Loading...
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 there are this many requests queue to be written by
50 * the raid1 thread, we become 'congested' to provide back-pressure
51 * for writeback.
52 */
53static int max_queued_requests = 1024;
54
55static void allow_barrier(struct r1conf *conf);
56static void lower_barrier(struct r1conf *conf);
57
58static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
59{
60 struct pool_info *pi = data;
61 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
62
63 /* allocate a r1bio with room for raid_disks entries in the bios array */
64 return kzalloc(size, gfp_flags);
65}
66
67static void r1bio_pool_free(void *r1_bio, void *data)
68{
69 kfree(r1_bio);
70}
71
72#define RESYNC_BLOCK_SIZE (64*1024)
73//#define RESYNC_BLOCK_SIZE PAGE_SIZE
74#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
75#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
76#define RESYNC_WINDOW (2048*1024)
77
78static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
79{
80 struct pool_info *pi = data;
81 struct page *page;
82 struct r1bio *r1_bio;
83 struct bio *bio;
84 int i, j;
85
86 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
87 if (!r1_bio)
88 return NULL;
89
90 /*
91 * Allocate bios : 1 for reading, n-1 for writing
92 */
93 for (j = pi->raid_disks ; j-- ; ) {
94 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
95 if (!bio)
96 goto out_free_bio;
97 r1_bio->bios[j] = bio;
98 }
99 /*
100 * Allocate RESYNC_PAGES data pages and attach them to
101 * the first bio.
102 * If this is a user-requested check/repair, allocate
103 * RESYNC_PAGES for each bio.
104 */
105 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
106 j = pi->raid_disks;
107 else
108 j = 1;
109 while(j--) {
110 bio = r1_bio->bios[j];
111 for (i = 0; i < RESYNC_PAGES; i++) {
112 page = alloc_page(gfp_flags);
113 if (unlikely(!page))
114 goto out_free_pages;
115
116 bio->bi_io_vec[i].bv_page = page;
117 bio->bi_vcnt = i+1;
118 }
119 }
120 /* If not user-requests, copy the page pointers to all bios */
121 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
122 for (i=0; i<RESYNC_PAGES ; i++)
123 for (j=1; j<pi->raid_disks; j++)
124 r1_bio->bios[j]->bi_io_vec[i].bv_page =
125 r1_bio->bios[0]->bi_io_vec[i].bv_page;
126 }
127
128 r1_bio->master_bio = NULL;
129
130 return r1_bio;
131
132out_free_pages:
133 for (j=0 ; j < pi->raid_disks; j++)
134 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
135 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
136 j = -1;
137out_free_bio:
138 while (++j < pi->raid_disks)
139 bio_put(r1_bio->bios[j]);
140 r1bio_pool_free(r1_bio, data);
141 return NULL;
142}
143
144static void r1buf_pool_free(void *__r1_bio, void *data)
145{
146 struct pool_info *pi = data;
147 int i,j;
148 struct r1bio *r1bio = __r1_bio;
149
150 for (i = 0; i < RESYNC_PAGES; i++)
151 for (j = pi->raid_disks; j-- ;) {
152 if (j == 0 ||
153 r1bio->bios[j]->bi_io_vec[i].bv_page !=
154 r1bio->bios[0]->bi_io_vec[i].bv_page)
155 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
156 }
157 for (i=0 ; i < pi->raid_disks; i++)
158 bio_put(r1bio->bios[i]);
159
160 r1bio_pool_free(r1bio, data);
161}
162
163static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
164{
165 int i;
166
167 for (i = 0; i < conf->raid_disks * 2; i++) {
168 struct bio **bio = r1_bio->bios + i;
169 if (!BIO_SPECIAL(*bio))
170 bio_put(*bio);
171 *bio = NULL;
172 }
173}
174
175static void free_r1bio(struct r1bio *r1_bio)
176{
177 struct r1conf *conf = r1_bio->mddev->private;
178
179 put_all_bios(conf, r1_bio);
180 mempool_free(r1_bio, conf->r1bio_pool);
181}
182
183static void put_buf(struct r1bio *r1_bio)
184{
185 struct r1conf *conf = r1_bio->mddev->private;
186 int i;
187
188 for (i = 0; i < conf->raid_disks * 2; i++) {
189 struct bio *bio = r1_bio->bios[i];
190 if (bio->bi_end_io)
191 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
192 }
193
194 mempool_free(r1_bio, conf->r1buf_pool);
195
196 lower_barrier(conf);
197}
198
199static void reschedule_retry(struct r1bio *r1_bio)
200{
201 unsigned long flags;
202 struct mddev *mddev = r1_bio->mddev;
203 struct r1conf *conf = mddev->private;
204
205 spin_lock_irqsave(&conf->device_lock, flags);
206 list_add(&r1_bio->retry_list, &conf->retry_list);
207 conf->nr_queued ++;
208 spin_unlock_irqrestore(&conf->device_lock, flags);
209
210 wake_up(&conf->wait_barrier);
211 md_wakeup_thread(mddev->thread);
212}
213
214/*
215 * raid_end_bio_io() is called when we have finished servicing a mirrored
216 * operation and are ready to return a success/failure code to the buffer
217 * cache layer.
218 */
219static void call_bio_endio(struct r1bio *r1_bio)
220{
221 struct bio *bio = r1_bio->master_bio;
222 int done;
223 struct r1conf *conf = r1_bio->mddev->private;
224
225 if (bio->bi_phys_segments) {
226 unsigned long flags;
227 spin_lock_irqsave(&conf->device_lock, flags);
228 bio->bi_phys_segments--;
229 done = (bio->bi_phys_segments == 0);
230 spin_unlock_irqrestore(&conf->device_lock, flags);
231 } else
232 done = 1;
233
234 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
235 clear_bit(BIO_UPTODATE, &bio->bi_flags);
236 if (done) {
237 bio_endio(bio, 0);
238 /*
239 * Wake up any possible resync thread that waits for the device
240 * to go idle.
241 */
242 allow_barrier(conf);
243 }
244}
245
246static void raid_end_bio_io(struct r1bio *r1_bio)
247{
248 struct bio *bio = r1_bio->master_bio;
249
250 /* if nobody has done the final endio yet, do it now */
251 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
252 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
253 (bio_data_dir(bio) == WRITE) ? "write" : "read",
254 (unsigned long long) bio->bi_sector,
255 (unsigned long long) bio->bi_sector +
256 (bio->bi_size >> 9) - 1);
257
258 call_bio_endio(r1_bio);
259 }
260 free_r1bio(r1_bio);
261}
262
263/*
264 * Update disk head position estimator based on IRQ completion info.
265 */
266static inline void update_head_pos(int disk, struct r1bio *r1_bio)
267{
268 struct r1conf *conf = r1_bio->mddev->private;
269
270 conf->mirrors[disk].head_position =
271 r1_bio->sector + (r1_bio->sectors);
272}
273
274/*
275 * Find the disk number which triggered given bio
276 */
277static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
278{
279 int mirror;
280 struct r1conf *conf = r1_bio->mddev->private;
281 int raid_disks = conf->raid_disks;
282
283 for (mirror = 0; mirror < raid_disks * 2; mirror++)
284 if (r1_bio->bios[mirror] == bio)
285 break;
286
287 BUG_ON(mirror == raid_disks * 2);
288 update_head_pos(mirror, r1_bio);
289
290 return mirror;
291}
292
293static void raid1_end_read_request(struct bio *bio, int error)
294{
295 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
296 struct r1bio *r1_bio = bio->bi_private;
297 int mirror;
298 struct r1conf *conf = r1_bio->mddev->private;
299
300 mirror = r1_bio->read_disk;
301 /*
302 * this branch is our 'one mirror IO has finished' event handler:
303 */
304 update_head_pos(mirror, r1_bio);
305
306 if (uptodate)
307 set_bit(R1BIO_Uptodate, &r1_bio->state);
308 else {
309 /* If all other devices have failed, we want to return
310 * the error upwards rather than fail the last device.
311 * Here we redefine "uptodate" to mean "Don't want to retry"
312 */
313 unsigned long flags;
314 spin_lock_irqsave(&conf->device_lock, flags);
315 if (r1_bio->mddev->degraded == conf->raid_disks ||
316 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
317 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
318 uptodate = 1;
319 spin_unlock_irqrestore(&conf->device_lock, flags);
320 }
321
322 if (uptodate)
323 raid_end_bio_io(r1_bio);
324 else {
325 /*
326 * oops, read error:
327 */
328 char b[BDEVNAME_SIZE];
329 printk_ratelimited(
330 KERN_ERR "md/raid1:%s: %s: "
331 "rescheduling sector %llu\n",
332 mdname(conf->mddev),
333 bdevname(conf->mirrors[mirror].rdev->bdev,
334 b),
335 (unsigned long long)r1_bio->sector);
336 set_bit(R1BIO_ReadError, &r1_bio->state);
337 reschedule_retry(r1_bio);
338 }
339
340 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
341}
342
343static void close_write(struct r1bio *r1_bio)
344{
345 /* it really is the end of this request */
346 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
347 /* free extra copy of the data pages */
348 int i = r1_bio->behind_page_count;
349 while (i--)
350 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
351 kfree(r1_bio->behind_bvecs);
352 r1_bio->behind_bvecs = NULL;
353 }
354 /* clear the bitmap if all writes complete successfully */
355 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
356 r1_bio->sectors,
357 !test_bit(R1BIO_Degraded, &r1_bio->state),
358 test_bit(R1BIO_BehindIO, &r1_bio->state));
359 md_write_end(r1_bio->mddev);
360}
361
362static void r1_bio_write_done(struct r1bio *r1_bio)
363{
364 if (!atomic_dec_and_test(&r1_bio->remaining))
365 return;
366
367 if (test_bit(R1BIO_WriteError, &r1_bio->state))
368 reschedule_retry(r1_bio);
369 else {
370 close_write(r1_bio);
371 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
372 reschedule_retry(r1_bio);
373 else
374 raid_end_bio_io(r1_bio);
375 }
376}
377
378static void raid1_end_write_request(struct bio *bio, int error)
379{
380 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
381 struct r1bio *r1_bio = bio->bi_private;
382 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
383 struct r1conf *conf = r1_bio->mddev->private;
384 struct bio *to_put = NULL;
385
386 mirror = find_bio_disk(r1_bio, bio);
387
388 /*
389 * 'one mirror IO has finished' event handler:
390 */
391 if (!uptodate) {
392 set_bit(WriteErrorSeen,
393 &conf->mirrors[mirror].rdev->flags);
394 if (!test_and_set_bit(WantReplacement,
395 &conf->mirrors[mirror].rdev->flags))
396 set_bit(MD_RECOVERY_NEEDED, &
397 conf->mddev->recovery);
398
399 set_bit(R1BIO_WriteError, &r1_bio->state);
400 } else {
401 /*
402 * Set R1BIO_Uptodate in our master bio, so that we
403 * will return a good error code for to the higher
404 * levels even if IO on some other mirrored buffer
405 * fails.
406 *
407 * The 'master' represents the composite IO operation
408 * to user-side. So if something waits for IO, then it
409 * will wait for the 'master' bio.
410 */
411 sector_t first_bad;
412 int bad_sectors;
413
414 r1_bio->bios[mirror] = NULL;
415 to_put = bio;
416 set_bit(R1BIO_Uptodate, &r1_bio->state);
417
418 /* Maybe we can clear some bad blocks. */
419 if (is_badblock(conf->mirrors[mirror].rdev,
420 r1_bio->sector, r1_bio->sectors,
421 &first_bad, &bad_sectors)) {
422 r1_bio->bios[mirror] = IO_MADE_GOOD;
423 set_bit(R1BIO_MadeGood, &r1_bio->state);
424 }
425 }
426
427 if (behind) {
428 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
429 atomic_dec(&r1_bio->behind_remaining);
430
431 /*
432 * In behind mode, we ACK the master bio once the I/O
433 * has safely reached all non-writemostly
434 * disks. Setting the Returned bit ensures that this
435 * gets done only once -- we don't ever want to return
436 * -EIO here, instead we'll wait
437 */
438 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
439 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
440 /* Maybe we can return now */
441 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
442 struct bio *mbio = r1_bio->master_bio;
443 pr_debug("raid1: behind end write sectors"
444 " %llu-%llu\n",
445 (unsigned long long) mbio->bi_sector,
446 (unsigned long long) mbio->bi_sector +
447 (mbio->bi_size >> 9) - 1);
448 call_bio_endio(r1_bio);
449 }
450 }
451 }
452 if (r1_bio->bios[mirror] == NULL)
453 rdev_dec_pending(conf->mirrors[mirror].rdev,
454 conf->mddev);
455
456 /*
457 * Let's see if all mirrored write operations have finished
458 * already.
459 */
460 r1_bio_write_done(r1_bio);
461
462 if (to_put)
463 bio_put(to_put);
464}
465
466
467/*
468 * This routine returns the disk from which the requested read should
469 * be done. There is a per-array 'next expected sequential IO' sector
470 * number - if this matches on the next IO then we use the last disk.
471 * There is also a per-disk 'last know head position' sector that is
472 * maintained from IRQ contexts, both the normal and the resync IO
473 * completion handlers update this position correctly. If there is no
474 * perfect sequential match then we pick the disk whose head is closest.
475 *
476 * If there are 2 mirrors in the same 2 devices, performance degrades
477 * because position is mirror, not device based.
478 *
479 * The rdev for the device selected will have nr_pending incremented.
480 */
481static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
482{
483 const sector_t this_sector = r1_bio->sector;
484 int sectors;
485 int best_good_sectors;
486 int start_disk;
487 int best_disk;
488 int i;
489 sector_t best_dist;
490 struct md_rdev *rdev;
491 int choose_first;
492
493 rcu_read_lock();
494 /*
495 * Check if we can balance. We can balance on the whole
496 * device if no resync is going on, or below the resync window.
497 * We take the first readable disk when above the resync window.
498 */
499 retry:
500 sectors = r1_bio->sectors;
501 best_disk = -1;
502 best_dist = MaxSector;
503 best_good_sectors = 0;
504
505 if (conf->mddev->recovery_cp < MaxSector &&
506 (this_sector + sectors >= conf->next_resync)) {
507 choose_first = 1;
508 start_disk = 0;
509 } else {
510 choose_first = 0;
511 start_disk = conf->last_used;
512 }
513
514 for (i = 0 ; i < conf->raid_disks * 2 ; i++) {
515 sector_t dist;
516 sector_t first_bad;
517 int bad_sectors;
518
519 int disk = start_disk + i;
520 if (disk >= conf->raid_disks * 2)
521 disk -= conf->raid_disks * 2;
522
523 rdev = rcu_dereference(conf->mirrors[disk].rdev);
524 if (r1_bio->bios[disk] == IO_BLOCKED
525 || rdev == NULL
526 || test_bit(Unmerged, &rdev->flags)
527 || test_bit(Faulty, &rdev->flags))
528 continue;
529 if (!test_bit(In_sync, &rdev->flags) &&
530 rdev->recovery_offset < this_sector + sectors)
531 continue;
532 if (test_bit(WriteMostly, &rdev->flags)) {
533 /* Don't balance among write-mostly, just
534 * use the first as a last resort */
535 if (best_disk < 0) {
536 if (is_badblock(rdev, this_sector, sectors,
537 &first_bad, &bad_sectors)) {
538 if (first_bad < this_sector)
539 /* Cannot use this */
540 continue;
541 best_good_sectors = first_bad - this_sector;
542 } else
543 best_good_sectors = sectors;
544 best_disk = disk;
545 }
546 continue;
547 }
548 /* This is a reasonable device to use. It might
549 * even be best.
550 */
551 if (is_badblock(rdev, this_sector, sectors,
552 &first_bad, &bad_sectors)) {
553 if (best_dist < MaxSector)
554 /* already have a better device */
555 continue;
556 if (first_bad <= this_sector) {
557 /* cannot read here. If this is the 'primary'
558 * device, then we must not read beyond
559 * bad_sectors from another device..
560 */
561 bad_sectors -= (this_sector - first_bad);
562 if (choose_first && sectors > bad_sectors)
563 sectors = bad_sectors;
564 if (best_good_sectors > sectors)
565 best_good_sectors = sectors;
566
567 } else {
568 sector_t good_sectors = first_bad - this_sector;
569 if (good_sectors > best_good_sectors) {
570 best_good_sectors = good_sectors;
571 best_disk = disk;
572 }
573 if (choose_first)
574 break;
575 }
576 continue;
577 } else
578 best_good_sectors = sectors;
579
580 dist = abs(this_sector - conf->mirrors[disk].head_position);
581 if (choose_first
582 /* Don't change to another disk for sequential reads */
583 || conf->next_seq_sect == this_sector
584 || dist == 0
585 /* If device is idle, use it */
586 || atomic_read(&rdev->nr_pending) == 0) {
587 best_disk = disk;
588 break;
589 }
590 if (dist < best_dist) {
591 best_dist = dist;
592 best_disk = disk;
593 }
594 }
595
596 if (best_disk >= 0) {
597 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
598 if (!rdev)
599 goto retry;
600 atomic_inc(&rdev->nr_pending);
601 if (test_bit(Faulty, &rdev->flags)) {
602 /* cannot risk returning a device that failed
603 * before we inc'ed nr_pending
604 */
605 rdev_dec_pending(rdev, conf->mddev);
606 goto retry;
607 }
608 sectors = best_good_sectors;
609 conf->next_seq_sect = this_sector + sectors;
610 conf->last_used = best_disk;
611 }
612 rcu_read_unlock();
613 *max_sectors = sectors;
614
615 return best_disk;
616}
617
618static int raid1_mergeable_bvec(struct request_queue *q,
619 struct bvec_merge_data *bvm,
620 struct bio_vec *biovec)
621{
622 struct mddev *mddev = q->queuedata;
623 struct r1conf *conf = mddev->private;
624 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
625 int max = biovec->bv_len;
626
627 if (mddev->merge_check_needed) {
628 int disk;
629 rcu_read_lock();
630 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
631 struct md_rdev *rdev = rcu_dereference(
632 conf->mirrors[disk].rdev);
633 if (rdev && !test_bit(Faulty, &rdev->flags)) {
634 struct request_queue *q =
635 bdev_get_queue(rdev->bdev);
636 if (q->merge_bvec_fn) {
637 bvm->bi_sector = sector +
638 rdev->data_offset;
639 bvm->bi_bdev = rdev->bdev;
640 max = min(max, q->merge_bvec_fn(
641 q, bvm, biovec));
642 }
643 }
644 }
645 rcu_read_unlock();
646 }
647 return max;
648
649}
650
651int md_raid1_congested(struct mddev *mddev, int bits)
652{
653 struct r1conf *conf = mddev->private;
654 int i, ret = 0;
655
656 if ((bits & (1 << BDI_async_congested)) &&
657 conf->pending_count >= max_queued_requests)
658 return 1;
659
660 rcu_read_lock();
661 for (i = 0; i < conf->raid_disks * 2; i++) {
662 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
663 if (rdev && !test_bit(Faulty, &rdev->flags)) {
664 struct request_queue *q = bdev_get_queue(rdev->bdev);
665
666 BUG_ON(!q);
667
668 /* Note the '|| 1' - when read_balance prefers
669 * non-congested targets, it can be removed
670 */
671 if ((bits & (1<<BDI_async_congested)) || 1)
672 ret |= bdi_congested(&q->backing_dev_info, bits);
673 else
674 ret &= bdi_congested(&q->backing_dev_info, bits);
675 }
676 }
677 rcu_read_unlock();
678 return ret;
679}
680EXPORT_SYMBOL_GPL(md_raid1_congested);
681
682static int raid1_congested(void *data, int bits)
683{
684 struct mddev *mddev = data;
685
686 return mddev_congested(mddev, bits) ||
687 md_raid1_congested(mddev, bits);
688}
689
690static void flush_pending_writes(struct r1conf *conf)
691{
692 /* Any writes that have been queued but are awaiting
693 * bitmap updates get flushed here.
694 */
695 spin_lock_irq(&conf->device_lock);
696
697 if (conf->pending_bio_list.head) {
698 struct bio *bio;
699 bio = bio_list_get(&conf->pending_bio_list);
700 conf->pending_count = 0;
701 spin_unlock_irq(&conf->device_lock);
702 /* flush any pending bitmap writes to
703 * disk before proceeding w/ I/O */
704 bitmap_unplug(conf->mddev->bitmap);
705 wake_up(&conf->wait_barrier);
706
707 while (bio) { /* submit pending writes */
708 struct bio *next = bio->bi_next;
709 bio->bi_next = NULL;
710 generic_make_request(bio);
711 bio = next;
712 }
713 } else
714 spin_unlock_irq(&conf->device_lock);
715}
716
717/* Barriers....
718 * Sometimes we need to suspend IO while we do something else,
719 * either some resync/recovery, or reconfigure the array.
720 * To do this we raise a 'barrier'.
721 * The 'barrier' is a counter that can be raised multiple times
722 * to count how many activities are happening which preclude
723 * normal IO.
724 * We can only raise the barrier if there is no pending IO.
725 * i.e. if nr_pending == 0.
726 * We choose only to raise the barrier if no-one is waiting for the
727 * barrier to go down. This means that as soon as an IO request
728 * is ready, no other operations which require a barrier will start
729 * until the IO request has had a chance.
730 *
731 * So: regular IO calls 'wait_barrier'. When that returns there
732 * is no backgroup IO happening, It must arrange to call
733 * allow_barrier when it has finished its IO.
734 * backgroup IO calls must call raise_barrier. Once that returns
735 * there is no normal IO happeing. It must arrange to call
736 * lower_barrier when the particular background IO completes.
737 */
738#define RESYNC_DEPTH 32
739
740static void raise_barrier(struct r1conf *conf)
741{
742 spin_lock_irq(&conf->resync_lock);
743
744 /* Wait until no block IO is waiting */
745 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
746 conf->resync_lock, );
747
748 /* block any new IO from starting */
749 conf->barrier++;
750
751 /* Now wait for all pending IO to complete */
752 wait_event_lock_irq(conf->wait_barrier,
753 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
754 conf->resync_lock, );
755
756 spin_unlock_irq(&conf->resync_lock);
757}
758
759static void lower_barrier(struct r1conf *conf)
760{
761 unsigned long flags;
762 BUG_ON(conf->barrier <= 0);
763 spin_lock_irqsave(&conf->resync_lock, flags);
764 conf->barrier--;
765 spin_unlock_irqrestore(&conf->resync_lock, flags);
766 wake_up(&conf->wait_barrier);
767}
768
769static void wait_barrier(struct r1conf *conf)
770{
771 spin_lock_irq(&conf->resync_lock);
772 if (conf->barrier) {
773 conf->nr_waiting++;
774 /* Wait for the barrier to drop.
775 * However if there are already pending
776 * requests (preventing the barrier from
777 * rising completely), and the
778 * pre-process bio queue isn't empty,
779 * then don't wait, as we need to empty
780 * that queue to get the nr_pending
781 * count down.
782 */
783 wait_event_lock_irq(conf->wait_barrier,
784 !conf->barrier ||
785 (conf->nr_pending &&
786 current->bio_list &&
787 !bio_list_empty(current->bio_list)),
788 conf->resync_lock,
789 );
790 conf->nr_waiting--;
791 }
792 conf->nr_pending++;
793 spin_unlock_irq(&conf->resync_lock);
794}
795
796static void allow_barrier(struct r1conf *conf)
797{
798 unsigned long flags;
799 spin_lock_irqsave(&conf->resync_lock, flags);
800 conf->nr_pending--;
801 spin_unlock_irqrestore(&conf->resync_lock, flags);
802 wake_up(&conf->wait_barrier);
803}
804
805static void freeze_array(struct r1conf *conf)
806{
807 /* stop syncio and normal IO and wait for everything to
808 * go quite.
809 * We increment barrier and nr_waiting, and then
810 * wait until nr_pending match nr_queued+1
811 * This is called in the context of one normal IO request
812 * that has failed. Thus any sync request that might be pending
813 * will be blocked by nr_pending, and we need to wait for
814 * pending IO requests to complete or be queued for re-try.
815 * Thus the number queued (nr_queued) plus this request (1)
816 * must match the number of pending IOs (nr_pending) before
817 * we continue.
818 */
819 spin_lock_irq(&conf->resync_lock);
820 conf->barrier++;
821 conf->nr_waiting++;
822 wait_event_lock_irq(conf->wait_barrier,
823 conf->nr_pending == conf->nr_queued+1,
824 conf->resync_lock,
825 flush_pending_writes(conf));
826 spin_unlock_irq(&conf->resync_lock);
827}
828static void unfreeze_array(struct r1conf *conf)
829{
830 /* reverse the effect of the freeze */
831 spin_lock_irq(&conf->resync_lock);
832 conf->barrier--;
833 conf->nr_waiting--;
834 wake_up(&conf->wait_barrier);
835 spin_unlock_irq(&conf->resync_lock);
836}
837
838
839/* duplicate the data pages for behind I/O
840 */
841static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
842{
843 int i;
844 struct bio_vec *bvec;
845 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
846 GFP_NOIO);
847 if (unlikely(!bvecs))
848 return;
849
850 bio_for_each_segment(bvec, bio, i) {
851 bvecs[i] = *bvec;
852 bvecs[i].bv_page = alloc_page(GFP_NOIO);
853 if (unlikely(!bvecs[i].bv_page))
854 goto do_sync_io;
855 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
856 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
857 kunmap(bvecs[i].bv_page);
858 kunmap(bvec->bv_page);
859 }
860 r1_bio->behind_bvecs = bvecs;
861 r1_bio->behind_page_count = bio->bi_vcnt;
862 set_bit(R1BIO_BehindIO, &r1_bio->state);
863 return;
864
865do_sync_io:
866 for (i = 0; i < bio->bi_vcnt; i++)
867 if (bvecs[i].bv_page)
868 put_page(bvecs[i].bv_page);
869 kfree(bvecs);
870 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
871}
872
873static void make_request(struct mddev *mddev, struct bio * bio)
874{
875 struct r1conf *conf = mddev->private;
876 struct mirror_info *mirror;
877 struct r1bio *r1_bio;
878 struct bio *read_bio;
879 int i, disks;
880 struct bitmap *bitmap;
881 unsigned long flags;
882 const int rw = bio_data_dir(bio);
883 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
884 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
885 struct md_rdev *blocked_rdev;
886 int first_clone;
887 int sectors_handled;
888 int max_sectors;
889
890 /*
891 * Register the new request and wait if the reconstruction
892 * thread has put up a bar for new requests.
893 * Continue immediately if no resync is active currently.
894 */
895
896 md_write_start(mddev, bio); /* wait on superblock update early */
897
898 if (bio_data_dir(bio) == WRITE &&
899 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
900 bio->bi_sector < mddev->suspend_hi) {
901 /* As the suspend_* range is controlled by
902 * userspace, we want an interruptible
903 * wait.
904 */
905 DEFINE_WAIT(w);
906 for (;;) {
907 flush_signals(current);
908 prepare_to_wait(&conf->wait_barrier,
909 &w, TASK_INTERRUPTIBLE);
910 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
911 bio->bi_sector >= mddev->suspend_hi)
912 break;
913 schedule();
914 }
915 finish_wait(&conf->wait_barrier, &w);
916 }
917
918 wait_barrier(conf);
919
920 bitmap = mddev->bitmap;
921
922 /*
923 * make_request() can abort the operation when READA is being
924 * used and no empty request is available.
925 *
926 */
927 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
928
929 r1_bio->master_bio = bio;
930 r1_bio->sectors = bio->bi_size >> 9;
931 r1_bio->state = 0;
932 r1_bio->mddev = mddev;
933 r1_bio->sector = bio->bi_sector;
934
935 /* We might need to issue multiple reads to different
936 * devices if there are bad blocks around, so we keep
937 * track of the number of reads in bio->bi_phys_segments.
938 * If this is 0, there is only one r1_bio and no locking
939 * will be needed when requests complete. If it is
940 * non-zero, then it is the number of not-completed requests.
941 */
942 bio->bi_phys_segments = 0;
943 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
944
945 if (rw == READ) {
946 /*
947 * read balancing logic:
948 */
949 int rdisk;
950
951read_again:
952 rdisk = read_balance(conf, r1_bio, &max_sectors);
953
954 if (rdisk < 0) {
955 /* couldn't find anywhere to read from */
956 raid_end_bio_io(r1_bio);
957 return;
958 }
959 mirror = conf->mirrors + rdisk;
960
961 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
962 bitmap) {
963 /* Reading from a write-mostly device must
964 * take care not to over-take any writes
965 * that are 'behind'
966 */
967 wait_event(bitmap->behind_wait,
968 atomic_read(&bitmap->behind_writes) == 0);
969 }
970 r1_bio->read_disk = rdisk;
971
972 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
973 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
974 max_sectors);
975
976 r1_bio->bios[rdisk] = read_bio;
977
978 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
979 read_bio->bi_bdev = mirror->rdev->bdev;
980 read_bio->bi_end_io = raid1_end_read_request;
981 read_bio->bi_rw = READ | do_sync;
982 read_bio->bi_private = r1_bio;
983
984 if (max_sectors < r1_bio->sectors) {
985 /* could not read all from this device, so we will
986 * need another r1_bio.
987 */
988
989 sectors_handled = (r1_bio->sector + max_sectors
990 - bio->bi_sector);
991 r1_bio->sectors = max_sectors;
992 spin_lock_irq(&conf->device_lock);
993 if (bio->bi_phys_segments == 0)
994 bio->bi_phys_segments = 2;
995 else
996 bio->bi_phys_segments++;
997 spin_unlock_irq(&conf->device_lock);
998 /* Cannot call generic_make_request directly
999 * as that will be queued in __make_request
1000 * and subsequent mempool_alloc might block waiting
1001 * for it. So hand bio over to raid1d.
1002 */
1003 reschedule_retry(r1_bio);
1004
1005 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1006
1007 r1_bio->master_bio = bio;
1008 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1009 r1_bio->state = 0;
1010 r1_bio->mddev = mddev;
1011 r1_bio->sector = bio->bi_sector + sectors_handled;
1012 goto read_again;
1013 } else
1014 generic_make_request(read_bio);
1015 return;
1016 }
1017
1018 /*
1019 * WRITE:
1020 */
1021 if (conf->pending_count >= max_queued_requests) {
1022 md_wakeup_thread(mddev->thread);
1023 wait_event(conf->wait_barrier,
1024 conf->pending_count < max_queued_requests);
1025 }
1026 /* first select target devices under rcu_lock and
1027 * inc refcount on their rdev. Record them by setting
1028 * bios[x] to bio
1029 * If there are known/acknowledged bad blocks on any device on
1030 * which we have seen a write error, we want to avoid writing those
1031 * blocks.
1032 * This potentially requires several writes to write around
1033 * the bad blocks. Each set of writes gets it's own r1bio
1034 * with a set of bios attached.
1035 */
1036
1037 disks = conf->raid_disks * 2;
1038 retry_write:
1039 blocked_rdev = NULL;
1040 rcu_read_lock();
1041 max_sectors = r1_bio->sectors;
1042 for (i = 0; i < disks; i++) {
1043 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1044 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1045 atomic_inc(&rdev->nr_pending);
1046 blocked_rdev = rdev;
1047 break;
1048 }
1049 r1_bio->bios[i] = NULL;
1050 if (!rdev || test_bit(Faulty, &rdev->flags)
1051 || test_bit(Unmerged, &rdev->flags)) {
1052 if (i < conf->raid_disks)
1053 set_bit(R1BIO_Degraded, &r1_bio->state);
1054 continue;
1055 }
1056
1057 atomic_inc(&rdev->nr_pending);
1058 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1059 sector_t first_bad;
1060 int bad_sectors;
1061 int is_bad;
1062
1063 is_bad = is_badblock(rdev, r1_bio->sector,
1064 max_sectors,
1065 &first_bad, &bad_sectors);
1066 if (is_bad < 0) {
1067 /* mustn't write here until the bad block is
1068 * acknowledged*/
1069 set_bit(BlockedBadBlocks, &rdev->flags);
1070 blocked_rdev = rdev;
1071 break;
1072 }
1073 if (is_bad && first_bad <= r1_bio->sector) {
1074 /* Cannot write here at all */
1075 bad_sectors -= (r1_bio->sector - first_bad);
1076 if (bad_sectors < max_sectors)
1077 /* mustn't write more than bad_sectors
1078 * to other devices yet
1079 */
1080 max_sectors = bad_sectors;
1081 rdev_dec_pending(rdev, mddev);
1082 /* We don't set R1BIO_Degraded as that
1083 * only applies if the disk is
1084 * missing, so it might be re-added,
1085 * and we want to know to recover this
1086 * chunk.
1087 * In this case the device is here,
1088 * and the fact that this chunk is not
1089 * in-sync is recorded in the bad
1090 * block log
1091 */
1092 continue;
1093 }
1094 if (is_bad) {
1095 int good_sectors = first_bad - r1_bio->sector;
1096 if (good_sectors < max_sectors)
1097 max_sectors = good_sectors;
1098 }
1099 }
1100 r1_bio->bios[i] = bio;
1101 }
1102 rcu_read_unlock();
1103
1104 if (unlikely(blocked_rdev)) {
1105 /* Wait for this device to become unblocked */
1106 int j;
1107
1108 for (j = 0; j < i; j++)
1109 if (r1_bio->bios[j])
1110 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1111 r1_bio->state = 0;
1112 allow_barrier(conf);
1113 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1114 wait_barrier(conf);
1115 goto retry_write;
1116 }
1117
1118 if (max_sectors < r1_bio->sectors) {
1119 /* We are splitting this write into multiple parts, so
1120 * we need to prepare for allocating another r1_bio.
1121 */
1122 r1_bio->sectors = max_sectors;
1123 spin_lock_irq(&conf->device_lock);
1124 if (bio->bi_phys_segments == 0)
1125 bio->bi_phys_segments = 2;
1126 else
1127 bio->bi_phys_segments++;
1128 spin_unlock_irq(&conf->device_lock);
1129 }
1130 sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1131
1132 atomic_set(&r1_bio->remaining, 1);
1133 atomic_set(&r1_bio->behind_remaining, 0);
1134
1135 first_clone = 1;
1136 for (i = 0; i < disks; i++) {
1137 struct bio *mbio;
1138 if (!r1_bio->bios[i])
1139 continue;
1140
1141 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1142 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1143
1144 if (first_clone) {
1145 /* do behind I/O ?
1146 * Not if there are too many, or cannot
1147 * allocate memory, or a reader on WriteMostly
1148 * is waiting for behind writes to flush */
1149 if (bitmap &&
1150 (atomic_read(&bitmap->behind_writes)
1151 < mddev->bitmap_info.max_write_behind) &&
1152 !waitqueue_active(&bitmap->behind_wait))
1153 alloc_behind_pages(mbio, r1_bio);
1154
1155 bitmap_startwrite(bitmap, r1_bio->sector,
1156 r1_bio->sectors,
1157 test_bit(R1BIO_BehindIO,
1158 &r1_bio->state));
1159 first_clone = 0;
1160 }
1161 if (r1_bio->behind_bvecs) {
1162 struct bio_vec *bvec;
1163 int j;
1164
1165 /* Yes, I really want the '__' version so that
1166 * we clear any unused pointer in the io_vec, rather
1167 * than leave them unchanged. This is important
1168 * because when we come to free the pages, we won't
1169 * know the original bi_idx, so we just free
1170 * them all
1171 */
1172 __bio_for_each_segment(bvec, mbio, j, 0)
1173 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1174 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1175 atomic_inc(&r1_bio->behind_remaining);
1176 }
1177
1178 r1_bio->bios[i] = mbio;
1179
1180 mbio->bi_sector = (r1_bio->sector +
1181 conf->mirrors[i].rdev->data_offset);
1182 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1183 mbio->bi_end_io = raid1_end_write_request;
1184 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1185 mbio->bi_private = r1_bio;
1186
1187 atomic_inc(&r1_bio->remaining);
1188 spin_lock_irqsave(&conf->device_lock, flags);
1189 bio_list_add(&conf->pending_bio_list, mbio);
1190 conf->pending_count++;
1191 spin_unlock_irqrestore(&conf->device_lock, flags);
1192 if (!mddev_check_plugged(mddev))
1193 md_wakeup_thread(mddev->thread);
1194 }
1195 /* Mustn't call r1_bio_write_done before this next test,
1196 * as it could result in the bio being freed.
1197 */
1198 if (sectors_handled < (bio->bi_size >> 9)) {
1199 r1_bio_write_done(r1_bio);
1200 /* We need another r1_bio. It has already been counted
1201 * in bio->bi_phys_segments
1202 */
1203 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1204 r1_bio->master_bio = bio;
1205 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1206 r1_bio->state = 0;
1207 r1_bio->mddev = mddev;
1208 r1_bio->sector = bio->bi_sector + sectors_handled;
1209 goto retry_write;
1210 }
1211
1212 r1_bio_write_done(r1_bio);
1213
1214 /* In case raid1d snuck in to freeze_array */
1215 wake_up(&conf->wait_barrier);
1216}
1217
1218static void status(struct seq_file *seq, struct mddev *mddev)
1219{
1220 struct r1conf *conf = mddev->private;
1221 int i;
1222
1223 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1224 conf->raid_disks - mddev->degraded);
1225 rcu_read_lock();
1226 for (i = 0; i < conf->raid_disks; i++) {
1227 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1228 seq_printf(seq, "%s",
1229 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1230 }
1231 rcu_read_unlock();
1232 seq_printf(seq, "]");
1233}
1234
1235
1236static void error(struct mddev *mddev, struct md_rdev *rdev)
1237{
1238 char b[BDEVNAME_SIZE];
1239 struct r1conf *conf = mddev->private;
1240
1241 /*
1242 * If it is not operational, then we have already marked it as dead
1243 * else if it is the last working disks, ignore the error, let the
1244 * next level up know.
1245 * else mark the drive as failed
1246 */
1247 if (test_bit(In_sync, &rdev->flags)
1248 && (conf->raid_disks - mddev->degraded) == 1) {
1249 /*
1250 * Don't fail the drive, act as though we were just a
1251 * normal single drive.
1252 * However don't try a recovery from this drive as
1253 * it is very likely to fail.
1254 */
1255 conf->recovery_disabled = mddev->recovery_disabled;
1256 return;
1257 }
1258 set_bit(Blocked, &rdev->flags);
1259 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1260 unsigned long flags;
1261 spin_lock_irqsave(&conf->device_lock, flags);
1262 mddev->degraded++;
1263 set_bit(Faulty, &rdev->flags);
1264 spin_unlock_irqrestore(&conf->device_lock, flags);
1265 /*
1266 * if recovery is running, make sure it aborts.
1267 */
1268 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1269 } else
1270 set_bit(Faulty, &rdev->flags);
1271 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1272 printk(KERN_ALERT
1273 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1274 "md/raid1:%s: Operation continuing on %d devices.\n",
1275 mdname(mddev), bdevname(rdev->bdev, b),
1276 mdname(mddev), conf->raid_disks - mddev->degraded);
1277}
1278
1279static void print_conf(struct r1conf *conf)
1280{
1281 int i;
1282
1283 printk(KERN_DEBUG "RAID1 conf printout:\n");
1284 if (!conf) {
1285 printk(KERN_DEBUG "(!conf)\n");
1286 return;
1287 }
1288 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1289 conf->raid_disks);
1290
1291 rcu_read_lock();
1292 for (i = 0; i < conf->raid_disks; i++) {
1293 char b[BDEVNAME_SIZE];
1294 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1295 if (rdev)
1296 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1297 i, !test_bit(In_sync, &rdev->flags),
1298 !test_bit(Faulty, &rdev->flags),
1299 bdevname(rdev->bdev,b));
1300 }
1301 rcu_read_unlock();
1302}
1303
1304static void close_sync(struct r1conf *conf)
1305{
1306 wait_barrier(conf);
1307 allow_barrier(conf);
1308
1309 mempool_destroy(conf->r1buf_pool);
1310 conf->r1buf_pool = NULL;
1311}
1312
1313static int raid1_spare_active(struct mddev *mddev)
1314{
1315 int i;
1316 struct r1conf *conf = mddev->private;
1317 int count = 0;
1318 unsigned long flags;
1319
1320 /*
1321 * Find all failed disks within the RAID1 configuration
1322 * and mark them readable.
1323 * Called under mddev lock, so rcu protection not needed.
1324 */
1325 for (i = 0; i < conf->raid_disks; i++) {
1326 struct md_rdev *rdev = conf->mirrors[i].rdev;
1327 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1328 if (repl
1329 && repl->recovery_offset == MaxSector
1330 && !test_bit(Faulty, &repl->flags)
1331 && !test_and_set_bit(In_sync, &repl->flags)) {
1332 /* replacement has just become active */
1333 if (!rdev ||
1334 !test_and_clear_bit(In_sync, &rdev->flags))
1335 count++;
1336 if (rdev) {
1337 /* Replaced device not technically
1338 * faulty, but we need to be sure
1339 * it gets removed and never re-added
1340 */
1341 set_bit(Faulty, &rdev->flags);
1342 sysfs_notify_dirent_safe(
1343 rdev->sysfs_state);
1344 }
1345 }
1346 if (rdev
1347 && !test_bit(Faulty, &rdev->flags)
1348 && !test_and_set_bit(In_sync, &rdev->flags)) {
1349 count++;
1350 sysfs_notify_dirent_safe(rdev->sysfs_state);
1351 }
1352 }
1353 spin_lock_irqsave(&conf->device_lock, flags);
1354 mddev->degraded -= count;
1355 spin_unlock_irqrestore(&conf->device_lock, flags);
1356
1357 print_conf(conf);
1358 return count;
1359}
1360
1361
1362static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1363{
1364 struct r1conf *conf = mddev->private;
1365 int err = -EEXIST;
1366 int mirror = 0;
1367 struct mirror_info *p;
1368 int first = 0;
1369 int last = conf->raid_disks - 1;
1370 struct request_queue *q = bdev_get_queue(rdev->bdev);
1371
1372 if (mddev->recovery_disabled == conf->recovery_disabled)
1373 return -EBUSY;
1374
1375 if (rdev->raid_disk >= 0)
1376 first = last = rdev->raid_disk;
1377
1378 if (q->merge_bvec_fn) {
1379 set_bit(Unmerged, &rdev->flags);
1380 mddev->merge_check_needed = 1;
1381 }
1382
1383 for (mirror = first; mirror <= last; mirror++) {
1384 p = conf->mirrors+mirror;
1385 if (!p->rdev) {
1386
1387 disk_stack_limits(mddev->gendisk, rdev->bdev,
1388 rdev->data_offset << 9);
1389
1390 p->head_position = 0;
1391 rdev->raid_disk = mirror;
1392 err = 0;
1393 /* As all devices are equivalent, we don't need a full recovery
1394 * if this was recently any drive of the array
1395 */
1396 if (rdev->saved_raid_disk < 0)
1397 conf->fullsync = 1;
1398 rcu_assign_pointer(p->rdev, rdev);
1399 break;
1400 }
1401 if (test_bit(WantReplacement, &p->rdev->flags) &&
1402 p[conf->raid_disks].rdev == NULL) {
1403 /* Add this device as a replacement */
1404 clear_bit(In_sync, &rdev->flags);
1405 set_bit(Replacement, &rdev->flags);
1406 rdev->raid_disk = mirror;
1407 err = 0;
1408 conf->fullsync = 1;
1409 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1410 break;
1411 }
1412 }
1413 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1414 /* Some requests might not have seen this new
1415 * merge_bvec_fn. We must wait for them to complete
1416 * before merging the device fully.
1417 * First we make sure any code which has tested
1418 * our function has submitted the request, then
1419 * we wait for all outstanding requests to complete.
1420 */
1421 synchronize_sched();
1422 raise_barrier(conf);
1423 lower_barrier(conf);
1424 clear_bit(Unmerged, &rdev->flags);
1425 }
1426 md_integrity_add_rdev(rdev, mddev);
1427 print_conf(conf);
1428 return err;
1429}
1430
1431static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1432{
1433 struct r1conf *conf = mddev->private;
1434 int err = 0;
1435 int number = rdev->raid_disk;
1436 struct mirror_info *p = conf->mirrors+ number;
1437
1438 if (rdev != p->rdev)
1439 p = conf->mirrors + conf->raid_disks + number;
1440
1441 print_conf(conf);
1442 if (rdev == p->rdev) {
1443 if (test_bit(In_sync, &rdev->flags) ||
1444 atomic_read(&rdev->nr_pending)) {
1445 err = -EBUSY;
1446 goto abort;
1447 }
1448 /* Only remove non-faulty devices if recovery
1449 * is not possible.
1450 */
1451 if (!test_bit(Faulty, &rdev->flags) &&
1452 mddev->recovery_disabled != conf->recovery_disabled &&
1453 mddev->degraded < conf->raid_disks) {
1454 err = -EBUSY;
1455 goto abort;
1456 }
1457 p->rdev = NULL;
1458 synchronize_rcu();
1459 if (atomic_read(&rdev->nr_pending)) {
1460 /* lost the race, try later */
1461 err = -EBUSY;
1462 p->rdev = rdev;
1463 goto abort;
1464 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1465 /* We just removed a device that is being replaced.
1466 * Move down the replacement. We drain all IO before
1467 * doing this to avoid confusion.
1468 */
1469 struct md_rdev *repl =
1470 conf->mirrors[conf->raid_disks + number].rdev;
1471 raise_barrier(conf);
1472 clear_bit(Replacement, &repl->flags);
1473 p->rdev = repl;
1474 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1475 lower_barrier(conf);
1476 clear_bit(WantReplacement, &rdev->flags);
1477 } else
1478 clear_bit(WantReplacement, &rdev->flags);
1479 err = md_integrity_register(mddev);
1480 }
1481abort:
1482
1483 print_conf(conf);
1484 return err;
1485}
1486
1487
1488static void end_sync_read(struct bio *bio, int error)
1489{
1490 struct r1bio *r1_bio = bio->bi_private;
1491
1492 update_head_pos(r1_bio->read_disk, r1_bio);
1493
1494 /*
1495 * we have read a block, now it needs to be re-written,
1496 * or re-read if the read failed.
1497 * We don't do much here, just schedule handling by raid1d
1498 */
1499 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1500 set_bit(R1BIO_Uptodate, &r1_bio->state);
1501
1502 if (atomic_dec_and_test(&r1_bio->remaining))
1503 reschedule_retry(r1_bio);
1504}
1505
1506static void end_sync_write(struct bio *bio, int error)
1507{
1508 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1509 struct r1bio *r1_bio = bio->bi_private;
1510 struct mddev *mddev = r1_bio->mddev;
1511 struct r1conf *conf = mddev->private;
1512 int mirror=0;
1513 sector_t first_bad;
1514 int bad_sectors;
1515
1516 mirror = find_bio_disk(r1_bio, bio);
1517
1518 if (!uptodate) {
1519 sector_t sync_blocks = 0;
1520 sector_t s = r1_bio->sector;
1521 long sectors_to_go = r1_bio->sectors;
1522 /* make sure these bits doesn't get cleared. */
1523 do {
1524 bitmap_end_sync(mddev->bitmap, s,
1525 &sync_blocks, 1);
1526 s += sync_blocks;
1527 sectors_to_go -= sync_blocks;
1528 } while (sectors_to_go > 0);
1529 set_bit(WriteErrorSeen,
1530 &conf->mirrors[mirror].rdev->flags);
1531 if (!test_and_set_bit(WantReplacement,
1532 &conf->mirrors[mirror].rdev->flags))
1533 set_bit(MD_RECOVERY_NEEDED, &
1534 mddev->recovery);
1535 set_bit(R1BIO_WriteError, &r1_bio->state);
1536 } else if (is_badblock(conf->mirrors[mirror].rdev,
1537 r1_bio->sector,
1538 r1_bio->sectors,
1539 &first_bad, &bad_sectors) &&
1540 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1541 r1_bio->sector,
1542 r1_bio->sectors,
1543 &first_bad, &bad_sectors)
1544 )
1545 set_bit(R1BIO_MadeGood, &r1_bio->state);
1546
1547 if (atomic_dec_and_test(&r1_bio->remaining)) {
1548 int s = r1_bio->sectors;
1549 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1550 test_bit(R1BIO_WriteError, &r1_bio->state))
1551 reschedule_retry(r1_bio);
1552 else {
1553 put_buf(r1_bio);
1554 md_done_sync(mddev, s, uptodate);
1555 }
1556 }
1557}
1558
1559static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1560 int sectors, struct page *page, int rw)
1561{
1562 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1563 /* success */
1564 return 1;
1565 if (rw == WRITE) {
1566 set_bit(WriteErrorSeen, &rdev->flags);
1567 if (!test_and_set_bit(WantReplacement,
1568 &rdev->flags))
1569 set_bit(MD_RECOVERY_NEEDED, &
1570 rdev->mddev->recovery);
1571 }
1572 /* need to record an error - either for the block or the device */
1573 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1574 md_error(rdev->mddev, rdev);
1575 return 0;
1576}
1577
1578static int fix_sync_read_error(struct r1bio *r1_bio)
1579{
1580 /* Try some synchronous reads of other devices to get
1581 * good data, much like with normal read errors. Only
1582 * read into the pages we already have so we don't
1583 * need to re-issue the read request.
1584 * We don't need to freeze the array, because being in an
1585 * active sync request, there is no normal IO, and
1586 * no overlapping syncs.
1587 * We don't need to check is_badblock() again as we
1588 * made sure that anything with a bad block in range
1589 * will have bi_end_io clear.
1590 */
1591 struct mddev *mddev = r1_bio->mddev;
1592 struct r1conf *conf = mddev->private;
1593 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1594 sector_t sect = r1_bio->sector;
1595 int sectors = r1_bio->sectors;
1596 int idx = 0;
1597
1598 while(sectors) {
1599 int s = sectors;
1600 int d = r1_bio->read_disk;
1601 int success = 0;
1602 struct md_rdev *rdev;
1603 int start;
1604
1605 if (s > (PAGE_SIZE>>9))
1606 s = PAGE_SIZE >> 9;
1607 do {
1608 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1609 /* No rcu protection needed here devices
1610 * can only be removed when no resync is
1611 * active, and resync is currently active
1612 */
1613 rdev = conf->mirrors[d].rdev;
1614 if (sync_page_io(rdev, sect, s<<9,
1615 bio->bi_io_vec[idx].bv_page,
1616 READ, false)) {
1617 success = 1;
1618 break;
1619 }
1620 }
1621 d++;
1622 if (d == conf->raid_disks * 2)
1623 d = 0;
1624 } while (!success && d != r1_bio->read_disk);
1625
1626 if (!success) {
1627 char b[BDEVNAME_SIZE];
1628 int abort = 0;
1629 /* Cannot read from anywhere, this block is lost.
1630 * Record a bad block on each device. If that doesn't
1631 * work just disable and interrupt the recovery.
1632 * Don't fail devices as that won't really help.
1633 */
1634 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1635 " for block %llu\n",
1636 mdname(mddev),
1637 bdevname(bio->bi_bdev, b),
1638 (unsigned long long)r1_bio->sector);
1639 for (d = 0; d < conf->raid_disks * 2; d++) {
1640 rdev = conf->mirrors[d].rdev;
1641 if (!rdev || test_bit(Faulty, &rdev->flags))
1642 continue;
1643 if (!rdev_set_badblocks(rdev, sect, s, 0))
1644 abort = 1;
1645 }
1646 if (abort) {
1647 conf->recovery_disabled =
1648 mddev->recovery_disabled;
1649 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1650 md_done_sync(mddev, r1_bio->sectors, 0);
1651 put_buf(r1_bio);
1652 return 0;
1653 }
1654 /* Try next page */
1655 sectors -= s;
1656 sect += s;
1657 idx++;
1658 continue;
1659 }
1660
1661 start = d;
1662 /* write it back and re-read */
1663 while (d != r1_bio->read_disk) {
1664 if (d == 0)
1665 d = conf->raid_disks * 2;
1666 d--;
1667 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1668 continue;
1669 rdev = conf->mirrors[d].rdev;
1670 if (r1_sync_page_io(rdev, sect, s,
1671 bio->bi_io_vec[idx].bv_page,
1672 WRITE) == 0) {
1673 r1_bio->bios[d]->bi_end_io = NULL;
1674 rdev_dec_pending(rdev, mddev);
1675 }
1676 }
1677 d = start;
1678 while (d != r1_bio->read_disk) {
1679 if (d == 0)
1680 d = conf->raid_disks * 2;
1681 d--;
1682 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1683 continue;
1684 rdev = conf->mirrors[d].rdev;
1685 if (r1_sync_page_io(rdev, sect, s,
1686 bio->bi_io_vec[idx].bv_page,
1687 READ) != 0)
1688 atomic_add(s, &rdev->corrected_errors);
1689 }
1690 sectors -= s;
1691 sect += s;
1692 idx ++;
1693 }
1694 set_bit(R1BIO_Uptodate, &r1_bio->state);
1695 set_bit(BIO_UPTODATE, &bio->bi_flags);
1696 return 1;
1697}
1698
1699static int process_checks(struct r1bio *r1_bio)
1700{
1701 /* We have read all readable devices. If we haven't
1702 * got the block, then there is no hope left.
1703 * If we have, then we want to do a comparison
1704 * and skip the write if everything is the same.
1705 * If any blocks failed to read, then we need to
1706 * attempt an over-write
1707 */
1708 struct mddev *mddev = r1_bio->mddev;
1709 struct r1conf *conf = mddev->private;
1710 int primary;
1711 int i;
1712 int vcnt;
1713
1714 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1715 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1716 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1717 r1_bio->bios[primary]->bi_end_io = NULL;
1718 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1719 break;
1720 }
1721 r1_bio->read_disk = primary;
1722 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1723 for (i = 0; i < conf->raid_disks * 2; i++) {
1724 int j;
1725 struct bio *pbio = r1_bio->bios[primary];
1726 struct bio *sbio = r1_bio->bios[i];
1727 int size;
1728
1729 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1730 continue;
1731
1732 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1733 for (j = vcnt; j-- ; ) {
1734 struct page *p, *s;
1735 p = pbio->bi_io_vec[j].bv_page;
1736 s = sbio->bi_io_vec[j].bv_page;
1737 if (memcmp(page_address(p),
1738 page_address(s),
1739 sbio->bi_io_vec[j].bv_len))
1740 break;
1741 }
1742 } else
1743 j = 0;
1744 if (j >= 0)
1745 mddev->resync_mismatches += r1_bio->sectors;
1746 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1747 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1748 /* No need to write to this device. */
1749 sbio->bi_end_io = NULL;
1750 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1751 continue;
1752 }
1753 /* fixup the bio for reuse */
1754 sbio->bi_vcnt = vcnt;
1755 sbio->bi_size = r1_bio->sectors << 9;
1756 sbio->bi_idx = 0;
1757 sbio->bi_phys_segments = 0;
1758 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1759 sbio->bi_flags |= 1 << BIO_UPTODATE;
1760 sbio->bi_next = NULL;
1761 sbio->bi_sector = r1_bio->sector +
1762 conf->mirrors[i].rdev->data_offset;
1763 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1764 size = sbio->bi_size;
1765 for (j = 0; j < vcnt ; j++) {
1766 struct bio_vec *bi;
1767 bi = &sbio->bi_io_vec[j];
1768 bi->bv_offset = 0;
1769 if (size > PAGE_SIZE)
1770 bi->bv_len = PAGE_SIZE;
1771 else
1772 bi->bv_len = size;
1773 size -= PAGE_SIZE;
1774 memcpy(page_address(bi->bv_page),
1775 page_address(pbio->bi_io_vec[j].bv_page),
1776 PAGE_SIZE);
1777 }
1778 }
1779 return 0;
1780}
1781
1782static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1783{
1784 struct r1conf *conf = mddev->private;
1785 int i;
1786 int disks = conf->raid_disks * 2;
1787 struct bio *bio, *wbio;
1788
1789 bio = r1_bio->bios[r1_bio->read_disk];
1790
1791 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1792 /* ouch - failed to read all of that. */
1793 if (!fix_sync_read_error(r1_bio))
1794 return;
1795
1796 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1797 if (process_checks(r1_bio) < 0)
1798 return;
1799 /*
1800 * schedule writes
1801 */
1802 atomic_set(&r1_bio->remaining, 1);
1803 for (i = 0; i < disks ; i++) {
1804 wbio = r1_bio->bios[i];
1805 if (wbio->bi_end_io == NULL ||
1806 (wbio->bi_end_io == end_sync_read &&
1807 (i == r1_bio->read_disk ||
1808 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1809 continue;
1810
1811 wbio->bi_rw = WRITE;
1812 wbio->bi_end_io = end_sync_write;
1813 atomic_inc(&r1_bio->remaining);
1814 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1815
1816 generic_make_request(wbio);
1817 }
1818
1819 if (atomic_dec_and_test(&r1_bio->remaining)) {
1820 /* if we're here, all write(s) have completed, so clean up */
1821 int s = r1_bio->sectors;
1822 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1823 test_bit(R1BIO_WriteError, &r1_bio->state))
1824 reschedule_retry(r1_bio);
1825 else {
1826 put_buf(r1_bio);
1827 md_done_sync(mddev, s, 1);
1828 }
1829 }
1830}
1831
1832/*
1833 * This is a kernel thread which:
1834 *
1835 * 1. Retries failed read operations on working mirrors.
1836 * 2. Updates the raid superblock when problems encounter.
1837 * 3. Performs writes following reads for array synchronising.
1838 */
1839
1840static void fix_read_error(struct r1conf *conf, int read_disk,
1841 sector_t sect, int sectors)
1842{
1843 struct mddev *mddev = conf->mddev;
1844 while(sectors) {
1845 int s = sectors;
1846 int d = read_disk;
1847 int success = 0;
1848 int start;
1849 struct md_rdev *rdev;
1850
1851 if (s > (PAGE_SIZE>>9))
1852 s = PAGE_SIZE >> 9;
1853
1854 do {
1855 /* Note: no rcu protection needed here
1856 * as this is synchronous in the raid1d thread
1857 * which is the thread that might remove
1858 * a device. If raid1d ever becomes multi-threaded....
1859 */
1860 sector_t first_bad;
1861 int bad_sectors;
1862
1863 rdev = conf->mirrors[d].rdev;
1864 if (rdev &&
1865 (test_bit(In_sync, &rdev->flags) ||
1866 (!test_bit(Faulty, &rdev->flags) &&
1867 rdev->recovery_offset >= sect + s)) &&
1868 is_badblock(rdev, sect, s,
1869 &first_bad, &bad_sectors) == 0 &&
1870 sync_page_io(rdev, sect, s<<9,
1871 conf->tmppage, READ, false))
1872 success = 1;
1873 else {
1874 d++;
1875 if (d == conf->raid_disks * 2)
1876 d = 0;
1877 }
1878 } while (!success && d != read_disk);
1879
1880 if (!success) {
1881 /* Cannot read from anywhere - mark it bad */
1882 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
1883 if (!rdev_set_badblocks(rdev, sect, s, 0))
1884 md_error(mddev, rdev);
1885 break;
1886 }
1887 /* write it back and re-read */
1888 start = d;
1889 while (d != read_disk) {
1890 if (d==0)
1891 d = conf->raid_disks * 2;
1892 d--;
1893 rdev = conf->mirrors[d].rdev;
1894 if (rdev &&
1895 test_bit(In_sync, &rdev->flags))
1896 r1_sync_page_io(rdev, sect, s,
1897 conf->tmppage, WRITE);
1898 }
1899 d = start;
1900 while (d != read_disk) {
1901 char b[BDEVNAME_SIZE];
1902 if (d==0)
1903 d = conf->raid_disks * 2;
1904 d--;
1905 rdev = conf->mirrors[d].rdev;
1906 if (rdev &&
1907 test_bit(In_sync, &rdev->flags)) {
1908 if (r1_sync_page_io(rdev, sect, s,
1909 conf->tmppage, READ)) {
1910 atomic_add(s, &rdev->corrected_errors);
1911 printk(KERN_INFO
1912 "md/raid1:%s: read error corrected "
1913 "(%d sectors at %llu on %s)\n",
1914 mdname(mddev), s,
1915 (unsigned long long)(sect +
1916 rdev->data_offset),
1917 bdevname(rdev->bdev, b));
1918 }
1919 }
1920 }
1921 sectors -= s;
1922 sect += s;
1923 }
1924}
1925
1926static void bi_complete(struct bio *bio, int error)
1927{
1928 complete((struct completion *)bio->bi_private);
1929}
1930
1931static int submit_bio_wait(int rw, struct bio *bio)
1932{
1933 struct completion event;
1934 rw |= REQ_SYNC;
1935
1936 init_completion(&event);
1937 bio->bi_private = &event;
1938 bio->bi_end_io = bi_complete;
1939 submit_bio(rw, bio);
1940 wait_for_completion(&event);
1941
1942 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1943}
1944
1945static int narrow_write_error(struct r1bio *r1_bio, int i)
1946{
1947 struct mddev *mddev = r1_bio->mddev;
1948 struct r1conf *conf = mddev->private;
1949 struct md_rdev *rdev = conf->mirrors[i].rdev;
1950 int vcnt, idx;
1951 struct bio_vec *vec;
1952
1953 /* bio has the data to be written to device 'i' where
1954 * we just recently had a write error.
1955 * We repeatedly clone the bio and trim down to one block,
1956 * then try the write. Where the write fails we record
1957 * a bad block.
1958 * It is conceivable that the bio doesn't exactly align with
1959 * blocks. We must handle this somehow.
1960 *
1961 * We currently own a reference on the rdev.
1962 */
1963
1964 int block_sectors;
1965 sector_t sector;
1966 int sectors;
1967 int sect_to_write = r1_bio->sectors;
1968 int ok = 1;
1969
1970 if (rdev->badblocks.shift < 0)
1971 return 0;
1972
1973 block_sectors = 1 << rdev->badblocks.shift;
1974 sector = r1_bio->sector;
1975 sectors = ((sector + block_sectors)
1976 & ~(sector_t)(block_sectors - 1))
1977 - sector;
1978
1979 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1980 vcnt = r1_bio->behind_page_count;
1981 vec = r1_bio->behind_bvecs;
1982 idx = 0;
1983 while (vec[idx].bv_page == NULL)
1984 idx++;
1985 } else {
1986 vcnt = r1_bio->master_bio->bi_vcnt;
1987 vec = r1_bio->master_bio->bi_io_vec;
1988 idx = r1_bio->master_bio->bi_idx;
1989 }
1990 while (sect_to_write) {
1991 struct bio *wbio;
1992 if (sectors > sect_to_write)
1993 sectors = sect_to_write;
1994 /* Write at 'sector' for 'sectors'*/
1995
1996 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1997 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1998 wbio->bi_sector = r1_bio->sector;
1999 wbio->bi_rw = WRITE;
2000 wbio->bi_vcnt = vcnt;
2001 wbio->bi_size = r1_bio->sectors << 9;
2002 wbio->bi_idx = idx;
2003
2004 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2005 wbio->bi_sector += rdev->data_offset;
2006 wbio->bi_bdev = rdev->bdev;
2007 if (submit_bio_wait(WRITE, wbio) == 0)
2008 /* failure! */
2009 ok = rdev_set_badblocks(rdev, sector,
2010 sectors, 0)
2011 && ok;
2012
2013 bio_put(wbio);
2014 sect_to_write -= sectors;
2015 sector += sectors;
2016 sectors = block_sectors;
2017 }
2018 return ok;
2019}
2020
2021static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2022{
2023 int m;
2024 int s = r1_bio->sectors;
2025 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2026 struct md_rdev *rdev = conf->mirrors[m].rdev;
2027 struct bio *bio = r1_bio->bios[m];
2028 if (bio->bi_end_io == NULL)
2029 continue;
2030 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2031 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2032 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2033 }
2034 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2035 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2036 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2037 md_error(conf->mddev, rdev);
2038 }
2039 }
2040 put_buf(r1_bio);
2041 md_done_sync(conf->mddev, s, 1);
2042}
2043
2044static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2045{
2046 int m;
2047 for (m = 0; m < conf->raid_disks * 2 ; m++)
2048 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2049 struct md_rdev *rdev = conf->mirrors[m].rdev;
2050 rdev_clear_badblocks(rdev,
2051 r1_bio->sector,
2052 r1_bio->sectors, 0);
2053 rdev_dec_pending(rdev, conf->mddev);
2054 } else if (r1_bio->bios[m] != NULL) {
2055 /* This drive got a write error. We need to
2056 * narrow down and record precise write
2057 * errors.
2058 */
2059 if (!narrow_write_error(r1_bio, m)) {
2060 md_error(conf->mddev,
2061 conf->mirrors[m].rdev);
2062 /* an I/O failed, we can't clear the bitmap */
2063 set_bit(R1BIO_Degraded, &r1_bio->state);
2064 }
2065 rdev_dec_pending(conf->mirrors[m].rdev,
2066 conf->mddev);
2067 }
2068 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2069 close_write(r1_bio);
2070 raid_end_bio_io(r1_bio);
2071}
2072
2073static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2074{
2075 int disk;
2076 int max_sectors;
2077 struct mddev *mddev = conf->mddev;
2078 struct bio *bio;
2079 char b[BDEVNAME_SIZE];
2080 struct md_rdev *rdev;
2081
2082 clear_bit(R1BIO_ReadError, &r1_bio->state);
2083 /* we got a read error. Maybe the drive is bad. Maybe just
2084 * the block and we can fix it.
2085 * We freeze all other IO, and try reading the block from
2086 * other devices. When we find one, we re-write
2087 * and check it that fixes the read error.
2088 * This is all done synchronously while the array is
2089 * frozen
2090 */
2091 if (mddev->ro == 0) {
2092 freeze_array(conf);
2093 fix_read_error(conf, r1_bio->read_disk,
2094 r1_bio->sector, r1_bio->sectors);
2095 unfreeze_array(conf);
2096 } else
2097 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2098
2099 bio = r1_bio->bios[r1_bio->read_disk];
2100 bdevname(bio->bi_bdev, b);
2101read_more:
2102 disk = read_balance(conf, r1_bio, &max_sectors);
2103 if (disk == -1) {
2104 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2105 " read error for block %llu\n",
2106 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2107 raid_end_bio_io(r1_bio);
2108 } else {
2109 const unsigned long do_sync
2110 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2111 if (bio) {
2112 r1_bio->bios[r1_bio->read_disk] =
2113 mddev->ro ? IO_BLOCKED : NULL;
2114 bio_put(bio);
2115 }
2116 r1_bio->read_disk = disk;
2117 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2118 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2119 r1_bio->bios[r1_bio->read_disk] = bio;
2120 rdev = conf->mirrors[disk].rdev;
2121 printk_ratelimited(KERN_ERR
2122 "md/raid1:%s: redirecting sector %llu"
2123 " to other mirror: %s\n",
2124 mdname(mddev),
2125 (unsigned long long)r1_bio->sector,
2126 bdevname(rdev->bdev, b));
2127 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2128 bio->bi_bdev = rdev->bdev;
2129 bio->bi_end_io = raid1_end_read_request;
2130 bio->bi_rw = READ | do_sync;
2131 bio->bi_private = r1_bio;
2132 if (max_sectors < r1_bio->sectors) {
2133 /* Drat - have to split this up more */
2134 struct bio *mbio = r1_bio->master_bio;
2135 int sectors_handled = (r1_bio->sector + max_sectors
2136 - mbio->bi_sector);
2137 r1_bio->sectors = max_sectors;
2138 spin_lock_irq(&conf->device_lock);
2139 if (mbio->bi_phys_segments == 0)
2140 mbio->bi_phys_segments = 2;
2141 else
2142 mbio->bi_phys_segments++;
2143 spin_unlock_irq(&conf->device_lock);
2144 generic_make_request(bio);
2145 bio = NULL;
2146
2147 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2148
2149 r1_bio->master_bio = mbio;
2150 r1_bio->sectors = (mbio->bi_size >> 9)
2151 - sectors_handled;
2152 r1_bio->state = 0;
2153 set_bit(R1BIO_ReadError, &r1_bio->state);
2154 r1_bio->mddev = mddev;
2155 r1_bio->sector = mbio->bi_sector + sectors_handled;
2156
2157 goto read_more;
2158 } else
2159 generic_make_request(bio);
2160 }
2161}
2162
2163static void raid1d(struct mddev *mddev)
2164{
2165 struct r1bio *r1_bio;
2166 unsigned long flags;
2167 struct r1conf *conf = mddev->private;
2168 struct list_head *head = &conf->retry_list;
2169 struct blk_plug plug;
2170
2171 md_check_recovery(mddev);
2172
2173 blk_start_plug(&plug);
2174 for (;;) {
2175
2176 if (atomic_read(&mddev->plug_cnt) == 0)
2177 flush_pending_writes(conf);
2178
2179 spin_lock_irqsave(&conf->device_lock, flags);
2180 if (list_empty(head)) {
2181 spin_unlock_irqrestore(&conf->device_lock, flags);
2182 break;
2183 }
2184 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2185 list_del(head->prev);
2186 conf->nr_queued--;
2187 spin_unlock_irqrestore(&conf->device_lock, flags);
2188
2189 mddev = r1_bio->mddev;
2190 conf = mddev->private;
2191 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2192 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2193 test_bit(R1BIO_WriteError, &r1_bio->state))
2194 handle_sync_write_finished(conf, r1_bio);
2195 else
2196 sync_request_write(mddev, r1_bio);
2197 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2198 test_bit(R1BIO_WriteError, &r1_bio->state))
2199 handle_write_finished(conf, r1_bio);
2200 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2201 handle_read_error(conf, r1_bio);
2202 else
2203 /* just a partial read to be scheduled from separate
2204 * context
2205 */
2206 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2207
2208 cond_resched();
2209 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2210 md_check_recovery(mddev);
2211 }
2212 blk_finish_plug(&plug);
2213}
2214
2215
2216static int init_resync(struct r1conf *conf)
2217{
2218 int buffs;
2219
2220 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2221 BUG_ON(conf->r1buf_pool);
2222 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2223 conf->poolinfo);
2224 if (!conf->r1buf_pool)
2225 return -ENOMEM;
2226 conf->next_resync = 0;
2227 return 0;
2228}
2229
2230/*
2231 * perform a "sync" on one "block"
2232 *
2233 * We need to make sure that no normal I/O request - particularly write
2234 * requests - conflict with active sync requests.
2235 *
2236 * This is achieved by tracking pending requests and a 'barrier' concept
2237 * that can be installed to exclude normal IO requests.
2238 */
2239
2240static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2241{
2242 struct r1conf *conf = mddev->private;
2243 struct r1bio *r1_bio;
2244 struct bio *bio;
2245 sector_t max_sector, nr_sectors;
2246 int disk = -1;
2247 int i;
2248 int wonly = -1;
2249 int write_targets = 0, read_targets = 0;
2250 sector_t sync_blocks;
2251 int still_degraded = 0;
2252 int good_sectors = RESYNC_SECTORS;
2253 int min_bad = 0; /* number of sectors that are bad in all devices */
2254
2255 if (!conf->r1buf_pool)
2256 if (init_resync(conf))
2257 return 0;
2258
2259 max_sector = mddev->dev_sectors;
2260 if (sector_nr >= max_sector) {
2261 /* If we aborted, we need to abort the
2262 * sync on the 'current' bitmap chunk (there will
2263 * only be one in raid1 resync.
2264 * We can find the current addess in mddev->curr_resync
2265 */
2266 if (mddev->curr_resync < max_sector) /* aborted */
2267 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2268 &sync_blocks, 1);
2269 else /* completed sync */
2270 conf->fullsync = 0;
2271
2272 bitmap_close_sync(mddev->bitmap);
2273 close_sync(conf);
2274 return 0;
2275 }
2276
2277 if (mddev->bitmap == NULL &&
2278 mddev->recovery_cp == MaxSector &&
2279 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2280 conf->fullsync == 0) {
2281 *skipped = 1;
2282 return max_sector - sector_nr;
2283 }
2284 /* before building a request, check if we can skip these blocks..
2285 * This call the bitmap_start_sync doesn't actually record anything
2286 */
2287 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2288 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2289 /* We can skip this block, and probably several more */
2290 *skipped = 1;
2291 return sync_blocks;
2292 }
2293 /*
2294 * If there is non-resync activity waiting for a turn,
2295 * and resync is going fast enough,
2296 * then let it though before starting on this new sync request.
2297 */
2298 if (!go_faster && conf->nr_waiting)
2299 msleep_interruptible(1000);
2300
2301 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2302 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2303 raise_barrier(conf);
2304
2305 conf->next_resync = sector_nr;
2306
2307 rcu_read_lock();
2308 /*
2309 * If we get a correctably read error during resync or recovery,
2310 * we might want to read from a different device. So we
2311 * flag all drives that could conceivably be read from for READ,
2312 * and any others (which will be non-In_sync devices) for WRITE.
2313 * If a read fails, we try reading from something else for which READ
2314 * is OK.
2315 */
2316
2317 r1_bio->mddev = mddev;
2318 r1_bio->sector = sector_nr;
2319 r1_bio->state = 0;
2320 set_bit(R1BIO_IsSync, &r1_bio->state);
2321
2322 for (i = 0; i < conf->raid_disks * 2; i++) {
2323 struct md_rdev *rdev;
2324 bio = r1_bio->bios[i];
2325
2326 /* take from bio_init */
2327 bio->bi_next = NULL;
2328 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2329 bio->bi_flags |= 1 << BIO_UPTODATE;
2330 bio->bi_rw = READ;
2331 bio->bi_vcnt = 0;
2332 bio->bi_idx = 0;
2333 bio->bi_phys_segments = 0;
2334 bio->bi_size = 0;
2335 bio->bi_end_io = NULL;
2336 bio->bi_private = NULL;
2337
2338 rdev = rcu_dereference(conf->mirrors[i].rdev);
2339 if (rdev == NULL ||
2340 test_bit(Faulty, &rdev->flags)) {
2341 if (i < conf->raid_disks)
2342 still_degraded = 1;
2343 } else if (!test_bit(In_sync, &rdev->flags)) {
2344 bio->bi_rw = WRITE;
2345 bio->bi_end_io = end_sync_write;
2346 write_targets ++;
2347 } else {
2348 /* may need to read from here */
2349 sector_t first_bad = MaxSector;
2350 int bad_sectors;
2351
2352 if (is_badblock(rdev, sector_nr, good_sectors,
2353 &first_bad, &bad_sectors)) {
2354 if (first_bad > sector_nr)
2355 good_sectors = first_bad - sector_nr;
2356 else {
2357 bad_sectors -= (sector_nr - first_bad);
2358 if (min_bad == 0 ||
2359 min_bad > bad_sectors)
2360 min_bad = bad_sectors;
2361 }
2362 }
2363 if (sector_nr < first_bad) {
2364 if (test_bit(WriteMostly, &rdev->flags)) {
2365 if (wonly < 0)
2366 wonly = i;
2367 } else {
2368 if (disk < 0)
2369 disk = i;
2370 }
2371 bio->bi_rw = READ;
2372 bio->bi_end_io = end_sync_read;
2373 read_targets++;
2374 }
2375 }
2376 if (bio->bi_end_io) {
2377 atomic_inc(&rdev->nr_pending);
2378 bio->bi_sector = sector_nr + rdev->data_offset;
2379 bio->bi_bdev = rdev->bdev;
2380 bio->bi_private = r1_bio;
2381 }
2382 }
2383 rcu_read_unlock();
2384 if (disk < 0)
2385 disk = wonly;
2386 r1_bio->read_disk = disk;
2387
2388 if (read_targets == 0 && min_bad > 0) {
2389 /* These sectors are bad on all InSync devices, so we
2390 * need to mark them bad on all write targets
2391 */
2392 int ok = 1;
2393 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2394 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2395 struct md_rdev *rdev = conf->mirrors[i].rdev;
2396 ok = rdev_set_badblocks(rdev, sector_nr,
2397 min_bad, 0
2398 ) && ok;
2399 }
2400 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2401 *skipped = 1;
2402 put_buf(r1_bio);
2403
2404 if (!ok) {
2405 /* Cannot record the badblocks, so need to
2406 * abort the resync.
2407 * If there are multiple read targets, could just
2408 * fail the really bad ones ???
2409 */
2410 conf->recovery_disabled = mddev->recovery_disabled;
2411 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2412 return 0;
2413 } else
2414 return min_bad;
2415
2416 }
2417 if (min_bad > 0 && min_bad < good_sectors) {
2418 /* only resync enough to reach the next bad->good
2419 * transition */
2420 good_sectors = min_bad;
2421 }
2422
2423 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2424 /* extra read targets are also write targets */
2425 write_targets += read_targets-1;
2426
2427 if (write_targets == 0 || read_targets == 0) {
2428 /* There is nowhere to write, so all non-sync
2429 * drives must be failed - so we are finished
2430 */
2431 sector_t rv;
2432 if (min_bad > 0)
2433 max_sector = sector_nr + min_bad;
2434 rv = max_sector - sector_nr;
2435 *skipped = 1;
2436 put_buf(r1_bio);
2437 return rv;
2438 }
2439
2440 if (max_sector > mddev->resync_max)
2441 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2442 if (max_sector > sector_nr + good_sectors)
2443 max_sector = sector_nr + good_sectors;
2444 nr_sectors = 0;
2445 sync_blocks = 0;
2446 do {
2447 struct page *page;
2448 int len = PAGE_SIZE;
2449 if (sector_nr + (len>>9) > max_sector)
2450 len = (max_sector - sector_nr) << 9;
2451 if (len == 0)
2452 break;
2453 if (sync_blocks == 0) {
2454 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2455 &sync_blocks, still_degraded) &&
2456 !conf->fullsync &&
2457 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2458 break;
2459 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2460 if ((len >> 9) > sync_blocks)
2461 len = sync_blocks<<9;
2462 }
2463
2464 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2465 bio = r1_bio->bios[i];
2466 if (bio->bi_end_io) {
2467 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2468 if (bio_add_page(bio, page, len, 0) == 0) {
2469 /* stop here */
2470 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2471 while (i > 0) {
2472 i--;
2473 bio = r1_bio->bios[i];
2474 if (bio->bi_end_io==NULL)
2475 continue;
2476 /* remove last page from this bio */
2477 bio->bi_vcnt--;
2478 bio->bi_size -= len;
2479 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2480 }
2481 goto bio_full;
2482 }
2483 }
2484 }
2485 nr_sectors += len>>9;
2486 sector_nr += len>>9;
2487 sync_blocks -= (len>>9);
2488 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2489 bio_full:
2490 r1_bio->sectors = nr_sectors;
2491
2492 /* For a user-requested sync, we read all readable devices and do a
2493 * compare
2494 */
2495 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2496 atomic_set(&r1_bio->remaining, read_targets);
2497 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2498 bio = r1_bio->bios[i];
2499 if (bio->bi_end_io == end_sync_read) {
2500 read_targets--;
2501 md_sync_acct(bio->bi_bdev, nr_sectors);
2502 generic_make_request(bio);
2503 }
2504 }
2505 } else {
2506 atomic_set(&r1_bio->remaining, 1);
2507 bio = r1_bio->bios[r1_bio->read_disk];
2508 md_sync_acct(bio->bi_bdev, nr_sectors);
2509 generic_make_request(bio);
2510
2511 }
2512 return nr_sectors;
2513}
2514
2515static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2516{
2517 if (sectors)
2518 return sectors;
2519
2520 return mddev->dev_sectors;
2521}
2522
2523static struct r1conf *setup_conf(struct mddev *mddev)
2524{
2525 struct r1conf *conf;
2526 int i;
2527 struct mirror_info *disk;
2528 struct md_rdev *rdev;
2529 int err = -ENOMEM;
2530
2531 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2532 if (!conf)
2533 goto abort;
2534
2535 conf->mirrors = kzalloc(sizeof(struct mirror_info)
2536 * mddev->raid_disks * 2,
2537 GFP_KERNEL);
2538 if (!conf->mirrors)
2539 goto abort;
2540
2541 conf->tmppage = alloc_page(GFP_KERNEL);
2542 if (!conf->tmppage)
2543 goto abort;
2544
2545 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2546 if (!conf->poolinfo)
2547 goto abort;
2548 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2549 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2550 r1bio_pool_free,
2551 conf->poolinfo);
2552 if (!conf->r1bio_pool)
2553 goto abort;
2554
2555 conf->poolinfo->mddev = mddev;
2556
2557 err = -EINVAL;
2558 spin_lock_init(&conf->device_lock);
2559 rdev_for_each(rdev, mddev) {
2560 struct request_queue *q;
2561 int disk_idx = rdev->raid_disk;
2562 if (disk_idx >= mddev->raid_disks
2563 || disk_idx < 0)
2564 continue;
2565 if (test_bit(Replacement, &rdev->flags))
2566 disk = conf->mirrors + conf->raid_disks + disk_idx;
2567 else
2568 disk = conf->mirrors + disk_idx;
2569
2570 if (disk->rdev)
2571 goto abort;
2572 disk->rdev = rdev;
2573 q = bdev_get_queue(rdev->bdev);
2574 if (q->merge_bvec_fn)
2575 mddev->merge_check_needed = 1;
2576
2577 disk->head_position = 0;
2578 }
2579 conf->raid_disks = mddev->raid_disks;
2580 conf->mddev = mddev;
2581 INIT_LIST_HEAD(&conf->retry_list);
2582
2583 spin_lock_init(&conf->resync_lock);
2584 init_waitqueue_head(&conf->wait_barrier);
2585
2586 bio_list_init(&conf->pending_bio_list);
2587 conf->pending_count = 0;
2588 conf->recovery_disabled = mddev->recovery_disabled - 1;
2589
2590 err = -EIO;
2591 conf->last_used = -1;
2592 for (i = 0; i < conf->raid_disks * 2; i++) {
2593
2594 disk = conf->mirrors + i;
2595
2596 if (i < conf->raid_disks &&
2597 disk[conf->raid_disks].rdev) {
2598 /* This slot has a replacement. */
2599 if (!disk->rdev) {
2600 /* No original, just make the replacement
2601 * a recovering spare
2602 */
2603 disk->rdev =
2604 disk[conf->raid_disks].rdev;
2605 disk[conf->raid_disks].rdev = NULL;
2606 } else if (!test_bit(In_sync, &disk->rdev->flags))
2607 /* Original is not in_sync - bad */
2608 goto abort;
2609 }
2610
2611 if (!disk->rdev ||
2612 !test_bit(In_sync, &disk->rdev->flags)) {
2613 disk->head_position = 0;
2614 if (disk->rdev &&
2615 (disk->rdev->saved_raid_disk < 0))
2616 conf->fullsync = 1;
2617 } else if (conf->last_used < 0)
2618 /*
2619 * The first working device is used as a
2620 * starting point to read balancing.
2621 */
2622 conf->last_used = i;
2623 }
2624
2625 if (conf->last_used < 0) {
2626 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2627 mdname(mddev));
2628 goto abort;
2629 }
2630 err = -ENOMEM;
2631 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2632 if (!conf->thread) {
2633 printk(KERN_ERR
2634 "md/raid1:%s: couldn't allocate thread\n",
2635 mdname(mddev));
2636 goto abort;
2637 }
2638
2639 return conf;
2640
2641 abort:
2642 if (conf) {
2643 if (conf->r1bio_pool)
2644 mempool_destroy(conf->r1bio_pool);
2645 kfree(conf->mirrors);
2646 safe_put_page(conf->tmppage);
2647 kfree(conf->poolinfo);
2648 kfree(conf);
2649 }
2650 return ERR_PTR(err);
2651}
2652
2653static int stop(struct mddev *mddev);
2654static int run(struct mddev *mddev)
2655{
2656 struct r1conf *conf;
2657 int i;
2658 struct md_rdev *rdev;
2659 int ret;
2660
2661 if (mddev->level != 1) {
2662 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2663 mdname(mddev), mddev->level);
2664 return -EIO;
2665 }
2666 if (mddev->reshape_position != MaxSector) {
2667 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2668 mdname(mddev));
2669 return -EIO;
2670 }
2671 /*
2672 * copy the already verified devices into our private RAID1
2673 * bookkeeping area. [whatever we allocate in run(),
2674 * should be freed in stop()]
2675 */
2676 if (mddev->private == NULL)
2677 conf = setup_conf(mddev);
2678 else
2679 conf = mddev->private;
2680
2681 if (IS_ERR(conf))
2682 return PTR_ERR(conf);
2683
2684 rdev_for_each(rdev, mddev) {
2685 if (!mddev->gendisk)
2686 continue;
2687 disk_stack_limits(mddev->gendisk, rdev->bdev,
2688 rdev->data_offset << 9);
2689 }
2690
2691 mddev->degraded = 0;
2692 for (i=0; i < conf->raid_disks; i++)
2693 if (conf->mirrors[i].rdev == NULL ||
2694 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2695 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2696 mddev->degraded++;
2697
2698 if (conf->raid_disks - mddev->degraded == 1)
2699 mddev->recovery_cp = MaxSector;
2700
2701 if (mddev->recovery_cp != MaxSector)
2702 printk(KERN_NOTICE "md/raid1:%s: not clean"
2703 " -- starting background reconstruction\n",
2704 mdname(mddev));
2705 printk(KERN_INFO
2706 "md/raid1:%s: active with %d out of %d mirrors\n",
2707 mdname(mddev), mddev->raid_disks - mddev->degraded,
2708 mddev->raid_disks);
2709
2710 /*
2711 * Ok, everything is just fine now
2712 */
2713 mddev->thread = conf->thread;
2714 conf->thread = NULL;
2715 mddev->private = conf;
2716
2717 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2718
2719 if (mddev->queue) {
2720 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2721 mddev->queue->backing_dev_info.congested_data = mddev;
2722 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2723 }
2724
2725 ret = md_integrity_register(mddev);
2726 if (ret)
2727 stop(mddev);
2728 return ret;
2729}
2730
2731static int stop(struct mddev *mddev)
2732{
2733 struct r1conf *conf = mddev->private;
2734 struct bitmap *bitmap = mddev->bitmap;
2735
2736 /* wait for behind writes to complete */
2737 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2738 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2739 mdname(mddev));
2740 /* need to kick something here to make sure I/O goes? */
2741 wait_event(bitmap->behind_wait,
2742 atomic_read(&bitmap->behind_writes) == 0);
2743 }
2744
2745 raise_barrier(conf);
2746 lower_barrier(conf);
2747
2748 md_unregister_thread(&mddev->thread);
2749 if (conf->r1bio_pool)
2750 mempool_destroy(conf->r1bio_pool);
2751 kfree(conf->mirrors);
2752 kfree(conf->poolinfo);
2753 kfree(conf);
2754 mddev->private = NULL;
2755 return 0;
2756}
2757
2758static int raid1_resize(struct mddev *mddev, sector_t sectors)
2759{
2760 /* no resync is happening, and there is enough space
2761 * on all devices, so we can resize.
2762 * We need to make sure resync covers any new space.
2763 * If the array is shrinking we should possibly wait until
2764 * any io in the removed space completes, but it hardly seems
2765 * worth it.
2766 */
2767 sector_t newsize = raid1_size(mddev, sectors, 0);
2768 if (mddev->external_size &&
2769 mddev->array_sectors > newsize)
2770 return -EINVAL;
2771 if (mddev->bitmap) {
2772 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2773 if (ret)
2774 return ret;
2775 }
2776 md_set_array_sectors(mddev, newsize);
2777 set_capacity(mddev->gendisk, mddev->array_sectors);
2778 revalidate_disk(mddev->gendisk);
2779 if (sectors > mddev->dev_sectors &&
2780 mddev->recovery_cp > mddev->dev_sectors) {
2781 mddev->recovery_cp = mddev->dev_sectors;
2782 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2783 }
2784 mddev->dev_sectors = sectors;
2785 mddev->resync_max_sectors = sectors;
2786 return 0;
2787}
2788
2789static int raid1_reshape(struct mddev *mddev)
2790{
2791 /* We need to:
2792 * 1/ resize the r1bio_pool
2793 * 2/ resize conf->mirrors
2794 *
2795 * We allocate a new r1bio_pool if we can.
2796 * Then raise a device barrier and wait until all IO stops.
2797 * Then resize conf->mirrors and swap in the new r1bio pool.
2798 *
2799 * At the same time, we "pack" the devices so that all the missing
2800 * devices have the higher raid_disk numbers.
2801 */
2802 mempool_t *newpool, *oldpool;
2803 struct pool_info *newpoolinfo;
2804 struct mirror_info *newmirrors;
2805 struct r1conf *conf = mddev->private;
2806 int cnt, raid_disks;
2807 unsigned long flags;
2808 int d, d2, err;
2809
2810 /* Cannot change chunk_size, layout, or level */
2811 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2812 mddev->layout != mddev->new_layout ||
2813 mddev->level != mddev->new_level) {
2814 mddev->new_chunk_sectors = mddev->chunk_sectors;
2815 mddev->new_layout = mddev->layout;
2816 mddev->new_level = mddev->level;
2817 return -EINVAL;
2818 }
2819
2820 err = md_allow_write(mddev);
2821 if (err)
2822 return err;
2823
2824 raid_disks = mddev->raid_disks + mddev->delta_disks;
2825
2826 if (raid_disks < conf->raid_disks) {
2827 cnt=0;
2828 for (d= 0; d < conf->raid_disks; d++)
2829 if (conf->mirrors[d].rdev)
2830 cnt++;
2831 if (cnt > raid_disks)
2832 return -EBUSY;
2833 }
2834
2835 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2836 if (!newpoolinfo)
2837 return -ENOMEM;
2838 newpoolinfo->mddev = mddev;
2839 newpoolinfo->raid_disks = raid_disks * 2;
2840
2841 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2842 r1bio_pool_free, newpoolinfo);
2843 if (!newpool) {
2844 kfree(newpoolinfo);
2845 return -ENOMEM;
2846 }
2847 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks * 2,
2848 GFP_KERNEL);
2849 if (!newmirrors) {
2850 kfree(newpoolinfo);
2851 mempool_destroy(newpool);
2852 return -ENOMEM;
2853 }
2854
2855 raise_barrier(conf);
2856
2857 /* ok, everything is stopped */
2858 oldpool = conf->r1bio_pool;
2859 conf->r1bio_pool = newpool;
2860
2861 for (d = d2 = 0; d < conf->raid_disks; d++) {
2862 struct md_rdev *rdev = conf->mirrors[d].rdev;
2863 if (rdev && rdev->raid_disk != d2) {
2864 sysfs_unlink_rdev(mddev, rdev);
2865 rdev->raid_disk = d2;
2866 sysfs_unlink_rdev(mddev, rdev);
2867 if (sysfs_link_rdev(mddev, rdev))
2868 printk(KERN_WARNING
2869 "md/raid1:%s: cannot register rd%d\n",
2870 mdname(mddev), rdev->raid_disk);
2871 }
2872 if (rdev)
2873 newmirrors[d2++].rdev = rdev;
2874 }
2875 kfree(conf->mirrors);
2876 conf->mirrors = newmirrors;
2877 kfree(conf->poolinfo);
2878 conf->poolinfo = newpoolinfo;
2879
2880 spin_lock_irqsave(&conf->device_lock, flags);
2881 mddev->degraded += (raid_disks - conf->raid_disks);
2882 spin_unlock_irqrestore(&conf->device_lock, flags);
2883 conf->raid_disks = mddev->raid_disks = raid_disks;
2884 mddev->delta_disks = 0;
2885
2886 conf->last_used = 0; /* just make sure it is in-range */
2887 lower_barrier(conf);
2888
2889 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2890 md_wakeup_thread(mddev->thread);
2891
2892 mempool_destroy(oldpool);
2893 return 0;
2894}
2895
2896static void raid1_quiesce(struct mddev *mddev, int state)
2897{
2898 struct r1conf *conf = mddev->private;
2899
2900 switch(state) {
2901 case 2: /* wake for suspend */
2902 wake_up(&conf->wait_barrier);
2903 break;
2904 case 1:
2905 raise_barrier(conf);
2906 break;
2907 case 0:
2908 lower_barrier(conf);
2909 break;
2910 }
2911}
2912
2913static void *raid1_takeover(struct mddev *mddev)
2914{
2915 /* raid1 can take over:
2916 * raid5 with 2 devices, any layout or chunk size
2917 */
2918 if (mddev->level == 5 && mddev->raid_disks == 2) {
2919 struct r1conf *conf;
2920 mddev->new_level = 1;
2921 mddev->new_layout = 0;
2922 mddev->new_chunk_sectors = 0;
2923 conf = setup_conf(mddev);
2924 if (!IS_ERR(conf))
2925 conf->barrier = 1;
2926 return conf;
2927 }
2928 return ERR_PTR(-EINVAL);
2929}
2930
2931static struct md_personality raid1_personality =
2932{
2933 .name = "raid1",
2934 .level = 1,
2935 .owner = THIS_MODULE,
2936 .make_request = make_request,
2937 .run = run,
2938 .stop = stop,
2939 .status = status,
2940 .error_handler = error,
2941 .hot_add_disk = raid1_add_disk,
2942 .hot_remove_disk= raid1_remove_disk,
2943 .spare_active = raid1_spare_active,
2944 .sync_request = sync_request,
2945 .resize = raid1_resize,
2946 .size = raid1_size,
2947 .check_reshape = raid1_reshape,
2948 .quiesce = raid1_quiesce,
2949 .takeover = raid1_takeover,
2950};
2951
2952static int __init raid_init(void)
2953{
2954 return register_md_personality(&raid1_personality);
2955}
2956
2957static void raid_exit(void)
2958{
2959 unregister_md_personality(&raid1_personality);
2960}
2961
2962module_init(raid_init);
2963module_exit(raid_exit);
2964MODULE_LICENSE("GPL");
2965MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2966MODULE_ALIAS("md-personality-3"); /* RAID1 */
2967MODULE_ALIAS("md-raid1");
2968MODULE_ALIAS("md-level-1");
2969
2970module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid1.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 *
7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 *
9 * RAID-1 management functions.
10 *
11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 *
13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 *
16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
17 * bitmapped intelligence in resync:
18 *
19 * - bitmap marked during normal i/o
20 * - bitmap used to skip nondirty blocks during sync
21 *
22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
23 * - persistent bitmap code
24 */
25
26#include <linux/slab.h>
27#include <linux/delay.h>
28#include <linux/blkdev.h>
29#include <linux/module.h>
30#include <linux/seq_file.h>
31#include <linux/ratelimit.h>
32#include <linux/interval_tree_generic.h>
33
34#include <trace/events/block.h>
35
36#include "md.h"
37#include "raid1.h"
38#include "md-bitmap.h"
39
40#define UNSUPPORTED_MDDEV_FLAGS \
41 ((1L << MD_HAS_JOURNAL) | \
42 (1L << MD_JOURNAL_CLEAN) | \
43 (1L << MD_HAS_PPL) | \
44 (1L << MD_HAS_MULTIPLE_PPLS))
45
46static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
47static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
48
49#define raid1_log(md, fmt, args...) \
50 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
51
52#include "raid1-10.c"
53
54#define START(node) ((node)->start)
55#define LAST(node) ((node)->last)
56INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
57 START, LAST, static inline, raid1_rb);
58
59static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
60 struct serial_info *si, int idx)
61{
62 unsigned long flags;
63 int ret = 0;
64 sector_t lo = r1_bio->sector;
65 sector_t hi = lo + r1_bio->sectors;
66 struct serial_in_rdev *serial = &rdev->serial[idx];
67
68 spin_lock_irqsave(&serial->serial_lock, flags);
69 /* collision happened */
70 if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
71 ret = -EBUSY;
72 else {
73 si->start = lo;
74 si->last = hi;
75 raid1_rb_insert(si, &serial->serial_rb);
76 }
77 spin_unlock_irqrestore(&serial->serial_lock, flags);
78
79 return ret;
80}
81
82static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
83{
84 struct mddev *mddev = rdev->mddev;
85 struct serial_info *si;
86 int idx = sector_to_idx(r1_bio->sector);
87 struct serial_in_rdev *serial = &rdev->serial[idx];
88
89 if (WARN_ON(!mddev->serial_info_pool))
90 return;
91 si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
92 wait_event(serial->serial_io_wait,
93 check_and_add_serial(rdev, r1_bio, si, idx) == 0);
94}
95
96static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
97{
98 struct serial_info *si;
99 unsigned long flags;
100 int found = 0;
101 struct mddev *mddev = rdev->mddev;
102 int idx = sector_to_idx(lo);
103 struct serial_in_rdev *serial = &rdev->serial[idx];
104
105 spin_lock_irqsave(&serial->serial_lock, flags);
106 for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
107 si; si = raid1_rb_iter_next(si, lo, hi)) {
108 if (si->start == lo && si->last == hi) {
109 raid1_rb_remove(si, &serial->serial_rb);
110 mempool_free(si, mddev->serial_info_pool);
111 found = 1;
112 break;
113 }
114 }
115 if (!found)
116 WARN(1, "The write IO is not recorded for serialization\n");
117 spin_unlock_irqrestore(&serial->serial_lock, flags);
118 wake_up(&serial->serial_io_wait);
119}
120
121/*
122 * for resync bio, r1bio pointer can be retrieved from the per-bio
123 * 'struct resync_pages'.
124 */
125static inline struct r1bio *get_resync_r1bio(struct bio *bio)
126{
127 return get_resync_pages(bio)->raid_bio;
128}
129
130static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
131{
132 struct pool_info *pi = data;
133 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
134
135 /* allocate a r1bio with room for raid_disks entries in the bios array */
136 return kzalloc(size, gfp_flags);
137}
138
139#define RESYNC_DEPTH 32
140#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
141#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
142#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
143#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
144#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
145
146static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
147{
148 struct pool_info *pi = data;
149 struct r1bio *r1_bio;
150 struct bio *bio;
151 int need_pages;
152 int j;
153 struct resync_pages *rps;
154
155 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
156 if (!r1_bio)
157 return NULL;
158
159 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
160 gfp_flags);
161 if (!rps)
162 goto out_free_r1bio;
163
164 /*
165 * Allocate bios : 1 for reading, n-1 for writing
166 */
167 for (j = pi->raid_disks ; j-- ; ) {
168 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
169 if (!bio)
170 goto out_free_bio;
171 r1_bio->bios[j] = bio;
172 }
173 /*
174 * Allocate RESYNC_PAGES data pages and attach them to
175 * the first bio.
176 * If this is a user-requested check/repair, allocate
177 * RESYNC_PAGES for each bio.
178 */
179 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
180 need_pages = pi->raid_disks;
181 else
182 need_pages = 1;
183 for (j = 0; j < pi->raid_disks; j++) {
184 struct resync_pages *rp = &rps[j];
185
186 bio = r1_bio->bios[j];
187
188 if (j < need_pages) {
189 if (resync_alloc_pages(rp, gfp_flags))
190 goto out_free_pages;
191 } else {
192 memcpy(rp, &rps[0], sizeof(*rp));
193 resync_get_all_pages(rp);
194 }
195
196 rp->raid_bio = r1_bio;
197 bio->bi_private = rp;
198 }
199
200 r1_bio->master_bio = NULL;
201
202 return r1_bio;
203
204out_free_pages:
205 while (--j >= 0)
206 resync_free_pages(&rps[j]);
207
208out_free_bio:
209 while (++j < pi->raid_disks)
210 bio_put(r1_bio->bios[j]);
211 kfree(rps);
212
213out_free_r1bio:
214 rbio_pool_free(r1_bio, data);
215 return NULL;
216}
217
218static void r1buf_pool_free(void *__r1_bio, void *data)
219{
220 struct pool_info *pi = data;
221 int i;
222 struct r1bio *r1bio = __r1_bio;
223 struct resync_pages *rp = NULL;
224
225 for (i = pi->raid_disks; i--; ) {
226 rp = get_resync_pages(r1bio->bios[i]);
227 resync_free_pages(rp);
228 bio_put(r1bio->bios[i]);
229 }
230
231 /* resync pages array stored in the 1st bio's .bi_private */
232 kfree(rp);
233
234 rbio_pool_free(r1bio, data);
235}
236
237static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
238{
239 int i;
240
241 for (i = 0; i < conf->raid_disks * 2; i++) {
242 struct bio **bio = r1_bio->bios + i;
243 if (!BIO_SPECIAL(*bio))
244 bio_put(*bio);
245 *bio = NULL;
246 }
247}
248
249static void free_r1bio(struct r1bio *r1_bio)
250{
251 struct r1conf *conf = r1_bio->mddev->private;
252
253 put_all_bios(conf, r1_bio);
254 mempool_free(r1_bio, &conf->r1bio_pool);
255}
256
257static void put_buf(struct r1bio *r1_bio)
258{
259 struct r1conf *conf = r1_bio->mddev->private;
260 sector_t sect = r1_bio->sector;
261 int i;
262
263 for (i = 0; i < conf->raid_disks * 2; i++) {
264 struct bio *bio = r1_bio->bios[i];
265 if (bio->bi_end_io)
266 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
267 }
268
269 mempool_free(r1_bio, &conf->r1buf_pool);
270
271 lower_barrier(conf, sect);
272}
273
274static void reschedule_retry(struct r1bio *r1_bio)
275{
276 unsigned long flags;
277 struct mddev *mddev = r1_bio->mddev;
278 struct r1conf *conf = mddev->private;
279 int idx;
280
281 idx = sector_to_idx(r1_bio->sector);
282 spin_lock_irqsave(&conf->device_lock, flags);
283 list_add(&r1_bio->retry_list, &conf->retry_list);
284 atomic_inc(&conf->nr_queued[idx]);
285 spin_unlock_irqrestore(&conf->device_lock, flags);
286
287 wake_up(&conf->wait_barrier);
288 md_wakeup_thread(mddev->thread);
289}
290
291/*
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
294 * cache layer.
295 */
296static void call_bio_endio(struct r1bio *r1_bio)
297{
298 struct bio *bio = r1_bio->master_bio;
299
300 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
301 bio->bi_status = BLK_STS_IOERR;
302
303 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
304 bio_end_io_acct(bio, r1_bio->start_time);
305 bio_endio(bio);
306}
307
308static void raid_end_bio_io(struct r1bio *r1_bio)
309{
310 struct bio *bio = r1_bio->master_bio;
311 struct r1conf *conf = r1_bio->mddev->private;
312
313 /* if nobody has done the final endio yet, do it now */
314 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
315 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
316 (bio_data_dir(bio) == WRITE) ? "write" : "read",
317 (unsigned long long) bio->bi_iter.bi_sector,
318 (unsigned long long) bio_end_sector(bio) - 1);
319
320 call_bio_endio(r1_bio);
321 }
322 /*
323 * Wake up any possible resync thread that waits for the device
324 * to go idle. All I/Os, even write-behind writes, are done.
325 */
326 allow_barrier(conf, r1_bio->sector);
327
328 free_r1bio(r1_bio);
329}
330
331/*
332 * Update disk head position estimator based on IRQ completion info.
333 */
334static inline void update_head_pos(int disk, struct r1bio *r1_bio)
335{
336 struct r1conf *conf = r1_bio->mddev->private;
337
338 conf->mirrors[disk].head_position =
339 r1_bio->sector + (r1_bio->sectors);
340}
341
342/*
343 * Find the disk number which triggered given bio
344 */
345static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
346{
347 int mirror;
348 struct r1conf *conf = r1_bio->mddev->private;
349 int raid_disks = conf->raid_disks;
350
351 for (mirror = 0; mirror < raid_disks * 2; mirror++)
352 if (r1_bio->bios[mirror] == bio)
353 break;
354
355 BUG_ON(mirror == raid_disks * 2);
356 update_head_pos(mirror, r1_bio);
357
358 return mirror;
359}
360
361static void raid1_end_read_request(struct bio *bio)
362{
363 int uptodate = !bio->bi_status;
364 struct r1bio *r1_bio = bio->bi_private;
365 struct r1conf *conf = r1_bio->mddev->private;
366 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
367
368 /*
369 * this branch is our 'one mirror IO has finished' event handler:
370 */
371 update_head_pos(r1_bio->read_disk, r1_bio);
372
373 if (uptodate)
374 set_bit(R1BIO_Uptodate, &r1_bio->state);
375 else if (test_bit(FailFast, &rdev->flags) &&
376 test_bit(R1BIO_FailFast, &r1_bio->state))
377 /* This was a fail-fast read so we definitely
378 * want to retry */
379 ;
380 else {
381 /* If all other devices have failed, we want to return
382 * the error upwards rather than fail the last device.
383 * Here we redefine "uptodate" to mean "Don't want to retry"
384 */
385 unsigned long flags;
386 spin_lock_irqsave(&conf->device_lock, flags);
387 if (r1_bio->mddev->degraded == conf->raid_disks ||
388 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
389 test_bit(In_sync, &rdev->flags)))
390 uptodate = 1;
391 spin_unlock_irqrestore(&conf->device_lock, flags);
392 }
393
394 if (uptodate) {
395 raid_end_bio_io(r1_bio);
396 rdev_dec_pending(rdev, conf->mddev);
397 } else {
398 /*
399 * oops, read error:
400 */
401 char b[BDEVNAME_SIZE];
402 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
403 mdname(conf->mddev),
404 bdevname(rdev->bdev, b),
405 (unsigned long long)r1_bio->sector);
406 set_bit(R1BIO_ReadError, &r1_bio->state);
407 reschedule_retry(r1_bio);
408 /* don't drop the reference on read_disk yet */
409 }
410}
411
412static void close_write(struct r1bio *r1_bio)
413{
414 /* it really is the end of this request */
415 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
416 bio_free_pages(r1_bio->behind_master_bio);
417 bio_put(r1_bio->behind_master_bio);
418 r1_bio->behind_master_bio = NULL;
419 }
420 /* clear the bitmap if all writes complete successfully */
421 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
422 r1_bio->sectors,
423 !test_bit(R1BIO_Degraded, &r1_bio->state),
424 test_bit(R1BIO_BehindIO, &r1_bio->state));
425 md_write_end(r1_bio->mddev);
426}
427
428static void r1_bio_write_done(struct r1bio *r1_bio)
429{
430 if (!atomic_dec_and_test(&r1_bio->remaining))
431 return;
432
433 if (test_bit(R1BIO_WriteError, &r1_bio->state))
434 reschedule_retry(r1_bio);
435 else {
436 close_write(r1_bio);
437 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
438 reschedule_retry(r1_bio);
439 else
440 raid_end_bio_io(r1_bio);
441 }
442}
443
444static void raid1_end_write_request(struct bio *bio)
445{
446 struct r1bio *r1_bio = bio->bi_private;
447 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
448 struct r1conf *conf = r1_bio->mddev->private;
449 struct bio *to_put = NULL;
450 int mirror = find_bio_disk(r1_bio, bio);
451 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
452 bool discard_error;
453 sector_t lo = r1_bio->sector;
454 sector_t hi = r1_bio->sector + r1_bio->sectors;
455
456 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
457
458 /*
459 * 'one mirror IO has finished' event handler:
460 */
461 if (bio->bi_status && !discard_error) {
462 set_bit(WriteErrorSeen, &rdev->flags);
463 if (!test_and_set_bit(WantReplacement, &rdev->flags))
464 set_bit(MD_RECOVERY_NEEDED, &
465 conf->mddev->recovery);
466
467 if (test_bit(FailFast, &rdev->flags) &&
468 (bio->bi_opf & MD_FAILFAST) &&
469 /* We never try FailFast to WriteMostly devices */
470 !test_bit(WriteMostly, &rdev->flags)) {
471 md_error(r1_bio->mddev, rdev);
472 }
473
474 /*
475 * When the device is faulty, it is not necessary to
476 * handle write error.
477 */
478 if (!test_bit(Faulty, &rdev->flags))
479 set_bit(R1BIO_WriteError, &r1_bio->state);
480 else {
481 /* Fail the request */
482 set_bit(R1BIO_Degraded, &r1_bio->state);
483 /* Finished with this branch */
484 r1_bio->bios[mirror] = NULL;
485 to_put = bio;
486 }
487 } else {
488 /*
489 * Set R1BIO_Uptodate in our master bio, so that we
490 * will return a good error code for to the higher
491 * levels even if IO on some other mirrored buffer
492 * fails.
493 *
494 * The 'master' represents the composite IO operation
495 * to user-side. So if something waits for IO, then it
496 * will wait for the 'master' bio.
497 */
498 sector_t first_bad;
499 int bad_sectors;
500
501 r1_bio->bios[mirror] = NULL;
502 to_put = bio;
503 /*
504 * Do not set R1BIO_Uptodate if the current device is
505 * rebuilding or Faulty. This is because we cannot use
506 * such device for properly reading the data back (we could
507 * potentially use it, if the current write would have felt
508 * before rdev->recovery_offset, but for simplicity we don't
509 * check this here.
510 */
511 if (test_bit(In_sync, &rdev->flags) &&
512 !test_bit(Faulty, &rdev->flags))
513 set_bit(R1BIO_Uptodate, &r1_bio->state);
514
515 /* Maybe we can clear some bad blocks. */
516 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
517 &first_bad, &bad_sectors) && !discard_error) {
518 r1_bio->bios[mirror] = IO_MADE_GOOD;
519 set_bit(R1BIO_MadeGood, &r1_bio->state);
520 }
521 }
522
523 if (behind) {
524 if (test_bit(CollisionCheck, &rdev->flags))
525 remove_serial(rdev, lo, hi);
526 if (test_bit(WriteMostly, &rdev->flags))
527 atomic_dec(&r1_bio->behind_remaining);
528
529 /*
530 * In behind mode, we ACK the master bio once the I/O
531 * has safely reached all non-writemostly
532 * disks. Setting the Returned bit ensures that this
533 * gets done only once -- we don't ever want to return
534 * -EIO here, instead we'll wait
535 */
536 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
537 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
538 /* Maybe we can return now */
539 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
540 struct bio *mbio = r1_bio->master_bio;
541 pr_debug("raid1: behind end write sectors"
542 " %llu-%llu\n",
543 (unsigned long long) mbio->bi_iter.bi_sector,
544 (unsigned long long) bio_end_sector(mbio) - 1);
545 call_bio_endio(r1_bio);
546 }
547 }
548 } else if (rdev->mddev->serialize_policy)
549 remove_serial(rdev, lo, hi);
550 if (r1_bio->bios[mirror] == NULL)
551 rdev_dec_pending(rdev, conf->mddev);
552
553 /*
554 * Let's see if all mirrored write operations have finished
555 * already.
556 */
557 r1_bio_write_done(r1_bio);
558
559 if (to_put)
560 bio_put(to_put);
561}
562
563static sector_t align_to_barrier_unit_end(sector_t start_sector,
564 sector_t sectors)
565{
566 sector_t len;
567
568 WARN_ON(sectors == 0);
569 /*
570 * len is the number of sectors from start_sector to end of the
571 * barrier unit which start_sector belongs to.
572 */
573 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
574 start_sector;
575
576 if (len > sectors)
577 len = sectors;
578
579 return len;
580}
581
582/*
583 * This routine returns the disk from which the requested read should
584 * be done. There is a per-array 'next expected sequential IO' sector
585 * number - if this matches on the next IO then we use the last disk.
586 * There is also a per-disk 'last know head position' sector that is
587 * maintained from IRQ contexts, both the normal and the resync IO
588 * completion handlers update this position correctly. If there is no
589 * perfect sequential match then we pick the disk whose head is closest.
590 *
591 * If there are 2 mirrors in the same 2 devices, performance degrades
592 * because position is mirror, not device based.
593 *
594 * The rdev for the device selected will have nr_pending incremented.
595 */
596static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
597{
598 const sector_t this_sector = r1_bio->sector;
599 int sectors;
600 int best_good_sectors;
601 int best_disk, best_dist_disk, best_pending_disk;
602 int has_nonrot_disk;
603 int disk;
604 sector_t best_dist;
605 unsigned int min_pending;
606 struct md_rdev *rdev;
607 int choose_first;
608 int choose_next_idle;
609
610 rcu_read_lock();
611 /*
612 * Check if we can balance. We can balance on the whole
613 * device if no resync is going on, or below the resync window.
614 * We take the first readable disk when above the resync window.
615 */
616 retry:
617 sectors = r1_bio->sectors;
618 best_disk = -1;
619 best_dist_disk = -1;
620 best_dist = MaxSector;
621 best_pending_disk = -1;
622 min_pending = UINT_MAX;
623 best_good_sectors = 0;
624 has_nonrot_disk = 0;
625 choose_next_idle = 0;
626 clear_bit(R1BIO_FailFast, &r1_bio->state);
627
628 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
629 (mddev_is_clustered(conf->mddev) &&
630 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
631 this_sector + sectors)))
632 choose_first = 1;
633 else
634 choose_first = 0;
635
636 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
637 sector_t dist;
638 sector_t first_bad;
639 int bad_sectors;
640 unsigned int pending;
641 bool nonrot;
642
643 rdev = rcu_dereference(conf->mirrors[disk].rdev);
644 if (r1_bio->bios[disk] == IO_BLOCKED
645 || rdev == NULL
646 || test_bit(Faulty, &rdev->flags))
647 continue;
648 if (!test_bit(In_sync, &rdev->flags) &&
649 rdev->recovery_offset < this_sector + sectors)
650 continue;
651 if (test_bit(WriteMostly, &rdev->flags)) {
652 /* Don't balance among write-mostly, just
653 * use the first as a last resort */
654 if (best_dist_disk < 0) {
655 if (is_badblock(rdev, this_sector, sectors,
656 &first_bad, &bad_sectors)) {
657 if (first_bad <= this_sector)
658 /* Cannot use this */
659 continue;
660 best_good_sectors = first_bad - this_sector;
661 } else
662 best_good_sectors = sectors;
663 best_dist_disk = disk;
664 best_pending_disk = disk;
665 }
666 continue;
667 }
668 /* This is a reasonable device to use. It might
669 * even be best.
670 */
671 if (is_badblock(rdev, this_sector, sectors,
672 &first_bad, &bad_sectors)) {
673 if (best_dist < MaxSector)
674 /* already have a better device */
675 continue;
676 if (first_bad <= this_sector) {
677 /* cannot read here. If this is the 'primary'
678 * device, then we must not read beyond
679 * bad_sectors from another device..
680 */
681 bad_sectors -= (this_sector - first_bad);
682 if (choose_first && sectors > bad_sectors)
683 sectors = bad_sectors;
684 if (best_good_sectors > sectors)
685 best_good_sectors = sectors;
686
687 } else {
688 sector_t good_sectors = first_bad - this_sector;
689 if (good_sectors > best_good_sectors) {
690 best_good_sectors = good_sectors;
691 best_disk = disk;
692 }
693 if (choose_first)
694 break;
695 }
696 continue;
697 } else {
698 if ((sectors > best_good_sectors) && (best_disk >= 0))
699 best_disk = -1;
700 best_good_sectors = sectors;
701 }
702
703 if (best_disk >= 0)
704 /* At least two disks to choose from so failfast is OK */
705 set_bit(R1BIO_FailFast, &r1_bio->state);
706
707 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
708 has_nonrot_disk |= nonrot;
709 pending = atomic_read(&rdev->nr_pending);
710 dist = abs(this_sector - conf->mirrors[disk].head_position);
711 if (choose_first) {
712 best_disk = disk;
713 break;
714 }
715 /* Don't change to another disk for sequential reads */
716 if (conf->mirrors[disk].next_seq_sect == this_sector
717 || dist == 0) {
718 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
719 struct raid1_info *mirror = &conf->mirrors[disk];
720
721 best_disk = disk;
722 /*
723 * If buffered sequential IO size exceeds optimal
724 * iosize, check if there is idle disk. If yes, choose
725 * the idle disk. read_balance could already choose an
726 * idle disk before noticing it's a sequential IO in
727 * this disk. This doesn't matter because this disk
728 * will idle, next time it will be utilized after the
729 * first disk has IO size exceeds optimal iosize. In
730 * this way, iosize of the first disk will be optimal
731 * iosize at least. iosize of the second disk might be
732 * small, but not a big deal since when the second disk
733 * starts IO, the first disk is likely still busy.
734 */
735 if (nonrot && opt_iosize > 0 &&
736 mirror->seq_start != MaxSector &&
737 mirror->next_seq_sect > opt_iosize &&
738 mirror->next_seq_sect - opt_iosize >=
739 mirror->seq_start) {
740 choose_next_idle = 1;
741 continue;
742 }
743 break;
744 }
745
746 if (choose_next_idle)
747 continue;
748
749 if (min_pending > pending) {
750 min_pending = pending;
751 best_pending_disk = disk;
752 }
753
754 if (dist < best_dist) {
755 best_dist = dist;
756 best_dist_disk = disk;
757 }
758 }
759
760 /*
761 * If all disks are rotational, choose the closest disk. If any disk is
762 * non-rotational, choose the disk with less pending request even the
763 * disk is rotational, which might/might not be optimal for raids with
764 * mixed ratation/non-rotational disks depending on workload.
765 */
766 if (best_disk == -1) {
767 if (has_nonrot_disk || min_pending == 0)
768 best_disk = best_pending_disk;
769 else
770 best_disk = best_dist_disk;
771 }
772
773 if (best_disk >= 0) {
774 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
775 if (!rdev)
776 goto retry;
777 atomic_inc(&rdev->nr_pending);
778 sectors = best_good_sectors;
779
780 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
781 conf->mirrors[best_disk].seq_start = this_sector;
782
783 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
784 }
785 rcu_read_unlock();
786 *max_sectors = sectors;
787
788 return best_disk;
789}
790
791static void flush_bio_list(struct r1conf *conf, struct bio *bio)
792{
793 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
794 md_bitmap_unplug(conf->mddev->bitmap);
795 wake_up(&conf->wait_barrier);
796
797 while (bio) { /* submit pending writes */
798 struct bio *next = bio->bi_next;
799 struct md_rdev *rdev = (void *)bio->bi_bdev;
800 bio->bi_next = NULL;
801 bio_set_dev(bio, rdev->bdev);
802 if (test_bit(Faulty, &rdev->flags)) {
803 bio_io_error(bio);
804 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
805 !blk_queue_discard(bio->bi_bdev->bd_disk->queue)))
806 /* Just ignore it */
807 bio_endio(bio);
808 else
809 submit_bio_noacct(bio);
810 bio = next;
811 cond_resched();
812 }
813}
814
815static void flush_pending_writes(struct r1conf *conf)
816{
817 /* Any writes that have been queued but are awaiting
818 * bitmap updates get flushed here.
819 */
820 spin_lock_irq(&conf->device_lock);
821
822 if (conf->pending_bio_list.head) {
823 struct blk_plug plug;
824 struct bio *bio;
825
826 bio = bio_list_get(&conf->pending_bio_list);
827 conf->pending_count = 0;
828 spin_unlock_irq(&conf->device_lock);
829
830 /*
831 * As this is called in a wait_event() loop (see freeze_array),
832 * current->state might be TASK_UNINTERRUPTIBLE which will
833 * cause a warning when we prepare to wait again. As it is
834 * rare that this path is taken, it is perfectly safe to force
835 * us to go around the wait_event() loop again, so the warning
836 * is a false-positive. Silence the warning by resetting
837 * thread state
838 */
839 __set_current_state(TASK_RUNNING);
840 blk_start_plug(&plug);
841 flush_bio_list(conf, bio);
842 blk_finish_plug(&plug);
843 } else
844 spin_unlock_irq(&conf->device_lock);
845}
846
847/* Barriers....
848 * Sometimes we need to suspend IO while we do something else,
849 * either some resync/recovery, or reconfigure the array.
850 * To do this we raise a 'barrier'.
851 * The 'barrier' is a counter that can be raised multiple times
852 * to count how many activities are happening which preclude
853 * normal IO.
854 * We can only raise the barrier if there is no pending IO.
855 * i.e. if nr_pending == 0.
856 * We choose only to raise the barrier if no-one is waiting for the
857 * barrier to go down. This means that as soon as an IO request
858 * is ready, no other operations which require a barrier will start
859 * until the IO request has had a chance.
860 *
861 * So: regular IO calls 'wait_barrier'. When that returns there
862 * is no backgroup IO happening, It must arrange to call
863 * allow_barrier when it has finished its IO.
864 * backgroup IO calls must call raise_barrier. Once that returns
865 * there is no normal IO happeing. It must arrange to call
866 * lower_barrier when the particular background IO completes.
867 *
868 * If resync/recovery is interrupted, returns -EINTR;
869 * Otherwise, returns 0.
870 */
871static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
872{
873 int idx = sector_to_idx(sector_nr);
874
875 spin_lock_irq(&conf->resync_lock);
876
877 /* Wait until no block IO is waiting */
878 wait_event_lock_irq(conf->wait_barrier,
879 !atomic_read(&conf->nr_waiting[idx]),
880 conf->resync_lock);
881
882 /* block any new IO from starting */
883 atomic_inc(&conf->barrier[idx]);
884 /*
885 * In raise_barrier() we firstly increase conf->barrier[idx] then
886 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
887 * increase conf->nr_pending[idx] then check conf->barrier[idx].
888 * A memory barrier here to make sure conf->nr_pending[idx] won't
889 * be fetched before conf->barrier[idx] is increased. Otherwise
890 * there will be a race between raise_barrier() and _wait_barrier().
891 */
892 smp_mb__after_atomic();
893
894 /* For these conditions we must wait:
895 * A: while the array is in frozen state
896 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
897 * existing in corresponding I/O barrier bucket.
898 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
899 * max resync count which allowed on current I/O barrier bucket.
900 */
901 wait_event_lock_irq(conf->wait_barrier,
902 (!conf->array_frozen &&
903 !atomic_read(&conf->nr_pending[idx]) &&
904 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
905 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
906 conf->resync_lock);
907
908 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
909 atomic_dec(&conf->barrier[idx]);
910 spin_unlock_irq(&conf->resync_lock);
911 wake_up(&conf->wait_barrier);
912 return -EINTR;
913 }
914
915 atomic_inc(&conf->nr_sync_pending);
916 spin_unlock_irq(&conf->resync_lock);
917
918 return 0;
919}
920
921static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
922{
923 int idx = sector_to_idx(sector_nr);
924
925 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
926
927 atomic_dec(&conf->barrier[idx]);
928 atomic_dec(&conf->nr_sync_pending);
929 wake_up(&conf->wait_barrier);
930}
931
932static void _wait_barrier(struct r1conf *conf, int idx)
933{
934 /*
935 * We need to increase conf->nr_pending[idx] very early here,
936 * then raise_barrier() can be blocked when it waits for
937 * conf->nr_pending[idx] to be 0. Then we can avoid holding
938 * conf->resync_lock when there is no barrier raised in same
939 * barrier unit bucket. Also if the array is frozen, I/O
940 * should be blocked until array is unfrozen.
941 */
942 atomic_inc(&conf->nr_pending[idx]);
943 /*
944 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
945 * check conf->barrier[idx]. In raise_barrier() we firstly increase
946 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
947 * barrier is necessary here to make sure conf->barrier[idx] won't be
948 * fetched before conf->nr_pending[idx] is increased. Otherwise there
949 * will be a race between _wait_barrier() and raise_barrier().
950 */
951 smp_mb__after_atomic();
952
953 /*
954 * Don't worry about checking two atomic_t variables at same time
955 * here. If during we check conf->barrier[idx], the array is
956 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
957 * 0, it is safe to return and make the I/O continue. Because the
958 * array is frozen, all I/O returned here will eventually complete
959 * or be queued, no race will happen. See code comment in
960 * frozen_array().
961 */
962 if (!READ_ONCE(conf->array_frozen) &&
963 !atomic_read(&conf->barrier[idx]))
964 return;
965
966 /*
967 * After holding conf->resync_lock, conf->nr_pending[idx]
968 * should be decreased before waiting for barrier to drop.
969 * Otherwise, we may encounter a race condition because
970 * raise_barrer() might be waiting for conf->nr_pending[idx]
971 * to be 0 at same time.
972 */
973 spin_lock_irq(&conf->resync_lock);
974 atomic_inc(&conf->nr_waiting[idx]);
975 atomic_dec(&conf->nr_pending[idx]);
976 /*
977 * In case freeze_array() is waiting for
978 * get_unqueued_pending() == extra
979 */
980 wake_up(&conf->wait_barrier);
981 /* Wait for the barrier in same barrier unit bucket to drop. */
982 wait_event_lock_irq(conf->wait_barrier,
983 !conf->array_frozen &&
984 !atomic_read(&conf->barrier[idx]),
985 conf->resync_lock);
986 atomic_inc(&conf->nr_pending[idx]);
987 atomic_dec(&conf->nr_waiting[idx]);
988 spin_unlock_irq(&conf->resync_lock);
989}
990
991static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
992{
993 int idx = sector_to_idx(sector_nr);
994
995 /*
996 * Very similar to _wait_barrier(). The difference is, for read
997 * I/O we don't need wait for sync I/O, but if the whole array
998 * is frozen, the read I/O still has to wait until the array is
999 * unfrozen. Since there is no ordering requirement with
1000 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1001 */
1002 atomic_inc(&conf->nr_pending[idx]);
1003
1004 if (!READ_ONCE(conf->array_frozen))
1005 return;
1006
1007 spin_lock_irq(&conf->resync_lock);
1008 atomic_inc(&conf->nr_waiting[idx]);
1009 atomic_dec(&conf->nr_pending[idx]);
1010 /*
1011 * In case freeze_array() is waiting for
1012 * get_unqueued_pending() == extra
1013 */
1014 wake_up(&conf->wait_barrier);
1015 /* Wait for array to be unfrozen */
1016 wait_event_lock_irq(conf->wait_barrier,
1017 !conf->array_frozen,
1018 conf->resync_lock);
1019 atomic_inc(&conf->nr_pending[idx]);
1020 atomic_dec(&conf->nr_waiting[idx]);
1021 spin_unlock_irq(&conf->resync_lock);
1022}
1023
1024static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1025{
1026 int idx = sector_to_idx(sector_nr);
1027
1028 _wait_barrier(conf, idx);
1029}
1030
1031static void _allow_barrier(struct r1conf *conf, int idx)
1032{
1033 atomic_dec(&conf->nr_pending[idx]);
1034 wake_up(&conf->wait_barrier);
1035}
1036
1037static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1038{
1039 int idx = sector_to_idx(sector_nr);
1040
1041 _allow_barrier(conf, idx);
1042}
1043
1044/* conf->resync_lock should be held */
1045static int get_unqueued_pending(struct r1conf *conf)
1046{
1047 int idx, ret;
1048
1049 ret = atomic_read(&conf->nr_sync_pending);
1050 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1051 ret += atomic_read(&conf->nr_pending[idx]) -
1052 atomic_read(&conf->nr_queued[idx]);
1053
1054 return ret;
1055}
1056
1057static void freeze_array(struct r1conf *conf, int extra)
1058{
1059 /* Stop sync I/O and normal I/O and wait for everything to
1060 * go quiet.
1061 * This is called in two situations:
1062 * 1) management command handlers (reshape, remove disk, quiesce).
1063 * 2) one normal I/O request failed.
1064
1065 * After array_frozen is set to 1, new sync IO will be blocked at
1066 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1067 * or wait_read_barrier(). The flying I/Os will either complete or be
1068 * queued. When everything goes quite, there are only queued I/Os left.
1069
1070 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1071 * barrier bucket index which this I/O request hits. When all sync and
1072 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1073 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1074 * in handle_read_error(), we may call freeze_array() before trying to
1075 * fix the read error. In this case, the error read I/O is not queued,
1076 * so get_unqueued_pending() == 1.
1077 *
1078 * Therefore before this function returns, we need to wait until
1079 * get_unqueued_pendings(conf) gets equal to extra. For
1080 * normal I/O context, extra is 1, in rested situations extra is 0.
1081 */
1082 spin_lock_irq(&conf->resync_lock);
1083 conf->array_frozen = 1;
1084 raid1_log(conf->mddev, "wait freeze");
1085 wait_event_lock_irq_cmd(
1086 conf->wait_barrier,
1087 get_unqueued_pending(conf) == extra,
1088 conf->resync_lock,
1089 flush_pending_writes(conf));
1090 spin_unlock_irq(&conf->resync_lock);
1091}
1092static void unfreeze_array(struct r1conf *conf)
1093{
1094 /* reverse the effect of the freeze */
1095 spin_lock_irq(&conf->resync_lock);
1096 conf->array_frozen = 0;
1097 spin_unlock_irq(&conf->resync_lock);
1098 wake_up(&conf->wait_barrier);
1099}
1100
1101static void alloc_behind_master_bio(struct r1bio *r1_bio,
1102 struct bio *bio)
1103{
1104 int size = bio->bi_iter.bi_size;
1105 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1106 int i = 0;
1107 struct bio *behind_bio = NULL;
1108
1109 behind_bio = bio_alloc_bioset(GFP_NOIO, vcnt, &r1_bio->mddev->bio_set);
1110 if (!behind_bio)
1111 return;
1112
1113 /* discard op, we don't support writezero/writesame yet */
1114 if (!bio_has_data(bio)) {
1115 behind_bio->bi_iter.bi_size = size;
1116 goto skip_copy;
1117 }
1118
1119 behind_bio->bi_write_hint = bio->bi_write_hint;
1120
1121 while (i < vcnt && size) {
1122 struct page *page;
1123 int len = min_t(int, PAGE_SIZE, size);
1124
1125 page = alloc_page(GFP_NOIO);
1126 if (unlikely(!page))
1127 goto free_pages;
1128
1129 bio_add_page(behind_bio, page, len, 0);
1130
1131 size -= len;
1132 i++;
1133 }
1134
1135 bio_copy_data(behind_bio, bio);
1136skip_copy:
1137 r1_bio->behind_master_bio = behind_bio;
1138 set_bit(R1BIO_BehindIO, &r1_bio->state);
1139
1140 return;
1141
1142free_pages:
1143 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1144 bio->bi_iter.bi_size);
1145 bio_free_pages(behind_bio);
1146 bio_put(behind_bio);
1147}
1148
1149struct raid1_plug_cb {
1150 struct blk_plug_cb cb;
1151 struct bio_list pending;
1152 int pending_cnt;
1153};
1154
1155static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1156{
1157 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1158 cb);
1159 struct mddev *mddev = plug->cb.data;
1160 struct r1conf *conf = mddev->private;
1161 struct bio *bio;
1162
1163 if (from_schedule || current->bio_list) {
1164 spin_lock_irq(&conf->device_lock);
1165 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1166 conf->pending_count += plug->pending_cnt;
1167 spin_unlock_irq(&conf->device_lock);
1168 wake_up(&conf->wait_barrier);
1169 md_wakeup_thread(mddev->thread);
1170 kfree(plug);
1171 return;
1172 }
1173
1174 /* we aren't scheduling, so we can do the write-out directly. */
1175 bio = bio_list_get(&plug->pending);
1176 flush_bio_list(conf, bio);
1177 kfree(plug);
1178}
1179
1180static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1181{
1182 r1_bio->master_bio = bio;
1183 r1_bio->sectors = bio_sectors(bio);
1184 r1_bio->state = 0;
1185 r1_bio->mddev = mddev;
1186 r1_bio->sector = bio->bi_iter.bi_sector;
1187}
1188
1189static inline struct r1bio *
1190alloc_r1bio(struct mddev *mddev, struct bio *bio)
1191{
1192 struct r1conf *conf = mddev->private;
1193 struct r1bio *r1_bio;
1194
1195 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1196 /* Ensure no bio records IO_BLOCKED */
1197 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1198 init_r1bio(r1_bio, mddev, bio);
1199 return r1_bio;
1200}
1201
1202static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1203 int max_read_sectors, struct r1bio *r1_bio)
1204{
1205 struct r1conf *conf = mddev->private;
1206 struct raid1_info *mirror;
1207 struct bio *read_bio;
1208 struct bitmap *bitmap = mddev->bitmap;
1209 const int op = bio_op(bio);
1210 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1211 int max_sectors;
1212 int rdisk;
1213 bool r1bio_existed = !!r1_bio;
1214 char b[BDEVNAME_SIZE];
1215
1216 /*
1217 * If r1_bio is set, we are blocking the raid1d thread
1218 * so there is a tiny risk of deadlock. So ask for
1219 * emergency memory if needed.
1220 */
1221 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1222
1223 if (r1bio_existed) {
1224 /* Need to get the block device name carefully */
1225 struct md_rdev *rdev;
1226 rcu_read_lock();
1227 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1228 if (rdev)
1229 bdevname(rdev->bdev, b);
1230 else
1231 strcpy(b, "???");
1232 rcu_read_unlock();
1233 }
1234
1235 /*
1236 * Still need barrier for READ in case that whole
1237 * array is frozen.
1238 */
1239 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1240
1241 if (!r1_bio)
1242 r1_bio = alloc_r1bio(mddev, bio);
1243 else
1244 init_r1bio(r1_bio, mddev, bio);
1245 r1_bio->sectors = max_read_sectors;
1246
1247 /*
1248 * make_request() can abort the operation when read-ahead is being
1249 * used and no empty request is available.
1250 */
1251 rdisk = read_balance(conf, r1_bio, &max_sectors);
1252
1253 if (rdisk < 0) {
1254 /* couldn't find anywhere to read from */
1255 if (r1bio_existed) {
1256 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1257 mdname(mddev),
1258 b,
1259 (unsigned long long)r1_bio->sector);
1260 }
1261 raid_end_bio_io(r1_bio);
1262 return;
1263 }
1264 mirror = conf->mirrors + rdisk;
1265
1266 if (r1bio_existed)
1267 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1268 mdname(mddev),
1269 (unsigned long long)r1_bio->sector,
1270 bdevname(mirror->rdev->bdev, b));
1271
1272 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1273 bitmap) {
1274 /*
1275 * Reading from a write-mostly device must take care not to
1276 * over-take any writes that are 'behind'
1277 */
1278 raid1_log(mddev, "wait behind writes");
1279 wait_event(bitmap->behind_wait,
1280 atomic_read(&bitmap->behind_writes) == 0);
1281 }
1282
1283 if (max_sectors < bio_sectors(bio)) {
1284 struct bio *split = bio_split(bio, max_sectors,
1285 gfp, &conf->bio_split);
1286 bio_chain(split, bio);
1287 submit_bio_noacct(bio);
1288 bio = split;
1289 r1_bio->master_bio = bio;
1290 r1_bio->sectors = max_sectors;
1291 }
1292
1293 r1_bio->read_disk = rdisk;
1294
1295 if (!r1bio_existed && blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1296 r1_bio->start_time = bio_start_io_acct(bio);
1297
1298 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1299
1300 r1_bio->bios[rdisk] = read_bio;
1301
1302 read_bio->bi_iter.bi_sector = r1_bio->sector +
1303 mirror->rdev->data_offset;
1304 bio_set_dev(read_bio, mirror->rdev->bdev);
1305 read_bio->bi_end_io = raid1_end_read_request;
1306 bio_set_op_attrs(read_bio, op, do_sync);
1307 if (test_bit(FailFast, &mirror->rdev->flags) &&
1308 test_bit(R1BIO_FailFast, &r1_bio->state))
1309 read_bio->bi_opf |= MD_FAILFAST;
1310 read_bio->bi_private = r1_bio;
1311
1312 if (mddev->gendisk)
1313 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1314 r1_bio->sector);
1315
1316 submit_bio_noacct(read_bio);
1317}
1318
1319static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1320 int max_write_sectors)
1321{
1322 struct r1conf *conf = mddev->private;
1323 struct r1bio *r1_bio;
1324 int i, disks;
1325 struct bitmap *bitmap = mddev->bitmap;
1326 unsigned long flags;
1327 struct md_rdev *blocked_rdev;
1328 struct blk_plug_cb *cb;
1329 struct raid1_plug_cb *plug = NULL;
1330 int first_clone;
1331 int max_sectors;
1332 bool write_behind = false;
1333
1334 if (mddev_is_clustered(mddev) &&
1335 md_cluster_ops->area_resyncing(mddev, WRITE,
1336 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1337
1338 DEFINE_WAIT(w);
1339 for (;;) {
1340 prepare_to_wait(&conf->wait_barrier,
1341 &w, TASK_IDLE);
1342 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1343 bio->bi_iter.bi_sector,
1344 bio_end_sector(bio)))
1345 break;
1346 schedule();
1347 }
1348 finish_wait(&conf->wait_barrier, &w);
1349 }
1350
1351 /*
1352 * Register the new request and wait if the reconstruction
1353 * thread has put up a bar for new requests.
1354 * Continue immediately if no resync is active currently.
1355 */
1356 wait_barrier(conf, bio->bi_iter.bi_sector);
1357
1358 r1_bio = alloc_r1bio(mddev, bio);
1359 r1_bio->sectors = max_write_sectors;
1360
1361 if (conf->pending_count >= max_queued_requests) {
1362 md_wakeup_thread(mddev->thread);
1363 raid1_log(mddev, "wait queued");
1364 wait_event(conf->wait_barrier,
1365 conf->pending_count < max_queued_requests);
1366 }
1367 /* first select target devices under rcu_lock and
1368 * inc refcount on their rdev. Record them by setting
1369 * bios[x] to bio
1370 * If there are known/acknowledged bad blocks on any device on
1371 * which we have seen a write error, we want to avoid writing those
1372 * blocks.
1373 * This potentially requires several writes to write around
1374 * the bad blocks. Each set of writes gets it's own r1bio
1375 * with a set of bios attached.
1376 */
1377
1378 disks = conf->raid_disks * 2;
1379 retry_write:
1380 blocked_rdev = NULL;
1381 rcu_read_lock();
1382 max_sectors = r1_bio->sectors;
1383 for (i = 0; i < disks; i++) {
1384 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1385
1386 /*
1387 * The write-behind io is only attempted on drives marked as
1388 * write-mostly, which means we could allocate write behind
1389 * bio later.
1390 */
1391 if (rdev && test_bit(WriteMostly, &rdev->flags))
1392 write_behind = true;
1393
1394 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1395 atomic_inc(&rdev->nr_pending);
1396 blocked_rdev = rdev;
1397 break;
1398 }
1399 r1_bio->bios[i] = NULL;
1400 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1401 if (i < conf->raid_disks)
1402 set_bit(R1BIO_Degraded, &r1_bio->state);
1403 continue;
1404 }
1405
1406 atomic_inc(&rdev->nr_pending);
1407 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1408 sector_t first_bad;
1409 int bad_sectors;
1410 int is_bad;
1411
1412 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1413 &first_bad, &bad_sectors);
1414 if (is_bad < 0) {
1415 /* mustn't write here until the bad block is
1416 * acknowledged*/
1417 set_bit(BlockedBadBlocks, &rdev->flags);
1418 blocked_rdev = rdev;
1419 break;
1420 }
1421 if (is_bad && first_bad <= r1_bio->sector) {
1422 /* Cannot write here at all */
1423 bad_sectors -= (r1_bio->sector - first_bad);
1424 if (bad_sectors < max_sectors)
1425 /* mustn't write more than bad_sectors
1426 * to other devices yet
1427 */
1428 max_sectors = bad_sectors;
1429 rdev_dec_pending(rdev, mddev);
1430 /* We don't set R1BIO_Degraded as that
1431 * only applies if the disk is
1432 * missing, so it might be re-added,
1433 * and we want to know to recover this
1434 * chunk.
1435 * In this case the device is here,
1436 * and the fact that this chunk is not
1437 * in-sync is recorded in the bad
1438 * block log
1439 */
1440 continue;
1441 }
1442 if (is_bad) {
1443 int good_sectors = first_bad - r1_bio->sector;
1444 if (good_sectors < max_sectors)
1445 max_sectors = good_sectors;
1446 }
1447 }
1448 r1_bio->bios[i] = bio;
1449 }
1450 rcu_read_unlock();
1451
1452 if (unlikely(blocked_rdev)) {
1453 /* Wait for this device to become unblocked */
1454 int j;
1455
1456 for (j = 0; j < i; j++)
1457 if (r1_bio->bios[j])
1458 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1459 r1_bio->state = 0;
1460 allow_barrier(conf, bio->bi_iter.bi_sector);
1461 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1462 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1463 wait_barrier(conf, bio->bi_iter.bi_sector);
1464 goto retry_write;
1465 }
1466
1467 /*
1468 * When using a bitmap, we may call alloc_behind_master_bio below.
1469 * alloc_behind_master_bio allocates a copy of the data payload a page
1470 * at a time and thus needs a new bio that can fit the whole payload
1471 * this bio in page sized chunks.
1472 */
1473 if (write_behind && bitmap)
1474 max_sectors = min_t(int, max_sectors,
1475 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1476 if (max_sectors < bio_sectors(bio)) {
1477 struct bio *split = bio_split(bio, max_sectors,
1478 GFP_NOIO, &conf->bio_split);
1479 bio_chain(split, bio);
1480 submit_bio_noacct(bio);
1481 bio = split;
1482 r1_bio->master_bio = bio;
1483 r1_bio->sectors = max_sectors;
1484 }
1485
1486 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1487 r1_bio->start_time = bio_start_io_acct(bio);
1488 atomic_set(&r1_bio->remaining, 1);
1489 atomic_set(&r1_bio->behind_remaining, 0);
1490
1491 first_clone = 1;
1492
1493 for (i = 0; i < disks; i++) {
1494 struct bio *mbio = NULL;
1495 struct md_rdev *rdev = conf->mirrors[i].rdev;
1496 if (!r1_bio->bios[i])
1497 continue;
1498
1499 if (first_clone) {
1500 /* do behind I/O ?
1501 * Not if there are too many, or cannot
1502 * allocate memory, or a reader on WriteMostly
1503 * is waiting for behind writes to flush */
1504 if (bitmap &&
1505 (atomic_read(&bitmap->behind_writes)
1506 < mddev->bitmap_info.max_write_behind) &&
1507 !waitqueue_active(&bitmap->behind_wait)) {
1508 alloc_behind_master_bio(r1_bio, bio);
1509 }
1510
1511 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1512 test_bit(R1BIO_BehindIO, &r1_bio->state));
1513 first_clone = 0;
1514 }
1515
1516 if (r1_bio->behind_master_bio)
1517 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1518 GFP_NOIO, &mddev->bio_set);
1519 else
1520 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1521
1522 if (r1_bio->behind_master_bio) {
1523 if (test_bit(CollisionCheck, &rdev->flags))
1524 wait_for_serialization(rdev, r1_bio);
1525 if (test_bit(WriteMostly, &rdev->flags))
1526 atomic_inc(&r1_bio->behind_remaining);
1527 } else if (mddev->serialize_policy)
1528 wait_for_serialization(rdev, r1_bio);
1529
1530 r1_bio->bios[i] = mbio;
1531
1532 mbio->bi_iter.bi_sector = (r1_bio->sector +
1533 conf->mirrors[i].rdev->data_offset);
1534 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1535 mbio->bi_end_io = raid1_end_write_request;
1536 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1537 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1538 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1539 conf->raid_disks - mddev->degraded > 1)
1540 mbio->bi_opf |= MD_FAILFAST;
1541 mbio->bi_private = r1_bio;
1542
1543 atomic_inc(&r1_bio->remaining);
1544
1545 if (mddev->gendisk)
1546 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1547 r1_bio->sector);
1548 /* flush_pending_writes() needs access to the rdev so...*/
1549 mbio->bi_bdev = (void *)conf->mirrors[i].rdev;
1550
1551 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1552 if (cb)
1553 plug = container_of(cb, struct raid1_plug_cb, cb);
1554 else
1555 plug = NULL;
1556 if (plug) {
1557 bio_list_add(&plug->pending, mbio);
1558 plug->pending_cnt++;
1559 } else {
1560 spin_lock_irqsave(&conf->device_lock, flags);
1561 bio_list_add(&conf->pending_bio_list, mbio);
1562 conf->pending_count++;
1563 spin_unlock_irqrestore(&conf->device_lock, flags);
1564 md_wakeup_thread(mddev->thread);
1565 }
1566 }
1567
1568 r1_bio_write_done(r1_bio);
1569
1570 /* In case raid1d snuck in to freeze_array */
1571 wake_up(&conf->wait_barrier);
1572}
1573
1574static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1575{
1576 sector_t sectors;
1577
1578 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1579 && md_flush_request(mddev, bio))
1580 return true;
1581
1582 /*
1583 * There is a limit to the maximum size, but
1584 * the read/write handler might find a lower limit
1585 * due to bad blocks. To avoid multiple splits,
1586 * we pass the maximum number of sectors down
1587 * and let the lower level perform the split.
1588 */
1589 sectors = align_to_barrier_unit_end(
1590 bio->bi_iter.bi_sector, bio_sectors(bio));
1591
1592 if (bio_data_dir(bio) == READ)
1593 raid1_read_request(mddev, bio, sectors, NULL);
1594 else {
1595 if (!md_write_start(mddev,bio))
1596 return false;
1597 raid1_write_request(mddev, bio, sectors);
1598 }
1599 return true;
1600}
1601
1602static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1603{
1604 struct r1conf *conf = mddev->private;
1605 int i;
1606
1607 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1608 conf->raid_disks - mddev->degraded);
1609 rcu_read_lock();
1610 for (i = 0; i < conf->raid_disks; i++) {
1611 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1612 seq_printf(seq, "%s",
1613 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1614 }
1615 rcu_read_unlock();
1616 seq_printf(seq, "]");
1617}
1618
1619static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1620{
1621 char b[BDEVNAME_SIZE];
1622 struct r1conf *conf = mddev->private;
1623 unsigned long flags;
1624
1625 /*
1626 * If it is not operational, then we have already marked it as dead
1627 * else if it is the last working disks with "fail_last_dev == false",
1628 * ignore the error, let the next level up know.
1629 * else mark the drive as failed
1630 */
1631 spin_lock_irqsave(&conf->device_lock, flags);
1632 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1633 && (conf->raid_disks - mddev->degraded) == 1) {
1634 /*
1635 * Don't fail the drive, act as though we were just a
1636 * normal single drive.
1637 * However don't try a recovery from this drive as
1638 * it is very likely to fail.
1639 */
1640 conf->recovery_disabled = mddev->recovery_disabled;
1641 spin_unlock_irqrestore(&conf->device_lock, flags);
1642 return;
1643 }
1644 set_bit(Blocked, &rdev->flags);
1645 if (test_and_clear_bit(In_sync, &rdev->flags))
1646 mddev->degraded++;
1647 set_bit(Faulty, &rdev->flags);
1648 spin_unlock_irqrestore(&conf->device_lock, flags);
1649 /*
1650 * if recovery is running, make sure it aborts.
1651 */
1652 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1653 set_mask_bits(&mddev->sb_flags, 0,
1654 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1655 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1656 "md/raid1:%s: Operation continuing on %d devices.\n",
1657 mdname(mddev), bdevname(rdev->bdev, b),
1658 mdname(mddev), conf->raid_disks - mddev->degraded);
1659}
1660
1661static void print_conf(struct r1conf *conf)
1662{
1663 int i;
1664
1665 pr_debug("RAID1 conf printout:\n");
1666 if (!conf) {
1667 pr_debug("(!conf)\n");
1668 return;
1669 }
1670 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1671 conf->raid_disks);
1672
1673 rcu_read_lock();
1674 for (i = 0; i < conf->raid_disks; i++) {
1675 char b[BDEVNAME_SIZE];
1676 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1677 if (rdev)
1678 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1679 i, !test_bit(In_sync, &rdev->flags),
1680 !test_bit(Faulty, &rdev->flags),
1681 bdevname(rdev->bdev,b));
1682 }
1683 rcu_read_unlock();
1684}
1685
1686static void close_sync(struct r1conf *conf)
1687{
1688 int idx;
1689
1690 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1691 _wait_barrier(conf, idx);
1692 _allow_barrier(conf, idx);
1693 }
1694
1695 mempool_exit(&conf->r1buf_pool);
1696}
1697
1698static int raid1_spare_active(struct mddev *mddev)
1699{
1700 int i;
1701 struct r1conf *conf = mddev->private;
1702 int count = 0;
1703 unsigned long flags;
1704
1705 /*
1706 * Find all failed disks within the RAID1 configuration
1707 * and mark them readable.
1708 * Called under mddev lock, so rcu protection not needed.
1709 * device_lock used to avoid races with raid1_end_read_request
1710 * which expects 'In_sync' flags and ->degraded to be consistent.
1711 */
1712 spin_lock_irqsave(&conf->device_lock, flags);
1713 for (i = 0; i < conf->raid_disks; i++) {
1714 struct md_rdev *rdev = conf->mirrors[i].rdev;
1715 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1716 if (repl
1717 && !test_bit(Candidate, &repl->flags)
1718 && repl->recovery_offset == MaxSector
1719 && !test_bit(Faulty, &repl->flags)
1720 && !test_and_set_bit(In_sync, &repl->flags)) {
1721 /* replacement has just become active */
1722 if (!rdev ||
1723 !test_and_clear_bit(In_sync, &rdev->flags))
1724 count++;
1725 if (rdev) {
1726 /* Replaced device not technically
1727 * faulty, but we need to be sure
1728 * it gets removed and never re-added
1729 */
1730 set_bit(Faulty, &rdev->flags);
1731 sysfs_notify_dirent_safe(
1732 rdev->sysfs_state);
1733 }
1734 }
1735 if (rdev
1736 && rdev->recovery_offset == MaxSector
1737 && !test_bit(Faulty, &rdev->flags)
1738 && !test_and_set_bit(In_sync, &rdev->flags)) {
1739 count++;
1740 sysfs_notify_dirent_safe(rdev->sysfs_state);
1741 }
1742 }
1743 mddev->degraded -= count;
1744 spin_unlock_irqrestore(&conf->device_lock, flags);
1745
1746 print_conf(conf);
1747 return count;
1748}
1749
1750static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1751{
1752 struct r1conf *conf = mddev->private;
1753 int err = -EEXIST;
1754 int mirror = 0;
1755 struct raid1_info *p;
1756 int first = 0;
1757 int last = conf->raid_disks - 1;
1758
1759 if (mddev->recovery_disabled == conf->recovery_disabled)
1760 return -EBUSY;
1761
1762 if (md_integrity_add_rdev(rdev, mddev))
1763 return -ENXIO;
1764
1765 if (rdev->raid_disk >= 0)
1766 first = last = rdev->raid_disk;
1767
1768 /*
1769 * find the disk ... but prefer rdev->saved_raid_disk
1770 * if possible.
1771 */
1772 if (rdev->saved_raid_disk >= 0 &&
1773 rdev->saved_raid_disk >= first &&
1774 rdev->saved_raid_disk < conf->raid_disks &&
1775 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1776 first = last = rdev->saved_raid_disk;
1777
1778 for (mirror = first; mirror <= last; mirror++) {
1779 p = conf->mirrors + mirror;
1780 if (!p->rdev) {
1781 if (mddev->gendisk)
1782 disk_stack_limits(mddev->gendisk, rdev->bdev,
1783 rdev->data_offset << 9);
1784
1785 p->head_position = 0;
1786 rdev->raid_disk = mirror;
1787 err = 0;
1788 /* As all devices are equivalent, we don't need a full recovery
1789 * if this was recently any drive of the array
1790 */
1791 if (rdev->saved_raid_disk < 0)
1792 conf->fullsync = 1;
1793 rcu_assign_pointer(p->rdev, rdev);
1794 break;
1795 }
1796 if (test_bit(WantReplacement, &p->rdev->flags) &&
1797 p[conf->raid_disks].rdev == NULL) {
1798 /* Add this device as a replacement */
1799 clear_bit(In_sync, &rdev->flags);
1800 set_bit(Replacement, &rdev->flags);
1801 rdev->raid_disk = mirror;
1802 err = 0;
1803 conf->fullsync = 1;
1804 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1805 break;
1806 }
1807 }
1808 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1809 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1810 print_conf(conf);
1811 return err;
1812}
1813
1814static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1815{
1816 struct r1conf *conf = mddev->private;
1817 int err = 0;
1818 int number = rdev->raid_disk;
1819 struct raid1_info *p = conf->mirrors + number;
1820
1821 if (rdev != p->rdev)
1822 p = conf->mirrors + conf->raid_disks + number;
1823
1824 print_conf(conf);
1825 if (rdev == p->rdev) {
1826 if (test_bit(In_sync, &rdev->flags) ||
1827 atomic_read(&rdev->nr_pending)) {
1828 err = -EBUSY;
1829 goto abort;
1830 }
1831 /* Only remove non-faulty devices if recovery
1832 * is not possible.
1833 */
1834 if (!test_bit(Faulty, &rdev->flags) &&
1835 mddev->recovery_disabled != conf->recovery_disabled &&
1836 mddev->degraded < conf->raid_disks) {
1837 err = -EBUSY;
1838 goto abort;
1839 }
1840 p->rdev = NULL;
1841 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1842 synchronize_rcu();
1843 if (atomic_read(&rdev->nr_pending)) {
1844 /* lost the race, try later */
1845 err = -EBUSY;
1846 p->rdev = rdev;
1847 goto abort;
1848 }
1849 }
1850 if (conf->mirrors[conf->raid_disks + number].rdev) {
1851 /* We just removed a device that is being replaced.
1852 * Move down the replacement. We drain all IO before
1853 * doing this to avoid confusion.
1854 */
1855 struct md_rdev *repl =
1856 conf->mirrors[conf->raid_disks + number].rdev;
1857 freeze_array(conf, 0);
1858 if (atomic_read(&repl->nr_pending)) {
1859 /* It means that some queued IO of retry_list
1860 * hold repl. Thus, we cannot set replacement
1861 * as NULL, avoiding rdev NULL pointer
1862 * dereference in sync_request_write and
1863 * handle_write_finished.
1864 */
1865 err = -EBUSY;
1866 unfreeze_array(conf);
1867 goto abort;
1868 }
1869 clear_bit(Replacement, &repl->flags);
1870 p->rdev = repl;
1871 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1872 unfreeze_array(conf);
1873 }
1874
1875 clear_bit(WantReplacement, &rdev->flags);
1876 err = md_integrity_register(mddev);
1877 }
1878abort:
1879
1880 print_conf(conf);
1881 return err;
1882}
1883
1884static void end_sync_read(struct bio *bio)
1885{
1886 struct r1bio *r1_bio = get_resync_r1bio(bio);
1887
1888 update_head_pos(r1_bio->read_disk, r1_bio);
1889
1890 /*
1891 * we have read a block, now it needs to be re-written,
1892 * or re-read if the read failed.
1893 * We don't do much here, just schedule handling by raid1d
1894 */
1895 if (!bio->bi_status)
1896 set_bit(R1BIO_Uptodate, &r1_bio->state);
1897
1898 if (atomic_dec_and_test(&r1_bio->remaining))
1899 reschedule_retry(r1_bio);
1900}
1901
1902static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1903{
1904 sector_t sync_blocks = 0;
1905 sector_t s = r1_bio->sector;
1906 long sectors_to_go = r1_bio->sectors;
1907
1908 /* make sure these bits don't get cleared. */
1909 do {
1910 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1911 s += sync_blocks;
1912 sectors_to_go -= sync_blocks;
1913 } while (sectors_to_go > 0);
1914}
1915
1916static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1917{
1918 if (atomic_dec_and_test(&r1_bio->remaining)) {
1919 struct mddev *mddev = r1_bio->mddev;
1920 int s = r1_bio->sectors;
1921
1922 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1923 test_bit(R1BIO_WriteError, &r1_bio->state))
1924 reschedule_retry(r1_bio);
1925 else {
1926 put_buf(r1_bio);
1927 md_done_sync(mddev, s, uptodate);
1928 }
1929 }
1930}
1931
1932static void end_sync_write(struct bio *bio)
1933{
1934 int uptodate = !bio->bi_status;
1935 struct r1bio *r1_bio = get_resync_r1bio(bio);
1936 struct mddev *mddev = r1_bio->mddev;
1937 struct r1conf *conf = mddev->private;
1938 sector_t first_bad;
1939 int bad_sectors;
1940 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1941
1942 if (!uptodate) {
1943 abort_sync_write(mddev, r1_bio);
1944 set_bit(WriteErrorSeen, &rdev->flags);
1945 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1946 set_bit(MD_RECOVERY_NEEDED, &
1947 mddev->recovery);
1948 set_bit(R1BIO_WriteError, &r1_bio->state);
1949 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1950 &first_bad, &bad_sectors) &&
1951 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1952 r1_bio->sector,
1953 r1_bio->sectors,
1954 &first_bad, &bad_sectors)
1955 )
1956 set_bit(R1BIO_MadeGood, &r1_bio->state);
1957
1958 put_sync_write_buf(r1_bio, uptodate);
1959}
1960
1961static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1962 int sectors, struct page *page, int rw)
1963{
1964 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1965 /* success */
1966 return 1;
1967 if (rw == WRITE) {
1968 set_bit(WriteErrorSeen, &rdev->flags);
1969 if (!test_and_set_bit(WantReplacement,
1970 &rdev->flags))
1971 set_bit(MD_RECOVERY_NEEDED, &
1972 rdev->mddev->recovery);
1973 }
1974 /* need to record an error - either for the block or the device */
1975 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1976 md_error(rdev->mddev, rdev);
1977 return 0;
1978}
1979
1980static int fix_sync_read_error(struct r1bio *r1_bio)
1981{
1982 /* Try some synchronous reads of other devices to get
1983 * good data, much like with normal read errors. Only
1984 * read into the pages we already have so we don't
1985 * need to re-issue the read request.
1986 * We don't need to freeze the array, because being in an
1987 * active sync request, there is no normal IO, and
1988 * no overlapping syncs.
1989 * We don't need to check is_badblock() again as we
1990 * made sure that anything with a bad block in range
1991 * will have bi_end_io clear.
1992 */
1993 struct mddev *mddev = r1_bio->mddev;
1994 struct r1conf *conf = mddev->private;
1995 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1996 struct page **pages = get_resync_pages(bio)->pages;
1997 sector_t sect = r1_bio->sector;
1998 int sectors = r1_bio->sectors;
1999 int idx = 0;
2000 struct md_rdev *rdev;
2001
2002 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2003 if (test_bit(FailFast, &rdev->flags)) {
2004 /* Don't try recovering from here - just fail it
2005 * ... unless it is the last working device of course */
2006 md_error(mddev, rdev);
2007 if (test_bit(Faulty, &rdev->flags))
2008 /* Don't try to read from here, but make sure
2009 * put_buf does it's thing
2010 */
2011 bio->bi_end_io = end_sync_write;
2012 }
2013
2014 while(sectors) {
2015 int s = sectors;
2016 int d = r1_bio->read_disk;
2017 int success = 0;
2018 int start;
2019
2020 if (s > (PAGE_SIZE>>9))
2021 s = PAGE_SIZE >> 9;
2022 do {
2023 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2024 /* No rcu protection needed here devices
2025 * can only be removed when no resync is
2026 * active, and resync is currently active
2027 */
2028 rdev = conf->mirrors[d].rdev;
2029 if (sync_page_io(rdev, sect, s<<9,
2030 pages[idx],
2031 REQ_OP_READ, 0, false)) {
2032 success = 1;
2033 break;
2034 }
2035 }
2036 d++;
2037 if (d == conf->raid_disks * 2)
2038 d = 0;
2039 } while (!success && d != r1_bio->read_disk);
2040
2041 if (!success) {
2042 char b[BDEVNAME_SIZE];
2043 int abort = 0;
2044 /* Cannot read from anywhere, this block is lost.
2045 * Record a bad block on each device. If that doesn't
2046 * work just disable and interrupt the recovery.
2047 * Don't fail devices as that won't really help.
2048 */
2049 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2050 mdname(mddev), bio_devname(bio, b),
2051 (unsigned long long)r1_bio->sector);
2052 for (d = 0; d < conf->raid_disks * 2; d++) {
2053 rdev = conf->mirrors[d].rdev;
2054 if (!rdev || test_bit(Faulty, &rdev->flags))
2055 continue;
2056 if (!rdev_set_badblocks(rdev, sect, s, 0))
2057 abort = 1;
2058 }
2059 if (abort) {
2060 conf->recovery_disabled =
2061 mddev->recovery_disabled;
2062 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2063 md_done_sync(mddev, r1_bio->sectors, 0);
2064 put_buf(r1_bio);
2065 return 0;
2066 }
2067 /* Try next page */
2068 sectors -= s;
2069 sect += s;
2070 idx++;
2071 continue;
2072 }
2073
2074 start = d;
2075 /* write it back and re-read */
2076 while (d != r1_bio->read_disk) {
2077 if (d == 0)
2078 d = conf->raid_disks * 2;
2079 d--;
2080 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2081 continue;
2082 rdev = conf->mirrors[d].rdev;
2083 if (r1_sync_page_io(rdev, sect, s,
2084 pages[idx],
2085 WRITE) == 0) {
2086 r1_bio->bios[d]->bi_end_io = NULL;
2087 rdev_dec_pending(rdev, mddev);
2088 }
2089 }
2090 d = start;
2091 while (d != r1_bio->read_disk) {
2092 if (d == 0)
2093 d = conf->raid_disks * 2;
2094 d--;
2095 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2096 continue;
2097 rdev = conf->mirrors[d].rdev;
2098 if (r1_sync_page_io(rdev, sect, s,
2099 pages[idx],
2100 READ) != 0)
2101 atomic_add(s, &rdev->corrected_errors);
2102 }
2103 sectors -= s;
2104 sect += s;
2105 idx ++;
2106 }
2107 set_bit(R1BIO_Uptodate, &r1_bio->state);
2108 bio->bi_status = 0;
2109 return 1;
2110}
2111
2112static void process_checks(struct r1bio *r1_bio)
2113{
2114 /* We have read all readable devices. If we haven't
2115 * got the block, then there is no hope left.
2116 * If we have, then we want to do a comparison
2117 * and skip the write if everything is the same.
2118 * If any blocks failed to read, then we need to
2119 * attempt an over-write
2120 */
2121 struct mddev *mddev = r1_bio->mddev;
2122 struct r1conf *conf = mddev->private;
2123 int primary;
2124 int i;
2125 int vcnt;
2126
2127 /* Fix variable parts of all bios */
2128 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2129 for (i = 0; i < conf->raid_disks * 2; i++) {
2130 blk_status_t status;
2131 struct bio *b = r1_bio->bios[i];
2132 struct resync_pages *rp = get_resync_pages(b);
2133 if (b->bi_end_io != end_sync_read)
2134 continue;
2135 /* fixup the bio for reuse, but preserve errno */
2136 status = b->bi_status;
2137 bio_reset(b);
2138 b->bi_status = status;
2139 b->bi_iter.bi_sector = r1_bio->sector +
2140 conf->mirrors[i].rdev->data_offset;
2141 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2142 b->bi_end_io = end_sync_read;
2143 rp->raid_bio = r1_bio;
2144 b->bi_private = rp;
2145
2146 /* initialize bvec table again */
2147 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2148 }
2149 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2150 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2151 !r1_bio->bios[primary]->bi_status) {
2152 r1_bio->bios[primary]->bi_end_io = NULL;
2153 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2154 break;
2155 }
2156 r1_bio->read_disk = primary;
2157 for (i = 0; i < conf->raid_disks * 2; i++) {
2158 int j = 0;
2159 struct bio *pbio = r1_bio->bios[primary];
2160 struct bio *sbio = r1_bio->bios[i];
2161 blk_status_t status = sbio->bi_status;
2162 struct page **ppages = get_resync_pages(pbio)->pages;
2163 struct page **spages = get_resync_pages(sbio)->pages;
2164 struct bio_vec *bi;
2165 int page_len[RESYNC_PAGES] = { 0 };
2166 struct bvec_iter_all iter_all;
2167
2168 if (sbio->bi_end_io != end_sync_read)
2169 continue;
2170 /* Now we can 'fixup' the error value */
2171 sbio->bi_status = 0;
2172
2173 bio_for_each_segment_all(bi, sbio, iter_all)
2174 page_len[j++] = bi->bv_len;
2175
2176 if (!status) {
2177 for (j = vcnt; j-- ; ) {
2178 if (memcmp(page_address(ppages[j]),
2179 page_address(spages[j]),
2180 page_len[j]))
2181 break;
2182 }
2183 } else
2184 j = 0;
2185 if (j >= 0)
2186 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2187 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2188 && !status)) {
2189 /* No need to write to this device. */
2190 sbio->bi_end_io = NULL;
2191 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2192 continue;
2193 }
2194
2195 bio_copy_data(sbio, pbio);
2196 }
2197}
2198
2199static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2200{
2201 struct r1conf *conf = mddev->private;
2202 int i;
2203 int disks = conf->raid_disks * 2;
2204 struct bio *wbio;
2205
2206 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2207 /* ouch - failed to read all of that. */
2208 if (!fix_sync_read_error(r1_bio))
2209 return;
2210
2211 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2212 process_checks(r1_bio);
2213
2214 /*
2215 * schedule writes
2216 */
2217 atomic_set(&r1_bio->remaining, 1);
2218 for (i = 0; i < disks ; i++) {
2219 wbio = r1_bio->bios[i];
2220 if (wbio->bi_end_io == NULL ||
2221 (wbio->bi_end_io == end_sync_read &&
2222 (i == r1_bio->read_disk ||
2223 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2224 continue;
2225 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2226 abort_sync_write(mddev, r1_bio);
2227 continue;
2228 }
2229
2230 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2231 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2232 wbio->bi_opf |= MD_FAILFAST;
2233
2234 wbio->bi_end_io = end_sync_write;
2235 atomic_inc(&r1_bio->remaining);
2236 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2237
2238 submit_bio_noacct(wbio);
2239 }
2240
2241 put_sync_write_buf(r1_bio, 1);
2242}
2243
2244/*
2245 * This is a kernel thread which:
2246 *
2247 * 1. Retries failed read operations on working mirrors.
2248 * 2. Updates the raid superblock when problems encounter.
2249 * 3. Performs writes following reads for array synchronising.
2250 */
2251
2252static void fix_read_error(struct r1conf *conf, int read_disk,
2253 sector_t sect, int sectors)
2254{
2255 struct mddev *mddev = conf->mddev;
2256 while(sectors) {
2257 int s = sectors;
2258 int d = read_disk;
2259 int success = 0;
2260 int start;
2261 struct md_rdev *rdev;
2262
2263 if (s > (PAGE_SIZE>>9))
2264 s = PAGE_SIZE >> 9;
2265
2266 do {
2267 sector_t first_bad;
2268 int bad_sectors;
2269
2270 rcu_read_lock();
2271 rdev = rcu_dereference(conf->mirrors[d].rdev);
2272 if (rdev &&
2273 (test_bit(In_sync, &rdev->flags) ||
2274 (!test_bit(Faulty, &rdev->flags) &&
2275 rdev->recovery_offset >= sect + s)) &&
2276 is_badblock(rdev, sect, s,
2277 &first_bad, &bad_sectors) == 0) {
2278 atomic_inc(&rdev->nr_pending);
2279 rcu_read_unlock();
2280 if (sync_page_io(rdev, sect, s<<9,
2281 conf->tmppage, REQ_OP_READ, 0, false))
2282 success = 1;
2283 rdev_dec_pending(rdev, mddev);
2284 if (success)
2285 break;
2286 } else
2287 rcu_read_unlock();
2288 d++;
2289 if (d == conf->raid_disks * 2)
2290 d = 0;
2291 } while (!success && d != read_disk);
2292
2293 if (!success) {
2294 /* Cannot read from anywhere - mark it bad */
2295 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2296 if (!rdev_set_badblocks(rdev, sect, s, 0))
2297 md_error(mddev, rdev);
2298 break;
2299 }
2300 /* write it back and re-read */
2301 start = d;
2302 while (d != read_disk) {
2303 if (d==0)
2304 d = conf->raid_disks * 2;
2305 d--;
2306 rcu_read_lock();
2307 rdev = rcu_dereference(conf->mirrors[d].rdev);
2308 if (rdev &&
2309 !test_bit(Faulty, &rdev->flags)) {
2310 atomic_inc(&rdev->nr_pending);
2311 rcu_read_unlock();
2312 r1_sync_page_io(rdev, sect, s,
2313 conf->tmppage, WRITE);
2314 rdev_dec_pending(rdev, mddev);
2315 } else
2316 rcu_read_unlock();
2317 }
2318 d = start;
2319 while (d != read_disk) {
2320 char b[BDEVNAME_SIZE];
2321 if (d==0)
2322 d = conf->raid_disks * 2;
2323 d--;
2324 rcu_read_lock();
2325 rdev = rcu_dereference(conf->mirrors[d].rdev);
2326 if (rdev &&
2327 !test_bit(Faulty, &rdev->flags)) {
2328 atomic_inc(&rdev->nr_pending);
2329 rcu_read_unlock();
2330 if (r1_sync_page_io(rdev, sect, s,
2331 conf->tmppage, READ)) {
2332 atomic_add(s, &rdev->corrected_errors);
2333 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2334 mdname(mddev), s,
2335 (unsigned long long)(sect +
2336 rdev->data_offset),
2337 bdevname(rdev->bdev, b));
2338 }
2339 rdev_dec_pending(rdev, mddev);
2340 } else
2341 rcu_read_unlock();
2342 }
2343 sectors -= s;
2344 sect += s;
2345 }
2346}
2347
2348static int narrow_write_error(struct r1bio *r1_bio, int i)
2349{
2350 struct mddev *mddev = r1_bio->mddev;
2351 struct r1conf *conf = mddev->private;
2352 struct md_rdev *rdev = conf->mirrors[i].rdev;
2353
2354 /* bio has the data to be written to device 'i' where
2355 * we just recently had a write error.
2356 * We repeatedly clone the bio and trim down to one block,
2357 * then try the write. Where the write fails we record
2358 * a bad block.
2359 * It is conceivable that the bio doesn't exactly align with
2360 * blocks. We must handle this somehow.
2361 *
2362 * We currently own a reference on the rdev.
2363 */
2364
2365 int block_sectors;
2366 sector_t sector;
2367 int sectors;
2368 int sect_to_write = r1_bio->sectors;
2369 int ok = 1;
2370
2371 if (rdev->badblocks.shift < 0)
2372 return 0;
2373
2374 block_sectors = roundup(1 << rdev->badblocks.shift,
2375 bdev_logical_block_size(rdev->bdev) >> 9);
2376 sector = r1_bio->sector;
2377 sectors = ((sector + block_sectors)
2378 & ~(sector_t)(block_sectors - 1))
2379 - sector;
2380
2381 while (sect_to_write) {
2382 struct bio *wbio;
2383 if (sectors > sect_to_write)
2384 sectors = sect_to_write;
2385 /* Write at 'sector' for 'sectors'*/
2386
2387 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2388 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2389 GFP_NOIO,
2390 &mddev->bio_set);
2391 } else {
2392 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2393 &mddev->bio_set);
2394 }
2395
2396 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2397 wbio->bi_iter.bi_sector = r1_bio->sector;
2398 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2399
2400 bio_trim(wbio, sector - r1_bio->sector, sectors);
2401 wbio->bi_iter.bi_sector += rdev->data_offset;
2402 bio_set_dev(wbio, rdev->bdev);
2403
2404 if (submit_bio_wait(wbio) < 0)
2405 /* failure! */
2406 ok = rdev_set_badblocks(rdev, sector,
2407 sectors, 0)
2408 && ok;
2409
2410 bio_put(wbio);
2411 sect_to_write -= sectors;
2412 sector += sectors;
2413 sectors = block_sectors;
2414 }
2415 return ok;
2416}
2417
2418static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2419{
2420 int m;
2421 int s = r1_bio->sectors;
2422 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2423 struct md_rdev *rdev = conf->mirrors[m].rdev;
2424 struct bio *bio = r1_bio->bios[m];
2425 if (bio->bi_end_io == NULL)
2426 continue;
2427 if (!bio->bi_status &&
2428 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2429 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2430 }
2431 if (bio->bi_status &&
2432 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2433 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2434 md_error(conf->mddev, rdev);
2435 }
2436 }
2437 put_buf(r1_bio);
2438 md_done_sync(conf->mddev, s, 1);
2439}
2440
2441static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2442{
2443 int m, idx;
2444 bool fail = false;
2445
2446 for (m = 0; m < conf->raid_disks * 2 ; m++)
2447 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2448 struct md_rdev *rdev = conf->mirrors[m].rdev;
2449 rdev_clear_badblocks(rdev,
2450 r1_bio->sector,
2451 r1_bio->sectors, 0);
2452 rdev_dec_pending(rdev, conf->mddev);
2453 } else if (r1_bio->bios[m] != NULL) {
2454 /* This drive got a write error. We need to
2455 * narrow down and record precise write
2456 * errors.
2457 */
2458 fail = true;
2459 if (!narrow_write_error(r1_bio, m)) {
2460 md_error(conf->mddev,
2461 conf->mirrors[m].rdev);
2462 /* an I/O failed, we can't clear the bitmap */
2463 set_bit(R1BIO_Degraded, &r1_bio->state);
2464 }
2465 rdev_dec_pending(conf->mirrors[m].rdev,
2466 conf->mddev);
2467 }
2468 if (fail) {
2469 spin_lock_irq(&conf->device_lock);
2470 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2471 idx = sector_to_idx(r1_bio->sector);
2472 atomic_inc(&conf->nr_queued[idx]);
2473 spin_unlock_irq(&conf->device_lock);
2474 /*
2475 * In case freeze_array() is waiting for condition
2476 * get_unqueued_pending() == extra to be true.
2477 */
2478 wake_up(&conf->wait_barrier);
2479 md_wakeup_thread(conf->mddev->thread);
2480 } else {
2481 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2482 close_write(r1_bio);
2483 raid_end_bio_io(r1_bio);
2484 }
2485}
2486
2487static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2488{
2489 struct mddev *mddev = conf->mddev;
2490 struct bio *bio;
2491 struct md_rdev *rdev;
2492
2493 clear_bit(R1BIO_ReadError, &r1_bio->state);
2494 /* we got a read error. Maybe the drive is bad. Maybe just
2495 * the block and we can fix it.
2496 * We freeze all other IO, and try reading the block from
2497 * other devices. When we find one, we re-write
2498 * and check it that fixes the read error.
2499 * This is all done synchronously while the array is
2500 * frozen
2501 */
2502
2503 bio = r1_bio->bios[r1_bio->read_disk];
2504 bio_put(bio);
2505 r1_bio->bios[r1_bio->read_disk] = NULL;
2506
2507 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2508 if (mddev->ro == 0
2509 && !test_bit(FailFast, &rdev->flags)) {
2510 freeze_array(conf, 1);
2511 fix_read_error(conf, r1_bio->read_disk,
2512 r1_bio->sector, r1_bio->sectors);
2513 unfreeze_array(conf);
2514 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2515 md_error(mddev, rdev);
2516 } else {
2517 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2518 }
2519
2520 rdev_dec_pending(rdev, conf->mddev);
2521 allow_barrier(conf, r1_bio->sector);
2522 bio = r1_bio->master_bio;
2523
2524 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2525 r1_bio->state = 0;
2526 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2527}
2528
2529static void raid1d(struct md_thread *thread)
2530{
2531 struct mddev *mddev = thread->mddev;
2532 struct r1bio *r1_bio;
2533 unsigned long flags;
2534 struct r1conf *conf = mddev->private;
2535 struct list_head *head = &conf->retry_list;
2536 struct blk_plug plug;
2537 int idx;
2538
2539 md_check_recovery(mddev);
2540
2541 if (!list_empty_careful(&conf->bio_end_io_list) &&
2542 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2543 LIST_HEAD(tmp);
2544 spin_lock_irqsave(&conf->device_lock, flags);
2545 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2546 list_splice_init(&conf->bio_end_io_list, &tmp);
2547 spin_unlock_irqrestore(&conf->device_lock, flags);
2548 while (!list_empty(&tmp)) {
2549 r1_bio = list_first_entry(&tmp, struct r1bio,
2550 retry_list);
2551 list_del(&r1_bio->retry_list);
2552 idx = sector_to_idx(r1_bio->sector);
2553 atomic_dec(&conf->nr_queued[idx]);
2554 if (mddev->degraded)
2555 set_bit(R1BIO_Degraded, &r1_bio->state);
2556 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2557 close_write(r1_bio);
2558 raid_end_bio_io(r1_bio);
2559 }
2560 }
2561
2562 blk_start_plug(&plug);
2563 for (;;) {
2564
2565 flush_pending_writes(conf);
2566
2567 spin_lock_irqsave(&conf->device_lock, flags);
2568 if (list_empty(head)) {
2569 spin_unlock_irqrestore(&conf->device_lock, flags);
2570 break;
2571 }
2572 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2573 list_del(head->prev);
2574 idx = sector_to_idx(r1_bio->sector);
2575 atomic_dec(&conf->nr_queued[idx]);
2576 spin_unlock_irqrestore(&conf->device_lock, flags);
2577
2578 mddev = r1_bio->mddev;
2579 conf = mddev->private;
2580 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2581 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2582 test_bit(R1BIO_WriteError, &r1_bio->state))
2583 handle_sync_write_finished(conf, r1_bio);
2584 else
2585 sync_request_write(mddev, r1_bio);
2586 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2587 test_bit(R1BIO_WriteError, &r1_bio->state))
2588 handle_write_finished(conf, r1_bio);
2589 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2590 handle_read_error(conf, r1_bio);
2591 else
2592 WARN_ON_ONCE(1);
2593
2594 cond_resched();
2595 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2596 md_check_recovery(mddev);
2597 }
2598 blk_finish_plug(&plug);
2599}
2600
2601static int init_resync(struct r1conf *conf)
2602{
2603 int buffs;
2604
2605 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2606 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2607
2608 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2609 r1buf_pool_free, conf->poolinfo);
2610}
2611
2612static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2613{
2614 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2615 struct resync_pages *rps;
2616 struct bio *bio;
2617 int i;
2618
2619 for (i = conf->poolinfo->raid_disks; i--; ) {
2620 bio = r1bio->bios[i];
2621 rps = bio->bi_private;
2622 bio_reset(bio);
2623 bio->bi_private = rps;
2624 }
2625 r1bio->master_bio = NULL;
2626 return r1bio;
2627}
2628
2629/*
2630 * perform a "sync" on one "block"
2631 *
2632 * We need to make sure that no normal I/O request - particularly write
2633 * requests - conflict with active sync requests.
2634 *
2635 * This is achieved by tracking pending requests and a 'barrier' concept
2636 * that can be installed to exclude normal IO requests.
2637 */
2638
2639static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2640 int *skipped)
2641{
2642 struct r1conf *conf = mddev->private;
2643 struct r1bio *r1_bio;
2644 struct bio *bio;
2645 sector_t max_sector, nr_sectors;
2646 int disk = -1;
2647 int i;
2648 int wonly = -1;
2649 int write_targets = 0, read_targets = 0;
2650 sector_t sync_blocks;
2651 int still_degraded = 0;
2652 int good_sectors = RESYNC_SECTORS;
2653 int min_bad = 0; /* number of sectors that are bad in all devices */
2654 int idx = sector_to_idx(sector_nr);
2655 int page_idx = 0;
2656
2657 if (!mempool_initialized(&conf->r1buf_pool))
2658 if (init_resync(conf))
2659 return 0;
2660
2661 max_sector = mddev->dev_sectors;
2662 if (sector_nr >= max_sector) {
2663 /* If we aborted, we need to abort the
2664 * sync on the 'current' bitmap chunk (there will
2665 * only be one in raid1 resync.
2666 * We can find the current addess in mddev->curr_resync
2667 */
2668 if (mddev->curr_resync < max_sector) /* aborted */
2669 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2670 &sync_blocks, 1);
2671 else /* completed sync */
2672 conf->fullsync = 0;
2673
2674 md_bitmap_close_sync(mddev->bitmap);
2675 close_sync(conf);
2676
2677 if (mddev_is_clustered(mddev)) {
2678 conf->cluster_sync_low = 0;
2679 conf->cluster_sync_high = 0;
2680 }
2681 return 0;
2682 }
2683
2684 if (mddev->bitmap == NULL &&
2685 mddev->recovery_cp == MaxSector &&
2686 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2687 conf->fullsync == 0) {
2688 *skipped = 1;
2689 return max_sector - sector_nr;
2690 }
2691 /* before building a request, check if we can skip these blocks..
2692 * This call the bitmap_start_sync doesn't actually record anything
2693 */
2694 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2695 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2696 /* We can skip this block, and probably several more */
2697 *skipped = 1;
2698 return sync_blocks;
2699 }
2700
2701 /*
2702 * If there is non-resync activity waiting for a turn, then let it
2703 * though before starting on this new sync request.
2704 */
2705 if (atomic_read(&conf->nr_waiting[idx]))
2706 schedule_timeout_uninterruptible(1);
2707
2708 /* we are incrementing sector_nr below. To be safe, we check against
2709 * sector_nr + two times RESYNC_SECTORS
2710 */
2711
2712 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2713 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2714
2715
2716 if (raise_barrier(conf, sector_nr))
2717 return 0;
2718
2719 r1_bio = raid1_alloc_init_r1buf(conf);
2720
2721 rcu_read_lock();
2722 /*
2723 * If we get a correctably read error during resync or recovery,
2724 * we might want to read from a different device. So we
2725 * flag all drives that could conceivably be read from for READ,
2726 * and any others (which will be non-In_sync devices) for WRITE.
2727 * If a read fails, we try reading from something else for which READ
2728 * is OK.
2729 */
2730
2731 r1_bio->mddev = mddev;
2732 r1_bio->sector = sector_nr;
2733 r1_bio->state = 0;
2734 set_bit(R1BIO_IsSync, &r1_bio->state);
2735 /* make sure good_sectors won't go across barrier unit boundary */
2736 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2737
2738 for (i = 0; i < conf->raid_disks * 2; i++) {
2739 struct md_rdev *rdev;
2740 bio = r1_bio->bios[i];
2741
2742 rdev = rcu_dereference(conf->mirrors[i].rdev);
2743 if (rdev == NULL ||
2744 test_bit(Faulty, &rdev->flags)) {
2745 if (i < conf->raid_disks)
2746 still_degraded = 1;
2747 } else if (!test_bit(In_sync, &rdev->flags)) {
2748 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2749 bio->bi_end_io = end_sync_write;
2750 write_targets ++;
2751 } else {
2752 /* may need to read from here */
2753 sector_t first_bad = MaxSector;
2754 int bad_sectors;
2755
2756 if (is_badblock(rdev, sector_nr, good_sectors,
2757 &first_bad, &bad_sectors)) {
2758 if (first_bad > sector_nr)
2759 good_sectors = first_bad - sector_nr;
2760 else {
2761 bad_sectors -= (sector_nr - first_bad);
2762 if (min_bad == 0 ||
2763 min_bad > bad_sectors)
2764 min_bad = bad_sectors;
2765 }
2766 }
2767 if (sector_nr < first_bad) {
2768 if (test_bit(WriteMostly, &rdev->flags)) {
2769 if (wonly < 0)
2770 wonly = i;
2771 } else {
2772 if (disk < 0)
2773 disk = i;
2774 }
2775 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2776 bio->bi_end_io = end_sync_read;
2777 read_targets++;
2778 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2779 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2780 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2781 /*
2782 * The device is suitable for reading (InSync),
2783 * but has bad block(s) here. Let's try to correct them,
2784 * if we are doing resync or repair. Otherwise, leave
2785 * this device alone for this sync request.
2786 */
2787 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2788 bio->bi_end_io = end_sync_write;
2789 write_targets++;
2790 }
2791 }
2792 if (rdev && bio->bi_end_io) {
2793 atomic_inc(&rdev->nr_pending);
2794 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2795 bio_set_dev(bio, rdev->bdev);
2796 if (test_bit(FailFast, &rdev->flags))
2797 bio->bi_opf |= MD_FAILFAST;
2798 }
2799 }
2800 rcu_read_unlock();
2801 if (disk < 0)
2802 disk = wonly;
2803 r1_bio->read_disk = disk;
2804
2805 if (read_targets == 0 && min_bad > 0) {
2806 /* These sectors are bad on all InSync devices, so we
2807 * need to mark them bad on all write targets
2808 */
2809 int ok = 1;
2810 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2811 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2812 struct md_rdev *rdev = conf->mirrors[i].rdev;
2813 ok = rdev_set_badblocks(rdev, sector_nr,
2814 min_bad, 0
2815 ) && ok;
2816 }
2817 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2818 *skipped = 1;
2819 put_buf(r1_bio);
2820
2821 if (!ok) {
2822 /* Cannot record the badblocks, so need to
2823 * abort the resync.
2824 * If there are multiple read targets, could just
2825 * fail the really bad ones ???
2826 */
2827 conf->recovery_disabled = mddev->recovery_disabled;
2828 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2829 return 0;
2830 } else
2831 return min_bad;
2832
2833 }
2834 if (min_bad > 0 && min_bad < good_sectors) {
2835 /* only resync enough to reach the next bad->good
2836 * transition */
2837 good_sectors = min_bad;
2838 }
2839
2840 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2841 /* extra read targets are also write targets */
2842 write_targets += read_targets-1;
2843
2844 if (write_targets == 0 || read_targets == 0) {
2845 /* There is nowhere to write, so all non-sync
2846 * drives must be failed - so we are finished
2847 */
2848 sector_t rv;
2849 if (min_bad > 0)
2850 max_sector = sector_nr + min_bad;
2851 rv = max_sector - sector_nr;
2852 *skipped = 1;
2853 put_buf(r1_bio);
2854 return rv;
2855 }
2856
2857 if (max_sector > mddev->resync_max)
2858 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2859 if (max_sector > sector_nr + good_sectors)
2860 max_sector = sector_nr + good_sectors;
2861 nr_sectors = 0;
2862 sync_blocks = 0;
2863 do {
2864 struct page *page;
2865 int len = PAGE_SIZE;
2866 if (sector_nr + (len>>9) > max_sector)
2867 len = (max_sector - sector_nr) << 9;
2868 if (len == 0)
2869 break;
2870 if (sync_blocks == 0) {
2871 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2872 &sync_blocks, still_degraded) &&
2873 !conf->fullsync &&
2874 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2875 break;
2876 if ((len >> 9) > sync_blocks)
2877 len = sync_blocks<<9;
2878 }
2879
2880 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2881 struct resync_pages *rp;
2882
2883 bio = r1_bio->bios[i];
2884 rp = get_resync_pages(bio);
2885 if (bio->bi_end_io) {
2886 page = resync_fetch_page(rp, page_idx);
2887
2888 /*
2889 * won't fail because the vec table is big
2890 * enough to hold all these pages
2891 */
2892 bio_add_page(bio, page, len, 0);
2893 }
2894 }
2895 nr_sectors += len>>9;
2896 sector_nr += len>>9;
2897 sync_blocks -= (len>>9);
2898 } while (++page_idx < RESYNC_PAGES);
2899
2900 r1_bio->sectors = nr_sectors;
2901
2902 if (mddev_is_clustered(mddev) &&
2903 conf->cluster_sync_high < sector_nr + nr_sectors) {
2904 conf->cluster_sync_low = mddev->curr_resync_completed;
2905 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2906 /* Send resync message */
2907 md_cluster_ops->resync_info_update(mddev,
2908 conf->cluster_sync_low,
2909 conf->cluster_sync_high);
2910 }
2911
2912 /* For a user-requested sync, we read all readable devices and do a
2913 * compare
2914 */
2915 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2916 atomic_set(&r1_bio->remaining, read_targets);
2917 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2918 bio = r1_bio->bios[i];
2919 if (bio->bi_end_io == end_sync_read) {
2920 read_targets--;
2921 md_sync_acct_bio(bio, nr_sectors);
2922 if (read_targets == 1)
2923 bio->bi_opf &= ~MD_FAILFAST;
2924 submit_bio_noacct(bio);
2925 }
2926 }
2927 } else {
2928 atomic_set(&r1_bio->remaining, 1);
2929 bio = r1_bio->bios[r1_bio->read_disk];
2930 md_sync_acct_bio(bio, nr_sectors);
2931 if (read_targets == 1)
2932 bio->bi_opf &= ~MD_FAILFAST;
2933 submit_bio_noacct(bio);
2934 }
2935 return nr_sectors;
2936}
2937
2938static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2939{
2940 if (sectors)
2941 return sectors;
2942
2943 return mddev->dev_sectors;
2944}
2945
2946static struct r1conf *setup_conf(struct mddev *mddev)
2947{
2948 struct r1conf *conf;
2949 int i;
2950 struct raid1_info *disk;
2951 struct md_rdev *rdev;
2952 int err = -ENOMEM;
2953
2954 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2955 if (!conf)
2956 goto abort;
2957
2958 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2959 sizeof(atomic_t), GFP_KERNEL);
2960 if (!conf->nr_pending)
2961 goto abort;
2962
2963 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2964 sizeof(atomic_t), GFP_KERNEL);
2965 if (!conf->nr_waiting)
2966 goto abort;
2967
2968 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2969 sizeof(atomic_t), GFP_KERNEL);
2970 if (!conf->nr_queued)
2971 goto abort;
2972
2973 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2974 sizeof(atomic_t), GFP_KERNEL);
2975 if (!conf->barrier)
2976 goto abort;
2977
2978 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2979 mddev->raid_disks, 2),
2980 GFP_KERNEL);
2981 if (!conf->mirrors)
2982 goto abort;
2983
2984 conf->tmppage = alloc_page(GFP_KERNEL);
2985 if (!conf->tmppage)
2986 goto abort;
2987
2988 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2989 if (!conf->poolinfo)
2990 goto abort;
2991 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2992 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2993 rbio_pool_free, conf->poolinfo);
2994 if (err)
2995 goto abort;
2996
2997 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2998 if (err)
2999 goto abort;
3000
3001 conf->poolinfo->mddev = mddev;
3002
3003 err = -EINVAL;
3004 spin_lock_init(&conf->device_lock);
3005 rdev_for_each(rdev, mddev) {
3006 int disk_idx = rdev->raid_disk;
3007 if (disk_idx >= mddev->raid_disks
3008 || disk_idx < 0)
3009 continue;
3010 if (test_bit(Replacement, &rdev->flags))
3011 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3012 else
3013 disk = conf->mirrors + disk_idx;
3014
3015 if (disk->rdev)
3016 goto abort;
3017 disk->rdev = rdev;
3018 disk->head_position = 0;
3019 disk->seq_start = MaxSector;
3020 }
3021 conf->raid_disks = mddev->raid_disks;
3022 conf->mddev = mddev;
3023 INIT_LIST_HEAD(&conf->retry_list);
3024 INIT_LIST_HEAD(&conf->bio_end_io_list);
3025
3026 spin_lock_init(&conf->resync_lock);
3027 init_waitqueue_head(&conf->wait_barrier);
3028
3029 bio_list_init(&conf->pending_bio_list);
3030 conf->pending_count = 0;
3031 conf->recovery_disabled = mddev->recovery_disabled - 1;
3032
3033 err = -EIO;
3034 for (i = 0; i < conf->raid_disks * 2; i++) {
3035
3036 disk = conf->mirrors + i;
3037
3038 if (i < conf->raid_disks &&
3039 disk[conf->raid_disks].rdev) {
3040 /* This slot has a replacement. */
3041 if (!disk->rdev) {
3042 /* No original, just make the replacement
3043 * a recovering spare
3044 */
3045 disk->rdev =
3046 disk[conf->raid_disks].rdev;
3047 disk[conf->raid_disks].rdev = NULL;
3048 } else if (!test_bit(In_sync, &disk->rdev->flags))
3049 /* Original is not in_sync - bad */
3050 goto abort;
3051 }
3052
3053 if (!disk->rdev ||
3054 !test_bit(In_sync, &disk->rdev->flags)) {
3055 disk->head_position = 0;
3056 if (disk->rdev &&
3057 (disk->rdev->saved_raid_disk < 0))
3058 conf->fullsync = 1;
3059 }
3060 }
3061
3062 err = -ENOMEM;
3063 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3064 if (!conf->thread)
3065 goto abort;
3066
3067 return conf;
3068
3069 abort:
3070 if (conf) {
3071 mempool_exit(&conf->r1bio_pool);
3072 kfree(conf->mirrors);
3073 safe_put_page(conf->tmppage);
3074 kfree(conf->poolinfo);
3075 kfree(conf->nr_pending);
3076 kfree(conf->nr_waiting);
3077 kfree(conf->nr_queued);
3078 kfree(conf->barrier);
3079 bioset_exit(&conf->bio_split);
3080 kfree(conf);
3081 }
3082 return ERR_PTR(err);
3083}
3084
3085static void raid1_free(struct mddev *mddev, void *priv);
3086static int raid1_run(struct mddev *mddev)
3087{
3088 struct r1conf *conf;
3089 int i;
3090 struct md_rdev *rdev;
3091 int ret;
3092 bool discard_supported = false;
3093
3094 if (mddev->level != 1) {
3095 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3096 mdname(mddev), mddev->level);
3097 return -EIO;
3098 }
3099 if (mddev->reshape_position != MaxSector) {
3100 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3101 mdname(mddev));
3102 return -EIO;
3103 }
3104 if (mddev_init_writes_pending(mddev) < 0)
3105 return -ENOMEM;
3106 /*
3107 * copy the already verified devices into our private RAID1
3108 * bookkeeping area. [whatever we allocate in run(),
3109 * should be freed in raid1_free()]
3110 */
3111 if (mddev->private == NULL)
3112 conf = setup_conf(mddev);
3113 else
3114 conf = mddev->private;
3115
3116 if (IS_ERR(conf))
3117 return PTR_ERR(conf);
3118
3119 if (mddev->queue) {
3120 blk_queue_max_write_same_sectors(mddev->queue, 0);
3121 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3122 }
3123
3124 rdev_for_each(rdev, mddev) {
3125 if (!mddev->gendisk)
3126 continue;
3127 disk_stack_limits(mddev->gendisk, rdev->bdev,
3128 rdev->data_offset << 9);
3129 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3130 discard_supported = true;
3131 }
3132
3133 mddev->degraded = 0;
3134 for (i = 0; i < conf->raid_disks; i++)
3135 if (conf->mirrors[i].rdev == NULL ||
3136 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3137 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3138 mddev->degraded++;
3139 /*
3140 * RAID1 needs at least one disk in active
3141 */
3142 if (conf->raid_disks - mddev->degraded < 1) {
3143 ret = -EINVAL;
3144 goto abort;
3145 }
3146
3147 if (conf->raid_disks - mddev->degraded == 1)
3148 mddev->recovery_cp = MaxSector;
3149
3150 if (mddev->recovery_cp != MaxSector)
3151 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3152 mdname(mddev));
3153 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3154 mdname(mddev), mddev->raid_disks - mddev->degraded,
3155 mddev->raid_disks);
3156
3157 /*
3158 * Ok, everything is just fine now
3159 */
3160 mddev->thread = conf->thread;
3161 conf->thread = NULL;
3162 mddev->private = conf;
3163 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3164
3165 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3166
3167 if (mddev->queue) {
3168 if (discard_supported)
3169 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3170 mddev->queue);
3171 else
3172 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3173 mddev->queue);
3174 }
3175
3176 ret = md_integrity_register(mddev);
3177 if (ret) {
3178 md_unregister_thread(&mddev->thread);
3179 goto abort;
3180 }
3181 return 0;
3182
3183abort:
3184 raid1_free(mddev, conf);
3185 return ret;
3186}
3187
3188static void raid1_free(struct mddev *mddev, void *priv)
3189{
3190 struct r1conf *conf = priv;
3191
3192 mempool_exit(&conf->r1bio_pool);
3193 kfree(conf->mirrors);
3194 safe_put_page(conf->tmppage);
3195 kfree(conf->poolinfo);
3196 kfree(conf->nr_pending);
3197 kfree(conf->nr_waiting);
3198 kfree(conf->nr_queued);
3199 kfree(conf->barrier);
3200 bioset_exit(&conf->bio_split);
3201 kfree(conf);
3202}
3203
3204static int raid1_resize(struct mddev *mddev, sector_t sectors)
3205{
3206 /* no resync is happening, and there is enough space
3207 * on all devices, so we can resize.
3208 * We need to make sure resync covers any new space.
3209 * If the array is shrinking we should possibly wait until
3210 * any io in the removed space completes, but it hardly seems
3211 * worth it.
3212 */
3213 sector_t newsize = raid1_size(mddev, sectors, 0);
3214 if (mddev->external_size &&
3215 mddev->array_sectors > newsize)
3216 return -EINVAL;
3217 if (mddev->bitmap) {
3218 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3219 if (ret)
3220 return ret;
3221 }
3222 md_set_array_sectors(mddev, newsize);
3223 if (sectors > mddev->dev_sectors &&
3224 mddev->recovery_cp > mddev->dev_sectors) {
3225 mddev->recovery_cp = mddev->dev_sectors;
3226 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3227 }
3228 mddev->dev_sectors = sectors;
3229 mddev->resync_max_sectors = sectors;
3230 return 0;
3231}
3232
3233static int raid1_reshape(struct mddev *mddev)
3234{
3235 /* We need to:
3236 * 1/ resize the r1bio_pool
3237 * 2/ resize conf->mirrors
3238 *
3239 * We allocate a new r1bio_pool if we can.
3240 * Then raise a device barrier and wait until all IO stops.
3241 * Then resize conf->mirrors and swap in the new r1bio pool.
3242 *
3243 * At the same time, we "pack" the devices so that all the missing
3244 * devices have the higher raid_disk numbers.
3245 */
3246 mempool_t newpool, oldpool;
3247 struct pool_info *newpoolinfo;
3248 struct raid1_info *newmirrors;
3249 struct r1conf *conf = mddev->private;
3250 int cnt, raid_disks;
3251 unsigned long flags;
3252 int d, d2;
3253 int ret;
3254
3255 memset(&newpool, 0, sizeof(newpool));
3256 memset(&oldpool, 0, sizeof(oldpool));
3257
3258 /* Cannot change chunk_size, layout, or level */
3259 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3260 mddev->layout != mddev->new_layout ||
3261 mddev->level != mddev->new_level) {
3262 mddev->new_chunk_sectors = mddev->chunk_sectors;
3263 mddev->new_layout = mddev->layout;
3264 mddev->new_level = mddev->level;
3265 return -EINVAL;
3266 }
3267
3268 if (!mddev_is_clustered(mddev))
3269 md_allow_write(mddev);
3270
3271 raid_disks = mddev->raid_disks + mddev->delta_disks;
3272
3273 if (raid_disks < conf->raid_disks) {
3274 cnt=0;
3275 for (d= 0; d < conf->raid_disks; d++)
3276 if (conf->mirrors[d].rdev)
3277 cnt++;
3278 if (cnt > raid_disks)
3279 return -EBUSY;
3280 }
3281
3282 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3283 if (!newpoolinfo)
3284 return -ENOMEM;
3285 newpoolinfo->mddev = mddev;
3286 newpoolinfo->raid_disks = raid_disks * 2;
3287
3288 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3289 rbio_pool_free, newpoolinfo);
3290 if (ret) {
3291 kfree(newpoolinfo);
3292 return ret;
3293 }
3294 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3295 raid_disks, 2),
3296 GFP_KERNEL);
3297 if (!newmirrors) {
3298 kfree(newpoolinfo);
3299 mempool_exit(&newpool);
3300 return -ENOMEM;
3301 }
3302
3303 freeze_array(conf, 0);
3304
3305 /* ok, everything is stopped */
3306 oldpool = conf->r1bio_pool;
3307 conf->r1bio_pool = newpool;
3308
3309 for (d = d2 = 0; d < conf->raid_disks; d++) {
3310 struct md_rdev *rdev = conf->mirrors[d].rdev;
3311 if (rdev && rdev->raid_disk != d2) {
3312 sysfs_unlink_rdev(mddev, rdev);
3313 rdev->raid_disk = d2;
3314 sysfs_unlink_rdev(mddev, rdev);
3315 if (sysfs_link_rdev(mddev, rdev))
3316 pr_warn("md/raid1:%s: cannot register rd%d\n",
3317 mdname(mddev), rdev->raid_disk);
3318 }
3319 if (rdev)
3320 newmirrors[d2++].rdev = rdev;
3321 }
3322 kfree(conf->mirrors);
3323 conf->mirrors = newmirrors;
3324 kfree(conf->poolinfo);
3325 conf->poolinfo = newpoolinfo;
3326
3327 spin_lock_irqsave(&conf->device_lock, flags);
3328 mddev->degraded += (raid_disks - conf->raid_disks);
3329 spin_unlock_irqrestore(&conf->device_lock, flags);
3330 conf->raid_disks = mddev->raid_disks = raid_disks;
3331 mddev->delta_disks = 0;
3332
3333 unfreeze_array(conf);
3334
3335 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3336 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3337 md_wakeup_thread(mddev->thread);
3338
3339 mempool_exit(&oldpool);
3340 return 0;
3341}
3342
3343static void raid1_quiesce(struct mddev *mddev, int quiesce)
3344{
3345 struct r1conf *conf = mddev->private;
3346
3347 if (quiesce)
3348 freeze_array(conf, 0);
3349 else
3350 unfreeze_array(conf);
3351}
3352
3353static void *raid1_takeover(struct mddev *mddev)
3354{
3355 /* raid1 can take over:
3356 * raid5 with 2 devices, any layout or chunk size
3357 */
3358 if (mddev->level == 5 && mddev->raid_disks == 2) {
3359 struct r1conf *conf;
3360 mddev->new_level = 1;
3361 mddev->new_layout = 0;
3362 mddev->new_chunk_sectors = 0;
3363 conf = setup_conf(mddev);
3364 if (!IS_ERR(conf)) {
3365 /* Array must appear to be quiesced */
3366 conf->array_frozen = 1;
3367 mddev_clear_unsupported_flags(mddev,
3368 UNSUPPORTED_MDDEV_FLAGS);
3369 }
3370 return conf;
3371 }
3372 return ERR_PTR(-EINVAL);
3373}
3374
3375static struct md_personality raid1_personality =
3376{
3377 .name = "raid1",
3378 .level = 1,
3379 .owner = THIS_MODULE,
3380 .make_request = raid1_make_request,
3381 .run = raid1_run,
3382 .free = raid1_free,
3383 .status = raid1_status,
3384 .error_handler = raid1_error,
3385 .hot_add_disk = raid1_add_disk,
3386 .hot_remove_disk= raid1_remove_disk,
3387 .spare_active = raid1_spare_active,
3388 .sync_request = raid1_sync_request,
3389 .resize = raid1_resize,
3390 .size = raid1_size,
3391 .check_reshape = raid1_reshape,
3392 .quiesce = raid1_quiesce,
3393 .takeover = raid1_takeover,
3394};
3395
3396static int __init raid_init(void)
3397{
3398 return register_md_personality(&raid1_personality);
3399}
3400
3401static void raid_exit(void)
3402{
3403 unregister_md_personality(&raid1_personality);
3404}
3405
3406module_init(raid_init);
3407module_exit(raid_exit);
3408MODULE_LICENSE("GPL");
3409MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3410MODULE_ALIAS("md-personality-3"); /* RAID1 */
3411MODULE_ALIAS("md-raid1");
3412MODULE_ALIAS("md-level-1");
3413
3414module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);