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