, and there is enough space
3163 * on all devices, so we can resize.
3164 * We need to make sure resync covers any new space.
3165 * If the array is shrinking we should possibly wait until
3166 * any io in the removed space completes, but it hardly seems
3167 * worth it.
3168 */
3169 sector_t newsize = raid1_size(mddev, sectors, 0);
3170 if (mddev->external_size &&
3171 mddev->array_sectors > newsize)
3172 return -EINVAL;
3173 if (mddev->bitmap) {
3174 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3175 if (ret)
3176 return ret;
3177 }
3178 md_set_array_sectors(mddev, newsize);
3179 if (sectors > mddev->dev_sectors &&
3180 mddev->recovery_cp > mddev->dev_sectors) {
3181 mddev->recovery_cp = mddev->dev_sectors;
3182 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3183 }
3184 mddev->dev_sectors = sectors;
3185 mddev->resync_max_sectors = sectors;
3186 return 0;
3187}
3188
3189static int raid1_reshape(struct mddev *mddev)
3190{
3191 /* We need to:
3192 * 1/ resize the r1bio_pool
3193 * 2/ resize conf->mirrors
3194 *
3195 * We allocate a new r1bio_pool if we can.
3196 * Then raise a device barrier and wait until all IO stops.
3197 * Then resize conf->mirrors and swap in the new r1bio pool.
3198 *
3199 * At the same time, we "pack" the devices so that all the missing
3200 * devices have the higher raid_disk numbers.
3201 */
3202 mempool_t *newpool, *oldpool;
3203 struct pool_info *newpoolinfo;
3204 struct raid1_info *newmirrors;
3205 struct r1conf *conf = mddev->private;
3206 int cnt, raid_disks;
3207 unsigned long flags;
3208 int d, d2;
3209
3210 /* Cannot change chunk_size, layout, or level */
3211 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3212 mddev->layout != mddev->new_layout ||
3213 mddev->level != mddev->new_level) {
3214 mddev->new_chunk_sectors = mddev->chunk_sectors;
3215 mddev->new_layout = mddev->layout;
3216 mddev->new_level = mddev->level;
3217 return -EINVAL;
3218 }
3219
3220 if (!mddev_is_clustered(mddev))
3221 md_allow_write(mddev);
3222
3223 raid_disks = mddev->raid_disks + mddev->delta_disks;
3224
3225 if (raid_disks < conf->raid_disks) {
3226 cnt=0;
3227 for (d= 0; d < conf->raid_disks; d++)
3228 if (conf->mirrors[d].rdev)
3229 cnt++;
3230 if (cnt > raid_disks)
3231 return -EBUSY;
3232 }
3233
3234 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3235 if (!newpoolinfo)
3236 return -ENOMEM;
3237 newpoolinfo->mddev = mddev;
3238 newpoolinfo->raid_disks = raid_disks * 2;
3239
3240 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3241 r1bio_pool_free, newpoolinfo);
3242 if (!newpool) {
3243 kfree(newpoolinfo);
3244 return -ENOMEM;
3245 }
3246 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3247 GFP_KERNEL);
3248 if (!newmirrors) {
3249 kfree(newpoolinfo);
3250 mempool_destroy(newpool);
3251 return -ENOMEM;
3252 }
3253
3254 freeze_array(conf, 0);
3255
3256 /* ok, everything is stopped */
3257 oldpool = conf->r1bio_pool;
3258 conf->r1bio_pool = newpool;
3259
3260 for (d = d2 = 0; d < conf->raid_disks; d++) {
3261 struct md_rdev *rdev = conf->mirrors[d].rdev;
3262 if (rdev && rdev->raid_disk != d2) {
3263 sysfs_unlink_rdev(mddev, rdev);
3264 rdev->raid_disk = d2;
3265 sysfs_unlink_rdev(mddev, rdev);
3266 if (sysfs_link_rdev(mddev, rdev))
3267 pr_warn("md/raid1:%s: cannot register rd%d\n",
3268 mdname(mddev), rdev->raid_disk);
3269 }
3270 if (rdev)
3271 newmirrors[d2++].rdev = rdev;
3272 }
3273 kfree(conf->mirrors);
3274 conf->mirrors = newmirrors;
3275 kfree(conf->poolinfo);
3276 conf->poolinfo = newpoolinfo;
3277
3278 spin_lock_irqsave(&conf->device_lock, flags);
3279 mddev->degraded += (raid_disks - conf->raid_disks);
3280 spin_unlock_irqrestore(&conf->device_lock, flags);
3281 conf->raid_disks = mddev->raid_disks = raid_disks;
3282 mddev->delta_disks = 0;
3283
3284 unfreeze_array(conf);
3285
3286 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3287 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3288 md_wakeup_thread(mddev->thread);
3289
3290 mempool_destroy(oldpool);
3291 return 0;
3292}
3293
3294static void raid1_quiesce(struct mddev *mddev, int quiesce)
3295{
3296 struct r1conf *conf = mddev->private;
3297
3298 if (quiesce)
3299 freeze_array(conf, 0);
3300 else
3301 unfreeze_array(conf);
3302}
3303
3304static void *raid1_takeover(struct mddev *mddev)
3305{
3306 /* raid1 can take over:
3307 * raid5 with 2 devices, any layout or chunk size
3308 */
3309 if (mddev->level == 5 && mddev->raid_disks == 2) {
3310 struct r1conf *conf;
3311 mddev->new_level = 1;
3312 mddev->new_layout = 0;
3313 mddev->new_chunk_sectors = 0;
3314 conf = setup_conf(mddev);
3315 if (!IS_ERR(conf)) {
3316 /* Array must appear to be quiesced */
3317 conf->array_frozen = 1;
3318 mddev_clear_unsupported_flags(mddev,
3319 UNSUPPORTED_MDDEV_FLAGS);
3320 }
3321 return conf;
3322 }
3323 return ERR_PTR(-EINVAL);
3324}
3325
3326static struct md_personality raid1_personality =
3327{
3328 .name = "raid1",
3329 .level = 1,
3330 .owner = THIS_MODULE,
3331 .make_request = raid1_make_request,
3332 .run = raid1_run,
3333 .free = raid1_free,
3334 .status = raid1_status,
3335 .error_handler = raid1_error,
3336 .hot_add_disk = raid1_add_disk,
3337 .hot_remove_disk= raid1_remove_disk,
3338 .spare_active = raid1_spare_active,
3339 .sync_request = raid1_sync_request,
3340 .resize = raid1_resize,
3341 .size = raid1_size,
3342 .check_reshape = raid1_reshape,
3343 .quiesce = raid1_quiesce,
3344 .takeover = raid1_takeover,
3345 .congested = raid1_congested,
3346};
3347
3348static int __init raid_init(void)
3349{
3350 return register_md_personality(&raid1_personality);
3351}
3352
3353static void raid_exit(void)
3354{
3355 unregister_md_personality(&raid1_personality);
3356}
3357
3358module_init(raid_init);
3359module_exit(raid_exit);
3360MODULE_LICENSE("GPL");
3361MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3362MODULE_ALIAS("md-personality-3"); /* RAID1 */
3363MODULE_ALIAS("md-raid1");
3364MODULE_ALIAS("md-level-1");
3365
3366module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);