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