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 we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
53 */
54#define IO_BLOCKED ((struct bio *)1)
55/* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
58 */
59#define IO_MADE_GOOD ((struct bio *)2)
60
61#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
62
63/* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
65 * for writeback.
66 */
67static int max_queued_requests = 1024;
68
69static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
70 sector_t bi_sector);
71static void lower_barrier(struct r1conf *conf);
72
73static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
74{
75 struct pool_info *pi = data;
76 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
77
78 /* allocate a r1bio with room for raid_disks entries in the bios array */
79 return kzalloc(size, gfp_flags);
80}
81
82static void r1bio_pool_free(void *r1_bio, void *data)
83{
84 kfree(r1_bio);
85}
86
87#define RESYNC_BLOCK_SIZE (64*1024)
88#define RESYNC_DEPTH 32
89#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
90#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
91#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
92#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
93#define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
94
95static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
96{
97 struct pool_info *pi = data;
98 struct r1bio *r1_bio;
99 struct bio *bio;
100 int need_pages;
101 int i, j;
102
103 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
104 if (!r1_bio)
105 return NULL;
106
107 /*
108 * Allocate bios : 1 for reading, n-1 for writing
109 */
110 for (j = pi->raid_disks ; j-- ; ) {
111 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
112 if (!bio)
113 goto out_free_bio;
114 r1_bio->bios[j] = bio;
115 }
116 /*
117 * Allocate RESYNC_PAGES data pages and attach them to
118 * the first bio.
119 * If this is a user-requested check/repair, allocate
120 * RESYNC_PAGES for each bio.
121 */
122 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
123 need_pages = pi->raid_disks;
124 else
125 need_pages = 1;
126 for (j = 0; j < need_pages; j++) {
127 bio = r1_bio->bios[j];
128 bio->bi_vcnt = RESYNC_PAGES;
129
130 if (bio_alloc_pages(bio, gfp_flags))
131 goto out_free_pages;
132 }
133 /* If not user-requests, copy the page pointers to all bios */
134 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
135 for (i=0; i<RESYNC_PAGES ; i++)
136 for (j=1; j<pi->raid_disks; j++)
137 r1_bio->bios[j]->bi_io_vec[i].bv_page =
138 r1_bio->bios[0]->bi_io_vec[i].bv_page;
139 }
140
141 r1_bio->master_bio = NULL;
142
143 return r1_bio;
144
145out_free_pages:
146 while (--j >= 0) {
147 struct bio_vec *bv;
148
149 bio_for_each_segment_all(bv, r1_bio->bios[j], i)
150 __free_page(bv->bv_page);
151 }
152
153out_free_bio:
154 while (++j < pi->raid_disks)
155 bio_put(r1_bio->bios[j]);
156 r1bio_pool_free(r1_bio, data);
157 return NULL;
158}
159
160static void r1buf_pool_free(void *__r1_bio, void *data)
161{
162 struct pool_info *pi = data;
163 int i,j;
164 struct r1bio *r1bio = __r1_bio;
165
166 for (i = 0; i < RESYNC_PAGES; i++)
167 for (j = pi->raid_disks; j-- ;) {
168 if (j == 0 ||
169 r1bio->bios[j]->bi_io_vec[i].bv_page !=
170 r1bio->bios[0]->bi_io_vec[i].bv_page)
171 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
172 }
173 for (i=0 ; i < pi->raid_disks; i++)
174 bio_put(r1bio->bios[i]);
175
176 r1bio_pool_free(r1bio, data);
177}
178
179static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
180{
181 int i;
182
183 for (i = 0; i < conf->raid_disks * 2; i++) {
184 struct bio **bio = r1_bio->bios + i;
185 if (!BIO_SPECIAL(*bio))
186 bio_put(*bio);
187 *bio = NULL;
188 }
189}
190
191static void free_r1bio(struct r1bio *r1_bio)
192{
193 struct r1conf *conf = r1_bio->mddev->private;
194
195 put_all_bios(conf, r1_bio);
196 mempool_free(r1_bio, conf->r1bio_pool);
197}
198
199static void put_buf(struct r1bio *r1_bio)
200{
201 struct r1conf *conf = r1_bio->mddev->private;
202 int i;
203
204 for (i = 0; i < conf->raid_disks * 2; i++) {
205 struct bio *bio = r1_bio->bios[i];
206 if (bio->bi_end_io)
207 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
208 }
209
210 mempool_free(r1_bio, conf->r1buf_pool);
211
212 lower_barrier(conf);
213}
214
215static void reschedule_retry(struct r1bio *r1_bio)
216{
217 unsigned long flags;
218 struct mddev *mddev = r1_bio->mddev;
219 struct r1conf *conf = mddev->private;
220
221 spin_lock_irqsave(&conf->device_lock, flags);
222 list_add(&r1_bio->retry_list, &conf->retry_list);
223 conf->nr_queued ++;
224 spin_unlock_irqrestore(&conf->device_lock, flags);
225
226 wake_up(&conf->wait_barrier);
227 md_wakeup_thread(mddev->thread);
228}
229
230/*
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
233 * cache layer.
234 */
235static void call_bio_endio(struct r1bio *r1_bio)
236{
237 struct bio *bio = r1_bio->master_bio;
238 int done;
239 struct r1conf *conf = r1_bio->mddev->private;
240 sector_t start_next_window = r1_bio->start_next_window;
241 sector_t bi_sector = bio->bi_iter.bi_sector;
242
243 if (bio->bi_phys_segments) {
244 unsigned long flags;
245 spin_lock_irqsave(&conf->device_lock, flags);
246 bio->bi_phys_segments--;
247 done = (bio->bi_phys_segments == 0);
248 spin_unlock_irqrestore(&conf->device_lock, flags);
249 /*
250 * make_request() might be waiting for
251 * bi_phys_segments to decrease
252 */
253 wake_up(&conf->wait_barrier);
254 } else
255 done = 1;
256
257 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
258 clear_bit(BIO_UPTODATE, &bio->bi_flags);
259 if (done) {
260 bio_endio(bio, 0);
261 /*
262 * Wake up any possible resync thread that waits for the device
263 * to go idle.
264 */
265 allow_barrier(conf, start_next_window, bi_sector);
266 }
267}
268
269static void raid_end_bio_io(struct r1bio *r1_bio)
270{
271 struct bio *bio = r1_bio->master_bio;
272
273 /* if nobody has done the final endio yet, do it now */
274 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
275 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
276 (bio_data_dir(bio) == WRITE) ? "write" : "read",
277 (unsigned long long) bio->bi_iter.bi_sector,
278 (unsigned long long) bio_end_sector(bio) - 1);
279
280 call_bio_endio(r1_bio);
281 }
282 free_r1bio(r1_bio);
283}
284
285/*
286 * Update disk head position estimator based on IRQ completion info.
287 */
288static inline void update_head_pos(int disk, struct r1bio *r1_bio)
289{
290 struct r1conf *conf = r1_bio->mddev->private;
291
292 conf->mirrors[disk].head_position =
293 r1_bio->sector + (r1_bio->sectors);
294}
295
296/*
297 * Find the disk number which triggered given bio
298 */
299static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
300{
301 int mirror;
302 struct r1conf *conf = r1_bio->mddev->private;
303 int raid_disks = conf->raid_disks;
304
305 for (mirror = 0; mirror < raid_disks * 2; mirror++)
306 if (r1_bio->bios[mirror] == bio)
307 break;
308
309 BUG_ON(mirror == raid_disks * 2);
310 update_head_pos(mirror, r1_bio);
311
312 return mirror;
313}
314
315static void raid1_end_read_request(struct bio *bio, int error)
316{
317 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
318 struct r1bio *r1_bio = bio->bi_private;
319 int mirror;
320 struct r1conf *conf = r1_bio->mddev->private;
321
322 mirror = r1_bio->read_disk;
323 /*
324 * this branch is our 'one mirror IO has finished' event handler:
325 */
326 update_head_pos(mirror, r1_bio);
327
328 if (uptodate)
329 set_bit(R1BIO_Uptodate, &r1_bio->state);
330 else {
331 /* If all other devices have failed, we want to return
332 * the error upwards rather than fail the last device.
333 * Here we redefine "uptodate" to mean "Don't want to retry"
334 */
335 unsigned long flags;
336 spin_lock_irqsave(&conf->device_lock, flags);
337 if (r1_bio->mddev->degraded == conf->raid_disks ||
338 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
339 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
340 uptodate = 1;
341 spin_unlock_irqrestore(&conf->device_lock, flags);
342 }
343
344 if (uptodate) {
345 raid_end_bio_io(r1_bio);
346 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
347 } else {
348 /*
349 * oops, read error:
350 */
351 char b[BDEVNAME_SIZE];
352 printk_ratelimited(
353 KERN_ERR "md/raid1:%s: %s: "
354 "rescheduling sector %llu\n",
355 mdname(conf->mddev),
356 bdevname(conf->mirrors[mirror].rdev->bdev,
357 b),
358 (unsigned long long)r1_bio->sector);
359 set_bit(R1BIO_ReadError, &r1_bio->state);
360 reschedule_retry(r1_bio);
361 /* don't drop the reference on read_disk yet */
362 }
363}
364
365static void close_write(struct r1bio *r1_bio)
366{
367 /* it really is the end of this request */
368 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
369 /* free extra copy of the data pages */
370 int i = r1_bio->behind_page_count;
371 while (i--)
372 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
373 kfree(r1_bio->behind_bvecs);
374 r1_bio->behind_bvecs = NULL;
375 }
376 /* clear the bitmap if all writes complete successfully */
377 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
378 r1_bio->sectors,
379 !test_bit(R1BIO_Degraded, &r1_bio->state),
380 test_bit(R1BIO_BehindIO, &r1_bio->state));
381 md_write_end(r1_bio->mddev);
382}
383
384static void r1_bio_write_done(struct r1bio *r1_bio)
385{
386 if (!atomic_dec_and_test(&r1_bio->remaining))
387 return;
388
389 if (test_bit(R1BIO_WriteError, &r1_bio->state))
390 reschedule_retry(r1_bio);
391 else {
392 close_write(r1_bio);
393 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
394 reschedule_retry(r1_bio);
395 else
396 raid_end_bio_io(r1_bio);
397 }
398}
399
400static void raid1_end_write_request(struct bio *bio, int error)
401{
402 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
403 struct r1bio *r1_bio = bio->bi_private;
404 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
405 struct r1conf *conf = r1_bio->mddev->private;
406 struct bio *to_put = NULL;
407
408 mirror = find_bio_disk(r1_bio, bio);
409
410 /*
411 * 'one mirror IO has finished' event handler:
412 */
413 if (!uptodate) {
414 set_bit(WriteErrorSeen,
415 &conf->mirrors[mirror].rdev->flags);
416 if (!test_and_set_bit(WantReplacement,
417 &conf->mirrors[mirror].rdev->flags))
418 set_bit(MD_RECOVERY_NEEDED, &
419 conf->mddev->recovery);
420
421 set_bit(R1BIO_WriteError, &r1_bio->state);
422 } else {
423 /*
424 * Set R1BIO_Uptodate in our master bio, so that we
425 * will return a good error code for to the higher
426 * levels even if IO on some other mirrored buffer
427 * fails.
428 *
429 * The 'master' represents the composite IO operation
430 * to user-side. So if something waits for IO, then it
431 * will wait for the 'master' bio.
432 */
433 sector_t first_bad;
434 int bad_sectors;
435
436 r1_bio->bios[mirror] = NULL;
437 to_put = bio;
438 /*
439 * Do not set R1BIO_Uptodate if the current device is
440 * rebuilding or Faulty. This is because we cannot use
441 * such device for properly reading the data back (we could
442 * potentially use it, if the current write would have felt
443 * before rdev->recovery_offset, but for simplicity we don't
444 * check this here.
445 */
446 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
447 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
448 set_bit(R1BIO_Uptodate, &r1_bio->state);
449
450 /* Maybe we can clear some bad blocks. */
451 if (is_badblock(conf->mirrors[mirror].rdev,
452 r1_bio->sector, r1_bio->sectors,
453 &first_bad, &bad_sectors)) {
454 r1_bio->bios[mirror] = IO_MADE_GOOD;
455 set_bit(R1BIO_MadeGood, &r1_bio->state);
456 }
457 }
458
459 if (behind) {
460 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
461 atomic_dec(&r1_bio->behind_remaining);
462
463 /*
464 * In behind mode, we ACK the master bio once the I/O
465 * has safely reached all non-writemostly
466 * disks. Setting the Returned bit ensures that this
467 * gets done only once -- we don't ever want to return
468 * -EIO here, instead we'll wait
469 */
470 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
471 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
472 /* Maybe we can return now */
473 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
474 struct bio *mbio = r1_bio->master_bio;
475 pr_debug("raid1: behind end write sectors"
476 " %llu-%llu\n",
477 (unsigned long long) mbio->bi_iter.bi_sector,
478 (unsigned long long) bio_end_sector(mbio) - 1);
479 call_bio_endio(r1_bio);
480 }
481 }
482 }
483 if (r1_bio->bios[mirror] == NULL)
484 rdev_dec_pending(conf->mirrors[mirror].rdev,
485 conf->mddev);
486
487 /*
488 * Let's see if all mirrored write operations have finished
489 * already.
490 */
491 r1_bio_write_done(r1_bio);
492
493 if (to_put)
494 bio_put(to_put);
495}
496
497
498/*
499 * This routine returns the disk from which the requested read should
500 * be done. There is a per-array 'next expected sequential IO' sector
501 * number - if this matches on the next IO then we use the last disk.
502 * There is also a per-disk 'last know head position' sector that is
503 * maintained from IRQ contexts, both the normal and the resync IO
504 * completion handlers update this position correctly. If there is no
505 * perfect sequential match then we pick the disk whose head is closest.
506 *
507 * If there are 2 mirrors in the same 2 devices, performance degrades
508 * because position is mirror, not device based.
509 *
510 * The rdev for the device selected will have nr_pending incremented.
511 */
512static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
513{
514 const sector_t this_sector = r1_bio->sector;
515 int sectors;
516 int best_good_sectors;
517 int best_disk, best_dist_disk, best_pending_disk;
518 int has_nonrot_disk;
519 int disk;
520 sector_t best_dist;
521 unsigned int min_pending;
522 struct md_rdev *rdev;
523 int choose_first;
524 int choose_next_idle;
525
526 rcu_read_lock();
527 /*
528 * Check if we can balance. We can balance on the whole
529 * device if no resync is going on, or below the resync window.
530 * We take the first readable disk when above the resync window.
531 */
532 retry:
533 sectors = r1_bio->sectors;
534 best_disk = -1;
535 best_dist_disk = -1;
536 best_dist = MaxSector;
537 best_pending_disk = -1;
538 min_pending = UINT_MAX;
539 best_good_sectors = 0;
540 has_nonrot_disk = 0;
541 choose_next_idle = 0;
542
543 if (conf->mddev->recovery_cp < MaxSector &&
544 (this_sector + sectors >= conf->next_resync))
545 choose_first = 1;
546 else
547 choose_first = 0;
548
549 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
550 sector_t dist;
551 sector_t first_bad;
552 int bad_sectors;
553 unsigned int pending;
554 bool nonrot;
555
556 rdev = rcu_dereference(conf->mirrors[disk].rdev);
557 if (r1_bio->bios[disk] == IO_BLOCKED
558 || rdev == NULL
559 || test_bit(Unmerged, &rdev->flags)
560 || test_bit(Faulty, &rdev->flags))
561 continue;
562 if (!test_bit(In_sync, &rdev->flags) &&
563 rdev->recovery_offset < this_sector + sectors)
564 continue;
565 if (test_bit(WriteMostly, &rdev->flags)) {
566 /* Don't balance among write-mostly, just
567 * use the first as a last resort */
568 if (best_disk < 0) {
569 if (is_badblock(rdev, this_sector, sectors,
570 &first_bad, &bad_sectors)) {
571 if (first_bad < this_sector)
572 /* Cannot use this */
573 continue;
574 best_good_sectors = first_bad - this_sector;
575 } else
576 best_good_sectors = sectors;
577 best_disk = disk;
578 }
579 continue;
580 }
581 /* This is a reasonable device to use. It might
582 * even be best.
583 */
584 if (is_badblock(rdev, this_sector, sectors,
585 &first_bad, &bad_sectors)) {
586 if (best_dist < MaxSector)
587 /* already have a better device */
588 continue;
589 if (first_bad <= this_sector) {
590 /* cannot read here. If this is the 'primary'
591 * device, then we must not read beyond
592 * bad_sectors from another device..
593 */
594 bad_sectors -= (this_sector - first_bad);
595 if (choose_first && sectors > bad_sectors)
596 sectors = bad_sectors;
597 if (best_good_sectors > sectors)
598 best_good_sectors = sectors;
599
600 } else {
601 sector_t good_sectors = first_bad - this_sector;
602 if (good_sectors > best_good_sectors) {
603 best_good_sectors = good_sectors;
604 best_disk = disk;
605 }
606 if (choose_first)
607 break;
608 }
609 continue;
610 } else
611 best_good_sectors = sectors;
612
613 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
614 has_nonrot_disk |= nonrot;
615 pending = atomic_read(&rdev->nr_pending);
616 dist = abs(this_sector - conf->mirrors[disk].head_position);
617 if (choose_first) {
618 best_disk = disk;
619 break;
620 }
621 /* Don't change to another disk for sequential reads */
622 if (conf->mirrors[disk].next_seq_sect == this_sector
623 || dist == 0) {
624 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
625 struct raid1_info *mirror = &conf->mirrors[disk];
626
627 best_disk = disk;
628 /*
629 * If buffered sequential IO size exceeds optimal
630 * iosize, check if there is idle disk. If yes, choose
631 * the idle disk. read_balance could already choose an
632 * idle disk before noticing it's a sequential IO in
633 * this disk. This doesn't matter because this disk
634 * will idle, next time it will be utilized after the
635 * first disk has IO size exceeds optimal iosize. In
636 * this way, iosize of the first disk will be optimal
637 * iosize at least. iosize of the second disk might be
638 * small, but not a big deal since when the second disk
639 * starts IO, the first disk is likely still busy.
640 */
641 if (nonrot && opt_iosize > 0 &&
642 mirror->seq_start != MaxSector &&
643 mirror->next_seq_sect > opt_iosize &&
644 mirror->next_seq_sect - opt_iosize >=
645 mirror->seq_start) {
646 choose_next_idle = 1;
647 continue;
648 }
649 break;
650 }
651 /* If device is idle, use it */
652 if (pending == 0) {
653 best_disk = disk;
654 break;
655 }
656
657 if (choose_next_idle)
658 continue;
659
660 if (min_pending > pending) {
661 min_pending = pending;
662 best_pending_disk = disk;
663 }
664
665 if (dist < best_dist) {
666 best_dist = dist;
667 best_dist_disk = disk;
668 }
669 }
670
671 /*
672 * If all disks are rotational, choose the closest disk. If any disk is
673 * non-rotational, choose the disk with less pending request even the
674 * disk is rotational, which might/might not be optimal for raids with
675 * mixed ratation/non-rotational disks depending on workload.
676 */
677 if (best_disk == -1) {
678 if (has_nonrot_disk)
679 best_disk = best_pending_disk;
680 else
681 best_disk = best_dist_disk;
682 }
683
684 if (best_disk >= 0) {
685 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
686 if (!rdev)
687 goto retry;
688 atomic_inc(&rdev->nr_pending);
689 if (test_bit(Faulty, &rdev->flags)) {
690 /* cannot risk returning a device that failed
691 * before we inc'ed nr_pending
692 */
693 rdev_dec_pending(rdev, conf->mddev);
694 goto retry;
695 }
696 sectors = best_good_sectors;
697
698 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
699 conf->mirrors[best_disk].seq_start = this_sector;
700
701 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
702 }
703 rcu_read_unlock();
704 *max_sectors = sectors;
705
706 return best_disk;
707}
708
709static int raid1_mergeable_bvec(struct request_queue *q,
710 struct bvec_merge_data *bvm,
711 struct bio_vec *biovec)
712{
713 struct mddev *mddev = q->queuedata;
714 struct r1conf *conf = mddev->private;
715 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
716 int max = biovec->bv_len;
717
718 if (mddev->merge_check_needed) {
719 int disk;
720 rcu_read_lock();
721 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
722 struct md_rdev *rdev = rcu_dereference(
723 conf->mirrors[disk].rdev);
724 if (rdev && !test_bit(Faulty, &rdev->flags)) {
725 struct request_queue *q =
726 bdev_get_queue(rdev->bdev);
727 if (q->merge_bvec_fn) {
728 bvm->bi_sector = sector +
729 rdev->data_offset;
730 bvm->bi_bdev = rdev->bdev;
731 max = min(max, q->merge_bvec_fn(
732 q, bvm, biovec));
733 }
734 }
735 }
736 rcu_read_unlock();
737 }
738 return max;
739
740}
741
742int md_raid1_congested(struct mddev *mddev, int bits)
743{
744 struct r1conf *conf = mddev->private;
745 int i, ret = 0;
746
747 if ((bits & (1 << BDI_async_congested)) &&
748 conf->pending_count >= max_queued_requests)
749 return 1;
750
751 rcu_read_lock();
752 for (i = 0; i < conf->raid_disks * 2; i++) {
753 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
754 if (rdev && !test_bit(Faulty, &rdev->flags)) {
755 struct request_queue *q = bdev_get_queue(rdev->bdev);
756
757 BUG_ON(!q);
758
759 /* Note the '|| 1' - when read_balance prefers
760 * non-congested targets, it can be removed
761 */
762 if ((bits & (1<<BDI_async_congested)) || 1)
763 ret |= bdi_congested(&q->backing_dev_info, bits);
764 else
765 ret &= bdi_congested(&q->backing_dev_info, bits);
766 }
767 }
768 rcu_read_unlock();
769 return ret;
770}
771EXPORT_SYMBOL_GPL(md_raid1_congested);
772
773static int raid1_congested(void *data, int bits)
774{
775 struct mddev *mddev = data;
776
777 return mddev_congested(mddev, bits) ||
778 md_raid1_congested(mddev, bits);
779}
780
781static void flush_pending_writes(struct r1conf *conf)
782{
783 /* Any writes that have been queued but are awaiting
784 * bitmap updates get flushed here.
785 */
786 spin_lock_irq(&conf->device_lock);
787
788 if (conf->pending_bio_list.head) {
789 struct bio *bio;
790 bio = bio_list_get(&conf->pending_bio_list);
791 conf->pending_count = 0;
792 spin_unlock_irq(&conf->device_lock);
793 /* flush any pending bitmap writes to
794 * disk before proceeding w/ I/O */
795 bitmap_unplug(conf->mddev->bitmap);
796 wake_up(&conf->wait_barrier);
797
798 while (bio) { /* submit pending writes */
799 struct bio *next = bio->bi_next;
800 bio->bi_next = NULL;
801 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
802 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
803 /* Just ignore it */
804 bio_endio(bio, 0);
805 else
806 generic_make_request(bio);
807 bio = next;
808 }
809 } else
810 spin_unlock_irq(&conf->device_lock);
811}
812
813/* Barriers....
814 * Sometimes we need to suspend IO while we do something else,
815 * either some resync/recovery, or reconfigure the array.
816 * To do this we raise a 'barrier'.
817 * The 'barrier' is a counter that can be raised multiple times
818 * to count how many activities are happening which preclude
819 * normal IO.
820 * We can only raise the barrier if there is no pending IO.
821 * i.e. if nr_pending == 0.
822 * We choose only to raise the barrier if no-one is waiting for the
823 * barrier to go down. This means that as soon as an IO request
824 * is ready, no other operations which require a barrier will start
825 * until the IO request has had a chance.
826 *
827 * So: regular IO calls 'wait_barrier'. When that returns there
828 * is no backgroup IO happening, It must arrange to call
829 * allow_barrier when it has finished its IO.
830 * backgroup IO calls must call raise_barrier. Once that returns
831 * there is no normal IO happeing. It must arrange to call
832 * lower_barrier when the particular background IO completes.
833 */
834static void raise_barrier(struct r1conf *conf)
835{
836 spin_lock_irq(&conf->resync_lock);
837
838 /* Wait until no block IO is waiting */
839 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
840 conf->resync_lock);
841
842 /* block any new IO from starting */
843 conf->barrier++;
844
845 /* For these conditions we must wait:
846 * A: while the array is in frozen state
847 * B: while barrier >= RESYNC_DEPTH, meaning resync reach
848 * the max count which allowed.
849 * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
850 * next resync will reach to the window which normal bios are
851 * handling.
852 */
853 wait_event_lock_irq(conf->wait_barrier,
854 !conf->array_frozen &&
855 conf->barrier < RESYNC_DEPTH &&
856 (conf->start_next_window >=
857 conf->next_resync + RESYNC_SECTORS),
858 conf->resync_lock);
859
860 spin_unlock_irq(&conf->resync_lock);
861}
862
863static void lower_barrier(struct r1conf *conf)
864{
865 unsigned long flags;
866 BUG_ON(conf->barrier <= 0);
867 spin_lock_irqsave(&conf->resync_lock, flags);
868 conf->barrier--;
869 spin_unlock_irqrestore(&conf->resync_lock, flags);
870 wake_up(&conf->wait_barrier);
871}
872
873static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
874{
875 bool wait = false;
876
877 if (conf->array_frozen || !bio)
878 wait = true;
879 else if (conf->barrier && bio_data_dir(bio) == WRITE) {
880 if (conf->next_resync < RESYNC_WINDOW_SECTORS)
881 wait = true;
882 else if ((conf->next_resync - RESYNC_WINDOW_SECTORS
883 >= bio_end_sector(bio)) ||
884 (conf->next_resync + NEXT_NORMALIO_DISTANCE
885 <= bio->bi_iter.bi_sector))
886 wait = false;
887 else
888 wait = true;
889 }
890
891 return wait;
892}
893
894static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
895{
896 sector_t sector = 0;
897
898 spin_lock_irq(&conf->resync_lock);
899 if (need_to_wait_for_sync(conf, bio)) {
900 conf->nr_waiting++;
901 /* Wait for the barrier to drop.
902 * However if there are already pending
903 * requests (preventing the barrier from
904 * rising completely), and the
905 * pre-process bio queue isn't empty,
906 * then don't wait, as we need to empty
907 * that queue to get the nr_pending
908 * count down.
909 */
910 wait_event_lock_irq(conf->wait_barrier,
911 !conf->array_frozen &&
912 (!conf->barrier ||
913 ((conf->start_next_window <
914 conf->next_resync + RESYNC_SECTORS) &&
915 current->bio_list &&
916 !bio_list_empty(current->bio_list))),
917 conf->resync_lock);
918 conf->nr_waiting--;
919 }
920
921 if (bio && bio_data_dir(bio) == WRITE) {
922 if (conf->next_resync + NEXT_NORMALIO_DISTANCE
923 <= bio->bi_iter.bi_sector) {
924 if (conf->start_next_window == MaxSector)
925 conf->start_next_window =
926 conf->next_resync +
927 NEXT_NORMALIO_DISTANCE;
928
929 if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
930 <= bio->bi_iter.bi_sector)
931 conf->next_window_requests++;
932 else
933 conf->current_window_requests++;
934 sector = conf->start_next_window;
935 }
936 }
937
938 conf->nr_pending++;
939 spin_unlock_irq(&conf->resync_lock);
940 return sector;
941}
942
943static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
944 sector_t bi_sector)
945{
946 unsigned long flags;
947
948 spin_lock_irqsave(&conf->resync_lock, flags);
949 conf->nr_pending--;
950 if (start_next_window) {
951 if (start_next_window == conf->start_next_window) {
952 if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
953 <= bi_sector)
954 conf->next_window_requests--;
955 else
956 conf->current_window_requests--;
957 } else
958 conf->current_window_requests--;
959
960 if (!conf->current_window_requests) {
961 if (conf->next_window_requests) {
962 conf->current_window_requests =
963 conf->next_window_requests;
964 conf->next_window_requests = 0;
965 conf->start_next_window +=
966 NEXT_NORMALIO_DISTANCE;
967 } else
968 conf->start_next_window = MaxSector;
969 }
970 }
971 spin_unlock_irqrestore(&conf->resync_lock, flags);
972 wake_up(&conf->wait_barrier);
973}
974
975static void freeze_array(struct r1conf *conf, int extra)
976{
977 /* stop syncio and normal IO and wait for everything to
978 * go quite.
979 * We wait until nr_pending match nr_queued+extra
980 * This is called in the context of one normal IO request
981 * that has failed. Thus any sync request that might be pending
982 * will be blocked by nr_pending, and we need to wait for
983 * pending IO requests to complete or be queued for re-try.
984 * Thus the number queued (nr_queued) plus this request (extra)
985 * must match the number of pending IOs (nr_pending) before
986 * we continue.
987 */
988 spin_lock_irq(&conf->resync_lock);
989 conf->array_frozen = 1;
990 wait_event_lock_irq_cmd(conf->wait_barrier,
991 conf->nr_pending == conf->nr_queued+extra,
992 conf->resync_lock,
993 flush_pending_writes(conf));
994 spin_unlock_irq(&conf->resync_lock);
995}
996static void unfreeze_array(struct r1conf *conf)
997{
998 /* reverse the effect of the freeze */
999 spin_lock_irq(&conf->resync_lock);
1000 conf->array_frozen = 0;
1001 wake_up(&conf->wait_barrier);
1002 spin_unlock_irq(&conf->resync_lock);
1003}
1004
1005
1006/* duplicate the data pages for behind I/O
1007 */
1008static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
1009{
1010 int i;
1011 struct bio_vec *bvec;
1012 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
1013 GFP_NOIO);
1014 if (unlikely(!bvecs))
1015 return;
1016
1017 bio_for_each_segment_all(bvec, bio, i) {
1018 bvecs[i] = *bvec;
1019 bvecs[i].bv_page = alloc_page(GFP_NOIO);
1020 if (unlikely(!bvecs[i].bv_page))
1021 goto do_sync_io;
1022 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
1023 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
1024 kunmap(bvecs[i].bv_page);
1025 kunmap(bvec->bv_page);
1026 }
1027 r1_bio->behind_bvecs = bvecs;
1028 r1_bio->behind_page_count = bio->bi_vcnt;
1029 set_bit(R1BIO_BehindIO, &r1_bio->state);
1030 return;
1031
1032do_sync_io:
1033 for (i = 0; i < bio->bi_vcnt; i++)
1034 if (bvecs[i].bv_page)
1035 put_page(bvecs[i].bv_page);
1036 kfree(bvecs);
1037 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1038 bio->bi_iter.bi_size);
1039}
1040
1041struct raid1_plug_cb {
1042 struct blk_plug_cb cb;
1043 struct bio_list pending;
1044 int pending_cnt;
1045};
1046
1047static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1048{
1049 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1050 cb);
1051 struct mddev *mddev = plug->cb.data;
1052 struct r1conf *conf = mddev->private;
1053 struct bio *bio;
1054
1055 if (from_schedule || current->bio_list) {
1056 spin_lock_irq(&conf->device_lock);
1057 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1058 conf->pending_count += plug->pending_cnt;
1059 spin_unlock_irq(&conf->device_lock);
1060 wake_up(&conf->wait_barrier);
1061 md_wakeup_thread(mddev->thread);
1062 kfree(plug);
1063 return;
1064 }
1065
1066 /* we aren't scheduling, so we can do the write-out directly. */
1067 bio = bio_list_get(&plug->pending);
1068 bitmap_unplug(mddev->bitmap);
1069 wake_up(&conf->wait_barrier);
1070
1071 while (bio) { /* submit pending writes */
1072 struct bio *next = bio->bi_next;
1073 bio->bi_next = NULL;
1074 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1075 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1076 /* Just ignore it */
1077 bio_endio(bio, 0);
1078 else
1079 generic_make_request(bio);
1080 bio = next;
1081 }
1082 kfree(plug);
1083}
1084
1085static void make_request(struct mddev *mddev, struct bio * bio)
1086{
1087 struct r1conf *conf = mddev->private;
1088 struct raid1_info *mirror;
1089 struct r1bio *r1_bio;
1090 struct bio *read_bio;
1091 int i, disks;
1092 struct bitmap *bitmap;
1093 unsigned long flags;
1094 const int rw = bio_data_dir(bio);
1095 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1096 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1097 const unsigned long do_discard = (bio->bi_rw
1098 & (REQ_DISCARD | REQ_SECURE));
1099 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1100 struct md_rdev *blocked_rdev;
1101 struct blk_plug_cb *cb;
1102 struct raid1_plug_cb *plug = NULL;
1103 int first_clone;
1104 int sectors_handled;
1105 int max_sectors;
1106 sector_t start_next_window;
1107
1108 /*
1109 * Register the new request and wait if the reconstruction
1110 * thread has put up a bar for new requests.
1111 * Continue immediately if no resync is active currently.
1112 */
1113
1114 md_write_start(mddev, bio); /* wait on superblock update early */
1115
1116 if (bio_data_dir(bio) == WRITE &&
1117 bio_end_sector(bio) > mddev->suspend_lo &&
1118 bio->bi_iter.bi_sector < mddev->suspend_hi) {
1119 /* As the suspend_* range is controlled by
1120 * userspace, we want an interruptible
1121 * wait.
1122 */
1123 DEFINE_WAIT(w);
1124 for (;;) {
1125 flush_signals(current);
1126 prepare_to_wait(&conf->wait_barrier,
1127 &w, TASK_INTERRUPTIBLE);
1128 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1129 bio->bi_iter.bi_sector >= mddev->suspend_hi)
1130 break;
1131 schedule();
1132 }
1133 finish_wait(&conf->wait_barrier, &w);
1134 }
1135
1136 start_next_window = wait_barrier(conf, bio);
1137
1138 bitmap = mddev->bitmap;
1139
1140 /*
1141 * make_request() can abort the operation when READA is being
1142 * used and no empty request is available.
1143 *
1144 */
1145 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1146
1147 r1_bio->master_bio = bio;
1148 r1_bio->sectors = bio_sectors(bio);
1149 r1_bio->state = 0;
1150 r1_bio->mddev = mddev;
1151 r1_bio->sector = bio->bi_iter.bi_sector;
1152
1153 /* We might need to issue multiple reads to different
1154 * devices if there are bad blocks around, so we keep
1155 * track of the number of reads in bio->bi_phys_segments.
1156 * If this is 0, there is only one r1_bio and no locking
1157 * will be needed when requests complete. If it is
1158 * non-zero, then it is the number of not-completed requests.
1159 */
1160 bio->bi_phys_segments = 0;
1161 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1162
1163 if (rw == READ) {
1164 /*
1165 * read balancing logic:
1166 */
1167 int rdisk;
1168
1169read_again:
1170 rdisk = read_balance(conf, r1_bio, &max_sectors);
1171
1172 if (rdisk < 0) {
1173 /* couldn't find anywhere to read from */
1174 raid_end_bio_io(r1_bio);
1175 return;
1176 }
1177 mirror = conf->mirrors + rdisk;
1178
1179 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1180 bitmap) {
1181 /* Reading from a write-mostly device must
1182 * take care not to over-take any writes
1183 * that are 'behind'
1184 */
1185 wait_event(bitmap->behind_wait,
1186 atomic_read(&bitmap->behind_writes) == 0);
1187 }
1188 r1_bio->read_disk = rdisk;
1189
1190 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1191 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1192 max_sectors);
1193
1194 r1_bio->bios[rdisk] = read_bio;
1195
1196 read_bio->bi_iter.bi_sector = r1_bio->sector +
1197 mirror->rdev->data_offset;
1198 read_bio->bi_bdev = mirror->rdev->bdev;
1199 read_bio->bi_end_io = raid1_end_read_request;
1200 read_bio->bi_rw = READ | do_sync;
1201 read_bio->bi_private = r1_bio;
1202
1203 if (max_sectors < r1_bio->sectors) {
1204 /* could not read all from this device, so we will
1205 * need another r1_bio.
1206 */
1207
1208 sectors_handled = (r1_bio->sector + max_sectors
1209 - bio->bi_iter.bi_sector);
1210 r1_bio->sectors = max_sectors;
1211 spin_lock_irq(&conf->device_lock);
1212 if (bio->bi_phys_segments == 0)
1213 bio->bi_phys_segments = 2;
1214 else
1215 bio->bi_phys_segments++;
1216 spin_unlock_irq(&conf->device_lock);
1217 /* Cannot call generic_make_request directly
1218 * as that will be queued in __make_request
1219 * and subsequent mempool_alloc might block waiting
1220 * for it. So hand bio over to raid1d.
1221 */
1222 reschedule_retry(r1_bio);
1223
1224 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1225
1226 r1_bio->master_bio = bio;
1227 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1228 r1_bio->state = 0;
1229 r1_bio->mddev = mddev;
1230 r1_bio->sector = bio->bi_iter.bi_sector +
1231 sectors_handled;
1232 goto read_again;
1233 } else
1234 generic_make_request(read_bio);
1235 return;
1236 }
1237
1238 /*
1239 * WRITE:
1240 */
1241 if (conf->pending_count >= max_queued_requests) {
1242 md_wakeup_thread(mddev->thread);
1243 wait_event(conf->wait_barrier,
1244 conf->pending_count < max_queued_requests);
1245 }
1246 /* first select target devices under rcu_lock and
1247 * inc refcount on their rdev. Record them by setting
1248 * bios[x] to bio
1249 * If there are known/acknowledged bad blocks on any device on
1250 * which we have seen a write error, we want to avoid writing those
1251 * blocks.
1252 * This potentially requires several writes to write around
1253 * the bad blocks. Each set of writes gets it's own r1bio
1254 * with a set of bios attached.
1255 */
1256
1257 disks = conf->raid_disks * 2;
1258 retry_write:
1259 r1_bio->start_next_window = start_next_window;
1260 blocked_rdev = NULL;
1261 rcu_read_lock();
1262 max_sectors = r1_bio->sectors;
1263 for (i = 0; i < disks; i++) {
1264 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1265 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1266 atomic_inc(&rdev->nr_pending);
1267 blocked_rdev = rdev;
1268 break;
1269 }
1270 r1_bio->bios[i] = NULL;
1271 if (!rdev || test_bit(Faulty, &rdev->flags)
1272 || test_bit(Unmerged, &rdev->flags)) {
1273 if (i < conf->raid_disks)
1274 set_bit(R1BIO_Degraded, &r1_bio->state);
1275 continue;
1276 }
1277
1278 atomic_inc(&rdev->nr_pending);
1279 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1280 sector_t first_bad;
1281 int bad_sectors;
1282 int is_bad;
1283
1284 is_bad = is_badblock(rdev, r1_bio->sector,
1285 max_sectors,
1286 &first_bad, &bad_sectors);
1287 if (is_bad < 0) {
1288 /* mustn't write here until the bad block is
1289 * acknowledged*/
1290 set_bit(BlockedBadBlocks, &rdev->flags);
1291 blocked_rdev = rdev;
1292 break;
1293 }
1294 if (is_bad && first_bad <= r1_bio->sector) {
1295 /* Cannot write here at all */
1296 bad_sectors -= (r1_bio->sector - first_bad);
1297 if (bad_sectors < max_sectors)
1298 /* mustn't write more than bad_sectors
1299 * to other devices yet
1300 */
1301 max_sectors = bad_sectors;
1302 rdev_dec_pending(rdev, mddev);
1303 /* We don't set R1BIO_Degraded as that
1304 * only applies if the disk is
1305 * missing, so it might be re-added,
1306 * and we want to know to recover this
1307 * chunk.
1308 * In this case the device is here,
1309 * and the fact that this chunk is not
1310 * in-sync is recorded in the bad
1311 * block log
1312 */
1313 continue;
1314 }
1315 if (is_bad) {
1316 int good_sectors = first_bad - r1_bio->sector;
1317 if (good_sectors < max_sectors)
1318 max_sectors = good_sectors;
1319 }
1320 }
1321 r1_bio->bios[i] = bio;
1322 }
1323 rcu_read_unlock();
1324
1325 if (unlikely(blocked_rdev)) {
1326 /* Wait for this device to become unblocked */
1327 int j;
1328 sector_t old = start_next_window;
1329
1330 for (j = 0; j < i; j++)
1331 if (r1_bio->bios[j])
1332 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1333 r1_bio->state = 0;
1334 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1335 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1336 start_next_window = wait_barrier(conf, bio);
1337 /*
1338 * We must make sure the multi r1bios of bio have
1339 * the same value of bi_phys_segments
1340 */
1341 if (bio->bi_phys_segments && old &&
1342 old != start_next_window)
1343 /* Wait for the former r1bio(s) to complete */
1344 wait_event(conf->wait_barrier,
1345 bio->bi_phys_segments == 1);
1346 goto retry_write;
1347 }
1348
1349 if (max_sectors < r1_bio->sectors) {
1350 /* We are splitting this write into multiple parts, so
1351 * we need to prepare for allocating another r1_bio.
1352 */
1353 r1_bio->sectors = max_sectors;
1354 spin_lock_irq(&conf->device_lock);
1355 if (bio->bi_phys_segments == 0)
1356 bio->bi_phys_segments = 2;
1357 else
1358 bio->bi_phys_segments++;
1359 spin_unlock_irq(&conf->device_lock);
1360 }
1361 sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1362
1363 atomic_set(&r1_bio->remaining, 1);
1364 atomic_set(&r1_bio->behind_remaining, 0);
1365
1366 first_clone = 1;
1367 for (i = 0; i < disks; i++) {
1368 struct bio *mbio;
1369 if (!r1_bio->bios[i])
1370 continue;
1371
1372 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1373 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1374
1375 if (first_clone) {
1376 /* do behind I/O ?
1377 * Not if there are too many, or cannot
1378 * allocate memory, or a reader on WriteMostly
1379 * is waiting for behind writes to flush */
1380 if (bitmap &&
1381 (atomic_read(&bitmap->behind_writes)
1382 < mddev->bitmap_info.max_write_behind) &&
1383 !waitqueue_active(&bitmap->behind_wait))
1384 alloc_behind_pages(mbio, r1_bio);
1385
1386 bitmap_startwrite(bitmap, r1_bio->sector,
1387 r1_bio->sectors,
1388 test_bit(R1BIO_BehindIO,
1389 &r1_bio->state));
1390 first_clone = 0;
1391 }
1392 if (r1_bio->behind_bvecs) {
1393 struct bio_vec *bvec;
1394 int j;
1395
1396 /*
1397 * We trimmed the bio, so _all is legit
1398 */
1399 bio_for_each_segment_all(bvec, mbio, j)
1400 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1401 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1402 atomic_inc(&r1_bio->behind_remaining);
1403 }
1404
1405 r1_bio->bios[i] = mbio;
1406
1407 mbio->bi_iter.bi_sector = (r1_bio->sector +
1408 conf->mirrors[i].rdev->data_offset);
1409 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1410 mbio->bi_end_io = raid1_end_write_request;
1411 mbio->bi_rw =
1412 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1413 mbio->bi_private = r1_bio;
1414
1415 atomic_inc(&r1_bio->remaining);
1416
1417 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1418 if (cb)
1419 plug = container_of(cb, struct raid1_plug_cb, cb);
1420 else
1421 plug = NULL;
1422 spin_lock_irqsave(&conf->device_lock, flags);
1423 if (plug) {
1424 bio_list_add(&plug->pending, mbio);
1425 plug->pending_cnt++;
1426 } else {
1427 bio_list_add(&conf->pending_bio_list, mbio);
1428 conf->pending_count++;
1429 }
1430 spin_unlock_irqrestore(&conf->device_lock, flags);
1431 if (!plug)
1432 md_wakeup_thread(mddev->thread);
1433 }
1434 /* Mustn't call r1_bio_write_done before this next test,
1435 * as it could result in the bio being freed.
1436 */
1437 if (sectors_handled < bio_sectors(bio)) {
1438 r1_bio_write_done(r1_bio);
1439 /* We need another r1_bio. It has already been counted
1440 * in bio->bi_phys_segments
1441 */
1442 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1443 r1_bio->master_bio = bio;
1444 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1445 r1_bio->state = 0;
1446 r1_bio->mddev = mddev;
1447 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1448 goto retry_write;
1449 }
1450
1451 r1_bio_write_done(r1_bio);
1452
1453 /* In case raid1d snuck in to freeze_array */
1454 wake_up(&conf->wait_barrier);
1455}
1456
1457static void status(struct seq_file *seq, struct mddev *mddev)
1458{
1459 struct r1conf *conf = mddev->private;
1460 int i;
1461
1462 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1463 conf->raid_disks - mddev->degraded);
1464 rcu_read_lock();
1465 for (i = 0; i < conf->raid_disks; i++) {
1466 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1467 seq_printf(seq, "%s",
1468 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1469 }
1470 rcu_read_unlock();
1471 seq_printf(seq, "]");
1472}
1473
1474
1475static void error(struct mddev *mddev, struct md_rdev *rdev)
1476{
1477 char b[BDEVNAME_SIZE];
1478 struct r1conf *conf = mddev->private;
1479
1480 /*
1481 * If it is not operational, then we have already marked it as dead
1482 * else if it is the last working disks, ignore the error, let the
1483 * next level up know.
1484 * else mark the drive as failed
1485 */
1486 if (test_bit(In_sync, &rdev->flags)
1487 && (conf->raid_disks - mddev->degraded) == 1) {
1488 /*
1489 * Don't fail the drive, act as though we were just a
1490 * normal single drive.
1491 * However don't try a recovery from this drive as
1492 * it is very likely to fail.
1493 */
1494 conf->recovery_disabled = mddev->recovery_disabled;
1495 return;
1496 }
1497 set_bit(Blocked, &rdev->flags);
1498 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1499 unsigned long flags;
1500 spin_lock_irqsave(&conf->device_lock, flags);
1501 mddev->degraded++;
1502 set_bit(Faulty, &rdev->flags);
1503 spin_unlock_irqrestore(&conf->device_lock, flags);
1504 /*
1505 * if recovery is running, make sure it aborts.
1506 */
1507 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1508 } else
1509 set_bit(Faulty, &rdev->flags);
1510 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1511 printk(KERN_ALERT
1512 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1513 "md/raid1:%s: Operation continuing on %d devices.\n",
1514 mdname(mddev), bdevname(rdev->bdev, b),
1515 mdname(mddev), conf->raid_disks - mddev->degraded);
1516}
1517
1518static void print_conf(struct r1conf *conf)
1519{
1520 int i;
1521
1522 printk(KERN_DEBUG "RAID1 conf printout:\n");
1523 if (!conf) {
1524 printk(KERN_DEBUG "(!conf)\n");
1525 return;
1526 }
1527 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1528 conf->raid_disks);
1529
1530 rcu_read_lock();
1531 for (i = 0; i < conf->raid_disks; i++) {
1532 char b[BDEVNAME_SIZE];
1533 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1534 if (rdev)
1535 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1536 i, !test_bit(In_sync, &rdev->flags),
1537 !test_bit(Faulty, &rdev->flags),
1538 bdevname(rdev->bdev,b));
1539 }
1540 rcu_read_unlock();
1541}
1542
1543static void close_sync(struct r1conf *conf)
1544{
1545 wait_barrier(conf, NULL);
1546 allow_barrier(conf, 0, 0);
1547
1548 mempool_destroy(conf->r1buf_pool);
1549 conf->r1buf_pool = NULL;
1550
1551 conf->next_resync = 0;
1552 conf->start_next_window = MaxSector;
1553}
1554
1555static int raid1_spare_active(struct mddev *mddev)
1556{
1557 int i;
1558 struct r1conf *conf = mddev->private;
1559 int count = 0;
1560 unsigned long flags;
1561
1562 /*
1563 * Find all failed disks within the RAID1 configuration
1564 * and mark them readable.
1565 * Called under mddev lock, so rcu protection not needed.
1566 */
1567 for (i = 0; i < conf->raid_disks; i++) {
1568 struct md_rdev *rdev = conf->mirrors[i].rdev;
1569 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1570 if (repl
1571 && repl->recovery_offset == MaxSector
1572 && !test_bit(Faulty, &repl->flags)
1573 && !test_and_set_bit(In_sync, &repl->flags)) {
1574 /* replacement has just become active */
1575 if (!rdev ||
1576 !test_and_clear_bit(In_sync, &rdev->flags))
1577 count++;
1578 if (rdev) {
1579 /* Replaced device not technically
1580 * faulty, but we need to be sure
1581 * it gets removed and never re-added
1582 */
1583 set_bit(Faulty, &rdev->flags);
1584 sysfs_notify_dirent_safe(
1585 rdev->sysfs_state);
1586 }
1587 }
1588 if (rdev
1589 && rdev->recovery_offset == MaxSector
1590 && !test_bit(Faulty, &rdev->flags)
1591 && !test_and_set_bit(In_sync, &rdev->flags)) {
1592 count++;
1593 sysfs_notify_dirent_safe(rdev->sysfs_state);
1594 }
1595 }
1596 spin_lock_irqsave(&conf->device_lock, flags);
1597 mddev->degraded -= count;
1598 spin_unlock_irqrestore(&conf->device_lock, flags);
1599
1600 print_conf(conf);
1601 return count;
1602}
1603
1604
1605static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1606{
1607 struct r1conf *conf = mddev->private;
1608 int err = -EEXIST;
1609 int mirror = 0;
1610 struct raid1_info *p;
1611 int first = 0;
1612 int last = conf->raid_disks - 1;
1613 struct request_queue *q = bdev_get_queue(rdev->bdev);
1614
1615 if (mddev->recovery_disabled == conf->recovery_disabled)
1616 return -EBUSY;
1617
1618 if (rdev->raid_disk >= 0)
1619 first = last = rdev->raid_disk;
1620
1621 if (q->merge_bvec_fn) {
1622 set_bit(Unmerged, &rdev->flags);
1623 mddev->merge_check_needed = 1;
1624 }
1625
1626 for (mirror = first; mirror <= last; mirror++) {
1627 p = conf->mirrors+mirror;
1628 if (!p->rdev) {
1629
1630 if (mddev->gendisk)
1631 disk_stack_limits(mddev->gendisk, rdev->bdev,
1632 rdev->data_offset << 9);
1633
1634 p->head_position = 0;
1635 rdev->raid_disk = mirror;
1636 err = 0;
1637 /* As all devices are equivalent, we don't need a full recovery
1638 * if this was recently any drive of the array
1639 */
1640 if (rdev->saved_raid_disk < 0)
1641 conf->fullsync = 1;
1642 rcu_assign_pointer(p->rdev, rdev);
1643 break;
1644 }
1645 if (test_bit(WantReplacement, &p->rdev->flags) &&
1646 p[conf->raid_disks].rdev == NULL) {
1647 /* Add this device as a replacement */
1648 clear_bit(In_sync, &rdev->flags);
1649 set_bit(Replacement, &rdev->flags);
1650 rdev->raid_disk = mirror;
1651 err = 0;
1652 conf->fullsync = 1;
1653 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1654 break;
1655 }
1656 }
1657 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1658 /* Some requests might not have seen this new
1659 * merge_bvec_fn. We must wait for them to complete
1660 * before merging the device fully.
1661 * First we make sure any code which has tested
1662 * our function has submitted the request, then
1663 * we wait for all outstanding requests to complete.
1664 */
1665 synchronize_sched();
1666 freeze_array(conf, 0);
1667 unfreeze_array(conf);
1668 clear_bit(Unmerged, &rdev->flags);
1669 }
1670 md_integrity_add_rdev(rdev, mddev);
1671 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1672 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1673 print_conf(conf);
1674 return err;
1675}
1676
1677static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1678{
1679 struct r1conf *conf = mddev->private;
1680 int err = 0;
1681 int number = rdev->raid_disk;
1682 struct raid1_info *p = conf->mirrors + number;
1683
1684 if (rdev != p->rdev)
1685 p = conf->mirrors + conf->raid_disks + number;
1686
1687 print_conf(conf);
1688 if (rdev == p->rdev) {
1689 if (test_bit(In_sync, &rdev->flags) ||
1690 atomic_read(&rdev->nr_pending)) {
1691 err = -EBUSY;
1692 goto abort;
1693 }
1694 /* Only remove non-faulty devices if recovery
1695 * is not possible.
1696 */
1697 if (!test_bit(Faulty, &rdev->flags) &&
1698 mddev->recovery_disabled != conf->recovery_disabled &&
1699 mddev->degraded < conf->raid_disks) {
1700 err = -EBUSY;
1701 goto abort;
1702 }
1703 p->rdev = NULL;
1704 synchronize_rcu();
1705 if (atomic_read(&rdev->nr_pending)) {
1706 /* lost the race, try later */
1707 err = -EBUSY;
1708 p->rdev = rdev;
1709 goto abort;
1710 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1711 /* We just removed a device that is being replaced.
1712 * Move down the replacement. We drain all IO before
1713 * doing this to avoid confusion.
1714 */
1715 struct md_rdev *repl =
1716 conf->mirrors[conf->raid_disks + number].rdev;
1717 freeze_array(conf, 0);
1718 clear_bit(Replacement, &repl->flags);
1719 p->rdev = repl;
1720 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1721 unfreeze_array(conf);
1722 clear_bit(WantReplacement, &rdev->flags);
1723 } else
1724 clear_bit(WantReplacement, &rdev->flags);
1725 err = md_integrity_register(mddev);
1726 }
1727abort:
1728
1729 print_conf(conf);
1730 return err;
1731}
1732
1733
1734static void end_sync_read(struct bio *bio, int error)
1735{
1736 struct r1bio *r1_bio = bio->bi_private;
1737
1738 update_head_pos(r1_bio->read_disk, r1_bio);
1739
1740 /*
1741 * we have read a block, now it needs to be re-written,
1742 * or re-read if the read failed.
1743 * We don't do much here, just schedule handling by raid1d
1744 */
1745 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1746 set_bit(R1BIO_Uptodate, &r1_bio->state);
1747
1748 if (atomic_dec_and_test(&r1_bio->remaining))
1749 reschedule_retry(r1_bio);
1750}
1751
1752static void end_sync_write(struct bio *bio, int error)
1753{
1754 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1755 struct r1bio *r1_bio = bio->bi_private;
1756 struct mddev *mddev = r1_bio->mddev;
1757 struct r1conf *conf = mddev->private;
1758 int mirror=0;
1759 sector_t first_bad;
1760 int bad_sectors;
1761
1762 mirror = find_bio_disk(r1_bio, bio);
1763
1764 if (!uptodate) {
1765 sector_t sync_blocks = 0;
1766 sector_t s = r1_bio->sector;
1767 long sectors_to_go = r1_bio->sectors;
1768 /* make sure these bits doesn't get cleared. */
1769 do {
1770 bitmap_end_sync(mddev->bitmap, s,
1771 &sync_blocks, 1);
1772 s += sync_blocks;
1773 sectors_to_go -= sync_blocks;
1774 } while (sectors_to_go > 0);
1775 set_bit(WriteErrorSeen,
1776 &conf->mirrors[mirror].rdev->flags);
1777 if (!test_and_set_bit(WantReplacement,
1778 &conf->mirrors[mirror].rdev->flags))
1779 set_bit(MD_RECOVERY_NEEDED, &
1780 mddev->recovery);
1781 set_bit(R1BIO_WriteError, &r1_bio->state);
1782 } else if (is_badblock(conf->mirrors[mirror].rdev,
1783 r1_bio->sector,
1784 r1_bio->sectors,
1785 &first_bad, &bad_sectors) &&
1786 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1787 r1_bio->sector,
1788 r1_bio->sectors,
1789 &first_bad, &bad_sectors)
1790 )
1791 set_bit(R1BIO_MadeGood, &r1_bio->state);
1792
1793 if (atomic_dec_and_test(&r1_bio->remaining)) {
1794 int s = r1_bio->sectors;
1795 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1796 test_bit(R1BIO_WriteError, &r1_bio->state))
1797 reschedule_retry(r1_bio);
1798 else {
1799 put_buf(r1_bio);
1800 md_done_sync(mddev, s, uptodate);
1801 }
1802 }
1803}
1804
1805static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1806 int sectors, struct page *page, int rw)
1807{
1808 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1809 /* success */
1810 return 1;
1811 if (rw == WRITE) {
1812 set_bit(WriteErrorSeen, &rdev->flags);
1813 if (!test_and_set_bit(WantReplacement,
1814 &rdev->flags))
1815 set_bit(MD_RECOVERY_NEEDED, &
1816 rdev->mddev->recovery);
1817 }
1818 /* need to record an error - either for the block or the device */
1819 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1820 md_error(rdev->mddev, rdev);
1821 return 0;
1822}
1823
1824static int fix_sync_read_error(struct r1bio *r1_bio)
1825{
1826 /* Try some synchronous reads of other devices to get
1827 * good data, much like with normal read errors. Only
1828 * read into the pages we already have so we don't
1829 * need to re-issue the read request.
1830 * We don't need to freeze the array, because being in an
1831 * active sync request, there is no normal IO, and
1832 * no overlapping syncs.
1833 * We don't need to check is_badblock() again as we
1834 * made sure that anything with a bad block in range
1835 * will have bi_end_io clear.
1836 */
1837 struct mddev *mddev = r1_bio->mddev;
1838 struct r1conf *conf = mddev->private;
1839 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1840 sector_t sect = r1_bio->sector;
1841 int sectors = r1_bio->sectors;
1842 int idx = 0;
1843
1844 while(sectors) {
1845 int s = sectors;
1846 int d = r1_bio->read_disk;
1847 int success = 0;
1848 struct md_rdev *rdev;
1849 int start;
1850
1851 if (s > (PAGE_SIZE>>9))
1852 s = PAGE_SIZE >> 9;
1853 do {
1854 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1855 /* No rcu protection needed here devices
1856 * can only be removed when no resync is
1857 * active, and resync is currently active
1858 */
1859 rdev = conf->mirrors[d].rdev;
1860 if (sync_page_io(rdev, sect, s<<9,
1861 bio->bi_io_vec[idx].bv_page,
1862 READ, false)) {
1863 success = 1;
1864 break;
1865 }
1866 }
1867 d++;
1868 if (d == conf->raid_disks * 2)
1869 d = 0;
1870 } while (!success && d != r1_bio->read_disk);
1871
1872 if (!success) {
1873 char b[BDEVNAME_SIZE];
1874 int abort = 0;
1875 /* Cannot read from anywhere, this block is lost.
1876 * Record a bad block on each device. If that doesn't
1877 * work just disable and interrupt the recovery.
1878 * Don't fail devices as that won't really help.
1879 */
1880 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1881 " for block %llu\n",
1882 mdname(mddev),
1883 bdevname(bio->bi_bdev, b),
1884 (unsigned long long)r1_bio->sector);
1885 for (d = 0; d < conf->raid_disks * 2; d++) {
1886 rdev = conf->mirrors[d].rdev;
1887 if (!rdev || test_bit(Faulty, &rdev->flags))
1888 continue;
1889 if (!rdev_set_badblocks(rdev, sect, s, 0))
1890 abort = 1;
1891 }
1892 if (abort) {
1893 conf->recovery_disabled =
1894 mddev->recovery_disabled;
1895 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1896 md_done_sync(mddev, r1_bio->sectors, 0);
1897 put_buf(r1_bio);
1898 return 0;
1899 }
1900 /* Try next page */
1901 sectors -= s;
1902 sect += s;
1903 idx++;
1904 continue;
1905 }
1906
1907 start = d;
1908 /* write it back and re-read */
1909 while (d != r1_bio->read_disk) {
1910 if (d == 0)
1911 d = conf->raid_disks * 2;
1912 d--;
1913 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1914 continue;
1915 rdev = conf->mirrors[d].rdev;
1916 if (r1_sync_page_io(rdev, sect, s,
1917 bio->bi_io_vec[idx].bv_page,
1918 WRITE) == 0) {
1919 r1_bio->bios[d]->bi_end_io = NULL;
1920 rdev_dec_pending(rdev, mddev);
1921 }
1922 }
1923 d = start;
1924 while (d != r1_bio->read_disk) {
1925 if (d == 0)
1926 d = conf->raid_disks * 2;
1927 d--;
1928 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1929 continue;
1930 rdev = conf->mirrors[d].rdev;
1931 if (r1_sync_page_io(rdev, sect, s,
1932 bio->bi_io_vec[idx].bv_page,
1933 READ) != 0)
1934 atomic_add(s, &rdev->corrected_errors);
1935 }
1936 sectors -= s;
1937 sect += s;
1938 idx ++;
1939 }
1940 set_bit(R1BIO_Uptodate, &r1_bio->state);
1941 set_bit(BIO_UPTODATE, &bio->bi_flags);
1942 return 1;
1943}
1944
1945static int process_checks(struct r1bio *r1_bio)
1946{
1947 /* We have read all readable devices. If we haven't
1948 * got the block, then there is no hope left.
1949 * If we have, then we want to do a comparison
1950 * and skip the write if everything is the same.
1951 * If any blocks failed to read, then we need to
1952 * attempt an over-write
1953 */
1954 struct mddev *mddev = r1_bio->mddev;
1955 struct r1conf *conf = mddev->private;
1956 int primary;
1957 int i;
1958 int vcnt;
1959
1960 /* Fix variable parts of all bios */
1961 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1962 for (i = 0; i < conf->raid_disks * 2; i++) {
1963 int j;
1964 int size;
1965 int uptodate;
1966 struct bio *b = r1_bio->bios[i];
1967 if (b->bi_end_io != end_sync_read)
1968 continue;
1969 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1970 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags);
1971 bio_reset(b);
1972 if (!uptodate)
1973 clear_bit(BIO_UPTODATE, &b->bi_flags);
1974 b->bi_vcnt = vcnt;
1975 b->bi_iter.bi_size = r1_bio->sectors << 9;
1976 b->bi_iter.bi_sector = r1_bio->sector +
1977 conf->mirrors[i].rdev->data_offset;
1978 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1979 b->bi_end_io = end_sync_read;
1980 b->bi_private = r1_bio;
1981
1982 size = b->bi_iter.bi_size;
1983 for (j = 0; j < vcnt ; j++) {
1984 struct bio_vec *bi;
1985 bi = &b->bi_io_vec[j];
1986 bi->bv_offset = 0;
1987 if (size > PAGE_SIZE)
1988 bi->bv_len = PAGE_SIZE;
1989 else
1990 bi->bv_len = size;
1991 size -= PAGE_SIZE;
1992 }
1993 }
1994 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1995 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1996 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1997 r1_bio->bios[primary]->bi_end_io = NULL;
1998 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1999 break;
2000 }
2001 r1_bio->read_disk = primary;
2002 for (i = 0; i < conf->raid_disks * 2; i++) {
2003 int j;
2004 struct bio *pbio = r1_bio->bios[primary];
2005 struct bio *sbio = r1_bio->bios[i];
2006 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags);
2007
2008 if (sbio->bi_end_io != end_sync_read)
2009 continue;
2010 /* Now we can 'fixup' the BIO_UPTODATE flag */
2011 set_bit(BIO_UPTODATE, &sbio->bi_flags);
2012
2013 if (uptodate) {
2014 for (j = vcnt; j-- ; ) {
2015 struct page *p, *s;
2016 p = pbio->bi_io_vec[j].bv_page;
2017 s = sbio->bi_io_vec[j].bv_page;
2018 if (memcmp(page_address(p),
2019 page_address(s),
2020 sbio->bi_io_vec[j].bv_len))
2021 break;
2022 }
2023 } else
2024 j = 0;
2025 if (j >= 0)
2026 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2027 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2028 && uptodate)) {
2029 /* No need to write to this device. */
2030 sbio->bi_end_io = NULL;
2031 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2032 continue;
2033 }
2034
2035 bio_copy_data(sbio, pbio);
2036 }
2037 return 0;
2038}
2039
2040static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2041{
2042 struct r1conf *conf = mddev->private;
2043 int i;
2044 int disks = conf->raid_disks * 2;
2045 struct bio *bio, *wbio;
2046
2047 bio = r1_bio->bios[r1_bio->read_disk];
2048
2049 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2050 /* ouch - failed to read all of that. */
2051 if (!fix_sync_read_error(r1_bio))
2052 return;
2053
2054 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2055 if (process_checks(r1_bio) < 0)
2056 return;
2057 /*
2058 * schedule writes
2059 */
2060 atomic_set(&r1_bio->remaining, 1);
2061 for (i = 0; i < disks ; i++) {
2062 wbio = r1_bio->bios[i];
2063 if (wbio->bi_end_io == NULL ||
2064 (wbio->bi_end_io == end_sync_read &&
2065 (i == r1_bio->read_disk ||
2066 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2067 continue;
2068
2069 wbio->bi_rw = WRITE;
2070 wbio->bi_end_io = end_sync_write;
2071 atomic_inc(&r1_bio->remaining);
2072 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2073
2074 generic_make_request(wbio);
2075 }
2076
2077 if (atomic_dec_and_test(&r1_bio->remaining)) {
2078 /* if we're here, all write(s) have completed, so clean up */
2079 int s = r1_bio->sectors;
2080 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2081 test_bit(R1BIO_WriteError, &r1_bio->state))
2082 reschedule_retry(r1_bio);
2083 else {
2084 put_buf(r1_bio);
2085 md_done_sync(mddev, s, 1);
2086 }
2087 }
2088}
2089
2090/*
2091 * This is a kernel thread which:
2092 *
2093 * 1. Retries failed read operations on working mirrors.
2094 * 2. Updates the raid superblock when problems encounter.
2095 * 3. Performs writes following reads for array synchronising.
2096 */
2097
2098static void fix_read_error(struct r1conf *conf, int read_disk,
2099 sector_t sect, int sectors)
2100{
2101 struct mddev *mddev = conf->mddev;
2102 while(sectors) {
2103 int s = sectors;
2104 int d = read_disk;
2105 int success = 0;
2106 int start;
2107 struct md_rdev *rdev;
2108
2109 if (s > (PAGE_SIZE>>9))
2110 s = PAGE_SIZE >> 9;
2111
2112 do {
2113 /* Note: no rcu protection needed here
2114 * as this is synchronous in the raid1d thread
2115 * which is the thread that might remove
2116 * a device. If raid1d ever becomes multi-threaded....
2117 */
2118 sector_t first_bad;
2119 int bad_sectors;
2120
2121 rdev = conf->mirrors[d].rdev;
2122 if (rdev &&
2123 (test_bit(In_sync, &rdev->flags) ||
2124 (!test_bit(Faulty, &rdev->flags) &&
2125 rdev->recovery_offset >= sect + s)) &&
2126 is_badblock(rdev, sect, s,
2127 &first_bad, &bad_sectors) == 0 &&
2128 sync_page_io(rdev, sect, s<<9,
2129 conf->tmppage, READ, false))
2130 success = 1;
2131 else {
2132 d++;
2133 if (d == conf->raid_disks * 2)
2134 d = 0;
2135 }
2136 } while (!success && d != read_disk);
2137
2138 if (!success) {
2139 /* Cannot read from anywhere - mark it bad */
2140 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2141 if (!rdev_set_badblocks(rdev, sect, s, 0))
2142 md_error(mddev, rdev);
2143 break;
2144 }
2145 /* write it back and re-read */
2146 start = d;
2147 while (d != read_disk) {
2148 if (d==0)
2149 d = conf->raid_disks * 2;
2150 d--;
2151 rdev = conf->mirrors[d].rdev;
2152 if (rdev &&
2153 test_bit(In_sync, &rdev->flags))
2154 r1_sync_page_io(rdev, sect, s,
2155 conf->tmppage, WRITE);
2156 }
2157 d = start;
2158 while (d != read_disk) {
2159 char b[BDEVNAME_SIZE];
2160 if (d==0)
2161 d = conf->raid_disks * 2;
2162 d--;
2163 rdev = conf->mirrors[d].rdev;
2164 if (rdev &&
2165 test_bit(In_sync, &rdev->flags)) {
2166 if (r1_sync_page_io(rdev, sect, s,
2167 conf->tmppage, READ)) {
2168 atomic_add(s, &rdev->corrected_errors);
2169 printk(KERN_INFO
2170 "md/raid1:%s: read error corrected "
2171 "(%d sectors at %llu on %s)\n",
2172 mdname(mddev), s,
2173 (unsigned long long)(sect +
2174 rdev->data_offset),
2175 bdevname(rdev->bdev, b));
2176 }
2177 }
2178 }
2179 sectors -= s;
2180 sect += s;
2181 }
2182}
2183
2184static int narrow_write_error(struct r1bio *r1_bio, int i)
2185{
2186 struct mddev *mddev = r1_bio->mddev;
2187 struct r1conf *conf = mddev->private;
2188 struct md_rdev *rdev = conf->mirrors[i].rdev;
2189
2190 /* bio has the data to be written to device 'i' where
2191 * we just recently had a write error.
2192 * We repeatedly clone the bio and trim down to one block,
2193 * then try the write. Where the write fails we record
2194 * a bad block.
2195 * It is conceivable that the bio doesn't exactly align with
2196 * blocks. We must handle this somehow.
2197 *
2198 * We currently own a reference on the rdev.
2199 */
2200
2201 int block_sectors;
2202 sector_t sector;
2203 int sectors;
2204 int sect_to_write = r1_bio->sectors;
2205 int ok = 1;
2206
2207 if (rdev->badblocks.shift < 0)
2208 return 0;
2209
2210 block_sectors = 1 << rdev->badblocks.shift;
2211 sector = r1_bio->sector;
2212 sectors = ((sector + block_sectors)
2213 & ~(sector_t)(block_sectors - 1))
2214 - sector;
2215
2216 while (sect_to_write) {
2217 struct bio *wbio;
2218 if (sectors > sect_to_write)
2219 sectors = sect_to_write;
2220 /* Write at 'sector' for 'sectors'*/
2221
2222 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2223 unsigned vcnt = r1_bio->behind_page_count;
2224 struct bio_vec *vec = r1_bio->behind_bvecs;
2225
2226 while (!vec->bv_page) {
2227 vec++;
2228 vcnt--;
2229 }
2230
2231 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2232 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2233
2234 wbio->bi_vcnt = vcnt;
2235 } else {
2236 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2237 }
2238
2239 wbio->bi_rw = WRITE;
2240 wbio->bi_iter.bi_sector = r1_bio->sector;
2241 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2242
2243 bio_trim(wbio, sector - r1_bio->sector, sectors);
2244 wbio->bi_iter.bi_sector += rdev->data_offset;
2245 wbio->bi_bdev = rdev->bdev;
2246 if (submit_bio_wait(WRITE, wbio) == 0)
2247 /* failure! */
2248 ok = rdev_set_badblocks(rdev, sector,
2249 sectors, 0)
2250 && ok;
2251
2252 bio_put(wbio);
2253 sect_to_write -= sectors;
2254 sector += sectors;
2255 sectors = block_sectors;
2256 }
2257 return ok;
2258}
2259
2260static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2261{
2262 int m;
2263 int s = r1_bio->sectors;
2264 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2265 struct md_rdev *rdev = conf->mirrors[m].rdev;
2266 struct bio *bio = r1_bio->bios[m];
2267 if (bio->bi_end_io == NULL)
2268 continue;
2269 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2270 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2271 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2272 }
2273 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2274 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2275 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2276 md_error(conf->mddev, rdev);
2277 }
2278 }
2279 put_buf(r1_bio);
2280 md_done_sync(conf->mddev, s, 1);
2281}
2282
2283static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2284{
2285 int m;
2286 for (m = 0; m < conf->raid_disks * 2 ; m++)
2287 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2288 struct md_rdev *rdev = conf->mirrors[m].rdev;
2289 rdev_clear_badblocks(rdev,
2290 r1_bio->sector,
2291 r1_bio->sectors, 0);
2292 rdev_dec_pending(rdev, conf->mddev);
2293 } else if (r1_bio->bios[m] != NULL) {
2294 /* This drive got a write error. We need to
2295 * narrow down and record precise write
2296 * errors.
2297 */
2298 if (!narrow_write_error(r1_bio, m)) {
2299 md_error(conf->mddev,
2300 conf->mirrors[m].rdev);
2301 /* an I/O failed, we can't clear the bitmap */
2302 set_bit(R1BIO_Degraded, &r1_bio->state);
2303 }
2304 rdev_dec_pending(conf->mirrors[m].rdev,
2305 conf->mddev);
2306 }
2307 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2308 close_write(r1_bio);
2309 raid_end_bio_io(r1_bio);
2310}
2311
2312static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2313{
2314 int disk;
2315 int max_sectors;
2316 struct mddev *mddev = conf->mddev;
2317 struct bio *bio;
2318 char b[BDEVNAME_SIZE];
2319 struct md_rdev *rdev;
2320
2321 clear_bit(R1BIO_ReadError, &r1_bio->state);
2322 /* we got a read error. Maybe the drive is bad. Maybe just
2323 * the block and we can fix it.
2324 * We freeze all other IO, and try reading the block from
2325 * other devices. When we find one, we re-write
2326 * and check it that fixes the read error.
2327 * This is all done synchronously while the array is
2328 * frozen
2329 */
2330 if (mddev->ro == 0) {
2331 freeze_array(conf, 1);
2332 fix_read_error(conf, r1_bio->read_disk,
2333 r1_bio->sector, r1_bio->sectors);
2334 unfreeze_array(conf);
2335 } else
2336 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2337 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2338
2339 bio = r1_bio->bios[r1_bio->read_disk];
2340 bdevname(bio->bi_bdev, b);
2341read_more:
2342 disk = read_balance(conf, r1_bio, &max_sectors);
2343 if (disk == -1) {
2344 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2345 " read error for block %llu\n",
2346 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2347 raid_end_bio_io(r1_bio);
2348 } else {
2349 const unsigned long do_sync
2350 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2351 if (bio) {
2352 r1_bio->bios[r1_bio->read_disk] =
2353 mddev->ro ? IO_BLOCKED : NULL;
2354 bio_put(bio);
2355 }
2356 r1_bio->read_disk = disk;
2357 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2358 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2359 max_sectors);
2360 r1_bio->bios[r1_bio->read_disk] = bio;
2361 rdev = conf->mirrors[disk].rdev;
2362 printk_ratelimited(KERN_ERR
2363 "md/raid1:%s: redirecting sector %llu"
2364 " to other mirror: %s\n",
2365 mdname(mddev),
2366 (unsigned long long)r1_bio->sector,
2367 bdevname(rdev->bdev, b));
2368 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2369 bio->bi_bdev = rdev->bdev;
2370 bio->bi_end_io = raid1_end_read_request;
2371 bio->bi_rw = READ | do_sync;
2372 bio->bi_private = r1_bio;
2373 if (max_sectors < r1_bio->sectors) {
2374 /* Drat - have to split this up more */
2375 struct bio *mbio = r1_bio->master_bio;
2376 int sectors_handled = (r1_bio->sector + max_sectors
2377 - mbio->bi_iter.bi_sector);
2378 r1_bio->sectors = max_sectors;
2379 spin_lock_irq(&conf->device_lock);
2380 if (mbio->bi_phys_segments == 0)
2381 mbio->bi_phys_segments = 2;
2382 else
2383 mbio->bi_phys_segments++;
2384 spin_unlock_irq(&conf->device_lock);
2385 generic_make_request(bio);
2386 bio = NULL;
2387
2388 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2389
2390 r1_bio->master_bio = mbio;
2391 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2392 r1_bio->state = 0;
2393 set_bit(R1BIO_ReadError, &r1_bio->state);
2394 r1_bio->mddev = mddev;
2395 r1_bio->sector = mbio->bi_iter.bi_sector +
2396 sectors_handled;
2397
2398 goto read_more;
2399 } else
2400 generic_make_request(bio);
2401 }
2402}
2403
2404static void raid1d(struct md_thread *thread)
2405{
2406 struct mddev *mddev = thread->mddev;
2407 struct r1bio *r1_bio;
2408 unsigned long flags;
2409 struct r1conf *conf = mddev->private;
2410 struct list_head *head = &conf->retry_list;
2411 struct blk_plug plug;
2412
2413 md_check_recovery(mddev);
2414
2415 blk_start_plug(&plug);
2416 for (;;) {
2417
2418 flush_pending_writes(conf);
2419
2420 spin_lock_irqsave(&conf->device_lock, flags);
2421 if (list_empty(head)) {
2422 spin_unlock_irqrestore(&conf->device_lock, flags);
2423 break;
2424 }
2425 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2426 list_del(head->prev);
2427 conf->nr_queued--;
2428 spin_unlock_irqrestore(&conf->device_lock, flags);
2429
2430 mddev = r1_bio->mddev;
2431 conf = mddev->private;
2432 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2433 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2434 test_bit(R1BIO_WriteError, &r1_bio->state))
2435 handle_sync_write_finished(conf, r1_bio);
2436 else
2437 sync_request_write(mddev, r1_bio);
2438 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2439 test_bit(R1BIO_WriteError, &r1_bio->state))
2440 handle_write_finished(conf, r1_bio);
2441 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2442 handle_read_error(conf, r1_bio);
2443 else
2444 /* just a partial read to be scheduled from separate
2445 * context
2446 */
2447 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2448
2449 cond_resched();
2450 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2451 md_check_recovery(mddev);
2452 }
2453 blk_finish_plug(&plug);
2454}
2455
2456
2457static int init_resync(struct r1conf *conf)
2458{
2459 int buffs;
2460
2461 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2462 BUG_ON(conf->r1buf_pool);
2463 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2464 conf->poolinfo);
2465 if (!conf->r1buf_pool)
2466 return -ENOMEM;
2467 conf->next_resync = 0;
2468 return 0;
2469}
2470
2471/*
2472 * perform a "sync" on one "block"
2473 *
2474 * We need to make sure that no normal I/O request - particularly write
2475 * requests - conflict with active sync requests.
2476 *
2477 * This is achieved by tracking pending requests and a 'barrier' concept
2478 * that can be installed to exclude normal IO requests.
2479 */
2480
2481static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2482{
2483 struct r1conf *conf = mddev->private;
2484 struct r1bio *r1_bio;
2485 struct bio *bio;
2486 sector_t max_sector, nr_sectors;
2487 int disk = -1;
2488 int i;
2489 int wonly = -1;
2490 int write_targets = 0, read_targets = 0;
2491 sector_t sync_blocks;
2492 int still_degraded = 0;
2493 int good_sectors = RESYNC_SECTORS;
2494 int min_bad = 0; /* number of sectors that are bad in all devices */
2495
2496 if (!conf->r1buf_pool)
2497 if (init_resync(conf))
2498 return 0;
2499
2500 max_sector = mddev->dev_sectors;
2501 if (sector_nr >= max_sector) {
2502 /* If we aborted, we need to abort the
2503 * sync on the 'current' bitmap chunk (there will
2504 * only be one in raid1 resync.
2505 * We can find the current addess in mddev->curr_resync
2506 */
2507 if (mddev->curr_resync < max_sector) /* aborted */
2508 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2509 &sync_blocks, 1);
2510 else /* completed sync */
2511 conf->fullsync = 0;
2512
2513 bitmap_close_sync(mddev->bitmap);
2514 close_sync(conf);
2515 return 0;
2516 }
2517
2518 if (mddev->bitmap == NULL &&
2519 mddev->recovery_cp == MaxSector &&
2520 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2521 conf->fullsync == 0) {
2522 *skipped = 1;
2523 return max_sector - sector_nr;
2524 }
2525 /* before building a request, check if we can skip these blocks..
2526 * This call the bitmap_start_sync doesn't actually record anything
2527 */
2528 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2529 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2530 /* We can skip this block, and probably several more */
2531 *skipped = 1;
2532 return sync_blocks;
2533 }
2534 /*
2535 * If there is non-resync activity waiting for a turn,
2536 * and resync is going fast enough,
2537 * then let it though before starting on this new sync request.
2538 */
2539 if (!go_faster && conf->nr_waiting)
2540 msleep_interruptible(1000);
2541
2542 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2543 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2544 raise_barrier(conf);
2545
2546 conf->next_resync = sector_nr;
2547
2548 rcu_read_lock();
2549 /*
2550 * If we get a correctably read error during resync or recovery,
2551 * we might want to read from a different device. So we
2552 * flag all drives that could conceivably be read from for READ,
2553 * and any others (which will be non-In_sync devices) for WRITE.
2554 * If a read fails, we try reading from something else for which READ
2555 * is OK.
2556 */
2557
2558 r1_bio->mddev = mddev;
2559 r1_bio->sector = sector_nr;
2560 r1_bio->state = 0;
2561 set_bit(R1BIO_IsSync, &r1_bio->state);
2562
2563 for (i = 0; i < conf->raid_disks * 2; i++) {
2564 struct md_rdev *rdev;
2565 bio = r1_bio->bios[i];
2566 bio_reset(bio);
2567
2568 rdev = rcu_dereference(conf->mirrors[i].rdev);
2569 if (rdev == NULL ||
2570 test_bit(Faulty, &rdev->flags)) {
2571 if (i < conf->raid_disks)
2572 still_degraded = 1;
2573 } else if (!test_bit(In_sync, &rdev->flags)) {
2574 bio->bi_rw = WRITE;
2575 bio->bi_end_io = end_sync_write;
2576 write_targets ++;
2577 } else {
2578 /* may need to read from here */
2579 sector_t first_bad = MaxSector;
2580 int bad_sectors;
2581
2582 if (is_badblock(rdev, sector_nr, good_sectors,
2583 &first_bad, &bad_sectors)) {
2584 if (first_bad > sector_nr)
2585 good_sectors = first_bad - sector_nr;
2586 else {
2587 bad_sectors -= (sector_nr - first_bad);
2588 if (min_bad == 0 ||
2589 min_bad > bad_sectors)
2590 min_bad = bad_sectors;
2591 }
2592 }
2593 if (sector_nr < first_bad) {
2594 if (test_bit(WriteMostly, &rdev->flags)) {
2595 if (wonly < 0)
2596 wonly = i;
2597 } else {
2598 if (disk < 0)
2599 disk = i;
2600 }
2601 bio->bi_rw = READ;
2602 bio->bi_end_io = end_sync_read;
2603 read_targets++;
2604 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2605 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2606 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2607 /*
2608 * The device is suitable for reading (InSync),
2609 * but has bad block(s) here. Let's try to correct them,
2610 * if we are doing resync or repair. Otherwise, leave
2611 * this device alone for this sync request.
2612 */
2613 bio->bi_rw = WRITE;
2614 bio->bi_end_io = end_sync_write;
2615 write_targets++;
2616 }
2617 }
2618 if (bio->bi_end_io) {
2619 atomic_inc(&rdev->nr_pending);
2620 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2621 bio->bi_bdev = rdev->bdev;
2622 bio->bi_private = r1_bio;
2623 }
2624 }
2625 rcu_read_unlock();
2626 if (disk < 0)
2627 disk = wonly;
2628 r1_bio->read_disk = disk;
2629
2630 if (read_targets == 0 && min_bad > 0) {
2631 /* These sectors are bad on all InSync devices, so we
2632 * need to mark them bad on all write targets
2633 */
2634 int ok = 1;
2635 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2636 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2637 struct md_rdev *rdev = conf->mirrors[i].rdev;
2638 ok = rdev_set_badblocks(rdev, sector_nr,
2639 min_bad, 0
2640 ) && ok;
2641 }
2642 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2643 *skipped = 1;
2644 put_buf(r1_bio);
2645
2646 if (!ok) {
2647 /* Cannot record the badblocks, so need to
2648 * abort the resync.
2649 * If there are multiple read targets, could just
2650 * fail the really bad ones ???
2651 */
2652 conf->recovery_disabled = mddev->recovery_disabled;
2653 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2654 return 0;
2655 } else
2656 return min_bad;
2657
2658 }
2659 if (min_bad > 0 && min_bad < good_sectors) {
2660 /* only resync enough to reach the next bad->good
2661 * transition */
2662 good_sectors = min_bad;
2663 }
2664
2665 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2666 /* extra read targets are also write targets */
2667 write_targets += read_targets-1;
2668
2669 if (write_targets == 0 || read_targets == 0) {
2670 /* There is nowhere to write, so all non-sync
2671 * drives must be failed - so we are finished
2672 */
2673 sector_t rv;
2674 if (min_bad > 0)
2675 max_sector = sector_nr + min_bad;
2676 rv = max_sector - sector_nr;
2677 *skipped = 1;
2678 put_buf(r1_bio);
2679 return rv;
2680 }
2681
2682 if (max_sector > mddev->resync_max)
2683 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2684 if (max_sector > sector_nr + good_sectors)
2685 max_sector = sector_nr + good_sectors;
2686 nr_sectors = 0;
2687 sync_blocks = 0;
2688 do {
2689 struct page *page;
2690 int len = PAGE_SIZE;
2691 if (sector_nr + (len>>9) > max_sector)
2692 len = (max_sector - sector_nr) << 9;
2693 if (len == 0)
2694 break;
2695 if (sync_blocks == 0) {
2696 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2697 &sync_blocks, still_degraded) &&
2698 !conf->fullsync &&
2699 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2700 break;
2701 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2702 if ((len >> 9) > sync_blocks)
2703 len = sync_blocks<<9;
2704 }
2705
2706 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2707 bio = r1_bio->bios[i];
2708 if (bio->bi_end_io) {
2709 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2710 if (bio_add_page(bio, page, len, 0) == 0) {
2711 /* stop here */
2712 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2713 while (i > 0) {
2714 i--;
2715 bio = r1_bio->bios[i];
2716 if (bio->bi_end_io==NULL)
2717 continue;
2718 /* remove last page from this bio */
2719 bio->bi_vcnt--;
2720 bio->bi_iter.bi_size -= len;
2721 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2722 }
2723 goto bio_full;
2724 }
2725 }
2726 }
2727 nr_sectors += len>>9;
2728 sector_nr += len>>9;
2729 sync_blocks -= (len>>9);
2730 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2731 bio_full:
2732 r1_bio->sectors = nr_sectors;
2733
2734 /* For a user-requested sync, we read all readable devices and do a
2735 * compare
2736 */
2737 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2738 atomic_set(&r1_bio->remaining, read_targets);
2739 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2740 bio = r1_bio->bios[i];
2741 if (bio->bi_end_io == end_sync_read) {
2742 read_targets--;
2743 md_sync_acct(bio->bi_bdev, nr_sectors);
2744 generic_make_request(bio);
2745 }
2746 }
2747 } else {
2748 atomic_set(&r1_bio->remaining, 1);
2749 bio = r1_bio->bios[r1_bio->read_disk];
2750 md_sync_acct(bio->bi_bdev, nr_sectors);
2751 generic_make_request(bio);
2752
2753 }
2754 return nr_sectors;
2755}
2756
2757static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2758{
2759 if (sectors)
2760 return sectors;
2761
2762 return mddev->dev_sectors;
2763}
2764
2765static struct r1conf *setup_conf(struct mddev *mddev)
2766{
2767 struct r1conf *conf;
2768 int i;
2769 struct raid1_info *disk;
2770 struct md_rdev *rdev;
2771 int err = -ENOMEM;
2772
2773 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2774 if (!conf)
2775 goto abort;
2776
2777 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2778 * mddev->raid_disks * 2,
2779 GFP_KERNEL);
2780 if (!conf->mirrors)
2781 goto abort;
2782
2783 conf->tmppage = alloc_page(GFP_KERNEL);
2784 if (!conf->tmppage)
2785 goto abort;
2786
2787 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2788 if (!conf->poolinfo)
2789 goto abort;
2790 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2791 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2792 r1bio_pool_free,
2793 conf->poolinfo);
2794 if (!conf->r1bio_pool)
2795 goto abort;
2796
2797 conf->poolinfo->mddev = mddev;
2798
2799 err = -EINVAL;
2800 spin_lock_init(&conf->device_lock);
2801 rdev_for_each(rdev, mddev) {
2802 struct request_queue *q;
2803 int disk_idx = rdev->raid_disk;
2804 if (disk_idx >= mddev->raid_disks
2805 || disk_idx < 0)
2806 continue;
2807 if (test_bit(Replacement, &rdev->flags))
2808 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2809 else
2810 disk = conf->mirrors + disk_idx;
2811
2812 if (disk->rdev)
2813 goto abort;
2814 disk->rdev = rdev;
2815 q = bdev_get_queue(rdev->bdev);
2816 if (q->merge_bvec_fn)
2817 mddev->merge_check_needed = 1;
2818
2819 disk->head_position = 0;
2820 disk->seq_start = MaxSector;
2821 }
2822 conf->raid_disks = mddev->raid_disks;
2823 conf->mddev = mddev;
2824 INIT_LIST_HEAD(&conf->retry_list);
2825
2826 spin_lock_init(&conf->resync_lock);
2827 init_waitqueue_head(&conf->wait_barrier);
2828
2829 bio_list_init(&conf->pending_bio_list);
2830 conf->pending_count = 0;
2831 conf->recovery_disabled = mddev->recovery_disabled - 1;
2832
2833 conf->start_next_window = MaxSector;
2834 conf->current_window_requests = conf->next_window_requests = 0;
2835
2836 err = -EIO;
2837 for (i = 0; i < conf->raid_disks * 2; i++) {
2838
2839 disk = conf->mirrors + i;
2840
2841 if (i < conf->raid_disks &&
2842 disk[conf->raid_disks].rdev) {
2843 /* This slot has a replacement. */
2844 if (!disk->rdev) {
2845 /* No original, just make the replacement
2846 * a recovering spare
2847 */
2848 disk->rdev =
2849 disk[conf->raid_disks].rdev;
2850 disk[conf->raid_disks].rdev = NULL;
2851 } else if (!test_bit(In_sync, &disk->rdev->flags))
2852 /* Original is not in_sync - bad */
2853 goto abort;
2854 }
2855
2856 if (!disk->rdev ||
2857 !test_bit(In_sync, &disk->rdev->flags)) {
2858 disk->head_position = 0;
2859 if (disk->rdev &&
2860 (disk->rdev->saved_raid_disk < 0))
2861 conf->fullsync = 1;
2862 }
2863 }
2864
2865 err = -ENOMEM;
2866 conf->thread = md_register_thread(raid1d, mddev, "raid1");
2867 if (!conf->thread) {
2868 printk(KERN_ERR
2869 "md/raid1:%s: couldn't allocate thread\n",
2870 mdname(mddev));
2871 goto abort;
2872 }
2873
2874 return conf;
2875
2876 abort:
2877 if (conf) {
2878 if (conf->r1bio_pool)
2879 mempool_destroy(conf->r1bio_pool);
2880 kfree(conf->mirrors);
2881 safe_put_page(conf->tmppage);
2882 kfree(conf->poolinfo);
2883 kfree(conf);
2884 }
2885 return ERR_PTR(err);
2886}
2887
2888static int stop(struct mddev *mddev);
2889static int run(struct mddev *mddev)
2890{
2891 struct r1conf *conf;
2892 int i;
2893 struct md_rdev *rdev;
2894 int ret;
2895 bool discard_supported = false;
2896
2897 if (mddev->level != 1) {
2898 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2899 mdname(mddev), mddev->level);
2900 return -EIO;
2901 }
2902 if (mddev->reshape_position != MaxSector) {
2903 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2904 mdname(mddev));
2905 return -EIO;
2906 }
2907 /*
2908 * copy the already verified devices into our private RAID1
2909 * bookkeeping area. [whatever we allocate in run(),
2910 * should be freed in stop()]
2911 */
2912 if (mddev->private == NULL)
2913 conf = setup_conf(mddev);
2914 else
2915 conf = mddev->private;
2916
2917 if (IS_ERR(conf))
2918 return PTR_ERR(conf);
2919
2920 if (mddev->queue)
2921 blk_queue_max_write_same_sectors(mddev->queue, 0);
2922
2923 rdev_for_each(rdev, mddev) {
2924 if (!mddev->gendisk)
2925 continue;
2926 disk_stack_limits(mddev->gendisk, rdev->bdev,
2927 rdev->data_offset << 9);
2928 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2929 discard_supported = true;
2930 }
2931
2932 mddev->degraded = 0;
2933 for (i=0; i < conf->raid_disks; i++)
2934 if (conf->mirrors[i].rdev == NULL ||
2935 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2936 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2937 mddev->degraded++;
2938
2939 if (conf->raid_disks - mddev->degraded == 1)
2940 mddev->recovery_cp = MaxSector;
2941
2942 if (mddev->recovery_cp != MaxSector)
2943 printk(KERN_NOTICE "md/raid1:%s: not clean"
2944 " -- starting background reconstruction\n",
2945 mdname(mddev));
2946 printk(KERN_INFO
2947 "md/raid1:%s: active with %d out of %d mirrors\n",
2948 mdname(mddev), mddev->raid_disks - mddev->degraded,
2949 mddev->raid_disks);
2950
2951 /*
2952 * Ok, everything is just fine now
2953 */
2954 mddev->thread = conf->thread;
2955 conf->thread = NULL;
2956 mddev->private = conf;
2957
2958 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2959
2960 if (mddev->queue) {
2961 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2962 mddev->queue->backing_dev_info.congested_data = mddev;
2963 blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2964
2965 if (discard_supported)
2966 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2967 mddev->queue);
2968 else
2969 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2970 mddev->queue);
2971 }
2972
2973 ret = md_integrity_register(mddev);
2974 if (ret)
2975 stop(mddev);
2976 return ret;
2977}
2978
2979static int stop(struct mddev *mddev)
2980{
2981 struct r1conf *conf = mddev->private;
2982 struct bitmap *bitmap = mddev->bitmap;
2983
2984 /* wait for behind writes to complete */
2985 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2986 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2987 mdname(mddev));
2988 /* need to kick something here to make sure I/O goes? */
2989 wait_event(bitmap->behind_wait,
2990 atomic_read(&bitmap->behind_writes) == 0);
2991 }
2992
2993 freeze_array(conf, 0);
2994 unfreeze_array(conf);
2995
2996 md_unregister_thread(&mddev->thread);
2997 if (conf->r1bio_pool)
2998 mempool_destroy(conf->r1bio_pool);
2999 kfree(conf->mirrors);
3000 safe_put_page(conf->tmppage);
3001 kfree(conf->poolinfo);
3002 kfree(conf);
3003 mddev->private = NULL;
3004 return 0;
3005}
3006
3007static int raid1_resize(struct mddev *mddev, sector_t sectors)
3008{
3009 /* no resync is happening, and there is enough space
3010 * on all devices, so we can resize.
3011 * We need to make sure resync covers any new space.
3012 * If the array is shrinking we should possibly wait until
3013 * any io in the removed space completes, but it hardly seems
3014 * worth it.
3015 */
3016 sector_t newsize = raid1_size(mddev, sectors, 0);
3017 if (mddev->external_size &&
3018 mddev->array_sectors > newsize)
3019 return -EINVAL;
3020 if (mddev->bitmap) {
3021 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3022 if (ret)
3023 return ret;
3024 }
3025 md_set_array_sectors(mddev, newsize);
3026 set_capacity(mddev->gendisk, mddev->array_sectors);
3027 revalidate_disk(mddev->gendisk);
3028 if (sectors > mddev->dev_sectors &&
3029 mddev->recovery_cp > mddev->dev_sectors) {
3030 mddev->recovery_cp = mddev->dev_sectors;
3031 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3032 }
3033 mddev->dev_sectors = sectors;
3034 mddev->resync_max_sectors = sectors;
3035 return 0;
3036}
3037
3038static int raid1_reshape(struct mddev *mddev)
3039{
3040 /* We need to:
3041 * 1/ resize the r1bio_pool
3042 * 2/ resize conf->mirrors
3043 *
3044 * We allocate a new r1bio_pool if we can.
3045 * Then raise a device barrier and wait until all IO stops.
3046 * Then resize conf->mirrors and swap in the new r1bio pool.
3047 *
3048 * At the same time, we "pack" the devices so that all the missing
3049 * devices have the higher raid_disk numbers.
3050 */
3051 mempool_t *newpool, *oldpool;
3052 struct pool_info *newpoolinfo;
3053 struct raid1_info *newmirrors;
3054 struct r1conf *conf = mddev->private;
3055 int cnt, raid_disks;
3056 unsigned long flags;
3057 int d, d2, err;
3058
3059 /* Cannot change chunk_size, layout, or level */
3060 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3061 mddev->layout != mddev->new_layout ||
3062 mddev->level != mddev->new_level) {
3063 mddev->new_chunk_sectors = mddev->chunk_sectors;
3064 mddev->new_layout = mddev->layout;
3065 mddev->new_level = mddev->level;
3066 return -EINVAL;
3067 }
3068
3069 err = md_allow_write(mddev);
3070 if (err)
3071 return err;
3072
3073 raid_disks = mddev->raid_disks + mddev->delta_disks;
3074
3075 if (raid_disks < conf->raid_disks) {
3076 cnt=0;
3077 for (d= 0; d < conf->raid_disks; d++)
3078 if (conf->mirrors[d].rdev)
3079 cnt++;
3080 if (cnt > raid_disks)
3081 return -EBUSY;
3082 }
3083
3084 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3085 if (!newpoolinfo)
3086 return -ENOMEM;
3087 newpoolinfo->mddev = mddev;
3088 newpoolinfo->raid_disks = raid_disks * 2;
3089
3090 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3091 r1bio_pool_free, newpoolinfo);
3092 if (!newpool) {
3093 kfree(newpoolinfo);
3094 return -ENOMEM;
3095 }
3096 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3097 GFP_KERNEL);
3098 if (!newmirrors) {
3099 kfree(newpoolinfo);
3100 mempool_destroy(newpool);
3101 return -ENOMEM;
3102 }
3103
3104 freeze_array(conf, 0);
3105
3106 /* ok, everything is stopped */
3107 oldpool = conf->r1bio_pool;
3108 conf->r1bio_pool = newpool;
3109
3110 for (d = d2 = 0; d < conf->raid_disks; d++) {
3111 struct md_rdev *rdev = conf->mirrors[d].rdev;
3112 if (rdev && rdev->raid_disk != d2) {
3113 sysfs_unlink_rdev(mddev, rdev);
3114 rdev->raid_disk = d2;
3115 sysfs_unlink_rdev(mddev, rdev);
3116 if (sysfs_link_rdev(mddev, rdev))
3117 printk(KERN_WARNING
3118 "md/raid1:%s: cannot register rd%d\n",
3119 mdname(mddev), rdev->raid_disk);
3120 }
3121 if (rdev)
3122 newmirrors[d2++].rdev = rdev;
3123 }
3124 kfree(conf->mirrors);
3125 conf->mirrors = newmirrors;
3126 kfree(conf->poolinfo);
3127 conf->poolinfo = newpoolinfo;
3128
3129 spin_lock_irqsave(&conf->device_lock, flags);
3130 mddev->degraded += (raid_disks - conf->raid_disks);
3131 spin_unlock_irqrestore(&conf->device_lock, flags);
3132 conf->raid_disks = mddev->raid_disks = raid_disks;
3133 mddev->delta_disks = 0;
3134
3135 unfreeze_array(conf);
3136
3137 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3138 md_wakeup_thread(mddev->thread);
3139
3140 mempool_destroy(oldpool);
3141 return 0;
3142}
3143
3144static void raid1_quiesce(struct mddev *mddev, int state)
3145{
3146 struct r1conf *conf = mddev->private;
3147
3148 switch(state) {
3149 case 2: /* wake for suspend */
3150 wake_up(&conf->wait_barrier);
3151 break;
3152 case 1:
3153 freeze_array(conf, 0);
3154 break;
3155 case 0:
3156 unfreeze_array(conf);
3157 break;
3158 }
3159}
3160
3161static void *raid1_takeover(struct mddev *mddev)
3162{
3163 /* raid1 can take over:
3164 * raid5 with 2 devices, any layout or chunk size
3165 */
3166 if (mddev->level == 5 && mddev->raid_disks == 2) {
3167 struct r1conf *conf;
3168 mddev->new_level = 1;
3169 mddev->new_layout = 0;
3170 mddev->new_chunk_sectors = 0;
3171 conf = setup_conf(mddev);
3172 if (!IS_ERR(conf))
3173 /* Array must appear to be quiesced */
3174 conf->array_frozen = 1;
3175 return conf;
3176 }
3177 return ERR_PTR(-EINVAL);
3178}
3179
3180static struct md_personality raid1_personality =
3181{
3182 .name = "raid1",
3183 .level = 1,
3184 .owner = THIS_MODULE,
3185 .make_request = make_request,
3186 .run = run,
3187 .stop = stop,
3188 .status = status,
3189 .error_handler = error,
3190 .hot_add_disk = raid1_add_disk,
3191 .hot_remove_disk= raid1_remove_disk,
3192 .spare_active = raid1_spare_active,
3193 .sync_request = sync_request,
3194 .resize = raid1_resize,
3195 .size = raid1_size,
3196 .check_reshape = raid1_reshape,
3197 .quiesce = raid1_quiesce,
3198 .takeover = raid1_takeover,
3199};
3200
3201static int __init raid_init(void)
3202{
3203 return register_md_personality(&raid1_personality);
3204}
3205
3206static void raid_exit(void)
3207{
3208 unregister_md_personality(&raid1_personality);
3209}
3210
3211module_init(raid_init);
3212module_exit(raid_exit);
3213MODULE_LICENSE("GPL");
3214MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3215MODULE_ALIAS("md-personality-3"); /* RAID1 */
3216MODULE_ALIAS("md-raid1");
3217MODULE_ALIAS("md-level-1");
3218
3219module_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
33#include <trace/events/block.h>
34
35#include "md.h"
36#include "raid1.h"
37#include "md-bitmap.h"
38
39#define UNSUPPORTED_MDDEV_FLAGS \
40 ((1L << MD_HAS_JOURNAL) | \
41 (1L << MD_JOURNAL_CLEAN) | \
42 (1L << MD_HAS_PPL) | \
43 (1L << MD_HAS_MULTIPLE_PPLS))
44
45static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
46static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
47
48#define raid1_log(md, fmt, args...) \
49 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
50
51#include "raid1-10.c"
52
53static int check_and_add_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
54{
55 struct wb_info *wi, *temp_wi;
56 unsigned long flags;
57 int ret = 0;
58 struct mddev *mddev = rdev->mddev;
59
60 wi = mempool_alloc(mddev->wb_info_pool, GFP_NOIO);
61
62 spin_lock_irqsave(&rdev->wb_list_lock, flags);
63 list_for_each_entry(temp_wi, &rdev->wb_list, list) {
64 /* collision happened */
65 if (hi > temp_wi->lo && lo < temp_wi->hi) {
66 ret = -EBUSY;
67 break;
68 }
69 }
70
71 if (!ret) {
72 wi->lo = lo;
73 wi->hi = hi;
74 list_add(&wi->list, &rdev->wb_list);
75 } else
76 mempool_free(wi, mddev->wb_info_pool);
77 spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
78
79 return ret;
80}
81
82static void remove_wb(struct md_rdev *rdev, sector_t lo, sector_t hi)
83{
84 struct wb_info *wi;
85 unsigned long flags;
86 int found = 0;
87 struct mddev *mddev = rdev->mddev;
88
89 spin_lock_irqsave(&rdev->wb_list_lock, flags);
90 list_for_each_entry(wi, &rdev->wb_list, list)
91 if (hi == wi->hi && lo == wi->lo) {
92 list_del(&wi->list);
93 mempool_free(wi, mddev->wb_info_pool);
94 found = 1;
95 break;
96 }
97
98 if (!found)
99 WARN(1, "The write behind IO is not recorded\n");
100 spin_unlock_irqrestore(&rdev->wb_list_lock, flags);
101 wake_up(&rdev->wb_io_wait);
102}
103
104/*
105 * for resync bio, r1bio pointer can be retrieved from the per-bio
106 * 'struct resync_pages'.
107 */
108static inline struct r1bio *get_resync_r1bio(struct bio *bio)
109{
110 return get_resync_pages(bio)->raid_bio;
111}
112
113static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
114{
115 struct pool_info *pi = data;
116 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
117
118 /* allocate a r1bio with room for raid_disks entries in the bios array */
119 return kzalloc(size, gfp_flags);
120}
121
122#define RESYNC_DEPTH 32
123#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
124#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
125#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
126#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
127#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
128
129static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
130{
131 struct pool_info *pi = data;
132 struct r1bio *r1_bio;
133 struct bio *bio;
134 int need_pages;
135 int j;
136 struct resync_pages *rps;
137
138 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
139 if (!r1_bio)
140 return NULL;
141
142 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
143 gfp_flags);
144 if (!rps)
145 goto out_free_r1bio;
146
147 /*
148 * Allocate bios : 1 for reading, n-1 for writing
149 */
150 for (j = pi->raid_disks ; j-- ; ) {
151 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
152 if (!bio)
153 goto out_free_bio;
154 r1_bio->bios[j] = bio;
155 }
156 /*
157 * Allocate RESYNC_PAGES data pages and attach them to
158 * the first bio.
159 * If this is a user-requested check/repair, allocate
160 * RESYNC_PAGES for each bio.
161 */
162 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
163 need_pages = pi->raid_disks;
164 else
165 need_pages = 1;
166 for (j = 0; j < pi->raid_disks; j++) {
167 struct resync_pages *rp = &rps[j];
168
169 bio = r1_bio->bios[j];
170
171 if (j < need_pages) {
172 if (resync_alloc_pages(rp, gfp_flags))
173 goto out_free_pages;
174 } else {
175 memcpy(rp, &rps[0], sizeof(*rp));
176 resync_get_all_pages(rp);
177 }
178
179 rp->raid_bio = r1_bio;
180 bio->bi_private = rp;
181 }
182
183 r1_bio->master_bio = NULL;
184
185 return r1_bio;
186
187out_free_pages:
188 while (--j >= 0)
189 resync_free_pages(&rps[j]);
190
191out_free_bio:
192 while (++j < pi->raid_disks)
193 bio_put(r1_bio->bios[j]);
194 kfree(rps);
195
196out_free_r1bio:
197 rbio_pool_free(r1_bio, data);
198 return NULL;
199}
200
201static void r1buf_pool_free(void *__r1_bio, void *data)
202{
203 struct pool_info *pi = data;
204 int i;
205 struct r1bio *r1bio = __r1_bio;
206 struct resync_pages *rp = NULL;
207
208 for (i = pi->raid_disks; i--; ) {
209 rp = get_resync_pages(r1bio->bios[i]);
210 resync_free_pages(rp);
211 bio_put(r1bio->bios[i]);
212 }
213
214 /* resync pages array stored in the 1st bio's .bi_private */
215 kfree(rp);
216
217 rbio_pool_free(r1bio, data);
218}
219
220static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
221{
222 int i;
223
224 for (i = 0; i < conf->raid_disks * 2; i++) {
225 struct bio **bio = r1_bio->bios + i;
226 if (!BIO_SPECIAL(*bio))
227 bio_put(*bio);
228 *bio = NULL;
229 }
230}
231
232static void free_r1bio(struct r1bio *r1_bio)
233{
234 struct r1conf *conf = r1_bio->mddev->private;
235
236 put_all_bios(conf, r1_bio);
237 mempool_free(r1_bio, &conf->r1bio_pool);
238}
239
240static void put_buf(struct r1bio *r1_bio)
241{
242 struct r1conf *conf = r1_bio->mddev->private;
243 sector_t sect = r1_bio->sector;
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->bi_end_io)
249 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
250 }
251
252 mempool_free(r1_bio, &conf->r1buf_pool);
253
254 lower_barrier(conf, sect);
255}
256
257static void reschedule_retry(struct r1bio *r1_bio)
258{
259 unsigned long flags;
260 struct mddev *mddev = r1_bio->mddev;
261 struct r1conf *conf = mddev->private;
262 int idx;
263
264 idx = sector_to_idx(r1_bio->sector);
265 spin_lock_irqsave(&conf->device_lock, flags);
266 list_add(&r1_bio->retry_list, &conf->retry_list);
267 atomic_inc(&conf->nr_queued[idx]);
268 spin_unlock_irqrestore(&conf->device_lock, flags);
269
270 wake_up(&conf->wait_barrier);
271 md_wakeup_thread(mddev->thread);
272}
273
274/*
275 * raid_end_bio_io() is called when we have finished servicing a mirrored
276 * operation and are ready to return a success/failure code to the buffer
277 * cache layer.
278 */
279static void call_bio_endio(struct r1bio *r1_bio)
280{
281 struct bio *bio = r1_bio->master_bio;
282 struct r1conf *conf = r1_bio->mddev->private;
283
284 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
285 bio->bi_status = BLK_STS_IOERR;
286
287 bio_endio(bio);
288 /*
289 * Wake up any possible resync thread that waits for the device
290 * to go idle.
291 */
292 allow_barrier(conf, r1_bio->sector);
293}
294
295static void raid_end_bio_io(struct r1bio *r1_bio)
296{
297 struct bio *bio = r1_bio->master_bio;
298
299 /* if nobody has done the final endio yet, do it now */
300 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
301 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
302 (bio_data_dir(bio) == WRITE) ? "write" : "read",
303 (unsigned long long) bio->bi_iter.bi_sector,
304 (unsigned long long) bio_end_sector(bio) - 1);
305
306 call_bio_endio(r1_bio);
307 }
308 free_r1bio(r1_bio);
309}
310
311/*
312 * Update disk head position estimator based on IRQ completion info.
313 */
314static inline void update_head_pos(int disk, struct r1bio *r1_bio)
315{
316 struct r1conf *conf = r1_bio->mddev->private;
317
318 conf->mirrors[disk].head_position =
319 r1_bio->sector + (r1_bio->sectors);
320}
321
322/*
323 * Find the disk number which triggered given bio
324 */
325static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
326{
327 int mirror;
328 struct r1conf *conf = r1_bio->mddev->private;
329 int raid_disks = conf->raid_disks;
330
331 for (mirror = 0; mirror < raid_disks * 2; mirror++)
332 if (r1_bio->bios[mirror] == bio)
333 break;
334
335 BUG_ON(mirror == raid_disks * 2);
336 update_head_pos(mirror, r1_bio);
337
338 return mirror;
339}
340
341static void raid1_end_read_request(struct bio *bio)
342{
343 int uptodate = !bio->bi_status;
344 struct r1bio *r1_bio = bio->bi_private;
345 struct r1conf *conf = r1_bio->mddev->private;
346 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
347
348 /*
349 * this branch is our 'one mirror IO has finished' event handler:
350 */
351 update_head_pos(r1_bio->read_disk, r1_bio);
352
353 if (uptodate)
354 set_bit(R1BIO_Uptodate, &r1_bio->state);
355 else if (test_bit(FailFast, &rdev->flags) &&
356 test_bit(R1BIO_FailFast, &r1_bio->state))
357 /* This was a fail-fast read so we definitely
358 * want to retry */
359 ;
360 else {
361 /* If all other devices have failed, we want to return
362 * the error upwards rather than fail the last device.
363 * Here we redefine "uptodate" to mean "Don't want to retry"
364 */
365 unsigned long flags;
366 spin_lock_irqsave(&conf->device_lock, flags);
367 if (r1_bio->mddev->degraded == conf->raid_disks ||
368 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
369 test_bit(In_sync, &rdev->flags)))
370 uptodate = 1;
371 spin_unlock_irqrestore(&conf->device_lock, flags);
372 }
373
374 if (uptodate) {
375 raid_end_bio_io(r1_bio);
376 rdev_dec_pending(rdev, conf->mddev);
377 } else {
378 /*
379 * oops, read error:
380 */
381 char b[BDEVNAME_SIZE];
382 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
383 mdname(conf->mddev),
384 bdevname(rdev->bdev, b),
385 (unsigned long long)r1_bio->sector);
386 set_bit(R1BIO_ReadError, &r1_bio->state);
387 reschedule_retry(r1_bio);
388 /* don't drop the reference on read_disk yet */
389 }
390}
391
392static void close_write(struct r1bio *r1_bio)
393{
394 /* it really is the end of this request */
395 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
396 bio_free_pages(r1_bio->behind_master_bio);
397 bio_put(r1_bio->behind_master_bio);
398 r1_bio->behind_master_bio = NULL;
399 }
400 /* clear the bitmap if all writes complete successfully */
401 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
402 r1_bio->sectors,
403 !test_bit(R1BIO_Degraded, &r1_bio->state),
404 test_bit(R1BIO_BehindIO, &r1_bio->state));
405 md_write_end(r1_bio->mddev);
406}
407
408static void r1_bio_write_done(struct r1bio *r1_bio)
409{
410 if (!atomic_dec_and_test(&r1_bio->remaining))
411 return;
412
413 if (test_bit(R1BIO_WriteError, &r1_bio->state))
414 reschedule_retry(r1_bio);
415 else {
416 close_write(r1_bio);
417 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
418 reschedule_retry(r1_bio);
419 else
420 raid_end_bio_io(r1_bio);
421 }
422}
423
424static void raid1_end_write_request(struct bio *bio)
425{
426 struct r1bio *r1_bio = bio->bi_private;
427 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
428 struct r1conf *conf = r1_bio->mddev->private;
429 struct bio *to_put = NULL;
430 int mirror = find_bio_disk(r1_bio, bio);
431 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
432 bool discard_error;
433
434 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
435
436 /*
437 * 'one mirror IO has finished' event handler:
438 */
439 if (bio->bi_status && !discard_error) {
440 set_bit(WriteErrorSeen, &rdev->flags);
441 if (!test_and_set_bit(WantReplacement, &rdev->flags))
442 set_bit(MD_RECOVERY_NEEDED, &
443 conf->mddev->recovery);
444
445 if (test_bit(FailFast, &rdev->flags) &&
446 (bio->bi_opf & MD_FAILFAST) &&
447 /* We never try FailFast to WriteMostly devices */
448 !test_bit(WriteMostly, &rdev->flags)) {
449 md_error(r1_bio->mddev, rdev);
450 }
451
452 /*
453 * When the device is faulty, it is not necessary to
454 * handle write error.
455 * For failfast, this is the only remaining device,
456 * We need to retry the write without FailFast.
457 */
458 if (!test_bit(Faulty, &rdev->flags))
459 set_bit(R1BIO_WriteError, &r1_bio->state);
460 else {
461 /* Finished with this branch */
462 r1_bio->bios[mirror] = NULL;
463 to_put = bio;
464 }
465 } else {
466 /*
467 * Set R1BIO_Uptodate in our master bio, so that we
468 * will return a good error code for to the higher
469 * levels even if IO on some other mirrored buffer
470 * fails.
471 *
472 * The 'master' represents the composite IO operation
473 * to user-side. So if something waits for IO, then it
474 * will wait for the 'master' bio.
475 */
476 sector_t first_bad;
477 int bad_sectors;
478
479 r1_bio->bios[mirror] = NULL;
480 to_put = bio;
481 /*
482 * Do not set R1BIO_Uptodate if the current device is
483 * rebuilding or Faulty. This is because we cannot use
484 * such device for properly reading the data back (we could
485 * potentially use it, if the current write would have felt
486 * before rdev->recovery_offset, but for simplicity we don't
487 * check this here.
488 */
489 if (test_bit(In_sync, &rdev->flags) &&
490 !test_bit(Faulty, &rdev->flags))
491 set_bit(R1BIO_Uptodate, &r1_bio->state);
492
493 /* Maybe we can clear some bad blocks. */
494 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
495 &first_bad, &bad_sectors) && !discard_error) {
496 r1_bio->bios[mirror] = IO_MADE_GOOD;
497 set_bit(R1BIO_MadeGood, &r1_bio->state);
498 }
499 }
500
501 if (behind) {
502 if (test_bit(WBCollisionCheck, &rdev->flags)) {
503 sector_t lo = r1_bio->sector;
504 sector_t hi = r1_bio->sector + r1_bio->sectors;
505
506 remove_wb(rdev, lo, hi);
507 }
508 if (test_bit(WriteMostly, &rdev->flags))
509 atomic_dec(&r1_bio->behind_remaining);
510
511 /*
512 * In behind mode, we ACK the master bio once the I/O
513 * has safely reached all non-writemostly
514 * disks. Setting the Returned bit ensures that this
515 * gets done only once -- we don't ever want to return
516 * -EIO here, instead we'll wait
517 */
518 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
519 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
520 /* Maybe we can return now */
521 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
522 struct bio *mbio = r1_bio->master_bio;
523 pr_debug("raid1: behind end write sectors"
524 " %llu-%llu\n",
525 (unsigned long long) mbio->bi_iter.bi_sector,
526 (unsigned long long) bio_end_sector(mbio) - 1);
527 call_bio_endio(r1_bio);
528 }
529 }
530 }
531 if (r1_bio->bios[mirror] == NULL)
532 rdev_dec_pending(rdev, conf->mddev);
533
534 /*
535 * Let's see if all mirrored write operations have finished
536 * already.
537 */
538 r1_bio_write_done(r1_bio);
539
540 if (to_put)
541 bio_put(to_put);
542}
543
544static sector_t align_to_barrier_unit_end(sector_t start_sector,
545 sector_t sectors)
546{
547 sector_t len;
548
549 WARN_ON(sectors == 0);
550 /*
551 * len is the number of sectors from start_sector to end of the
552 * barrier unit which start_sector belongs to.
553 */
554 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
555 start_sector;
556
557 if (len > sectors)
558 len = sectors;
559
560 return len;
561}
562
563/*
564 * This routine returns the disk from which the requested read should
565 * be done. There is a per-array 'next expected sequential IO' sector
566 * number - if this matches on the next IO then we use the last disk.
567 * There is also a per-disk 'last know head position' sector that is
568 * maintained from IRQ contexts, both the normal and the resync IO
569 * completion handlers update this position correctly. If there is no
570 * perfect sequential match then we pick the disk whose head is closest.
571 *
572 * If there are 2 mirrors in the same 2 devices, performance degrades
573 * because position is mirror, not device based.
574 *
575 * The rdev for the device selected will have nr_pending incremented.
576 */
577static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
578{
579 const sector_t this_sector = r1_bio->sector;
580 int sectors;
581 int best_good_sectors;
582 int best_disk, best_dist_disk, best_pending_disk;
583 int has_nonrot_disk;
584 int disk;
585 sector_t best_dist;
586 unsigned int min_pending;
587 struct md_rdev *rdev;
588 int choose_first;
589 int choose_next_idle;
590
591 rcu_read_lock();
592 /*
593 * Check if we can balance. We can balance on the whole
594 * device if no resync is going on, or below the resync window.
595 * We take the first readable disk when above the resync window.
596 */
597 retry:
598 sectors = r1_bio->sectors;
599 best_disk = -1;
600 best_dist_disk = -1;
601 best_dist = MaxSector;
602 best_pending_disk = -1;
603 min_pending = UINT_MAX;
604 best_good_sectors = 0;
605 has_nonrot_disk = 0;
606 choose_next_idle = 0;
607 clear_bit(R1BIO_FailFast, &r1_bio->state);
608
609 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
610 (mddev_is_clustered(conf->mddev) &&
611 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
612 this_sector + sectors)))
613 choose_first = 1;
614 else
615 choose_first = 0;
616
617 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
618 sector_t dist;
619 sector_t first_bad;
620 int bad_sectors;
621 unsigned int pending;
622 bool nonrot;
623
624 rdev = rcu_dereference(conf->mirrors[disk].rdev);
625 if (r1_bio->bios[disk] == IO_BLOCKED
626 || rdev == NULL
627 || test_bit(Faulty, &rdev->flags))
628 continue;
629 if (!test_bit(In_sync, &rdev->flags) &&
630 rdev->recovery_offset < this_sector + sectors)
631 continue;
632 if (test_bit(WriteMostly, &rdev->flags)) {
633 /* Don't balance among write-mostly, just
634 * use the first as a last resort */
635 if (best_dist_disk < 0) {
636 if (is_badblock(rdev, this_sector, sectors,
637 &first_bad, &bad_sectors)) {
638 if (first_bad <= this_sector)
639 /* Cannot use this */
640 continue;
641 best_good_sectors = first_bad - this_sector;
642 } else
643 best_good_sectors = sectors;
644 best_dist_disk = disk;
645 best_pending_disk = disk;
646 }
647 continue;
648 }
649 /* This is a reasonable device to use. It might
650 * even be best.
651 */
652 if (is_badblock(rdev, this_sector, sectors,
653 &first_bad, &bad_sectors)) {
654 if (best_dist < MaxSector)
655 /* already have a better device */
656 continue;
657 if (first_bad <= this_sector) {
658 /* cannot read here. If this is the 'primary'
659 * device, then we must not read beyond
660 * bad_sectors from another device..
661 */
662 bad_sectors -= (this_sector - first_bad);
663 if (choose_first && sectors > bad_sectors)
664 sectors = bad_sectors;
665 if (best_good_sectors > sectors)
666 best_good_sectors = sectors;
667
668 } else {
669 sector_t good_sectors = first_bad - this_sector;
670 if (good_sectors > best_good_sectors) {
671 best_good_sectors = good_sectors;
672 best_disk = disk;
673 }
674 if (choose_first)
675 break;
676 }
677 continue;
678 } else {
679 if ((sectors > best_good_sectors) && (best_disk >= 0))
680 best_disk = -1;
681 best_good_sectors = sectors;
682 }
683
684 if (best_disk >= 0)
685 /* At least two disks to choose from so failfast is OK */
686 set_bit(R1BIO_FailFast, &r1_bio->state);
687
688 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
689 has_nonrot_disk |= nonrot;
690 pending = atomic_read(&rdev->nr_pending);
691 dist = abs(this_sector - conf->mirrors[disk].head_position);
692 if (choose_first) {
693 best_disk = disk;
694 break;
695 }
696 /* Don't change to another disk for sequential reads */
697 if (conf->mirrors[disk].next_seq_sect == this_sector
698 || dist == 0) {
699 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
700 struct raid1_info *mirror = &conf->mirrors[disk];
701
702 best_disk = disk;
703 /*
704 * If buffered sequential IO size exceeds optimal
705 * iosize, check if there is idle disk. If yes, choose
706 * the idle disk. read_balance could already choose an
707 * idle disk before noticing it's a sequential IO in
708 * this disk. This doesn't matter because this disk
709 * will idle, next time it will be utilized after the
710 * first disk has IO size exceeds optimal iosize. In
711 * this way, iosize of the first disk will be optimal
712 * iosize at least. iosize of the second disk might be
713 * small, but not a big deal since when the second disk
714 * starts IO, the first disk is likely still busy.
715 */
716 if (nonrot && opt_iosize > 0 &&
717 mirror->seq_start != MaxSector &&
718 mirror->next_seq_sect > opt_iosize &&
719 mirror->next_seq_sect - opt_iosize >=
720 mirror->seq_start) {
721 choose_next_idle = 1;
722 continue;
723 }
724 break;
725 }
726
727 if (choose_next_idle)
728 continue;
729
730 if (min_pending > pending) {
731 min_pending = pending;
732 best_pending_disk = disk;
733 }
734
735 if (dist < best_dist) {
736 best_dist = dist;
737 best_dist_disk = disk;
738 }
739 }
740
741 /*
742 * If all disks are rotational, choose the closest disk. If any disk is
743 * non-rotational, choose the disk with less pending request even the
744 * disk is rotational, which might/might not be optimal for raids with
745 * mixed ratation/non-rotational disks depending on workload.
746 */
747 if (best_disk == -1) {
748 if (has_nonrot_disk || min_pending == 0)
749 best_disk = best_pending_disk;
750 else
751 best_disk = best_dist_disk;
752 }
753
754 if (best_disk >= 0) {
755 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
756 if (!rdev)
757 goto retry;
758 atomic_inc(&rdev->nr_pending);
759 sectors = best_good_sectors;
760
761 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
762 conf->mirrors[best_disk].seq_start = this_sector;
763
764 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
765 }
766 rcu_read_unlock();
767 *max_sectors = sectors;
768
769 return best_disk;
770}
771
772static int raid1_congested(struct mddev *mddev, int bits)
773{
774 struct r1conf *conf = mddev->private;
775 int i, ret = 0;
776
777 if ((bits & (1 << WB_async_congested)) &&
778 conf->pending_count >= max_queued_requests)
779 return 1;
780
781 rcu_read_lock();
782 for (i = 0; i < conf->raid_disks * 2; i++) {
783 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
784 if (rdev && !test_bit(Faulty, &rdev->flags)) {
785 struct request_queue *q = bdev_get_queue(rdev->bdev);
786
787 BUG_ON(!q);
788
789 /* Note the '|| 1' - when read_balance prefers
790 * non-congested targets, it can be removed
791 */
792 if ((bits & (1 << WB_async_congested)) || 1)
793 ret |= bdi_congested(q->backing_dev_info, bits);
794 else
795 ret &= bdi_congested(q->backing_dev_info, bits);
796 }
797 }
798 rcu_read_unlock();
799 return ret;
800}
801
802static void flush_bio_list(struct r1conf *conf, struct bio *bio)
803{
804 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
805 md_bitmap_unplug(conf->mddev->bitmap);
806 wake_up(&conf->wait_barrier);
807
808 while (bio) { /* submit pending writes */
809 struct bio *next = bio->bi_next;
810 struct md_rdev *rdev = (void *)bio->bi_disk;
811 bio->bi_next = NULL;
812 bio_set_dev(bio, rdev->bdev);
813 if (test_bit(Faulty, &rdev->flags)) {
814 bio_io_error(bio);
815 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
816 !blk_queue_discard(bio->bi_disk->queue)))
817 /* Just ignore it */
818 bio_endio(bio);
819 else
820 generic_make_request(bio);
821 bio = next;
822 }
823}
824
825static void flush_pending_writes(struct r1conf *conf)
826{
827 /* Any writes that have been queued but are awaiting
828 * bitmap updates get flushed here.
829 */
830 spin_lock_irq(&conf->device_lock);
831
832 if (conf->pending_bio_list.head) {
833 struct blk_plug plug;
834 struct bio *bio;
835
836 bio = bio_list_get(&conf->pending_bio_list);
837 conf->pending_count = 0;
838 spin_unlock_irq(&conf->device_lock);
839
840 /*
841 * As this is called in a wait_event() loop (see freeze_array),
842 * current->state might be TASK_UNINTERRUPTIBLE which will
843 * cause a warning when we prepare to wait again. As it is
844 * rare that this path is taken, it is perfectly safe to force
845 * us to go around the wait_event() loop again, so the warning
846 * is a false-positive. Silence the warning by resetting
847 * thread state
848 */
849 __set_current_state(TASK_RUNNING);
850 blk_start_plug(&plug);
851 flush_bio_list(conf, bio);
852 blk_finish_plug(&plug);
853 } else
854 spin_unlock_irq(&conf->device_lock);
855}
856
857/* Barriers....
858 * Sometimes we need to suspend IO while we do something else,
859 * either some resync/recovery, or reconfigure the array.
860 * To do this we raise a 'barrier'.
861 * The 'barrier' is a counter that can be raised multiple times
862 * to count how many activities are happening which preclude
863 * normal IO.
864 * We can only raise the barrier if there is no pending IO.
865 * i.e. if nr_pending == 0.
866 * We choose only to raise the barrier if no-one is waiting for the
867 * barrier to go down. This means that as soon as an IO request
868 * is ready, no other operations which require a barrier will start
869 * until the IO request has had a chance.
870 *
871 * So: regular IO calls 'wait_barrier'. When that returns there
872 * is no backgroup IO happening, It must arrange to call
873 * allow_barrier when it has finished its IO.
874 * backgroup IO calls must call raise_barrier. Once that returns
875 * there is no normal IO happeing. It must arrange to call
876 * lower_barrier when the particular background IO completes.
877 *
878 * If resync/recovery is interrupted, returns -EINTR;
879 * Otherwise, returns 0.
880 */
881static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
882{
883 int idx = sector_to_idx(sector_nr);
884
885 spin_lock_irq(&conf->resync_lock);
886
887 /* Wait until no block IO is waiting */
888 wait_event_lock_irq(conf->wait_barrier,
889 !atomic_read(&conf->nr_waiting[idx]),
890 conf->resync_lock);
891
892 /* block any new IO from starting */
893 atomic_inc(&conf->barrier[idx]);
894 /*
895 * In raise_barrier() we firstly increase conf->barrier[idx] then
896 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
897 * increase conf->nr_pending[idx] then check conf->barrier[idx].
898 * A memory barrier here to make sure conf->nr_pending[idx] won't
899 * be fetched before conf->barrier[idx] is increased. Otherwise
900 * there will be a race between raise_barrier() and _wait_barrier().
901 */
902 smp_mb__after_atomic();
903
904 /* For these conditions we must wait:
905 * A: while the array is in frozen state
906 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
907 * existing in corresponding I/O barrier bucket.
908 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
909 * max resync count which allowed on current I/O barrier bucket.
910 */
911 wait_event_lock_irq(conf->wait_barrier,
912 (!conf->array_frozen &&
913 !atomic_read(&conf->nr_pending[idx]) &&
914 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
915 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
916 conf->resync_lock);
917
918 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
919 atomic_dec(&conf->barrier[idx]);
920 spin_unlock_irq(&conf->resync_lock);
921 wake_up(&conf->wait_barrier);
922 return -EINTR;
923 }
924
925 atomic_inc(&conf->nr_sync_pending);
926 spin_unlock_irq(&conf->resync_lock);
927
928 return 0;
929}
930
931static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
932{
933 int idx = sector_to_idx(sector_nr);
934
935 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
936
937 atomic_dec(&conf->barrier[idx]);
938 atomic_dec(&conf->nr_sync_pending);
939 wake_up(&conf->wait_barrier);
940}
941
942static void _wait_barrier(struct r1conf *conf, int idx)
943{
944 /*
945 * We need to increase conf->nr_pending[idx] very early here,
946 * then raise_barrier() can be blocked when it waits for
947 * conf->nr_pending[idx] to be 0. Then we can avoid holding
948 * conf->resync_lock when there is no barrier raised in same
949 * barrier unit bucket. Also if the array is frozen, I/O
950 * should be blocked until array is unfrozen.
951 */
952 atomic_inc(&conf->nr_pending[idx]);
953 /*
954 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
955 * check conf->barrier[idx]. In raise_barrier() we firstly increase
956 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
957 * barrier is necessary here to make sure conf->barrier[idx] won't be
958 * fetched before conf->nr_pending[idx] is increased. Otherwise there
959 * will be a race between _wait_barrier() and raise_barrier().
960 */
961 smp_mb__after_atomic();
962
963 /*
964 * Don't worry about checking two atomic_t variables at same time
965 * here. If during we check conf->barrier[idx], the array is
966 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
967 * 0, it is safe to return and make the I/O continue. Because the
968 * array is frozen, all I/O returned here will eventually complete
969 * or be queued, no race will happen. See code comment in
970 * frozen_array().
971 */
972 if (!READ_ONCE(conf->array_frozen) &&
973 !atomic_read(&conf->barrier[idx]))
974 return;
975
976 /*
977 * After holding conf->resync_lock, conf->nr_pending[idx]
978 * should be decreased before waiting for barrier to drop.
979 * Otherwise, we may encounter a race condition because
980 * raise_barrer() might be waiting for conf->nr_pending[idx]
981 * to be 0 at same time.
982 */
983 spin_lock_irq(&conf->resync_lock);
984 atomic_inc(&conf->nr_waiting[idx]);
985 atomic_dec(&conf->nr_pending[idx]);
986 /*
987 * In case freeze_array() is waiting for
988 * get_unqueued_pending() == extra
989 */
990 wake_up(&conf->wait_barrier);
991 /* Wait for the barrier in same barrier unit bucket to drop. */
992 wait_event_lock_irq(conf->wait_barrier,
993 !conf->array_frozen &&
994 !atomic_read(&conf->barrier[idx]),
995 conf->resync_lock);
996 atomic_inc(&conf->nr_pending[idx]);
997 atomic_dec(&conf->nr_waiting[idx]);
998 spin_unlock_irq(&conf->resync_lock);
999}
1000
1001static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
1002{
1003 int idx = sector_to_idx(sector_nr);
1004
1005 /*
1006 * Very similar to _wait_barrier(). The difference is, for read
1007 * I/O we don't need wait for sync I/O, but if the whole array
1008 * is frozen, the read I/O still has to wait until the array is
1009 * unfrozen. Since there is no ordering requirement with
1010 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1011 */
1012 atomic_inc(&conf->nr_pending[idx]);
1013
1014 if (!READ_ONCE(conf->array_frozen))
1015 return;
1016
1017 spin_lock_irq(&conf->resync_lock);
1018 atomic_inc(&conf->nr_waiting[idx]);
1019 atomic_dec(&conf->nr_pending[idx]);
1020 /*
1021 * In case freeze_array() is waiting for
1022 * get_unqueued_pending() == extra
1023 */
1024 wake_up(&conf->wait_barrier);
1025 /* Wait for array to be unfrozen */
1026 wait_event_lock_irq(conf->wait_barrier,
1027 !conf->array_frozen,
1028 conf->resync_lock);
1029 atomic_inc(&conf->nr_pending[idx]);
1030 atomic_dec(&conf->nr_waiting[idx]);
1031 spin_unlock_irq(&conf->resync_lock);
1032}
1033
1034static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1035{
1036 int idx = sector_to_idx(sector_nr);
1037
1038 _wait_barrier(conf, idx);
1039}
1040
1041static void _allow_barrier(struct r1conf *conf, int idx)
1042{
1043 atomic_dec(&conf->nr_pending[idx]);
1044 wake_up(&conf->wait_barrier);
1045}
1046
1047static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1048{
1049 int idx = sector_to_idx(sector_nr);
1050
1051 _allow_barrier(conf, idx);
1052}
1053
1054/* conf->resync_lock should be held */
1055static int get_unqueued_pending(struct r1conf *conf)
1056{
1057 int idx, ret;
1058
1059 ret = atomic_read(&conf->nr_sync_pending);
1060 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1061 ret += atomic_read(&conf->nr_pending[idx]) -
1062 atomic_read(&conf->nr_queued[idx]);
1063
1064 return ret;
1065}
1066
1067static void freeze_array(struct r1conf *conf, int extra)
1068{
1069 /* Stop sync I/O and normal I/O and wait for everything to
1070 * go quiet.
1071 * This is called in two situations:
1072 * 1) management command handlers (reshape, remove disk, quiesce).
1073 * 2) one normal I/O request failed.
1074
1075 * After array_frozen is set to 1, new sync IO will be blocked at
1076 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1077 * or wait_read_barrier(). The flying I/Os will either complete or be
1078 * queued. When everything goes quite, there are only queued I/Os left.
1079
1080 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1081 * barrier bucket index which this I/O request hits. When all sync and
1082 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1083 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1084 * in handle_read_error(), we may call freeze_array() before trying to
1085 * fix the read error. In this case, the error read I/O is not queued,
1086 * so get_unqueued_pending() == 1.
1087 *
1088 * Therefore before this function returns, we need to wait until
1089 * get_unqueued_pendings(conf) gets equal to extra. For
1090 * normal I/O context, extra is 1, in rested situations extra is 0.
1091 */
1092 spin_lock_irq(&conf->resync_lock);
1093 conf->array_frozen = 1;
1094 raid1_log(conf->mddev, "wait freeze");
1095 wait_event_lock_irq_cmd(
1096 conf->wait_barrier,
1097 get_unqueued_pending(conf) == extra,
1098 conf->resync_lock,
1099 flush_pending_writes(conf));
1100 spin_unlock_irq(&conf->resync_lock);
1101}
1102static void unfreeze_array(struct r1conf *conf)
1103{
1104 /* reverse the effect of the freeze */
1105 spin_lock_irq(&conf->resync_lock);
1106 conf->array_frozen = 0;
1107 spin_unlock_irq(&conf->resync_lock);
1108 wake_up(&conf->wait_barrier);
1109}
1110
1111static void alloc_behind_master_bio(struct r1bio *r1_bio,
1112 struct bio *bio)
1113{
1114 int size = bio->bi_iter.bi_size;
1115 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1116 int i = 0;
1117 struct bio *behind_bio = NULL;
1118
1119 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1120 if (!behind_bio)
1121 return;
1122
1123 /* discard op, we don't support writezero/writesame yet */
1124 if (!bio_has_data(bio)) {
1125 behind_bio->bi_iter.bi_size = size;
1126 goto skip_copy;
1127 }
1128
1129 behind_bio->bi_write_hint = bio->bi_write_hint;
1130
1131 while (i < vcnt && size) {
1132 struct page *page;
1133 int len = min_t(int, PAGE_SIZE, size);
1134
1135 page = alloc_page(GFP_NOIO);
1136 if (unlikely(!page))
1137 goto free_pages;
1138
1139 bio_add_page(behind_bio, page, len, 0);
1140
1141 size -= len;
1142 i++;
1143 }
1144
1145 bio_copy_data(behind_bio, bio);
1146skip_copy:
1147 r1_bio->behind_master_bio = behind_bio;
1148 set_bit(R1BIO_BehindIO, &r1_bio->state);
1149
1150 return;
1151
1152free_pages:
1153 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1154 bio->bi_iter.bi_size);
1155 bio_free_pages(behind_bio);
1156 bio_put(behind_bio);
1157}
1158
1159struct raid1_plug_cb {
1160 struct blk_plug_cb cb;
1161 struct bio_list pending;
1162 int pending_cnt;
1163};
1164
1165static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1166{
1167 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1168 cb);
1169 struct mddev *mddev = plug->cb.data;
1170 struct r1conf *conf = mddev->private;
1171 struct bio *bio;
1172
1173 if (from_schedule || current->bio_list) {
1174 spin_lock_irq(&conf->device_lock);
1175 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1176 conf->pending_count += plug->pending_cnt;
1177 spin_unlock_irq(&conf->device_lock);
1178 wake_up(&conf->wait_barrier);
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 int op = bio_op(bio);
1220 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1221 int max_sectors;
1222 int rdisk;
1223 bool print_msg = !!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 (print_msg) {
1234 /* Need to get the block device name carefully */
1235 struct md_rdev *rdev;
1236 rcu_read_lock();
1237 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1238 if (rdev)
1239 bdevname(rdev->bdev, b);
1240 else
1241 strcpy(b, "???");
1242 rcu_read_unlock();
1243 }
1244
1245 /*
1246 * Still need barrier for READ in case that whole
1247 * array is frozen.
1248 */
1249 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1250
1251 if (!r1_bio)
1252 r1_bio = alloc_r1bio(mddev, bio);
1253 else
1254 init_r1bio(r1_bio, mddev, bio);
1255 r1_bio->sectors = max_read_sectors;
1256
1257 /*
1258 * make_request() can abort the operation when read-ahead is being
1259 * used and no empty request is available.
1260 */
1261 rdisk = read_balance(conf, r1_bio, &max_sectors);
1262
1263 if (rdisk < 0) {
1264 /* couldn't find anywhere to read from */
1265 if (print_msg) {
1266 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1267 mdname(mddev),
1268 b,
1269 (unsigned long long)r1_bio->sector);
1270 }
1271 raid_end_bio_io(r1_bio);
1272 return;
1273 }
1274 mirror = conf->mirrors + rdisk;
1275
1276 if (print_msg)
1277 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1278 mdname(mddev),
1279 (unsigned long long)r1_bio->sector,
1280 bdevname(mirror->rdev->bdev, b));
1281
1282 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1283 bitmap) {
1284 /*
1285 * Reading from a write-mostly device must take care not to
1286 * over-take any writes that are 'behind'
1287 */
1288 raid1_log(mddev, "wait behind writes");
1289 wait_event(bitmap->behind_wait,
1290 atomic_read(&bitmap->behind_writes) == 0);
1291 }
1292
1293 if (max_sectors < bio_sectors(bio)) {
1294 struct bio *split = bio_split(bio, max_sectors,
1295 gfp, &conf->bio_split);
1296 bio_chain(split, bio);
1297 generic_make_request(bio);
1298 bio = split;
1299 r1_bio->master_bio = bio;
1300 r1_bio->sectors = max_sectors;
1301 }
1302
1303 r1_bio->read_disk = rdisk;
1304
1305 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1306
1307 r1_bio->bios[rdisk] = read_bio;
1308
1309 read_bio->bi_iter.bi_sector = r1_bio->sector +
1310 mirror->rdev->data_offset;
1311 bio_set_dev(read_bio, mirror->rdev->bdev);
1312 read_bio->bi_end_io = raid1_end_read_request;
1313 bio_set_op_attrs(read_bio, op, do_sync);
1314 if (test_bit(FailFast, &mirror->rdev->flags) &&
1315 test_bit(R1BIO_FailFast, &r1_bio->state))
1316 read_bio->bi_opf |= MD_FAILFAST;
1317 read_bio->bi_private = r1_bio;
1318
1319 if (mddev->gendisk)
1320 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1321 disk_devt(mddev->gendisk), r1_bio->sector);
1322
1323 generic_make_request(read_bio);
1324}
1325
1326static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1327 int max_write_sectors)
1328{
1329 struct r1conf *conf = mddev->private;
1330 struct r1bio *r1_bio;
1331 int i, disks;
1332 struct bitmap *bitmap = mddev->bitmap;
1333 unsigned long flags;
1334 struct md_rdev *blocked_rdev;
1335 struct blk_plug_cb *cb;
1336 struct raid1_plug_cb *plug = NULL;
1337 int first_clone;
1338 int max_sectors;
1339
1340 if (mddev_is_clustered(mddev) &&
1341 md_cluster_ops->area_resyncing(mddev, WRITE,
1342 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1343
1344 DEFINE_WAIT(w);
1345 for (;;) {
1346 prepare_to_wait(&conf->wait_barrier,
1347 &w, TASK_IDLE);
1348 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1349 bio->bi_iter.bi_sector,
1350 bio_end_sector(bio)))
1351 break;
1352 schedule();
1353 }
1354 finish_wait(&conf->wait_barrier, &w);
1355 }
1356
1357 /*
1358 * Register the new request and wait if the reconstruction
1359 * thread has put up a bar for new requests.
1360 * Continue immediately if no resync is active currently.
1361 */
1362 wait_barrier(conf, bio->bi_iter.bi_sector);
1363
1364 r1_bio = alloc_r1bio(mddev, bio);
1365 r1_bio->sectors = max_write_sectors;
1366
1367 if (conf->pending_count >= max_queued_requests) {
1368 md_wakeup_thread(mddev->thread);
1369 raid1_log(mddev, "wait queued");
1370 wait_event(conf->wait_barrier,
1371 conf->pending_count < max_queued_requests);
1372 }
1373 /* first select target devices under rcu_lock and
1374 * inc refcount on their rdev. Record them by setting
1375 * bios[x] to bio
1376 * If there are known/acknowledged bad blocks on any device on
1377 * which we have seen a write error, we want to avoid writing those
1378 * blocks.
1379 * This potentially requires several writes to write around
1380 * the bad blocks. Each set of writes gets it's own r1bio
1381 * with a set of bios attached.
1382 */
1383
1384 disks = conf->raid_disks * 2;
1385 retry_write:
1386 blocked_rdev = NULL;
1387 rcu_read_lock();
1388 max_sectors = r1_bio->sectors;
1389 for (i = 0; i < disks; i++) {
1390 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1391 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1392 atomic_inc(&rdev->nr_pending);
1393 blocked_rdev = rdev;
1394 break;
1395 }
1396 r1_bio->bios[i] = NULL;
1397 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1398 if (i < conf->raid_disks)
1399 set_bit(R1BIO_Degraded, &r1_bio->state);
1400 continue;
1401 }
1402
1403 atomic_inc(&rdev->nr_pending);
1404 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1405 sector_t first_bad;
1406 int bad_sectors;
1407 int is_bad;
1408
1409 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1410 &first_bad, &bad_sectors);
1411 if (is_bad < 0) {
1412 /* mustn't write here until the bad block is
1413 * acknowledged*/
1414 set_bit(BlockedBadBlocks, &rdev->flags);
1415 blocked_rdev = rdev;
1416 break;
1417 }
1418 if (is_bad && first_bad <= r1_bio->sector) {
1419 /* Cannot write here at all */
1420 bad_sectors -= (r1_bio->sector - first_bad);
1421 if (bad_sectors < max_sectors)
1422 /* mustn't write more than bad_sectors
1423 * to other devices yet
1424 */
1425 max_sectors = bad_sectors;
1426 rdev_dec_pending(rdev, mddev);
1427 /* We don't set R1BIO_Degraded as that
1428 * only applies if the disk is
1429 * missing, so it might be re-added,
1430 * and we want to know to recover this
1431 * chunk.
1432 * In this case the device is here,
1433 * and the fact that this chunk is not
1434 * in-sync is recorded in the bad
1435 * block log
1436 */
1437 continue;
1438 }
1439 if (is_bad) {
1440 int good_sectors = first_bad - r1_bio->sector;
1441 if (good_sectors < max_sectors)
1442 max_sectors = good_sectors;
1443 }
1444 }
1445 r1_bio->bios[i] = bio;
1446 }
1447 rcu_read_unlock();
1448
1449 if (unlikely(blocked_rdev)) {
1450 /* Wait for this device to become unblocked */
1451 int j;
1452
1453 for (j = 0; j < i; j++)
1454 if (r1_bio->bios[j])
1455 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1456 r1_bio->state = 0;
1457 allow_barrier(conf, bio->bi_iter.bi_sector);
1458 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1459 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1460 wait_barrier(conf, bio->bi_iter.bi_sector);
1461 goto retry_write;
1462 }
1463
1464 if (max_sectors < bio_sectors(bio)) {
1465 struct bio *split = bio_split(bio, max_sectors,
1466 GFP_NOIO, &conf->bio_split);
1467 bio_chain(split, bio);
1468 generic_make_request(bio);
1469 bio = split;
1470 r1_bio->master_bio = bio;
1471 r1_bio->sectors = max_sectors;
1472 }
1473
1474 atomic_set(&r1_bio->remaining, 1);
1475 atomic_set(&r1_bio->behind_remaining, 0);
1476
1477 first_clone = 1;
1478
1479 for (i = 0; i < disks; i++) {
1480 struct bio *mbio = NULL;
1481 if (!r1_bio->bios[i])
1482 continue;
1483
1484 if (first_clone) {
1485 /* do behind I/O ?
1486 * Not if there are too many, or cannot
1487 * allocate memory, or a reader on WriteMostly
1488 * is waiting for behind writes to flush */
1489 if (bitmap &&
1490 (atomic_read(&bitmap->behind_writes)
1491 < mddev->bitmap_info.max_write_behind) &&
1492 !waitqueue_active(&bitmap->behind_wait)) {
1493 alloc_behind_master_bio(r1_bio, bio);
1494 }
1495
1496 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1497 test_bit(R1BIO_BehindIO, &r1_bio->state));
1498 first_clone = 0;
1499 }
1500
1501 if (r1_bio->behind_master_bio)
1502 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1503 GFP_NOIO, &mddev->bio_set);
1504 else
1505 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1506
1507 if (r1_bio->behind_master_bio) {
1508 struct md_rdev *rdev = conf->mirrors[i].rdev;
1509
1510 if (test_bit(WBCollisionCheck, &rdev->flags)) {
1511 sector_t lo = r1_bio->sector;
1512 sector_t hi = r1_bio->sector + r1_bio->sectors;
1513
1514 wait_event(rdev->wb_io_wait,
1515 check_and_add_wb(rdev, lo, hi) == 0);
1516 }
1517 if (test_bit(WriteMostly, &rdev->flags))
1518 atomic_inc(&r1_bio->behind_remaining);
1519 }
1520
1521 r1_bio->bios[i] = mbio;
1522
1523 mbio->bi_iter.bi_sector = (r1_bio->sector +
1524 conf->mirrors[i].rdev->data_offset);
1525 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1526 mbio->bi_end_io = raid1_end_write_request;
1527 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1528 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1529 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1530 conf->raid_disks - mddev->degraded > 1)
1531 mbio->bi_opf |= MD_FAILFAST;
1532 mbio->bi_private = r1_bio;
1533
1534 atomic_inc(&r1_bio->remaining);
1535
1536 if (mddev->gendisk)
1537 trace_block_bio_remap(mbio->bi_disk->queue,
1538 mbio, disk_devt(mddev->gendisk),
1539 r1_bio->sector);
1540 /* flush_pending_writes() needs access to the rdev so...*/
1541 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1542
1543 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1544 if (cb)
1545 plug = container_of(cb, struct raid1_plug_cb, cb);
1546 else
1547 plug = NULL;
1548 if (plug) {
1549 bio_list_add(&plug->pending, mbio);
1550 plug->pending_cnt++;
1551 } else {
1552 spin_lock_irqsave(&conf->device_lock, flags);
1553 bio_list_add(&conf->pending_bio_list, mbio);
1554 conf->pending_count++;
1555 spin_unlock_irqrestore(&conf->device_lock, flags);
1556 md_wakeup_thread(mddev->thread);
1557 }
1558 }
1559
1560 r1_bio_write_done(r1_bio);
1561
1562 /* In case raid1d snuck in to freeze_array */
1563 wake_up(&conf->wait_barrier);
1564}
1565
1566static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1567{
1568 sector_t sectors;
1569
1570 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1571 md_flush_request(mddev, bio);
1572 return true;
1573 }
1574
1575 /*
1576 * There is a limit to the maximum size, but
1577 * the read/write handler might find a lower limit
1578 * due to bad blocks. To avoid multiple splits,
1579 * we pass the maximum number of sectors down
1580 * and let the lower level perform the split.
1581 */
1582 sectors = align_to_barrier_unit_end(
1583 bio->bi_iter.bi_sector, bio_sectors(bio));
1584
1585 if (bio_data_dir(bio) == READ)
1586 raid1_read_request(mddev, bio, sectors, NULL);
1587 else {
1588 if (!md_write_start(mddev,bio))
1589 return false;
1590 raid1_write_request(mddev, bio, sectors);
1591 }
1592 return true;
1593}
1594
1595static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1596{
1597 struct r1conf *conf = mddev->private;
1598 int i;
1599
1600 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1601 conf->raid_disks - mddev->degraded);
1602 rcu_read_lock();
1603 for (i = 0; i < conf->raid_disks; i++) {
1604 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1605 seq_printf(seq, "%s",
1606 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1607 }
1608 rcu_read_unlock();
1609 seq_printf(seq, "]");
1610}
1611
1612static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1613{
1614 char b[BDEVNAME_SIZE];
1615 struct r1conf *conf = mddev->private;
1616 unsigned long flags;
1617
1618 /*
1619 * If it is not operational, then we have already marked it as dead
1620 * else if it is the last working disks with "fail_last_dev == false",
1621 * ignore the error, let the next level up know.
1622 * else mark the drive as failed
1623 */
1624 spin_lock_irqsave(&conf->device_lock, flags);
1625 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1626 && (conf->raid_disks - mddev->degraded) == 1) {
1627 /*
1628 * Don't fail the drive, act as though we were just a
1629 * normal single drive.
1630 * However don't try a recovery from this drive as
1631 * it is very likely to fail.
1632 */
1633 conf->recovery_disabled = mddev->recovery_disabled;
1634 spin_unlock_irqrestore(&conf->device_lock, flags);
1635 return;
1636 }
1637 set_bit(Blocked, &rdev->flags);
1638 if (test_and_clear_bit(In_sync, &rdev->flags))
1639 mddev->degraded++;
1640 set_bit(Faulty, &rdev->flags);
1641 spin_unlock_irqrestore(&conf->device_lock, flags);
1642 /*
1643 * if recovery is running, make sure it aborts.
1644 */
1645 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1646 set_mask_bits(&mddev->sb_flags, 0,
1647 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1648 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1649 "md/raid1:%s: Operation continuing on %d devices.\n",
1650 mdname(mddev), bdevname(rdev->bdev, b),
1651 mdname(mddev), conf->raid_disks - mddev->degraded);
1652}
1653
1654static void print_conf(struct r1conf *conf)
1655{
1656 int i;
1657
1658 pr_debug("RAID1 conf printout:\n");
1659 if (!conf) {
1660 pr_debug("(!conf)\n");
1661 return;
1662 }
1663 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1664 conf->raid_disks);
1665
1666 rcu_read_lock();
1667 for (i = 0; i < conf->raid_disks; i++) {
1668 char b[BDEVNAME_SIZE];
1669 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1670 if (rdev)
1671 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1672 i, !test_bit(In_sync, &rdev->flags),
1673 !test_bit(Faulty, &rdev->flags),
1674 bdevname(rdev->bdev,b));
1675 }
1676 rcu_read_unlock();
1677}
1678
1679static void close_sync(struct r1conf *conf)
1680{
1681 int idx;
1682
1683 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1684 _wait_barrier(conf, idx);
1685 _allow_barrier(conf, idx);
1686 }
1687
1688 mempool_exit(&conf->r1buf_pool);
1689}
1690
1691static int raid1_spare_active(struct mddev *mddev)
1692{
1693 int i;
1694 struct r1conf *conf = mddev->private;
1695 int count = 0;
1696 unsigned long flags;
1697
1698 /*
1699 * Find all failed disks within the RAID1 configuration
1700 * and mark them readable.
1701 * Called under mddev lock, so rcu protection not needed.
1702 * device_lock used to avoid races with raid1_end_read_request
1703 * which expects 'In_sync' flags and ->degraded to be consistent.
1704 */
1705 spin_lock_irqsave(&conf->device_lock, flags);
1706 for (i = 0; i < conf->raid_disks; i++) {
1707 struct md_rdev *rdev = conf->mirrors[i].rdev;
1708 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1709 if (repl
1710 && !test_bit(Candidate, &repl->flags)
1711 && repl->recovery_offset == MaxSector
1712 && !test_bit(Faulty, &repl->flags)
1713 && !test_and_set_bit(In_sync, &repl->flags)) {
1714 /* replacement has just become active */
1715 if (!rdev ||
1716 !test_and_clear_bit(In_sync, &rdev->flags))
1717 count++;
1718 if (rdev) {
1719 /* Replaced device not technically
1720 * faulty, but we need to be sure
1721 * it gets removed and never re-added
1722 */
1723 set_bit(Faulty, &rdev->flags);
1724 sysfs_notify_dirent_safe(
1725 rdev->sysfs_state);
1726 }
1727 }
1728 if (rdev
1729 && rdev->recovery_offset == MaxSector
1730 && !test_bit(Faulty, &rdev->flags)
1731 && !test_and_set_bit(In_sync, &rdev->flags)) {
1732 count++;
1733 sysfs_notify_dirent_safe(rdev->sysfs_state);
1734 }
1735 }
1736 mddev->degraded -= count;
1737 spin_unlock_irqrestore(&conf->device_lock, flags);
1738
1739 print_conf(conf);
1740 return count;
1741}
1742
1743static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1744{
1745 struct r1conf *conf = mddev->private;
1746 int err = -EEXIST;
1747 int mirror = 0;
1748 struct raid1_info *p;
1749 int first = 0;
1750 int last = conf->raid_disks - 1;
1751
1752 if (mddev->recovery_disabled == conf->recovery_disabled)
1753 return -EBUSY;
1754
1755 if (md_integrity_add_rdev(rdev, mddev))
1756 return -ENXIO;
1757
1758 if (rdev->raid_disk >= 0)
1759 first = last = rdev->raid_disk;
1760
1761 /*
1762 * find the disk ... but prefer rdev->saved_raid_disk
1763 * if possible.
1764 */
1765 if (rdev->saved_raid_disk >= 0 &&
1766 rdev->saved_raid_disk >= first &&
1767 rdev->saved_raid_disk < conf->raid_disks &&
1768 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1769 first = last = rdev->saved_raid_disk;
1770
1771 for (mirror = first; mirror <= last; mirror++) {
1772 p = conf->mirrors + mirror;
1773 if (!p->rdev) {
1774 if (mddev->gendisk)
1775 disk_stack_limits(mddev->gendisk, rdev->bdev,
1776 rdev->data_offset << 9);
1777
1778 p->head_position = 0;
1779 rdev->raid_disk = mirror;
1780 err = 0;
1781 /* As all devices are equivalent, we don't need a full recovery
1782 * if this was recently any drive of the array
1783 */
1784 if (rdev->saved_raid_disk < 0)
1785 conf->fullsync = 1;
1786 rcu_assign_pointer(p->rdev, rdev);
1787 break;
1788 }
1789 if (test_bit(WantReplacement, &p->rdev->flags) &&
1790 p[conf->raid_disks].rdev == NULL) {
1791 /* Add this device as a replacement */
1792 clear_bit(In_sync, &rdev->flags);
1793 set_bit(Replacement, &rdev->flags);
1794 rdev->raid_disk = mirror;
1795 err = 0;
1796 conf->fullsync = 1;
1797 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1798 break;
1799 }
1800 }
1801 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1802 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1803 print_conf(conf);
1804 return err;
1805}
1806
1807static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1808{
1809 struct r1conf *conf = mddev->private;
1810 int err = 0;
1811 int number = rdev->raid_disk;
1812 struct raid1_info *p = conf->mirrors + number;
1813
1814 if (rdev != p->rdev)
1815 p = conf->mirrors + conf->raid_disks + number;
1816
1817 print_conf(conf);
1818 if (rdev == p->rdev) {
1819 if (test_bit(In_sync, &rdev->flags) ||
1820 atomic_read(&rdev->nr_pending)) {
1821 err = -EBUSY;
1822 goto abort;
1823 }
1824 /* Only remove non-faulty devices if recovery
1825 * is not possible.
1826 */
1827 if (!test_bit(Faulty, &rdev->flags) &&
1828 mddev->recovery_disabled != conf->recovery_disabled &&
1829 mddev->degraded < conf->raid_disks) {
1830 err = -EBUSY;
1831 goto abort;
1832 }
1833 p->rdev = NULL;
1834 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1835 synchronize_rcu();
1836 if (atomic_read(&rdev->nr_pending)) {
1837 /* lost the race, try later */
1838 err = -EBUSY;
1839 p->rdev = rdev;
1840 goto abort;
1841 }
1842 }
1843 if (conf->mirrors[conf->raid_disks + number].rdev) {
1844 /* We just removed a device that is being replaced.
1845 * Move down the replacement. We drain all IO before
1846 * doing this to avoid confusion.
1847 */
1848 struct md_rdev *repl =
1849 conf->mirrors[conf->raid_disks + number].rdev;
1850 freeze_array(conf, 0);
1851 if (atomic_read(&repl->nr_pending)) {
1852 /* It means that some queued IO of retry_list
1853 * hold repl. Thus, we cannot set replacement
1854 * as NULL, avoiding rdev NULL pointer
1855 * dereference in sync_request_write and
1856 * handle_write_finished.
1857 */
1858 err = -EBUSY;
1859 unfreeze_array(conf);
1860 goto abort;
1861 }
1862 clear_bit(Replacement, &repl->flags);
1863 p->rdev = repl;
1864 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1865 unfreeze_array(conf);
1866 }
1867
1868 clear_bit(WantReplacement, &rdev->flags);
1869 err = md_integrity_register(mddev);
1870 }
1871abort:
1872
1873 print_conf(conf);
1874 return err;
1875}
1876
1877static void end_sync_read(struct bio *bio)
1878{
1879 struct r1bio *r1_bio = get_resync_r1bio(bio);
1880
1881 update_head_pos(r1_bio->read_disk, r1_bio);
1882
1883 /*
1884 * we have read a block, now it needs to be re-written,
1885 * or re-read if the read failed.
1886 * We don't do much here, just schedule handling by raid1d
1887 */
1888 if (!bio->bi_status)
1889 set_bit(R1BIO_Uptodate, &r1_bio->state);
1890
1891 if (atomic_dec_and_test(&r1_bio->remaining))
1892 reschedule_retry(r1_bio);
1893}
1894
1895static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1896{
1897 sector_t sync_blocks = 0;
1898 sector_t s = r1_bio->sector;
1899 long sectors_to_go = r1_bio->sectors;
1900
1901 /* make sure these bits don't get cleared. */
1902 do {
1903 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1904 s += sync_blocks;
1905 sectors_to_go -= sync_blocks;
1906 } while (sectors_to_go > 0);
1907}
1908
1909static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1910{
1911 if (atomic_dec_and_test(&r1_bio->remaining)) {
1912 struct mddev *mddev = r1_bio->mddev;
1913 int s = r1_bio->sectors;
1914
1915 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1916 test_bit(R1BIO_WriteError, &r1_bio->state))
1917 reschedule_retry(r1_bio);
1918 else {
1919 put_buf(r1_bio);
1920 md_done_sync(mddev, s, uptodate);
1921 }
1922 }
1923}
1924
1925static void end_sync_write(struct bio *bio)
1926{
1927 int uptodate = !bio->bi_status;
1928 struct r1bio *r1_bio = get_resync_r1bio(bio);
1929 struct mddev *mddev = r1_bio->mddev;
1930 struct r1conf *conf = mddev->private;
1931 sector_t first_bad;
1932 int bad_sectors;
1933 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1934
1935 if (!uptodate) {
1936 abort_sync_write(mddev, r1_bio);
1937 set_bit(WriteErrorSeen, &rdev->flags);
1938 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1939 set_bit(MD_RECOVERY_NEEDED, &
1940 mddev->recovery);
1941 set_bit(R1BIO_WriteError, &r1_bio->state);
1942 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1943 &first_bad, &bad_sectors) &&
1944 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1945 r1_bio->sector,
1946 r1_bio->sectors,
1947 &first_bad, &bad_sectors)
1948 )
1949 set_bit(R1BIO_MadeGood, &r1_bio->state);
1950
1951 put_sync_write_buf(r1_bio, uptodate);
1952}
1953
1954static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1955 int sectors, struct page *page, int rw)
1956{
1957 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1958 /* success */
1959 return 1;
1960 if (rw == WRITE) {
1961 set_bit(WriteErrorSeen, &rdev->flags);
1962 if (!test_and_set_bit(WantReplacement,
1963 &rdev->flags))
1964 set_bit(MD_RECOVERY_NEEDED, &
1965 rdev->mddev->recovery);
1966 }
1967 /* need to record an error - either for the block or the device */
1968 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1969 md_error(rdev->mddev, rdev);
1970 return 0;
1971}
1972
1973static int fix_sync_read_error(struct r1bio *r1_bio)
1974{
1975 /* Try some synchronous reads of other devices to get
1976 * good data, much like with normal read errors. Only
1977 * read into the pages we already have so we don't
1978 * need to re-issue the read request.
1979 * We don't need to freeze the array, because being in an
1980 * active sync request, there is no normal IO, and
1981 * no overlapping syncs.
1982 * We don't need to check is_badblock() again as we
1983 * made sure that anything with a bad block in range
1984 * will have bi_end_io clear.
1985 */
1986 struct mddev *mddev = r1_bio->mddev;
1987 struct r1conf *conf = mddev->private;
1988 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1989 struct page **pages = get_resync_pages(bio)->pages;
1990 sector_t sect = r1_bio->sector;
1991 int sectors = r1_bio->sectors;
1992 int idx = 0;
1993 struct md_rdev *rdev;
1994
1995 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1996 if (test_bit(FailFast, &rdev->flags)) {
1997 /* Don't try recovering from here - just fail it
1998 * ... unless it is the last working device of course */
1999 md_error(mddev, rdev);
2000 if (test_bit(Faulty, &rdev->flags))
2001 /* Don't try to read from here, but make sure
2002 * put_buf does it's thing
2003 */
2004 bio->bi_end_io = end_sync_write;
2005 }
2006
2007 while(sectors) {
2008 int s = sectors;
2009 int d = r1_bio->read_disk;
2010 int success = 0;
2011 int start;
2012
2013 if (s > (PAGE_SIZE>>9))
2014 s = PAGE_SIZE >> 9;
2015 do {
2016 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2017 /* No rcu protection needed here devices
2018 * can only be removed when no resync is
2019 * active, and resync is currently active
2020 */
2021 rdev = conf->mirrors[d].rdev;
2022 if (sync_page_io(rdev, sect, s<<9,
2023 pages[idx],
2024 REQ_OP_READ, 0, false)) {
2025 success = 1;
2026 break;
2027 }
2028 }
2029 d++;
2030 if (d == conf->raid_disks * 2)
2031 d = 0;
2032 } while (!success && d != r1_bio->read_disk);
2033
2034 if (!success) {
2035 char b[BDEVNAME_SIZE];
2036 int abort = 0;
2037 /* Cannot read from anywhere, this block is lost.
2038 * Record a bad block on each device. If that doesn't
2039 * work just disable and interrupt the recovery.
2040 * Don't fail devices as that won't really help.
2041 */
2042 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2043 mdname(mddev), bio_devname(bio, b),
2044 (unsigned long long)r1_bio->sector);
2045 for (d = 0; d < conf->raid_disks * 2; d++) {
2046 rdev = conf->mirrors[d].rdev;
2047 if (!rdev || test_bit(Faulty, &rdev->flags))
2048 continue;
2049 if (!rdev_set_badblocks(rdev, sect, s, 0))
2050 abort = 1;
2051 }
2052 if (abort) {
2053 conf->recovery_disabled =
2054 mddev->recovery_disabled;
2055 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2056 md_done_sync(mddev, r1_bio->sectors, 0);
2057 put_buf(r1_bio);
2058 return 0;
2059 }
2060 /* Try next page */
2061 sectors -= s;
2062 sect += s;
2063 idx++;
2064 continue;
2065 }
2066
2067 start = d;
2068 /* write it back and re-read */
2069 while (d != r1_bio->read_disk) {
2070 if (d == 0)
2071 d = conf->raid_disks * 2;
2072 d--;
2073 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2074 continue;
2075 rdev = conf->mirrors[d].rdev;
2076 if (r1_sync_page_io(rdev, sect, s,
2077 pages[idx],
2078 WRITE) == 0) {
2079 r1_bio->bios[d]->bi_end_io = NULL;
2080 rdev_dec_pending(rdev, mddev);
2081 }
2082 }
2083 d = start;
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 READ) != 0)
2094 atomic_add(s, &rdev->corrected_errors);
2095 }
2096 sectors -= s;
2097 sect += s;
2098 idx ++;
2099 }
2100 set_bit(R1BIO_Uptodate, &r1_bio->state);
2101 bio->bi_status = 0;
2102 return 1;
2103}
2104
2105static void process_checks(struct r1bio *r1_bio)
2106{
2107 /* We have read all readable devices. If we haven't
2108 * got the block, then there is no hope left.
2109 * If we have, then we want to do a comparison
2110 * and skip the write if everything is the same.
2111 * If any blocks failed to read, then we need to
2112 * attempt an over-write
2113 */
2114 struct mddev *mddev = r1_bio->mddev;
2115 struct r1conf *conf = mddev->private;
2116 int primary;
2117 int i;
2118 int vcnt;
2119
2120 /* Fix variable parts of all bios */
2121 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2122 for (i = 0; i < conf->raid_disks * 2; i++) {
2123 blk_status_t status;
2124 struct bio *b = r1_bio->bios[i];
2125 struct resync_pages *rp = get_resync_pages(b);
2126 if (b->bi_end_io != end_sync_read)
2127 continue;
2128 /* fixup the bio for reuse, but preserve errno */
2129 status = b->bi_status;
2130 bio_reset(b);
2131 b->bi_status = status;
2132 b->bi_iter.bi_sector = r1_bio->sector +
2133 conf->mirrors[i].rdev->data_offset;
2134 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2135 b->bi_end_io = end_sync_read;
2136 rp->raid_bio = r1_bio;
2137 b->bi_private = rp;
2138
2139 /* initialize bvec table again */
2140 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2141 }
2142 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2143 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2144 !r1_bio->bios[primary]->bi_status) {
2145 r1_bio->bios[primary]->bi_end_io = NULL;
2146 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2147 break;
2148 }
2149 r1_bio->read_disk = primary;
2150 for (i = 0; i < conf->raid_disks * 2; i++) {
2151 int j = 0;
2152 struct bio *pbio = r1_bio->bios[primary];
2153 struct bio *sbio = r1_bio->bios[i];
2154 blk_status_t status = sbio->bi_status;
2155 struct page **ppages = get_resync_pages(pbio)->pages;
2156 struct page **spages = get_resync_pages(sbio)->pages;
2157 struct bio_vec *bi;
2158 int page_len[RESYNC_PAGES] = { 0 };
2159 struct bvec_iter_all iter_all;
2160
2161 if (sbio->bi_end_io != end_sync_read)
2162 continue;
2163 /* Now we can 'fixup' the error value */
2164 sbio->bi_status = 0;
2165
2166 bio_for_each_segment_all(bi, sbio, iter_all)
2167 page_len[j++] = bi->bv_len;
2168
2169 if (!status) {
2170 for (j = vcnt; j-- ; ) {
2171 if (memcmp(page_address(ppages[j]),
2172 page_address(spages[j]),
2173 page_len[j]))
2174 break;
2175 }
2176 } else
2177 j = 0;
2178 if (j >= 0)
2179 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2180 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2181 && !status)) {
2182 /* No need to write to this device. */
2183 sbio->bi_end_io = NULL;
2184 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2185 continue;
2186 }
2187
2188 bio_copy_data(sbio, pbio);
2189 }
2190}
2191
2192static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2193{
2194 struct r1conf *conf = mddev->private;
2195 int i;
2196 int disks = conf->raid_disks * 2;
2197 struct bio *wbio;
2198
2199 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2200 /* ouch - failed to read all of that. */
2201 if (!fix_sync_read_error(r1_bio))
2202 return;
2203
2204 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2205 process_checks(r1_bio);
2206
2207 /*
2208 * schedule writes
2209 */
2210 atomic_set(&r1_bio->remaining, 1);
2211 for (i = 0; i < disks ; i++) {
2212 wbio = r1_bio->bios[i];
2213 if (wbio->bi_end_io == NULL ||
2214 (wbio->bi_end_io == end_sync_read &&
2215 (i == r1_bio->read_disk ||
2216 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2217 continue;
2218 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2219 abort_sync_write(mddev, r1_bio);
2220 continue;
2221 }
2222
2223 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2224 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2225 wbio->bi_opf |= MD_FAILFAST;
2226
2227 wbio->bi_end_io = end_sync_write;
2228 atomic_inc(&r1_bio->remaining);
2229 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2230
2231 generic_make_request(wbio);
2232 }
2233
2234 put_sync_write_buf(r1_bio, 1);
2235}
2236
2237/*
2238 * This is a kernel thread which:
2239 *
2240 * 1. Retries failed read operations on working mirrors.
2241 * 2. Updates the raid superblock when problems encounter.
2242 * 3. Performs writes following reads for array synchronising.
2243 */
2244
2245static void fix_read_error(struct r1conf *conf, int read_disk,
2246 sector_t sect, int sectors)
2247{
2248 struct mddev *mddev = conf->mddev;
2249 while(sectors) {
2250 int s = sectors;
2251 int d = read_disk;
2252 int success = 0;
2253 int start;
2254 struct md_rdev *rdev;
2255
2256 if (s > (PAGE_SIZE>>9))
2257 s = PAGE_SIZE >> 9;
2258
2259 do {
2260 sector_t first_bad;
2261 int bad_sectors;
2262
2263 rcu_read_lock();
2264 rdev = rcu_dereference(conf->mirrors[d].rdev);
2265 if (rdev &&
2266 (test_bit(In_sync, &rdev->flags) ||
2267 (!test_bit(Faulty, &rdev->flags) &&
2268 rdev->recovery_offset >= sect + s)) &&
2269 is_badblock(rdev, sect, s,
2270 &first_bad, &bad_sectors) == 0) {
2271 atomic_inc(&rdev->nr_pending);
2272 rcu_read_unlock();
2273 if (sync_page_io(rdev, sect, s<<9,
2274 conf->tmppage, REQ_OP_READ, 0, false))
2275 success = 1;
2276 rdev_dec_pending(rdev, mddev);
2277 if (success)
2278 break;
2279 } else
2280 rcu_read_unlock();
2281 d++;
2282 if (d == conf->raid_disks * 2)
2283 d = 0;
2284 } while (!success && d != read_disk);
2285
2286 if (!success) {
2287 /* Cannot read from anywhere - mark it bad */
2288 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2289 if (!rdev_set_badblocks(rdev, sect, s, 0))
2290 md_error(mddev, rdev);
2291 break;
2292 }
2293 /* write it back and re-read */
2294 start = d;
2295 while (d != read_disk) {
2296 if (d==0)
2297 d = conf->raid_disks * 2;
2298 d--;
2299 rcu_read_lock();
2300 rdev = rcu_dereference(conf->mirrors[d].rdev);
2301 if (rdev &&
2302 !test_bit(Faulty, &rdev->flags)) {
2303 atomic_inc(&rdev->nr_pending);
2304 rcu_read_unlock();
2305 r1_sync_page_io(rdev, sect, s,
2306 conf->tmppage, WRITE);
2307 rdev_dec_pending(rdev, mddev);
2308 } else
2309 rcu_read_unlock();
2310 }
2311 d = start;
2312 while (d != read_disk) {
2313 char b[BDEVNAME_SIZE];
2314 if (d==0)
2315 d = conf->raid_disks * 2;
2316 d--;
2317 rcu_read_lock();
2318 rdev = rcu_dereference(conf->mirrors[d].rdev);
2319 if (rdev &&
2320 !test_bit(Faulty, &rdev->flags)) {
2321 atomic_inc(&rdev->nr_pending);
2322 rcu_read_unlock();
2323 if (r1_sync_page_io(rdev, sect, s,
2324 conf->tmppage, READ)) {
2325 atomic_add(s, &rdev->corrected_errors);
2326 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2327 mdname(mddev), s,
2328 (unsigned long long)(sect +
2329 rdev->data_offset),
2330 bdevname(rdev->bdev, b));
2331 }
2332 rdev_dec_pending(rdev, mddev);
2333 } else
2334 rcu_read_unlock();
2335 }
2336 sectors -= s;
2337 sect += s;
2338 }
2339}
2340
2341static int narrow_write_error(struct r1bio *r1_bio, int i)
2342{
2343 struct mddev *mddev = r1_bio->mddev;
2344 struct r1conf *conf = mddev->private;
2345 struct md_rdev *rdev = conf->mirrors[i].rdev;
2346
2347 /* bio has the data to be written to device 'i' where
2348 * we just recently had a write error.
2349 * We repeatedly clone the bio and trim down to one block,
2350 * then try the write. Where the write fails we record
2351 * a bad block.
2352 * It is conceivable that the bio doesn't exactly align with
2353 * blocks. We must handle this somehow.
2354 *
2355 * We currently own a reference on the rdev.
2356 */
2357
2358 int block_sectors;
2359 sector_t sector;
2360 int sectors;
2361 int sect_to_write = r1_bio->sectors;
2362 int ok = 1;
2363
2364 if (rdev->badblocks.shift < 0)
2365 return 0;
2366
2367 block_sectors = roundup(1 << rdev->badblocks.shift,
2368 bdev_logical_block_size(rdev->bdev) >> 9);
2369 sector = r1_bio->sector;
2370 sectors = ((sector + block_sectors)
2371 & ~(sector_t)(block_sectors - 1))
2372 - sector;
2373
2374 while (sect_to_write) {
2375 struct bio *wbio;
2376 if (sectors > sect_to_write)
2377 sectors = sect_to_write;
2378 /* Write at 'sector' for 'sectors'*/
2379
2380 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2381 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2382 GFP_NOIO,
2383 &mddev->bio_set);
2384 } else {
2385 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2386 &mddev->bio_set);
2387 }
2388
2389 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2390 wbio->bi_iter.bi_sector = r1_bio->sector;
2391 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2392
2393 bio_trim(wbio, sector - r1_bio->sector, sectors);
2394 wbio->bi_iter.bi_sector += rdev->data_offset;
2395 bio_set_dev(wbio, rdev->bdev);
2396
2397 if (submit_bio_wait(wbio) < 0)
2398 /* failure! */
2399 ok = rdev_set_badblocks(rdev, sector,
2400 sectors, 0)
2401 && ok;
2402
2403 bio_put(wbio);
2404 sect_to_write -= sectors;
2405 sector += sectors;
2406 sectors = block_sectors;
2407 }
2408 return ok;
2409}
2410
2411static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2412{
2413 int m;
2414 int s = r1_bio->sectors;
2415 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2416 struct md_rdev *rdev = conf->mirrors[m].rdev;
2417 struct bio *bio = r1_bio->bios[m];
2418 if (bio->bi_end_io == NULL)
2419 continue;
2420 if (!bio->bi_status &&
2421 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2422 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2423 }
2424 if (bio->bi_status &&
2425 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2426 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2427 md_error(conf->mddev, rdev);
2428 }
2429 }
2430 put_buf(r1_bio);
2431 md_done_sync(conf->mddev, s, 1);
2432}
2433
2434static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2435{
2436 int m, idx;
2437 bool fail = false;
2438
2439 for (m = 0; m < conf->raid_disks * 2 ; m++)
2440 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2441 struct md_rdev *rdev = conf->mirrors[m].rdev;
2442 rdev_clear_badblocks(rdev,
2443 r1_bio->sector,
2444 r1_bio->sectors, 0);
2445 rdev_dec_pending(rdev, conf->mddev);
2446 } else if (r1_bio->bios[m] != NULL) {
2447 /* This drive got a write error. We need to
2448 * narrow down and record precise write
2449 * errors.
2450 */
2451 fail = true;
2452 if (!narrow_write_error(r1_bio, m)) {
2453 md_error(conf->mddev,
2454 conf->mirrors[m].rdev);
2455 /* an I/O failed, we can't clear the bitmap */
2456 set_bit(R1BIO_Degraded, &r1_bio->state);
2457 }
2458 rdev_dec_pending(conf->mirrors[m].rdev,
2459 conf->mddev);
2460 }
2461 if (fail) {
2462 spin_lock_irq(&conf->device_lock);
2463 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2464 idx = sector_to_idx(r1_bio->sector);
2465 atomic_inc(&conf->nr_queued[idx]);
2466 spin_unlock_irq(&conf->device_lock);
2467 /*
2468 * In case freeze_array() is waiting for condition
2469 * get_unqueued_pending() == extra to be true.
2470 */
2471 wake_up(&conf->wait_barrier);
2472 md_wakeup_thread(conf->mddev->thread);
2473 } else {
2474 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2475 close_write(r1_bio);
2476 raid_end_bio_io(r1_bio);
2477 }
2478}
2479
2480static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2481{
2482 struct mddev *mddev = conf->mddev;
2483 struct bio *bio;
2484 struct md_rdev *rdev;
2485
2486 clear_bit(R1BIO_ReadError, &r1_bio->state);
2487 /* we got a read error. Maybe the drive is bad. Maybe just
2488 * the block and we can fix it.
2489 * We freeze all other IO, and try reading the block from
2490 * other devices. When we find one, we re-write
2491 * and check it that fixes the read error.
2492 * This is all done synchronously while the array is
2493 * frozen
2494 */
2495
2496 bio = r1_bio->bios[r1_bio->read_disk];
2497 bio_put(bio);
2498 r1_bio->bios[r1_bio->read_disk] = NULL;
2499
2500 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2501 if (mddev->ro == 0
2502 && !test_bit(FailFast, &rdev->flags)) {
2503 freeze_array(conf, 1);
2504 fix_read_error(conf, r1_bio->read_disk,
2505 r1_bio->sector, r1_bio->sectors);
2506 unfreeze_array(conf);
2507 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2508 md_error(mddev, rdev);
2509 } else {
2510 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2511 }
2512
2513 rdev_dec_pending(rdev, conf->mddev);
2514 allow_barrier(conf, r1_bio->sector);
2515 bio = r1_bio->master_bio;
2516
2517 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2518 r1_bio->state = 0;
2519 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2520}
2521
2522static void raid1d(struct md_thread *thread)
2523{
2524 struct mddev *mddev = thread->mddev;
2525 struct r1bio *r1_bio;
2526 unsigned long flags;
2527 struct r1conf *conf = mddev->private;
2528 struct list_head *head = &conf->retry_list;
2529 struct blk_plug plug;
2530 int idx;
2531
2532 md_check_recovery(mddev);
2533
2534 if (!list_empty_careful(&conf->bio_end_io_list) &&
2535 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2536 LIST_HEAD(tmp);
2537 spin_lock_irqsave(&conf->device_lock, flags);
2538 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2539 list_splice_init(&conf->bio_end_io_list, &tmp);
2540 spin_unlock_irqrestore(&conf->device_lock, flags);
2541 while (!list_empty(&tmp)) {
2542 r1_bio = list_first_entry(&tmp, struct r1bio,
2543 retry_list);
2544 list_del(&r1_bio->retry_list);
2545 idx = sector_to_idx(r1_bio->sector);
2546 atomic_dec(&conf->nr_queued[idx]);
2547 if (mddev->degraded)
2548 set_bit(R1BIO_Degraded, &r1_bio->state);
2549 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2550 close_write(r1_bio);
2551 raid_end_bio_io(r1_bio);
2552 }
2553 }
2554
2555 blk_start_plug(&plug);
2556 for (;;) {
2557
2558 flush_pending_writes(conf);
2559
2560 spin_lock_irqsave(&conf->device_lock, flags);
2561 if (list_empty(head)) {
2562 spin_unlock_irqrestore(&conf->device_lock, flags);
2563 break;
2564 }
2565 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2566 list_del(head->prev);
2567 idx = sector_to_idx(r1_bio->sector);
2568 atomic_dec(&conf->nr_queued[idx]);
2569 spin_unlock_irqrestore(&conf->device_lock, flags);
2570
2571 mddev = r1_bio->mddev;
2572 conf = mddev->private;
2573 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2574 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2575 test_bit(R1BIO_WriteError, &r1_bio->state))
2576 handle_sync_write_finished(conf, r1_bio);
2577 else
2578 sync_request_write(mddev, r1_bio);
2579 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2580 test_bit(R1BIO_WriteError, &r1_bio->state))
2581 handle_write_finished(conf, r1_bio);
2582 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2583 handle_read_error(conf, r1_bio);
2584 else
2585 WARN_ON_ONCE(1);
2586
2587 cond_resched();
2588 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2589 md_check_recovery(mddev);
2590 }
2591 blk_finish_plug(&plug);
2592}
2593
2594static int init_resync(struct r1conf *conf)
2595{
2596 int buffs;
2597
2598 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2599 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2600
2601 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2602 r1buf_pool_free, conf->poolinfo);
2603}
2604
2605static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2606{
2607 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2608 struct resync_pages *rps;
2609 struct bio *bio;
2610 int i;
2611
2612 for (i = conf->poolinfo->raid_disks; i--; ) {
2613 bio = r1bio->bios[i];
2614 rps = bio->bi_private;
2615 bio_reset(bio);
2616 bio->bi_private = rps;
2617 }
2618 r1bio->master_bio = NULL;
2619 return r1bio;
2620}
2621
2622/*
2623 * perform a "sync" on one "block"
2624 *
2625 * We need to make sure that no normal I/O request - particularly write
2626 * requests - conflict with active sync requests.
2627 *
2628 * This is achieved by tracking pending requests and a 'barrier' concept
2629 * that can be installed to exclude normal IO requests.
2630 */
2631
2632static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2633 int *skipped)
2634{
2635 struct r1conf *conf = mddev->private;
2636 struct r1bio *r1_bio;
2637 struct bio *bio;
2638 sector_t max_sector, nr_sectors;
2639 int disk = -1;
2640 int i;
2641 int wonly = -1;
2642 int write_targets = 0, read_targets = 0;
2643 sector_t sync_blocks;
2644 int still_degraded = 0;
2645 int good_sectors = RESYNC_SECTORS;
2646 int min_bad = 0; /* number of sectors that are bad in all devices */
2647 int idx = sector_to_idx(sector_nr);
2648 int page_idx = 0;
2649
2650 if (!mempool_initialized(&conf->r1buf_pool))
2651 if (init_resync(conf))
2652 return 0;
2653
2654 max_sector = mddev->dev_sectors;
2655 if (sector_nr >= max_sector) {
2656 /* If we aborted, we need to abort the
2657 * sync on the 'current' bitmap chunk (there will
2658 * only be one in raid1 resync.
2659 * We can find the current addess in mddev->curr_resync
2660 */
2661 if (mddev->curr_resync < max_sector) /* aborted */
2662 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2663 &sync_blocks, 1);
2664 else /* completed sync */
2665 conf->fullsync = 0;
2666
2667 md_bitmap_close_sync(mddev->bitmap);
2668 close_sync(conf);
2669
2670 if (mddev_is_clustered(mddev)) {
2671 conf->cluster_sync_low = 0;
2672 conf->cluster_sync_high = 0;
2673 }
2674 return 0;
2675 }
2676
2677 if (mddev->bitmap == NULL &&
2678 mddev->recovery_cp == MaxSector &&
2679 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2680 conf->fullsync == 0) {
2681 *skipped = 1;
2682 return max_sector - sector_nr;
2683 }
2684 /* before building a request, check if we can skip these blocks..
2685 * This call the bitmap_start_sync doesn't actually record anything
2686 */
2687 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2688 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2689 /* We can skip this block, and probably several more */
2690 *skipped = 1;
2691 return sync_blocks;
2692 }
2693
2694 /*
2695 * If there is non-resync activity waiting for a turn, then let it
2696 * though before starting on this new sync request.
2697 */
2698 if (atomic_read(&conf->nr_waiting[idx]))
2699 schedule_timeout_uninterruptible(1);
2700
2701 /* we are incrementing sector_nr below. To be safe, we check against
2702 * sector_nr + two times RESYNC_SECTORS
2703 */
2704
2705 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2706 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2707
2708
2709 if (raise_barrier(conf, sector_nr))
2710 return 0;
2711
2712 r1_bio = raid1_alloc_init_r1buf(conf);
2713
2714 rcu_read_lock();
2715 /*
2716 * If we get a correctably read error during resync or recovery,
2717 * we might want to read from a different device. So we
2718 * flag all drives that could conceivably be read from for READ,
2719 * and any others (which will be non-In_sync devices) for WRITE.
2720 * If a read fails, we try reading from something else for which READ
2721 * is OK.
2722 */
2723
2724 r1_bio->mddev = mddev;
2725 r1_bio->sector = sector_nr;
2726 r1_bio->state = 0;
2727 set_bit(R1BIO_IsSync, &r1_bio->state);
2728 /* make sure good_sectors won't go across barrier unit boundary */
2729 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2730
2731 for (i = 0; i < conf->raid_disks * 2; i++) {
2732 struct md_rdev *rdev;
2733 bio = r1_bio->bios[i];
2734
2735 rdev = rcu_dereference(conf->mirrors[i].rdev);
2736 if (rdev == NULL ||
2737 test_bit(Faulty, &rdev->flags)) {
2738 if (i < conf->raid_disks)
2739 still_degraded = 1;
2740 } else if (!test_bit(In_sync, &rdev->flags)) {
2741 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2742 bio->bi_end_io = end_sync_write;
2743 write_targets ++;
2744 } else {
2745 /* may need to read from here */
2746 sector_t first_bad = MaxSector;
2747 int bad_sectors;
2748
2749 if (is_badblock(rdev, sector_nr, good_sectors,
2750 &first_bad, &bad_sectors)) {
2751 if (first_bad > sector_nr)
2752 good_sectors = first_bad - sector_nr;
2753 else {
2754 bad_sectors -= (sector_nr - first_bad);
2755 if (min_bad == 0 ||
2756 min_bad > bad_sectors)
2757 min_bad = bad_sectors;
2758 }
2759 }
2760 if (sector_nr < first_bad) {
2761 if (test_bit(WriteMostly, &rdev->flags)) {
2762 if (wonly < 0)
2763 wonly = i;
2764 } else {
2765 if (disk < 0)
2766 disk = i;
2767 }
2768 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2769 bio->bi_end_io = end_sync_read;
2770 read_targets++;
2771 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2772 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2773 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2774 /*
2775 * The device is suitable for reading (InSync),
2776 * but has bad block(s) here. Let's try to correct them,
2777 * if we are doing resync or repair. Otherwise, leave
2778 * this device alone for this sync request.
2779 */
2780 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2781 bio->bi_end_io = end_sync_write;
2782 write_targets++;
2783 }
2784 }
2785 if (bio->bi_end_io) {
2786 atomic_inc(&rdev->nr_pending);
2787 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2788 bio_set_dev(bio, rdev->bdev);
2789 if (test_bit(FailFast, &rdev->flags))
2790 bio->bi_opf |= MD_FAILFAST;
2791 }
2792 }
2793 rcu_read_unlock();
2794 if (disk < 0)
2795 disk = wonly;
2796 r1_bio->read_disk = disk;
2797
2798 if (read_targets == 0 && min_bad > 0) {
2799 /* These sectors are bad on all InSync devices, so we
2800 * need to mark them bad on all write targets
2801 */
2802 int ok = 1;
2803 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2804 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2805 struct md_rdev *rdev = conf->mirrors[i].rdev;
2806 ok = rdev_set_badblocks(rdev, sector_nr,
2807 min_bad, 0
2808 ) && ok;
2809 }
2810 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2811 *skipped = 1;
2812 put_buf(r1_bio);
2813
2814 if (!ok) {
2815 /* Cannot record the badblocks, so need to
2816 * abort the resync.
2817 * If there are multiple read targets, could just
2818 * fail the really bad ones ???
2819 */
2820 conf->recovery_disabled = mddev->recovery_disabled;
2821 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2822 return 0;
2823 } else
2824 return min_bad;
2825
2826 }
2827 if (min_bad > 0 && min_bad < good_sectors) {
2828 /* only resync enough to reach the next bad->good
2829 * transition */
2830 good_sectors = min_bad;
2831 }
2832
2833 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2834 /* extra read targets are also write targets */
2835 write_targets += read_targets-1;
2836
2837 if (write_targets == 0 || read_targets == 0) {
2838 /* There is nowhere to write, so all non-sync
2839 * drives must be failed - so we are finished
2840 */
2841 sector_t rv;
2842 if (min_bad > 0)
2843 max_sector = sector_nr + min_bad;
2844 rv = max_sector - sector_nr;
2845 *skipped = 1;
2846 put_buf(r1_bio);
2847 return rv;
2848 }
2849
2850 if (max_sector > mddev->resync_max)
2851 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2852 if (max_sector > sector_nr + good_sectors)
2853 max_sector = sector_nr + good_sectors;
2854 nr_sectors = 0;
2855 sync_blocks = 0;
2856 do {
2857 struct page *page;
2858 int len = PAGE_SIZE;
2859 if (sector_nr + (len>>9) > max_sector)
2860 len = (max_sector - sector_nr) << 9;
2861 if (len == 0)
2862 break;
2863 if (sync_blocks == 0) {
2864 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2865 &sync_blocks, still_degraded) &&
2866 !conf->fullsync &&
2867 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2868 break;
2869 if ((len >> 9) > sync_blocks)
2870 len = sync_blocks<<9;
2871 }
2872
2873 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2874 struct resync_pages *rp;
2875
2876 bio = r1_bio->bios[i];
2877 rp = get_resync_pages(bio);
2878 if (bio->bi_end_io) {
2879 page = resync_fetch_page(rp, page_idx);
2880
2881 /*
2882 * won't fail because the vec table is big
2883 * enough to hold all these pages
2884 */
2885 bio_add_page(bio, page, len, 0);
2886 }
2887 }
2888 nr_sectors += len>>9;
2889 sector_nr += len>>9;
2890 sync_blocks -= (len>>9);
2891 } while (++page_idx < RESYNC_PAGES);
2892
2893 r1_bio->sectors = nr_sectors;
2894
2895 if (mddev_is_clustered(mddev) &&
2896 conf->cluster_sync_high < sector_nr + nr_sectors) {
2897 conf->cluster_sync_low = mddev->curr_resync_completed;
2898 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2899 /* Send resync message */
2900 md_cluster_ops->resync_info_update(mddev,
2901 conf->cluster_sync_low,
2902 conf->cluster_sync_high);
2903 }
2904
2905 /* For a user-requested sync, we read all readable devices and do a
2906 * compare
2907 */
2908 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2909 atomic_set(&r1_bio->remaining, read_targets);
2910 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2911 bio = r1_bio->bios[i];
2912 if (bio->bi_end_io == end_sync_read) {
2913 read_targets--;
2914 md_sync_acct_bio(bio, nr_sectors);
2915 if (read_targets == 1)
2916 bio->bi_opf &= ~MD_FAILFAST;
2917 generic_make_request(bio);
2918 }
2919 }
2920 } else {
2921 atomic_set(&r1_bio->remaining, 1);
2922 bio = r1_bio->bios[r1_bio->read_disk];
2923 md_sync_acct_bio(bio, nr_sectors);
2924 if (read_targets == 1)
2925 bio->bi_opf &= ~MD_FAILFAST;
2926 generic_make_request(bio);
2927 }
2928 return nr_sectors;
2929}
2930
2931static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2932{
2933 if (sectors)
2934 return sectors;
2935
2936 return mddev->dev_sectors;
2937}
2938
2939static struct r1conf *setup_conf(struct mddev *mddev)
2940{
2941 struct r1conf *conf;
2942 int i;
2943 struct raid1_info *disk;
2944 struct md_rdev *rdev;
2945 int err = -ENOMEM;
2946
2947 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2948 if (!conf)
2949 goto abort;
2950
2951 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2952 sizeof(atomic_t), GFP_KERNEL);
2953 if (!conf->nr_pending)
2954 goto abort;
2955
2956 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2957 sizeof(atomic_t), GFP_KERNEL);
2958 if (!conf->nr_waiting)
2959 goto abort;
2960
2961 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2962 sizeof(atomic_t), GFP_KERNEL);
2963 if (!conf->nr_queued)
2964 goto abort;
2965
2966 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2967 sizeof(atomic_t), GFP_KERNEL);
2968 if (!conf->barrier)
2969 goto abort;
2970
2971 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2972 mddev->raid_disks, 2),
2973 GFP_KERNEL);
2974 if (!conf->mirrors)
2975 goto abort;
2976
2977 conf->tmppage = alloc_page(GFP_KERNEL);
2978 if (!conf->tmppage)
2979 goto abort;
2980
2981 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2982 if (!conf->poolinfo)
2983 goto abort;
2984 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2985 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2986 rbio_pool_free, conf->poolinfo);
2987 if (err)
2988 goto abort;
2989
2990 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
2991 if (err)
2992 goto abort;
2993
2994 conf->poolinfo->mddev = mddev;
2995
2996 err = -EINVAL;
2997 spin_lock_init(&conf->device_lock);
2998 rdev_for_each(rdev, mddev) {
2999 int disk_idx = rdev->raid_disk;
3000 if (disk_idx >= mddev->raid_disks
3001 || disk_idx < 0)
3002 continue;
3003 if (test_bit(Replacement, &rdev->flags))
3004 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3005 else
3006 disk = conf->mirrors + disk_idx;
3007
3008 if (disk->rdev)
3009 goto abort;
3010 disk->rdev = rdev;
3011 disk->head_position = 0;
3012 disk->seq_start = MaxSector;
3013 }
3014 conf->raid_disks = mddev->raid_disks;
3015 conf->mddev = mddev;
3016 INIT_LIST_HEAD(&conf->retry_list);
3017 INIT_LIST_HEAD(&conf->bio_end_io_list);
3018
3019 spin_lock_init(&conf->resync_lock);
3020 init_waitqueue_head(&conf->wait_barrier);
3021
3022 bio_list_init(&conf->pending_bio_list);
3023 conf->pending_count = 0;
3024 conf->recovery_disabled = mddev->recovery_disabled - 1;
3025
3026 err = -EIO;
3027 for (i = 0; i < conf->raid_disks * 2; i++) {
3028
3029 disk = conf->mirrors + i;
3030
3031 if (i < conf->raid_disks &&
3032 disk[conf->raid_disks].rdev) {
3033 /* This slot has a replacement. */
3034 if (!disk->rdev) {
3035 /* No original, just make the replacement
3036 * a recovering spare
3037 */
3038 disk->rdev =
3039 disk[conf->raid_disks].rdev;
3040 disk[conf->raid_disks].rdev = NULL;
3041 } else if (!test_bit(In_sync, &disk->rdev->flags))
3042 /* Original is not in_sync - bad */
3043 goto abort;
3044 }
3045
3046 if (!disk->rdev ||
3047 !test_bit(In_sync, &disk->rdev->flags)) {
3048 disk->head_position = 0;
3049 if (disk->rdev &&
3050 (disk->rdev->saved_raid_disk < 0))
3051 conf->fullsync = 1;
3052 }
3053 }
3054
3055 err = -ENOMEM;
3056 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3057 if (!conf->thread)
3058 goto abort;
3059
3060 return conf;
3061
3062 abort:
3063 if (conf) {
3064 mempool_exit(&conf->r1bio_pool);
3065 kfree(conf->mirrors);
3066 safe_put_page(conf->tmppage);
3067 kfree(conf->poolinfo);
3068 kfree(conf->nr_pending);
3069 kfree(conf->nr_waiting);
3070 kfree(conf->nr_queued);
3071 kfree(conf->barrier);
3072 bioset_exit(&conf->bio_split);
3073 kfree(conf);
3074 }
3075 return ERR_PTR(err);
3076}
3077
3078static void raid1_free(struct mddev *mddev, void *priv);
3079static int raid1_run(struct mddev *mddev)
3080{
3081 struct r1conf *conf;
3082 int i;
3083 struct md_rdev *rdev;
3084 int ret;
3085 bool discard_supported = false;
3086
3087 if (mddev->level != 1) {
3088 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3089 mdname(mddev), mddev->level);
3090 return -EIO;
3091 }
3092 if (mddev->reshape_position != MaxSector) {
3093 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3094 mdname(mddev));
3095 return -EIO;
3096 }
3097 if (mddev_init_writes_pending(mddev) < 0)
3098 return -ENOMEM;
3099 /*
3100 * copy the already verified devices into our private RAID1
3101 * bookkeeping area. [whatever we allocate in run(),
3102 * should be freed in raid1_free()]
3103 */
3104 if (mddev->private == NULL)
3105 conf = setup_conf(mddev);
3106 else
3107 conf = mddev->private;
3108
3109 if (IS_ERR(conf))
3110 return PTR_ERR(conf);
3111
3112 if (mddev->queue) {
3113 blk_queue_max_write_same_sectors(mddev->queue, 0);
3114 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3115 }
3116
3117 rdev_for_each(rdev, mddev) {
3118 if (!mddev->gendisk)
3119 continue;
3120 disk_stack_limits(mddev->gendisk, rdev->bdev,
3121 rdev->data_offset << 9);
3122 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3123 discard_supported = true;
3124 }
3125
3126 mddev->degraded = 0;
3127 for (i = 0; i < conf->raid_disks; i++)
3128 if (conf->mirrors[i].rdev == NULL ||
3129 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3130 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3131 mddev->degraded++;
3132 /*
3133 * RAID1 needs at least one disk in active
3134 */
3135 if (conf->raid_disks - mddev->degraded < 1) {
3136 ret = -EINVAL;
3137 goto abort;
3138 }
3139
3140 if (conf->raid_disks - mddev->degraded == 1)
3141 mddev->recovery_cp = MaxSector;
3142
3143 if (mddev->recovery_cp != MaxSector)
3144 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3145 mdname(mddev));
3146 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3147 mdname(mddev), mddev->raid_disks - mddev->degraded,
3148 mddev->raid_disks);
3149
3150 /*
3151 * Ok, everything is just fine now
3152 */
3153 mddev->thread = conf->thread;
3154 conf->thread = NULL;
3155 mddev->private = conf;
3156 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3157
3158 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3159
3160 if (mddev->queue) {
3161 if (discard_supported)
3162 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3163 mddev->queue);
3164 else
3165 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3166 mddev->queue);
3167 }
3168
3169 ret = md_integrity_register(mddev);
3170 if (ret) {
3171 md_unregister_thread(&mddev->thread);
3172 goto abort;
3173 }
3174 return 0;
3175
3176abort:
3177 raid1_free(mddev, conf);
3178 return ret;
3179}
3180
3181static void raid1_free(struct mddev *mddev, void *priv)
3182{
3183 struct r1conf *conf = priv;
3184
3185 mempool_exit(&conf->r1bio_pool);
3186 kfree(conf->mirrors);
3187 safe_put_page(conf->tmppage);
3188 kfree(conf->poolinfo);
3189 kfree(conf->nr_pending);
3190 kfree(conf->nr_waiting);
3191 kfree(conf->nr_queued);
3192 kfree(conf->barrier);
3193 bioset_exit(&conf->bio_split);
3194 kfree(conf);
3195}
3196
3197static int raid1_resize(struct mddev *mddev, sector_t sectors)
3198{
3199 /* no resync is happening, and there is enough space
3200 * on all devices, so we can resize.
3201 * We need to make sure resync covers any new space.
3202 * If the array is shrinking we should possibly wait until
3203 * any io in the removed space completes, but it hardly seems
3204 * worth it.
3205 */
3206 sector_t newsize = raid1_size(mddev, sectors, 0);
3207 if (mddev->external_size &&
3208 mddev->array_sectors > newsize)
3209 return -EINVAL;
3210 if (mddev->bitmap) {
3211 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3212 if (ret)
3213 return ret;
3214 }
3215 md_set_array_sectors(mddev, newsize);
3216 if (sectors > mddev->dev_sectors &&
3217 mddev->recovery_cp > mddev->dev_sectors) {
3218 mddev->recovery_cp = mddev->dev_sectors;
3219 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3220 }
3221 mddev->dev_sectors = sectors;
3222 mddev->resync_max_sectors = sectors;
3223 return 0;
3224}
3225
3226static int raid1_reshape(struct mddev *mddev)
3227{
3228 /* We need to:
3229 * 1/ resize the r1bio_pool
3230 * 2/ resize conf->mirrors
3231 *
3232 * We allocate a new r1bio_pool if we can.
3233 * Then raise a device barrier and wait until all IO stops.
3234 * Then resize conf->mirrors and swap in the new r1bio pool.
3235 *
3236 * At the same time, we "pack" the devices so that all the missing
3237 * devices have the higher raid_disk numbers.
3238 */
3239 mempool_t newpool, oldpool;
3240 struct pool_info *newpoolinfo;
3241 struct raid1_info *newmirrors;
3242 struct r1conf *conf = mddev->private;
3243 int cnt, raid_disks;
3244 unsigned long flags;
3245 int d, d2;
3246 int ret;
3247
3248 memset(&newpool, 0, sizeof(newpool));
3249 memset(&oldpool, 0, sizeof(oldpool));
3250
3251 /* Cannot change chunk_size, layout, or level */
3252 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3253 mddev->layout != mddev->new_layout ||
3254 mddev->level != mddev->new_level) {
3255 mddev->new_chunk_sectors = mddev->chunk_sectors;
3256 mddev->new_layout = mddev->layout;
3257 mddev->new_level = mddev->level;
3258 return -EINVAL;
3259 }
3260
3261 if (!mddev_is_clustered(mddev))
3262 md_allow_write(mddev);
3263
3264 raid_disks = mddev->raid_disks + mddev->delta_disks;
3265
3266 if (raid_disks < conf->raid_disks) {
3267 cnt=0;
3268 for (d= 0; d < conf->raid_disks; d++)
3269 if (conf->mirrors[d].rdev)
3270 cnt++;
3271 if (cnt > raid_disks)
3272 return -EBUSY;
3273 }
3274
3275 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3276 if (!newpoolinfo)
3277 return -ENOMEM;
3278 newpoolinfo->mddev = mddev;
3279 newpoolinfo->raid_disks = raid_disks * 2;
3280
3281 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3282 rbio_pool_free, newpoolinfo);
3283 if (ret) {
3284 kfree(newpoolinfo);
3285 return ret;
3286 }
3287 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3288 raid_disks, 2),
3289 GFP_KERNEL);
3290 if (!newmirrors) {
3291 kfree(newpoolinfo);
3292 mempool_exit(&newpool);
3293 return -ENOMEM;
3294 }
3295
3296 freeze_array(conf, 0);
3297
3298 /* ok, everything is stopped */
3299 oldpool = conf->r1bio_pool;
3300 conf->r1bio_pool = newpool;
3301
3302 for (d = d2 = 0; d < conf->raid_disks; d++) {
3303 struct md_rdev *rdev = conf->mirrors[d].rdev;
3304 if (rdev && rdev->raid_disk != d2) {
3305 sysfs_unlink_rdev(mddev, rdev);
3306 rdev->raid_disk = d2;
3307 sysfs_unlink_rdev(mddev, rdev);
3308 if (sysfs_link_rdev(mddev, rdev))
3309 pr_warn("md/raid1:%s: cannot register rd%d\n",
3310 mdname(mddev), rdev->raid_disk);
3311 }
3312 if (rdev)
3313 newmirrors[d2++].rdev = rdev;
3314 }
3315 kfree(conf->mirrors);
3316 conf->mirrors = newmirrors;
3317 kfree(conf->poolinfo);
3318 conf->poolinfo = newpoolinfo;
3319
3320 spin_lock_irqsave(&conf->device_lock, flags);
3321 mddev->degraded += (raid_disks - conf->raid_disks);
3322 spin_unlock_irqrestore(&conf->device_lock, flags);
3323 conf->raid_disks = mddev->raid_disks = raid_disks;
3324 mddev->delta_disks = 0;
3325
3326 unfreeze_array(conf);
3327
3328 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3329 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3330 md_wakeup_thread(mddev->thread);
3331
3332 mempool_exit(&oldpool);
3333 return 0;
3334}
3335
3336static void raid1_quiesce(struct mddev *mddev, int quiesce)
3337{
3338 struct r1conf *conf = mddev->private;
3339
3340 if (quiesce)
3341 freeze_array(conf, 0);
3342 else
3343 unfreeze_array(conf);
3344}
3345
3346static void *raid1_takeover(struct mddev *mddev)
3347{
3348 /* raid1 can take over:
3349 * raid5 with 2 devices, any layout or chunk size
3350 */
3351 if (mddev->level == 5 && mddev->raid_disks == 2) {
3352 struct r1conf *conf;
3353 mddev->new_level = 1;
3354 mddev->new_layout = 0;
3355 mddev->new_chunk_sectors = 0;
3356 conf = setup_conf(mddev);
3357 if (!IS_ERR(conf)) {
3358 /* Array must appear to be quiesced */
3359 conf->array_frozen = 1;
3360 mddev_clear_unsupported_flags(mddev,
3361 UNSUPPORTED_MDDEV_FLAGS);
3362 }
3363 return conf;
3364 }
3365 return ERR_PTR(-EINVAL);
3366}
3367
3368static struct md_personality raid1_personality =
3369{
3370 .name = "raid1",
3371 .level = 1,
3372 .owner = THIS_MODULE,
3373 .make_request = raid1_make_request,
3374 .run = raid1_run,
3375 .free = raid1_free,
3376 .status = raid1_status,
3377 .error_handler = raid1_error,
3378 .hot_add_disk = raid1_add_disk,
3379 .hot_remove_disk= raid1_remove_disk,
3380 .spare_active = raid1_spare_active,
3381 .sync_request = raid1_sync_request,
3382 .resize = raid1_resize,
3383 .size = raid1_size,
3384 .check_reshape = raid1_reshape,
3385 .quiesce = raid1_quiesce,
3386 .takeover = raid1_takeover,
3387 .congested = raid1_congested,
3388};
3389
3390static int __init raid_init(void)
3391{
3392 return register_md_personality(&raid1_personality);
3393}
3394
3395static void raid_exit(void)
3396{
3397 unregister_md_personality(&raid1_personality);
3398}
3399
3400module_init(raid_init);
3401module_exit(raid_exit);
3402MODULE_LICENSE("GPL");
3403MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3404MODULE_ALIAS("md-personality-3"); /* RAID1 */
3405MODULE_ALIAS("md-raid1");
3406MODULE_ALIAS("md-level-1");
3407
3408module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);