Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid1.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 *
7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 *
9 * RAID-1 management functions.
10 *
11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 *
13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 *
16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
17 * bitmapped intelligence in resync:
18 *
19 * - bitmap marked during normal i/o
20 * - bitmap used to skip nondirty blocks during sync
21 *
22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
23 * - persistent bitmap code
24 */
25
26#include <linux/slab.h>
27#include <linux/delay.h>
28#include <linux/blkdev.h>
29#include <linux/module.h>
30#include <linux/seq_file.h>
31#include <linux/ratelimit.h>
32#include <linux/interval_tree_generic.h>
33
34#include <trace/events/block.h>
35
36#include "md.h"
37#include "raid1.h"
38#include "md-bitmap.h"
39
40#define UNSUPPORTED_MDDEV_FLAGS \
41 ((1L << MD_HAS_JOURNAL) | \
42 (1L << MD_JOURNAL_CLEAN) | \
43 (1L << MD_HAS_PPL) | \
44 (1L << MD_HAS_MULTIPLE_PPLS))
45
46static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
47static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
48
49#define raid1_log(md, fmt, args...) \
50 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
51
52#define RAID_1_10_NAME "raid1"
53#include "raid1-10.c"
54
55#define START(node) ((node)->start)
56#define LAST(node) ((node)->last)
57INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
58 START, LAST, static inline, raid1_rb);
59
60static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
61 struct serial_info *si, int idx)
62{
63 unsigned long flags;
64 int ret = 0;
65 sector_t lo = r1_bio->sector;
66 sector_t hi = lo + r1_bio->sectors;
67 struct serial_in_rdev *serial = &rdev->serial[idx];
68
69 spin_lock_irqsave(&serial->serial_lock, flags);
70 /* collision happened */
71 if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
72 ret = -EBUSY;
73 else {
74 si->start = lo;
75 si->last = hi;
76 raid1_rb_insert(si, &serial->serial_rb);
77 }
78 spin_unlock_irqrestore(&serial->serial_lock, flags);
79
80 return ret;
81}
82
83static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
84{
85 struct mddev *mddev = rdev->mddev;
86 struct serial_info *si;
87 int idx = sector_to_idx(r1_bio->sector);
88 struct serial_in_rdev *serial = &rdev->serial[idx];
89
90 if (WARN_ON(!mddev->serial_info_pool))
91 return;
92 si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
93 wait_event(serial->serial_io_wait,
94 check_and_add_serial(rdev, r1_bio, si, idx) == 0);
95}
96
97static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
98{
99 struct serial_info *si;
100 unsigned long flags;
101 int found = 0;
102 struct mddev *mddev = rdev->mddev;
103 int idx = sector_to_idx(lo);
104 struct serial_in_rdev *serial = &rdev->serial[idx];
105
106 spin_lock_irqsave(&serial->serial_lock, flags);
107 for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
108 si; si = raid1_rb_iter_next(si, lo, hi)) {
109 if (si->start == lo && si->last == hi) {
110 raid1_rb_remove(si, &serial->serial_rb);
111 mempool_free(si, mddev->serial_info_pool);
112 found = 1;
113 break;
114 }
115 }
116 if (!found)
117 WARN(1, "The write IO is not recorded for serialization\n");
118 spin_unlock_irqrestore(&serial->serial_lock, flags);
119 wake_up(&serial->serial_io_wait);
120}
121
122/*
123 * for resync bio, r1bio pointer can be retrieved from the per-bio
124 * 'struct resync_pages'.
125 */
126static inline struct r1bio *get_resync_r1bio(struct bio *bio)
127{
128 return get_resync_pages(bio)->raid_bio;
129}
130
131static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
132{
133 struct pool_info *pi = data;
134 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
135
136 /* allocate a r1bio with room for raid_disks entries in the bios array */
137 return kzalloc(size, gfp_flags);
138}
139
140#define RESYNC_DEPTH 32
141#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
142#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
143#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
144#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
145#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
146
147static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
148{
149 struct pool_info *pi = data;
150 struct r1bio *r1_bio;
151 struct bio *bio;
152 int need_pages;
153 int j;
154 struct resync_pages *rps;
155
156 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
157 if (!r1_bio)
158 return NULL;
159
160 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
161 gfp_flags);
162 if (!rps)
163 goto out_free_r1bio;
164
165 /*
166 * Allocate bios : 1 for reading, n-1 for writing
167 */
168 for (j = pi->raid_disks ; j-- ; ) {
169 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
170 if (!bio)
171 goto out_free_bio;
172 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
173 r1_bio->bios[j] = bio;
174 }
175 /*
176 * Allocate RESYNC_PAGES data pages and attach them to
177 * the first bio.
178 * If this is a user-requested check/repair, allocate
179 * RESYNC_PAGES for each bio.
180 */
181 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
182 need_pages = pi->raid_disks;
183 else
184 need_pages = 1;
185 for (j = 0; j < pi->raid_disks; j++) {
186 struct resync_pages *rp = &rps[j];
187
188 bio = r1_bio->bios[j];
189
190 if (j < need_pages) {
191 if (resync_alloc_pages(rp, gfp_flags))
192 goto out_free_pages;
193 } else {
194 memcpy(rp, &rps[0], sizeof(*rp));
195 resync_get_all_pages(rp);
196 }
197
198 rp->raid_bio = r1_bio;
199 bio->bi_private = rp;
200 }
201
202 r1_bio->master_bio = NULL;
203
204 return r1_bio;
205
206out_free_pages:
207 while (--j >= 0)
208 resync_free_pages(&rps[j]);
209
210out_free_bio:
211 while (++j < pi->raid_disks) {
212 bio_uninit(r1_bio->bios[j]);
213 kfree(r1_bio->bios[j]);
214 }
215 kfree(rps);
216
217out_free_r1bio:
218 rbio_pool_free(r1_bio, data);
219 return NULL;
220}
221
222static void r1buf_pool_free(void *__r1_bio, void *data)
223{
224 struct pool_info *pi = data;
225 int i;
226 struct r1bio *r1bio = __r1_bio;
227 struct resync_pages *rp = NULL;
228
229 for (i = pi->raid_disks; i--; ) {
230 rp = get_resync_pages(r1bio->bios[i]);
231 resync_free_pages(rp);
232 bio_uninit(r1bio->bios[i]);
233 kfree(r1bio->bios[i]);
234 }
235
236 /* resync pages array stored in the 1st bio's .bi_private */
237 kfree(rp);
238
239 rbio_pool_free(r1bio, data);
240}
241
242static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
243{
244 int i;
245
246 for (i = 0; i < conf->raid_disks * 2; i++) {
247 struct bio **bio = r1_bio->bios + i;
248 if (!BIO_SPECIAL(*bio))
249 bio_put(*bio);
250 *bio = NULL;
251 }
252}
253
254static void free_r1bio(struct r1bio *r1_bio)
255{
256 struct r1conf *conf = r1_bio->mddev->private;
257
258 put_all_bios(conf, r1_bio);
259 mempool_free(r1_bio, &conf->r1bio_pool);
260}
261
262static void put_buf(struct r1bio *r1_bio)
263{
264 struct r1conf *conf = r1_bio->mddev->private;
265 sector_t sect = r1_bio->sector;
266 int i;
267
268 for (i = 0; i < conf->raid_disks * 2; i++) {
269 struct bio *bio = r1_bio->bios[i];
270 if (bio->bi_end_io)
271 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
272 }
273
274 mempool_free(r1_bio, &conf->r1buf_pool);
275
276 lower_barrier(conf, sect);
277}
278
279static void reschedule_retry(struct r1bio *r1_bio)
280{
281 unsigned long flags;
282 struct mddev *mddev = r1_bio->mddev;
283 struct r1conf *conf = mddev->private;
284 int idx;
285
286 idx = sector_to_idx(r1_bio->sector);
287 spin_lock_irqsave(&conf->device_lock, flags);
288 list_add(&r1_bio->retry_list, &conf->retry_list);
289 atomic_inc(&conf->nr_queued[idx]);
290 spin_unlock_irqrestore(&conf->device_lock, flags);
291
292 wake_up(&conf->wait_barrier);
293 md_wakeup_thread(mddev->thread);
294}
295
296/*
297 * raid_end_bio_io() is called when we have finished servicing a mirrored
298 * operation and are ready to return a success/failure code to the buffer
299 * cache layer.
300 */
301static void call_bio_endio(struct r1bio *r1_bio)
302{
303 struct bio *bio = r1_bio->master_bio;
304
305 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
306 bio->bi_status = BLK_STS_IOERR;
307
308 bio_endio(bio);
309}
310
311static void raid_end_bio_io(struct r1bio *r1_bio)
312{
313 struct bio *bio = r1_bio->master_bio;
314 struct r1conf *conf = r1_bio->mddev->private;
315 sector_t sector = r1_bio->sector;
316
317 /* if nobody has done the final endio yet, do it now */
318 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
319 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
320 (bio_data_dir(bio) == WRITE) ? "write" : "read",
321 (unsigned long long) bio->bi_iter.bi_sector,
322 (unsigned long long) bio_end_sector(bio) - 1);
323
324 call_bio_endio(r1_bio);
325 }
326
327 free_r1bio(r1_bio);
328 /*
329 * Wake up any possible resync thread that waits for the device
330 * to go idle. All I/Os, even write-behind writes, are done.
331 */
332 allow_barrier(conf, sector);
333}
334
335/*
336 * Update disk head position estimator based on IRQ completion info.
337 */
338static inline void update_head_pos(int disk, struct r1bio *r1_bio)
339{
340 struct r1conf *conf = r1_bio->mddev->private;
341
342 conf->mirrors[disk].head_position =
343 r1_bio->sector + (r1_bio->sectors);
344}
345
346/*
347 * Find the disk number which triggered given bio
348 */
349static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
350{
351 int mirror;
352 struct r1conf *conf = r1_bio->mddev->private;
353 int raid_disks = conf->raid_disks;
354
355 for (mirror = 0; mirror < raid_disks * 2; mirror++)
356 if (r1_bio->bios[mirror] == bio)
357 break;
358
359 BUG_ON(mirror == raid_disks * 2);
360 update_head_pos(mirror, r1_bio);
361
362 return mirror;
363}
364
365static void raid1_end_read_request(struct bio *bio)
366{
367 int uptodate = !bio->bi_status;
368 struct r1bio *r1_bio = bio->bi_private;
369 struct r1conf *conf = r1_bio->mddev->private;
370 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
371
372 /*
373 * this branch is our 'one mirror IO has finished' event handler:
374 */
375 update_head_pos(r1_bio->read_disk, r1_bio);
376
377 if (uptodate)
378 set_bit(R1BIO_Uptodate, &r1_bio->state);
379 else if (test_bit(FailFast, &rdev->flags) &&
380 test_bit(R1BIO_FailFast, &r1_bio->state))
381 /* This was a fail-fast read so we definitely
382 * want to retry */
383 ;
384 else {
385 /* If all other devices have failed, we want to return
386 * the error upwards rather than fail the last device.
387 * Here we redefine "uptodate" to mean "Don't want to retry"
388 */
389 unsigned long flags;
390 spin_lock_irqsave(&conf->device_lock, flags);
391 if (r1_bio->mddev->degraded == conf->raid_disks ||
392 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
393 test_bit(In_sync, &rdev->flags)))
394 uptodate = 1;
395 spin_unlock_irqrestore(&conf->device_lock, flags);
396 }
397
398 if (uptodate) {
399 raid_end_bio_io(r1_bio);
400 rdev_dec_pending(rdev, conf->mddev);
401 } else {
402 /*
403 * oops, read error:
404 */
405 pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
406 mdname(conf->mddev),
407 rdev->bdev,
408 (unsigned long long)r1_bio->sector);
409 set_bit(R1BIO_ReadError, &r1_bio->state);
410 reschedule_retry(r1_bio);
411 /* don't drop the reference on read_disk yet */
412 }
413}
414
415static void close_write(struct r1bio *r1_bio)
416{
417 /* it really is the end of this request */
418 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
419 bio_free_pages(r1_bio->behind_master_bio);
420 bio_put(r1_bio->behind_master_bio);
421 r1_bio->behind_master_bio = NULL;
422 }
423 /* clear the bitmap if all writes complete successfully */
424 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
425 r1_bio->sectors,
426 !test_bit(R1BIO_Degraded, &r1_bio->state),
427 test_bit(R1BIO_BehindIO, &r1_bio->state));
428 md_write_end(r1_bio->mddev);
429}
430
431static void r1_bio_write_done(struct r1bio *r1_bio)
432{
433 if (!atomic_dec_and_test(&r1_bio->remaining))
434 return;
435
436 if (test_bit(R1BIO_WriteError, &r1_bio->state))
437 reschedule_retry(r1_bio);
438 else {
439 close_write(r1_bio);
440 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
441 reschedule_retry(r1_bio);
442 else
443 raid_end_bio_io(r1_bio);
444 }
445}
446
447static void raid1_end_write_request(struct bio *bio)
448{
449 struct r1bio *r1_bio = bio->bi_private;
450 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
451 struct r1conf *conf = r1_bio->mddev->private;
452 struct bio *to_put = NULL;
453 int mirror = find_bio_disk(r1_bio, bio);
454 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
455 bool discard_error;
456 sector_t lo = r1_bio->sector;
457 sector_t hi = r1_bio->sector + r1_bio->sectors;
458
459 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
460
461 /*
462 * 'one mirror IO has finished' event handler:
463 */
464 if (bio->bi_status && !discard_error) {
465 set_bit(WriteErrorSeen, &rdev->flags);
466 if (!test_and_set_bit(WantReplacement, &rdev->flags))
467 set_bit(MD_RECOVERY_NEEDED, &
468 conf->mddev->recovery);
469
470 if (test_bit(FailFast, &rdev->flags) &&
471 (bio->bi_opf & MD_FAILFAST) &&
472 /* We never try FailFast to WriteMostly devices */
473 !test_bit(WriteMostly, &rdev->flags)) {
474 md_error(r1_bio->mddev, rdev);
475 }
476
477 /*
478 * When the device is faulty, it is not necessary to
479 * handle write error.
480 */
481 if (!test_bit(Faulty, &rdev->flags))
482 set_bit(R1BIO_WriteError, &r1_bio->state);
483 else {
484 /* Fail the request */
485 set_bit(R1BIO_Degraded, &r1_bio->state);
486 /* Finished with this branch */
487 r1_bio->bios[mirror] = NULL;
488 to_put = bio;
489 }
490 } else {
491 /*
492 * Set R1BIO_Uptodate in our master bio, so that we
493 * will return a good error code for to the higher
494 * levels even if IO on some other mirrored buffer
495 * fails.
496 *
497 * The 'master' represents the composite IO operation
498 * to user-side. So if something waits for IO, then it
499 * will wait for the 'master' bio.
500 */
501 sector_t first_bad;
502 int bad_sectors;
503
504 r1_bio->bios[mirror] = NULL;
505 to_put = bio;
506 /*
507 * Do not set R1BIO_Uptodate if the current device is
508 * rebuilding or Faulty. This is because we cannot use
509 * such device for properly reading the data back (we could
510 * potentially use it, if the current write would have felt
511 * before rdev->recovery_offset, but for simplicity we don't
512 * check this here.
513 */
514 if (test_bit(In_sync, &rdev->flags) &&
515 !test_bit(Faulty, &rdev->flags))
516 set_bit(R1BIO_Uptodate, &r1_bio->state);
517
518 /* Maybe we can clear some bad blocks. */
519 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
520 &first_bad, &bad_sectors) && !discard_error) {
521 r1_bio->bios[mirror] = IO_MADE_GOOD;
522 set_bit(R1BIO_MadeGood, &r1_bio->state);
523 }
524 }
525
526 if (behind) {
527 if (test_bit(CollisionCheck, &rdev->flags))
528 remove_serial(rdev, lo, hi);
529 if (test_bit(WriteMostly, &rdev->flags))
530 atomic_dec(&r1_bio->behind_remaining);
531
532 /*
533 * In behind mode, we ACK the master bio once the I/O
534 * has safely reached all non-writemostly
535 * disks. Setting the Returned bit ensures that this
536 * gets done only once -- we don't ever want to return
537 * -EIO here, instead we'll wait
538 */
539 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
540 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
541 /* Maybe we can return now */
542 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
543 struct bio *mbio = r1_bio->master_bio;
544 pr_debug("raid1: behind end write sectors"
545 " %llu-%llu\n",
546 (unsigned long long) mbio->bi_iter.bi_sector,
547 (unsigned long long) bio_end_sector(mbio) - 1);
548 call_bio_endio(r1_bio);
549 }
550 }
551 } else if (rdev->mddev->serialize_policy)
552 remove_serial(rdev, lo, hi);
553 if (r1_bio->bios[mirror] == NULL)
554 rdev_dec_pending(rdev, conf->mddev);
555
556 /*
557 * Let's see if all mirrored write operations have finished
558 * already.
559 */
560 r1_bio_write_done(r1_bio);
561
562 if (to_put)
563 bio_put(to_put);
564}
565
566static sector_t align_to_barrier_unit_end(sector_t start_sector,
567 sector_t sectors)
568{
569 sector_t len;
570
571 WARN_ON(sectors == 0);
572 /*
573 * len is the number of sectors from start_sector to end of the
574 * barrier unit which start_sector belongs to.
575 */
576 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
577 start_sector;
578
579 if (len > sectors)
580 len = sectors;
581
582 return len;
583}
584
585/*
586 * This routine returns the disk from which the requested read should
587 * be done. There is a per-array 'next expected sequential IO' sector
588 * number - if this matches on the next IO then we use the last disk.
589 * There is also a per-disk 'last know head position' sector that is
590 * maintained from IRQ contexts, both the normal and the resync IO
591 * completion handlers update this position correctly. If there is no
592 * perfect sequential match then we pick the disk whose head is closest.
593 *
594 * If there are 2 mirrors in the same 2 devices, performance degrades
595 * because position is mirror, not device based.
596 *
597 * The rdev for the device selected will have nr_pending incremented.
598 */
599static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
600{
601 const sector_t this_sector = r1_bio->sector;
602 int sectors;
603 int best_good_sectors;
604 int best_disk, best_dist_disk, best_pending_disk;
605 int has_nonrot_disk;
606 int disk;
607 sector_t best_dist;
608 unsigned int min_pending;
609 struct md_rdev *rdev;
610 int choose_first;
611 int choose_next_idle;
612
613 /*
614 * Check if we can balance. We can balance on the whole
615 * device if no resync is going on, or below the resync window.
616 * We take the first readable disk when above the resync window.
617 */
618 retry:
619 sectors = r1_bio->sectors;
620 best_disk = -1;
621 best_dist_disk = -1;
622 best_dist = MaxSector;
623 best_pending_disk = -1;
624 min_pending = UINT_MAX;
625 best_good_sectors = 0;
626 has_nonrot_disk = 0;
627 choose_next_idle = 0;
628 clear_bit(R1BIO_FailFast, &r1_bio->state);
629
630 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
631 (mddev_is_clustered(conf->mddev) &&
632 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
633 this_sector + sectors)))
634 choose_first = 1;
635 else
636 choose_first = 0;
637
638 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
639 sector_t dist;
640 sector_t first_bad;
641 int bad_sectors;
642 unsigned int pending;
643 bool nonrot;
644
645 rdev = conf->mirrors[disk].rdev;
646 if (r1_bio->bios[disk] == IO_BLOCKED
647 || rdev == NULL
648 || test_bit(Faulty, &rdev->flags))
649 continue;
650 if (!test_bit(In_sync, &rdev->flags) &&
651 rdev->recovery_offset < this_sector + sectors)
652 continue;
653 if (test_bit(WriteMostly, &rdev->flags)) {
654 /* Don't balance among write-mostly, just
655 * use the first as a last resort */
656 if (best_dist_disk < 0) {
657 if (is_badblock(rdev, this_sector, sectors,
658 &first_bad, &bad_sectors)) {
659 if (first_bad <= this_sector)
660 /* Cannot use this */
661 continue;
662 best_good_sectors = first_bad - this_sector;
663 } else
664 best_good_sectors = sectors;
665 best_dist_disk = disk;
666 best_pending_disk = disk;
667 }
668 continue;
669 }
670 /* This is a reasonable device to use. It might
671 * even be best.
672 */
673 if (is_badblock(rdev, this_sector, sectors,
674 &first_bad, &bad_sectors)) {
675 if (best_dist < MaxSector)
676 /* already have a better device */
677 continue;
678 if (first_bad <= this_sector) {
679 /* cannot read here. If this is the 'primary'
680 * device, then we must not read beyond
681 * bad_sectors from another device..
682 */
683 bad_sectors -= (this_sector - first_bad);
684 if (choose_first && sectors > bad_sectors)
685 sectors = bad_sectors;
686 if (best_good_sectors > sectors)
687 best_good_sectors = sectors;
688
689 } else {
690 sector_t good_sectors = first_bad - this_sector;
691 if (good_sectors > best_good_sectors) {
692 best_good_sectors = good_sectors;
693 best_disk = disk;
694 }
695 if (choose_first)
696 break;
697 }
698 continue;
699 } else {
700 if ((sectors > best_good_sectors) && (best_disk >= 0))
701 best_disk = -1;
702 best_good_sectors = sectors;
703 }
704
705 if (best_disk >= 0)
706 /* At least two disks to choose from so failfast is OK */
707 set_bit(R1BIO_FailFast, &r1_bio->state);
708
709 nonrot = bdev_nonrot(rdev->bdev);
710 has_nonrot_disk |= nonrot;
711 pending = atomic_read(&rdev->nr_pending);
712 dist = abs(this_sector - conf->mirrors[disk].head_position);
713 if (choose_first) {
714 best_disk = disk;
715 break;
716 }
717 /* Don't change to another disk for sequential reads */
718 if (conf->mirrors[disk].next_seq_sect == this_sector
719 || dist == 0) {
720 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
721 struct raid1_info *mirror = &conf->mirrors[disk];
722
723 best_disk = disk;
724 /*
725 * If buffered sequential IO size exceeds optimal
726 * iosize, check if there is idle disk. If yes, choose
727 * the idle disk. read_balance could already choose an
728 * idle disk before noticing it's a sequential IO in
729 * this disk. This doesn't matter because this disk
730 * will idle, next time it will be utilized after the
731 * first disk has IO size exceeds optimal iosize. In
732 * this way, iosize of the first disk will be optimal
733 * iosize at least. iosize of the second disk might be
734 * small, but not a big deal since when the second disk
735 * starts IO, the first disk is likely still busy.
736 */
737 if (nonrot && opt_iosize > 0 &&
738 mirror->seq_start != MaxSector &&
739 mirror->next_seq_sect > opt_iosize &&
740 mirror->next_seq_sect - opt_iosize >=
741 mirror->seq_start) {
742 choose_next_idle = 1;
743 continue;
744 }
745 break;
746 }
747
748 if (choose_next_idle)
749 continue;
750
751 if (min_pending > pending) {
752 min_pending = pending;
753 best_pending_disk = disk;
754 }
755
756 if (dist < best_dist) {
757 best_dist = dist;
758 best_dist_disk = disk;
759 }
760 }
761
762 /*
763 * If all disks are rotational, choose the closest disk. If any disk is
764 * non-rotational, choose the disk with less pending request even the
765 * disk is rotational, which might/might not be optimal for raids with
766 * mixed ratation/non-rotational disks depending on workload.
767 */
768 if (best_disk == -1) {
769 if (has_nonrot_disk || min_pending == 0)
770 best_disk = best_pending_disk;
771 else
772 best_disk = best_dist_disk;
773 }
774
775 if (best_disk >= 0) {
776 rdev = conf->mirrors[best_disk].rdev;
777 if (!rdev)
778 goto retry;
779 atomic_inc(&rdev->nr_pending);
780 sectors = best_good_sectors;
781
782 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
783 conf->mirrors[best_disk].seq_start = this_sector;
784
785 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
786 }
787 *max_sectors = sectors;
788
789 return best_disk;
790}
791
792static void wake_up_barrier(struct r1conf *conf)
793{
794 if (wq_has_sleeper(&conf->wait_barrier))
795 wake_up(&conf->wait_barrier);
796}
797
798static void flush_bio_list(struct r1conf *conf, struct bio *bio)
799{
800 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
801 raid1_prepare_flush_writes(conf->mddev->bitmap);
802 wake_up_barrier(conf);
803
804 while (bio) { /* submit pending writes */
805 struct bio *next = bio->bi_next;
806
807 raid1_submit_write(bio);
808 bio = next;
809 cond_resched();
810 }
811}
812
813static void flush_pending_writes(struct r1conf *conf)
814{
815 /* Any writes that have been queued but are awaiting
816 * bitmap updates get flushed here.
817 */
818 spin_lock_irq(&conf->device_lock);
819
820 if (conf->pending_bio_list.head) {
821 struct blk_plug plug;
822 struct bio *bio;
823
824 bio = bio_list_get(&conf->pending_bio_list);
825 spin_unlock_irq(&conf->device_lock);
826
827 /*
828 * As this is called in a wait_event() loop (see freeze_array),
829 * current->state might be TASK_UNINTERRUPTIBLE which will
830 * cause a warning when we prepare to wait again. As it is
831 * rare that this path is taken, it is perfectly safe to force
832 * us to go around the wait_event() loop again, so the warning
833 * is a false-positive. Silence the warning by resetting
834 * thread state
835 */
836 __set_current_state(TASK_RUNNING);
837 blk_start_plug(&plug);
838 flush_bio_list(conf, bio);
839 blk_finish_plug(&plug);
840 } else
841 spin_unlock_irq(&conf->device_lock);
842}
843
844/* Barriers....
845 * Sometimes we need to suspend IO while we do something else,
846 * either some resync/recovery, or reconfigure the array.
847 * To do this we raise a 'barrier'.
848 * The 'barrier' is a counter that can be raised multiple times
849 * to count how many activities are happening which preclude
850 * normal IO.
851 * We can only raise the barrier if there is no pending IO.
852 * i.e. if nr_pending == 0.
853 * We choose only to raise the barrier if no-one is waiting for the
854 * barrier to go down. This means that as soon as an IO request
855 * is ready, no other operations which require a barrier will start
856 * until the IO request has had a chance.
857 *
858 * So: regular IO calls 'wait_barrier'. When that returns there
859 * is no backgroup IO happening, It must arrange to call
860 * allow_barrier when it has finished its IO.
861 * backgroup IO calls must call raise_barrier. Once that returns
862 * there is no normal IO happeing. It must arrange to call
863 * lower_barrier when the particular background IO completes.
864 *
865 * If resync/recovery is interrupted, returns -EINTR;
866 * Otherwise, returns 0.
867 */
868static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
869{
870 int idx = sector_to_idx(sector_nr);
871
872 spin_lock_irq(&conf->resync_lock);
873
874 /* Wait until no block IO is waiting */
875 wait_event_lock_irq(conf->wait_barrier,
876 !atomic_read(&conf->nr_waiting[idx]),
877 conf->resync_lock);
878
879 /* block any new IO from starting */
880 atomic_inc(&conf->barrier[idx]);
881 /*
882 * In raise_barrier() we firstly increase conf->barrier[idx] then
883 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
884 * increase conf->nr_pending[idx] then check conf->barrier[idx].
885 * A memory barrier here to make sure conf->nr_pending[idx] won't
886 * be fetched before conf->barrier[idx] is increased. Otherwise
887 * there will be a race between raise_barrier() and _wait_barrier().
888 */
889 smp_mb__after_atomic();
890
891 /* For these conditions we must wait:
892 * A: while the array is in frozen state
893 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
894 * existing in corresponding I/O barrier bucket.
895 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
896 * max resync count which allowed on current I/O barrier bucket.
897 */
898 wait_event_lock_irq(conf->wait_barrier,
899 (!conf->array_frozen &&
900 !atomic_read(&conf->nr_pending[idx]) &&
901 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
902 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
903 conf->resync_lock);
904
905 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
906 atomic_dec(&conf->barrier[idx]);
907 spin_unlock_irq(&conf->resync_lock);
908 wake_up(&conf->wait_barrier);
909 return -EINTR;
910 }
911
912 atomic_inc(&conf->nr_sync_pending);
913 spin_unlock_irq(&conf->resync_lock);
914
915 return 0;
916}
917
918static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
919{
920 int idx = sector_to_idx(sector_nr);
921
922 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
923
924 atomic_dec(&conf->barrier[idx]);
925 atomic_dec(&conf->nr_sync_pending);
926 wake_up(&conf->wait_barrier);
927}
928
929static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
930{
931 bool ret = true;
932
933 /*
934 * We need to increase conf->nr_pending[idx] very early here,
935 * then raise_barrier() can be blocked when it waits for
936 * conf->nr_pending[idx] to be 0. Then we can avoid holding
937 * conf->resync_lock when there is no barrier raised in same
938 * barrier unit bucket. Also if the array is frozen, I/O
939 * should be blocked until array is unfrozen.
940 */
941 atomic_inc(&conf->nr_pending[idx]);
942 /*
943 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
944 * check conf->barrier[idx]. In raise_barrier() we firstly increase
945 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
946 * barrier is necessary here to make sure conf->barrier[idx] won't be
947 * fetched before conf->nr_pending[idx] is increased. Otherwise there
948 * will be a race between _wait_barrier() and raise_barrier().
949 */
950 smp_mb__after_atomic();
951
952 /*
953 * Don't worry about checking two atomic_t variables at same time
954 * here. If during we check conf->barrier[idx], the array is
955 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
956 * 0, it is safe to return and make the I/O continue. Because the
957 * array is frozen, all I/O returned here will eventually complete
958 * or be queued, no race will happen. See code comment in
959 * frozen_array().
960 */
961 if (!READ_ONCE(conf->array_frozen) &&
962 !atomic_read(&conf->barrier[idx]))
963 return ret;
964
965 /*
966 * After holding conf->resync_lock, conf->nr_pending[idx]
967 * should be decreased before waiting for barrier to drop.
968 * Otherwise, we may encounter a race condition because
969 * raise_barrer() might be waiting for conf->nr_pending[idx]
970 * to be 0 at same time.
971 */
972 spin_lock_irq(&conf->resync_lock);
973 atomic_inc(&conf->nr_waiting[idx]);
974 atomic_dec(&conf->nr_pending[idx]);
975 /*
976 * In case freeze_array() is waiting for
977 * get_unqueued_pending() == extra
978 */
979 wake_up_barrier(conf);
980 /* Wait for the barrier in same barrier unit bucket to drop. */
981
982 /* Return false when nowait flag is set */
983 if (nowait) {
984 ret = false;
985 } else {
986 wait_event_lock_irq(conf->wait_barrier,
987 !conf->array_frozen &&
988 !atomic_read(&conf->barrier[idx]),
989 conf->resync_lock);
990 atomic_inc(&conf->nr_pending[idx]);
991 }
992
993 atomic_dec(&conf->nr_waiting[idx]);
994 spin_unlock_irq(&conf->resync_lock);
995 return ret;
996}
997
998static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
999{
1000 int idx = sector_to_idx(sector_nr);
1001 bool ret = true;
1002
1003 /*
1004 * Very similar to _wait_barrier(). The difference is, for read
1005 * I/O we don't need wait for sync I/O, but if the whole array
1006 * is frozen, the read I/O still has to wait until the array is
1007 * unfrozen. Since there is no ordering requirement with
1008 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1009 */
1010 atomic_inc(&conf->nr_pending[idx]);
1011
1012 if (!READ_ONCE(conf->array_frozen))
1013 return ret;
1014
1015 spin_lock_irq(&conf->resync_lock);
1016 atomic_inc(&conf->nr_waiting[idx]);
1017 atomic_dec(&conf->nr_pending[idx]);
1018 /*
1019 * In case freeze_array() is waiting for
1020 * get_unqueued_pending() == extra
1021 */
1022 wake_up_barrier(conf);
1023 /* Wait for array to be unfrozen */
1024
1025 /* Return false when nowait flag is set */
1026 if (nowait) {
1027 /* Return false when nowait flag is set */
1028 ret = false;
1029 } else {
1030 wait_event_lock_irq(conf->wait_barrier,
1031 !conf->array_frozen,
1032 conf->resync_lock);
1033 atomic_inc(&conf->nr_pending[idx]);
1034 }
1035
1036 atomic_dec(&conf->nr_waiting[idx]);
1037 spin_unlock_irq(&conf->resync_lock);
1038 return ret;
1039}
1040
1041static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1042{
1043 int idx = sector_to_idx(sector_nr);
1044
1045 return _wait_barrier(conf, idx, nowait);
1046}
1047
1048static void _allow_barrier(struct r1conf *conf, int idx)
1049{
1050 atomic_dec(&conf->nr_pending[idx]);
1051 wake_up_barrier(conf);
1052}
1053
1054static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1055{
1056 int idx = sector_to_idx(sector_nr);
1057
1058 _allow_barrier(conf, idx);
1059}
1060
1061/* conf->resync_lock should be held */
1062static int get_unqueued_pending(struct r1conf *conf)
1063{
1064 int idx, ret;
1065
1066 ret = atomic_read(&conf->nr_sync_pending);
1067 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1068 ret += atomic_read(&conf->nr_pending[idx]) -
1069 atomic_read(&conf->nr_queued[idx]);
1070
1071 return ret;
1072}
1073
1074static void freeze_array(struct r1conf *conf, int extra)
1075{
1076 /* Stop sync I/O and normal I/O and wait for everything to
1077 * go quiet.
1078 * This is called in two situations:
1079 * 1) management command handlers (reshape, remove disk, quiesce).
1080 * 2) one normal I/O request failed.
1081
1082 * After array_frozen is set to 1, new sync IO will be blocked at
1083 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1084 * or wait_read_barrier(). The flying I/Os will either complete or be
1085 * queued. When everything goes quite, there are only queued I/Os left.
1086
1087 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1088 * barrier bucket index which this I/O request hits. When all sync and
1089 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1090 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1091 * in handle_read_error(), we may call freeze_array() before trying to
1092 * fix the read error. In this case, the error read I/O is not queued,
1093 * so get_unqueued_pending() == 1.
1094 *
1095 * Therefore before this function returns, we need to wait until
1096 * get_unqueued_pendings(conf) gets equal to extra. For
1097 * normal I/O context, extra is 1, in rested situations extra is 0.
1098 */
1099 spin_lock_irq(&conf->resync_lock);
1100 conf->array_frozen = 1;
1101 raid1_log(conf->mddev, "wait freeze");
1102 wait_event_lock_irq_cmd(
1103 conf->wait_barrier,
1104 get_unqueued_pending(conf) == extra,
1105 conf->resync_lock,
1106 flush_pending_writes(conf));
1107 spin_unlock_irq(&conf->resync_lock);
1108}
1109static void unfreeze_array(struct r1conf *conf)
1110{
1111 /* reverse the effect of the freeze */
1112 spin_lock_irq(&conf->resync_lock);
1113 conf->array_frozen = 0;
1114 spin_unlock_irq(&conf->resync_lock);
1115 wake_up(&conf->wait_barrier);
1116}
1117
1118static void alloc_behind_master_bio(struct r1bio *r1_bio,
1119 struct bio *bio)
1120{
1121 int size = bio->bi_iter.bi_size;
1122 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1123 int i = 0;
1124 struct bio *behind_bio = NULL;
1125
1126 behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1127 &r1_bio->mddev->bio_set);
1128
1129 /* discard op, we don't support writezero/writesame yet */
1130 if (!bio_has_data(bio)) {
1131 behind_bio->bi_iter.bi_size = size;
1132 goto skip_copy;
1133 }
1134
1135 while (i < vcnt && size) {
1136 struct page *page;
1137 int len = min_t(int, PAGE_SIZE, size);
1138
1139 page = alloc_page(GFP_NOIO);
1140 if (unlikely(!page))
1141 goto free_pages;
1142
1143 if (!bio_add_page(behind_bio, page, len, 0)) {
1144 put_page(page);
1145 goto free_pages;
1146 }
1147
1148 size -= len;
1149 i++;
1150 }
1151
1152 bio_copy_data(behind_bio, bio);
1153skip_copy:
1154 r1_bio->behind_master_bio = behind_bio;
1155 set_bit(R1BIO_BehindIO, &r1_bio->state);
1156
1157 return;
1158
1159free_pages:
1160 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1161 bio->bi_iter.bi_size);
1162 bio_free_pages(behind_bio);
1163 bio_put(behind_bio);
1164}
1165
1166static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1167{
1168 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1169 cb);
1170 struct mddev *mddev = plug->cb.data;
1171 struct r1conf *conf = mddev->private;
1172 struct bio *bio;
1173
1174 if (from_schedule) {
1175 spin_lock_irq(&conf->device_lock);
1176 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1177 spin_unlock_irq(&conf->device_lock);
1178 wake_up_barrier(conf);
1179 md_wakeup_thread(mddev->thread);
1180 kfree(plug);
1181 return;
1182 }
1183
1184 /* we aren't scheduling, so we can do the write-out directly. */
1185 bio = bio_list_get(&plug->pending);
1186 flush_bio_list(conf, bio);
1187 kfree(plug);
1188}
1189
1190static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1191{
1192 r1_bio->master_bio = bio;
1193 r1_bio->sectors = bio_sectors(bio);
1194 r1_bio->state = 0;
1195 r1_bio->mddev = mddev;
1196 r1_bio->sector = bio->bi_iter.bi_sector;
1197}
1198
1199static inline struct r1bio *
1200alloc_r1bio(struct mddev *mddev, struct bio *bio)
1201{
1202 struct r1conf *conf = mddev->private;
1203 struct r1bio *r1_bio;
1204
1205 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1206 /* Ensure no bio records IO_BLOCKED */
1207 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1208 init_r1bio(r1_bio, mddev, bio);
1209 return r1_bio;
1210}
1211
1212static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1213 int max_read_sectors, struct r1bio *r1_bio)
1214{
1215 struct r1conf *conf = mddev->private;
1216 struct raid1_info *mirror;
1217 struct bio *read_bio;
1218 struct bitmap *bitmap = mddev->bitmap;
1219 const enum req_op op = bio_op(bio);
1220 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1221 int max_sectors;
1222 int rdisk;
1223 bool r1bio_existed = !!r1_bio;
1224 char b[BDEVNAME_SIZE];
1225
1226 /*
1227 * If r1_bio is set, we are blocking the raid1d thread
1228 * so there is a tiny risk of deadlock. So ask for
1229 * emergency memory if needed.
1230 */
1231 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1232
1233 if (r1bio_existed) {
1234 /* Need to get the block device name carefully */
1235 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
1236
1237 if (rdev)
1238 snprintf(b, sizeof(b), "%pg", rdev->bdev);
1239 else
1240 strcpy(b, "???");
1241 }
1242
1243 /*
1244 * Still need barrier for READ in case that whole
1245 * array is frozen.
1246 */
1247 if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1248 bio->bi_opf & REQ_NOWAIT)) {
1249 bio_wouldblock_error(bio);
1250 return;
1251 }
1252
1253 if (!r1_bio)
1254 r1_bio = alloc_r1bio(mddev, bio);
1255 else
1256 init_r1bio(r1_bio, mddev, bio);
1257 r1_bio->sectors = max_read_sectors;
1258
1259 /*
1260 * make_request() can abort the operation when read-ahead is being
1261 * used and no empty request is available.
1262 */
1263 rdisk = read_balance(conf, r1_bio, &max_sectors);
1264
1265 if (rdisk < 0) {
1266 /* couldn't find anywhere to read from */
1267 if (r1bio_existed) {
1268 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1269 mdname(mddev),
1270 b,
1271 (unsigned long long)r1_bio->sector);
1272 }
1273 raid_end_bio_io(r1_bio);
1274 return;
1275 }
1276 mirror = conf->mirrors + rdisk;
1277
1278 if (r1bio_existed)
1279 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1280 mdname(mddev),
1281 (unsigned long long)r1_bio->sector,
1282 mirror->rdev->bdev);
1283
1284 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1285 bitmap) {
1286 /*
1287 * Reading from a write-mostly device must take care not to
1288 * over-take any writes that are 'behind'
1289 */
1290 raid1_log(mddev, "wait behind writes");
1291 wait_event(bitmap->behind_wait,
1292 atomic_read(&bitmap->behind_writes) == 0);
1293 }
1294
1295 if (max_sectors < bio_sectors(bio)) {
1296 struct bio *split = bio_split(bio, max_sectors,
1297 gfp, &conf->bio_split);
1298 bio_chain(split, bio);
1299 submit_bio_noacct(bio);
1300 bio = split;
1301 r1_bio->master_bio = bio;
1302 r1_bio->sectors = max_sectors;
1303 }
1304
1305 r1_bio->read_disk = rdisk;
1306 if (!r1bio_existed) {
1307 md_account_bio(mddev, &bio);
1308 r1_bio->master_bio = bio;
1309 }
1310 read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1311 &mddev->bio_set);
1312
1313 r1_bio->bios[rdisk] = read_bio;
1314
1315 read_bio->bi_iter.bi_sector = r1_bio->sector +
1316 mirror->rdev->data_offset;
1317 read_bio->bi_end_io = raid1_end_read_request;
1318 read_bio->bi_opf = op | do_sync;
1319 if (test_bit(FailFast, &mirror->rdev->flags) &&
1320 test_bit(R1BIO_FailFast, &r1_bio->state))
1321 read_bio->bi_opf |= MD_FAILFAST;
1322 read_bio->bi_private = r1_bio;
1323
1324 if (mddev->gendisk)
1325 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1326 r1_bio->sector);
1327
1328 submit_bio_noacct(read_bio);
1329}
1330
1331static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1332 int max_write_sectors)
1333{
1334 struct r1conf *conf = mddev->private;
1335 struct r1bio *r1_bio;
1336 int i, disks;
1337 struct bitmap *bitmap = mddev->bitmap;
1338 unsigned long flags;
1339 struct md_rdev *blocked_rdev;
1340 int first_clone;
1341 int max_sectors;
1342 bool write_behind = false;
1343 bool is_discard = (bio_op(bio) == REQ_OP_DISCARD);
1344
1345 if (mddev_is_clustered(mddev) &&
1346 md_cluster_ops->area_resyncing(mddev, WRITE,
1347 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1348
1349 DEFINE_WAIT(w);
1350 if (bio->bi_opf & REQ_NOWAIT) {
1351 bio_wouldblock_error(bio);
1352 return;
1353 }
1354 for (;;) {
1355 prepare_to_wait(&conf->wait_barrier,
1356 &w, TASK_IDLE);
1357 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1358 bio->bi_iter.bi_sector,
1359 bio_end_sector(bio)))
1360 break;
1361 schedule();
1362 }
1363 finish_wait(&conf->wait_barrier, &w);
1364 }
1365
1366 /*
1367 * Register the new request and wait if the reconstruction
1368 * thread has put up a bar for new requests.
1369 * Continue immediately if no resync is active currently.
1370 */
1371 if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1372 bio->bi_opf & REQ_NOWAIT)) {
1373 bio_wouldblock_error(bio);
1374 return;
1375 }
1376
1377 retry_write:
1378 r1_bio = alloc_r1bio(mddev, bio);
1379 r1_bio->sectors = max_write_sectors;
1380
1381 /* first select target devices under rcu_lock and
1382 * inc refcount on their rdev. Record them by setting
1383 * bios[x] to bio
1384 * If there are known/acknowledged bad blocks on any device on
1385 * which we have seen a write error, we want to avoid writing those
1386 * blocks.
1387 * This potentially requires several writes to write around
1388 * the bad blocks. Each set of writes gets it's own r1bio
1389 * with a set of bios attached.
1390 */
1391
1392 disks = conf->raid_disks * 2;
1393 blocked_rdev = NULL;
1394 max_sectors = r1_bio->sectors;
1395 for (i = 0; i < disks; i++) {
1396 struct md_rdev *rdev = conf->mirrors[i].rdev;
1397
1398 /*
1399 * The write-behind io is only attempted on drives marked as
1400 * write-mostly, which means we could allocate write behind
1401 * bio later.
1402 */
1403 if (!is_discard && rdev && test_bit(WriteMostly, &rdev->flags))
1404 write_behind = true;
1405
1406 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1407 atomic_inc(&rdev->nr_pending);
1408 blocked_rdev = rdev;
1409 break;
1410 }
1411 r1_bio->bios[i] = NULL;
1412 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1413 if (i < conf->raid_disks)
1414 set_bit(R1BIO_Degraded, &r1_bio->state);
1415 continue;
1416 }
1417
1418 atomic_inc(&rdev->nr_pending);
1419 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1420 sector_t first_bad;
1421 int bad_sectors;
1422 int is_bad;
1423
1424 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1425 &first_bad, &bad_sectors);
1426 if (is_bad < 0) {
1427 /* mustn't write here until the bad block is
1428 * acknowledged*/
1429 set_bit(BlockedBadBlocks, &rdev->flags);
1430 blocked_rdev = rdev;
1431 break;
1432 }
1433 if (is_bad && first_bad <= r1_bio->sector) {
1434 /* Cannot write here at all */
1435 bad_sectors -= (r1_bio->sector - first_bad);
1436 if (bad_sectors < max_sectors)
1437 /* mustn't write more than bad_sectors
1438 * to other devices yet
1439 */
1440 max_sectors = bad_sectors;
1441 rdev_dec_pending(rdev, mddev);
1442 /* We don't set R1BIO_Degraded as that
1443 * only applies if the disk is
1444 * missing, so it might be re-added,
1445 * and we want to know to recover this
1446 * chunk.
1447 * In this case the device is here,
1448 * and the fact that this chunk is not
1449 * in-sync is recorded in the bad
1450 * block log
1451 */
1452 continue;
1453 }
1454 if (is_bad) {
1455 int good_sectors = first_bad - r1_bio->sector;
1456 if (good_sectors < max_sectors)
1457 max_sectors = good_sectors;
1458 }
1459 }
1460 r1_bio->bios[i] = bio;
1461 }
1462
1463 if (unlikely(blocked_rdev)) {
1464 /* Wait for this device to become unblocked */
1465 int j;
1466
1467 for (j = 0; j < i; j++)
1468 if (r1_bio->bios[j])
1469 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1470 free_r1bio(r1_bio);
1471 allow_barrier(conf, bio->bi_iter.bi_sector);
1472
1473 if (bio->bi_opf & REQ_NOWAIT) {
1474 bio_wouldblock_error(bio);
1475 return;
1476 }
1477 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1478 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1479 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1480 goto retry_write;
1481 }
1482
1483 /*
1484 * When using a bitmap, we may call alloc_behind_master_bio below.
1485 * alloc_behind_master_bio allocates a copy of the data payload a page
1486 * at a time and thus needs a new bio that can fit the whole payload
1487 * this bio in page sized chunks.
1488 */
1489 if (write_behind && bitmap)
1490 max_sectors = min_t(int, max_sectors,
1491 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1492 if (max_sectors < bio_sectors(bio)) {
1493 struct bio *split = bio_split(bio, max_sectors,
1494 GFP_NOIO, &conf->bio_split);
1495 bio_chain(split, bio);
1496 submit_bio_noacct(bio);
1497 bio = split;
1498 r1_bio->master_bio = bio;
1499 r1_bio->sectors = max_sectors;
1500 }
1501
1502 md_account_bio(mddev, &bio);
1503 r1_bio->master_bio = bio;
1504 atomic_set(&r1_bio->remaining, 1);
1505 atomic_set(&r1_bio->behind_remaining, 0);
1506
1507 first_clone = 1;
1508
1509 for (i = 0; i < disks; i++) {
1510 struct bio *mbio = NULL;
1511 struct md_rdev *rdev = conf->mirrors[i].rdev;
1512 if (!r1_bio->bios[i])
1513 continue;
1514
1515 if (first_clone) {
1516 /* do behind I/O ?
1517 * Not if there are too many, or cannot
1518 * allocate memory, or a reader on WriteMostly
1519 * is waiting for behind writes to flush */
1520 if (bitmap && write_behind &&
1521 (atomic_read(&bitmap->behind_writes)
1522 < mddev->bitmap_info.max_write_behind) &&
1523 !waitqueue_active(&bitmap->behind_wait)) {
1524 alloc_behind_master_bio(r1_bio, bio);
1525 }
1526
1527 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1528 test_bit(R1BIO_BehindIO, &r1_bio->state));
1529 first_clone = 0;
1530 }
1531
1532 if (r1_bio->behind_master_bio) {
1533 mbio = bio_alloc_clone(rdev->bdev,
1534 r1_bio->behind_master_bio,
1535 GFP_NOIO, &mddev->bio_set);
1536 if (test_bit(CollisionCheck, &rdev->flags))
1537 wait_for_serialization(rdev, r1_bio);
1538 if (test_bit(WriteMostly, &rdev->flags))
1539 atomic_inc(&r1_bio->behind_remaining);
1540 } else {
1541 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1542 &mddev->bio_set);
1543
1544 if (mddev->serialize_policy)
1545 wait_for_serialization(rdev, r1_bio);
1546 }
1547
1548 r1_bio->bios[i] = mbio;
1549
1550 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1551 mbio->bi_end_io = raid1_end_write_request;
1552 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1553 if (test_bit(FailFast, &rdev->flags) &&
1554 !test_bit(WriteMostly, &rdev->flags) &&
1555 conf->raid_disks - mddev->degraded > 1)
1556 mbio->bi_opf |= MD_FAILFAST;
1557 mbio->bi_private = r1_bio;
1558
1559 atomic_inc(&r1_bio->remaining);
1560
1561 if (mddev->gendisk)
1562 trace_block_bio_remap(mbio, disk_devt(mddev->gendisk),
1563 r1_bio->sector);
1564 /* flush_pending_writes() needs access to the rdev so...*/
1565 mbio->bi_bdev = (void *)rdev;
1566 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1567 spin_lock_irqsave(&conf->device_lock, flags);
1568 bio_list_add(&conf->pending_bio_list, mbio);
1569 spin_unlock_irqrestore(&conf->device_lock, flags);
1570 md_wakeup_thread(mddev->thread);
1571 }
1572 }
1573
1574 r1_bio_write_done(r1_bio);
1575
1576 /* In case raid1d snuck in to freeze_array */
1577 wake_up_barrier(conf);
1578}
1579
1580static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1581{
1582 sector_t sectors;
1583
1584 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1585 && md_flush_request(mddev, bio))
1586 return true;
1587
1588 /*
1589 * There is a limit to the maximum size, but
1590 * the read/write handler might find a lower limit
1591 * due to bad blocks. To avoid multiple splits,
1592 * we pass the maximum number of sectors down
1593 * and let the lower level perform the split.
1594 */
1595 sectors = align_to_barrier_unit_end(
1596 bio->bi_iter.bi_sector, bio_sectors(bio));
1597
1598 if (bio_data_dir(bio) == READ)
1599 raid1_read_request(mddev, bio, sectors, NULL);
1600 else {
1601 if (!md_write_start(mddev,bio))
1602 return false;
1603 raid1_write_request(mddev, bio, sectors);
1604 }
1605 return true;
1606}
1607
1608static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1609{
1610 struct r1conf *conf = mddev->private;
1611 int i;
1612
1613 lockdep_assert_held(&mddev->lock);
1614
1615 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1616 conf->raid_disks - mddev->degraded);
1617 for (i = 0; i < conf->raid_disks; i++) {
1618 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1619
1620 seq_printf(seq, "%s",
1621 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1622 }
1623 seq_printf(seq, "]");
1624}
1625
1626/**
1627 * raid1_error() - RAID1 error handler.
1628 * @mddev: affected md device.
1629 * @rdev: member device to fail.
1630 *
1631 * The routine acknowledges &rdev failure and determines new @mddev state.
1632 * If it failed, then:
1633 * - &MD_BROKEN flag is set in &mddev->flags.
1634 * - recovery is disabled.
1635 * Otherwise, it must be degraded:
1636 * - recovery is interrupted.
1637 * - &mddev->degraded is bumped.
1638 *
1639 * @rdev is marked as &Faulty excluding case when array is failed and
1640 * &mddev->fail_last_dev is off.
1641 */
1642static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1643{
1644 struct r1conf *conf = mddev->private;
1645 unsigned long flags;
1646
1647 spin_lock_irqsave(&conf->device_lock, flags);
1648
1649 if (test_bit(In_sync, &rdev->flags) &&
1650 (conf->raid_disks - mddev->degraded) == 1) {
1651 set_bit(MD_BROKEN, &mddev->flags);
1652
1653 if (!mddev->fail_last_dev) {
1654 conf->recovery_disabled = mddev->recovery_disabled;
1655 spin_unlock_irqrestore(&conf->device_lock, flags);
1656 return;
1657 }
1658 }
1659 set_bit(Blocked, &rdev->flags);
1660 if (test_and_clear_bit(In_sync, &rdev->flags))
1661 mddev->degraded++;
1662 set_bit(Faulty, &rdev->flags);
1663 spin_unlock_irqrestore(&conf->device_lock, flags);
1664 /*
1665 * if recovery is running, make sure it aborts.
1666 */
1667 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1668 set_mask_bits(&mddev->sb_flags, 0,
1669 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1670 pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1671 "md/raid1:%s: Operation continuing on %d devices.\n",
1672 mdname(mddev), rdev->bdev,
1673 mdname(mddev), conf->raid_disks - mddev->degraded);
1674}
1675
1676static void print_conf(struct r1conf *conf)
1677{
1678 int i;
1679
1680 pr_debug("RAID1 conf printout:\n");
1681 if (!conf) {
1682 pr_debug("(!conf)\n");
1683 return;
1684 }
1685 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1686 conf->raid_disks);
1687
1688 lockdep_assert_held(&conf->mddev->reconfig_mutex);
1689 for (i = 0; i < conf->raid_disks; i++) {
1690 struct md_rdev *rdev = conf->mirrors[i].rdev;
1691 if (rdev)
1692 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1693 i, !test_bit(In_sync, &rdev->flags),
1694 !test_bit(Faulty, &rdev->flags),
1695 rdev->bdev);
1696 }
1697}
1698
1699static void close_sync(struct r1conf *conf)
1700{
1701 int idx;
1702
1703 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1704 _wait_barrier(conf, idx, false);
1705 _allow_barrier(conf, idx);
1706 }
1707
1708 mempool_exit(&conf->r1buf_pool);
1709}
1710
1711static int raid1_spare_active(struct mddev *mddev)
1712{
1713 int i;
1714 struct r1conf *conf = mddev->private;
1715 int count = 0;
1716 unsigned long flags;
1717
1718 /*
1719 * Find all failed disks within the RAID1 configuration
1720 * and mark them readable.
1721 * Called under mddev lock, so rcu protection not needed.
1722 * device_lock used to avoid races with raid1_end_read_request
1723 * which expects 'In_sync' flags and ->degraded to be consistent.
1724 */
1725 spin_lock_irqsave(&conf->device_lock, flags);
1726 for (i = 0; i < conf->raid_disks; i++) {
1727 struct md_rdev *rdev = conf->mirrors[i].rdev;
1728 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1729 if (repl
1730 && !test_bit(Candidate, &repl->flags)
1731 && repl->recovery_offset == MaxSector
1732 && !test_bit(Faulty, &repl->flags)
1733 && !test_and_set_bit(In_sync, &repl->flags)) {
1734 /* replacement has just become active */
1735 if (!rdev ||
1736 !test_and_clear_bit(In_sync, &rdev->flags))
1737 count++;
1738 if (rdev) {
1739 /* Replaced device not technically
1740 * faulty, but we need to be sure
1741 * it gets removed and never re-added
1742 */
1743 set_bit(Faulty, &rdev->flags);
1744 sysfs_notify_dirent_safe(
1745 rdev->sysfs_state);
1746 }
1747 }
1748 if (rdev
1749 && rdev->recovery_offset == MaxSector
1750 && !test_bit(Faulty, &rdev->flags)
1751 && !test_and_set_bit(In_sync, &rdev->flags)) {
1752 count++;
1753 sysfs_notify_dirent_safe(rdev->sysfs_state);
1754 }
1755 }
1756 mddev->degraded -= count;
1757 spin_unlock_irqrestore(&conf->device_lock, flags);
1758
1759 print_conf(conf);
1760 return count;
1761}
1762
1763static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1764{
1765 struct r1conf *conf = mddev->private;
1766 int err = -EEXIST;
1767 int mirror = 0, repl_slot = -1;
1768 struct raid1_info *p;
1769 int first = 0;
1770 int last = conf->raid_disks - 1;
1771
1772 if (mddev->recovery_disabled == conf->recovery_disabled)
1773 return -EBUSY;
1774
1775 if (md_integrity_add_rdev(rdev, mddev))
1776 return -ENXIO;
1777
1778 if (rdev->raid_disk >= 0)
1779 first = last = rdev->raid_disk;
1780
1781 /*
1782 * find the disk ... but prefer rdev->saved_raid_disk
1783 * if possible.
1784 */
1785 if (rdev->saved_raid_disk >= 0 &&
1786 rdev->saved_raid_disk >= first &&
1787 rdev->saved_raid_disk < conf->raid_disks &&
1788 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1789 first = last = rdev->saved_raid_disk;
1790
1791 for (mirror = first; mirror <= last; mirror++) {
1792 p = conf->mirrors + mirror;
1793 if (!p->rdev) {
1794 if (mddev->gendisk)
1795 disk_stack_limits(mddev->gendisk, rdev->bdev,
1796 rdev->data_offset << 9);
1797
1798 p->head_position = 0;
1799 rdev->raid_disk = mirror;
1800 err = 0;
1801 /* As all devices are equivalent, we don't need a full recovery
1802 * if this was recently any drive of the array
1803 */
1804 if (rdev->saved_raid_disk < 0)
1805 conf->fullsync = 1;
1806 WRITE_ONCE(p->rdev, rdev);
1807 break;
1808 }
1809 if (test_bit(WantReplacement, &p->rdev->flags) &&
1810 p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1811 repl_slot = mirror;
1812 }
1813
1814 if (err && repl_slot >= 0) {
1815 /* Add this device as a replacement */
1816 p = conf->mirrors + repl_slot;
1817 clear_bit(In_sync, &rdev->flags);
1818 set_bit(Replacement, &rdev->flags);
1819 rdev->raid_disk = repl_slot;
1820 err = 0;
1821 conf->fullsync = 1;
1822 WRITE_ONCE(p[conf->raid_disks].rdev, rdev);
1823 }
1824
1825 print_conf(conf);
1826 return err;
1827}
1828
1829static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1830{
1831 struct r1conf *conf = mddev->private;
1832 int err = 0;
1833 int number = rdev->raid_disk;
1834 struct raid1_info *p = conf->mirrors + number;
1835
1836 if (unlikely(number >= conf->raid_disks))
1837 goto abort;
1838
1839 if (rdev != p->rdev)
1840 p = conf->mirrors + conf->raid_disks + number;
1841
1842 print_conf(conf);
1843 if (rdev == p->rdev) {
1844 if (test_bit(In_sync, &rdev->flags) ||
1845 atomic_read(&rdev->nr_pending)) {
1846 err = -EBUSY;
1847 goto abort;
1848 }
1849 /* Only remove non-faulty devices if recovery
1850 * is not possible.
1851 */
1852 if (!test_bit(Faulty, &rdev->flags) &&
1853 mddev->recovery_disabled != conf->recovery_disabled &&
1854 mddev->degraded < conf->raid_disks) {
1855 err = -EBUSY;
1856 goto abort;
1857 }
1858 WRITE_ONCE(p->rdev, NULL);
1859 if (conf->mirrors[conf->raid_disks + number].rdev) {
1860 /* We just removed a device that is being replaced.
1861 * Move down the replacement. We drain all IO before
1862 * doing this to avoid confusion.
1863 */
1864 struct md_rdev *repl =
1865 conf->mirrors[conf->raid_disks + number].rdev;
1866 freeze_array(conf, 0);
1867 if (atomic_read(&repl->nr_pending)) {
1868 /* It means that some queued IO of retry_list
1869 * hold repl. Thus, we cannot set replacement
1870 * as NULL, avoiding rdev NULL pointer
1871 * dereference in sync_request_write and
1872 * handle_write_finished.
1873 */
1874 err = -EBUSY;
1875 unfreeze_array(conf);
1876 goto abort;
1877 }
1878 clear_bit(Replacement, &repl->flags);
1879 WRITE_ONCE(p->rdev, repl);
1880 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1881 unfreeze_array(conf);
1882 }
1883
1884 clear_bit(WantReplacement, &rdev->flags);
1885 err = md_integrity_register(mddev);
1886 }
1887abort:
1888
1889 print_conf(conf);
1890 return err;
1891}
1892
1893static void end_sync_read(struct bio *bio)
1894{
1895 struct r1bio *r1_bio = get_resync_r1bio(bio);
1896
1897 update_head_pos(r1_bio->read_disk, r1_bio);
1898
1899 /*
1900 * we have read a block, now it needs to be re-written,
1901 * or re-read if the read failed.
1902 * We don't do much here, just schedule handling by raid1d
1903 */
1904 if (!bio->bi_status)
1905 set_bit(R1BIO_Uptodate, &r1_bio->state);
1906
1907 if (atomic_dec_and_test(&r1_bio->remaining))
1908 reschedule_retry(r1_bio);
1909}
1910
1911static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
1912{
1913 sector_t sync_blocks = 0;
1914 sector_t s = r1_bio->sector;
1915 long sectors_to_go = r1_bio->sectors;
1916
1917 /* make sure these bits don't get cleared. */
1918 do {
1919 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
1920 s += sync_blocks;
1921 sectors_to_go -= sync_blocks;
1922 } while (sectors_to_go > 0);
1923}
1924
1925static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
1926{
1927 if (atomic_dec_and_test(&r1_bio->remaining)) {
1928 struct mddev *mddev = r1_bio->mddev;
1929 int s = r1_bio->sectors;
1930
1931 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1932 test_bit(R1BIO_WriteError, &r1_bio->state))
1933 reschedule_retry(r1_bio);
1934 else {
1935 put_buf(r1_bio);
1936 md_done_sync(mddev, s, uptodate);
1937 }
1938 }
1939}
1940
1941static void end_sync_write(struct bio *bio)
1942{
1943 int uptodate = !bio->bi_status;
1944 struct r1bio *r1_bio = get_resync_r1bio(bio);
1945 struct mddev *mddev = r1_bio->mddev;
1946 struct r1conf *conf = mddev->private;
1947 sector_t first_bad;
1948 int bad_sectors;
1949 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1950
1951 if (!uptodate) {
1952 abort_sync_write(mddev, r1_bio);
1953 set_bit(WriteErrorSeen, &rdev->flags);
1954 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1955 set_bit(MD_RECOVERY_NEEDED, &
1956 mddev->recovery);
1957 set_bit(R1BIO_WriteError, &r1_bio->state);
1958 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1959 &first_bad, &bad_sectors) &&
1960 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1961 r1_bio->sector,
1962 r1_bio->sectors,
1963 &first_bad, &bad_sectors)
1964 )
1965 set_bit(R1BIO_MadeGood, &r1_bio->state);
1966
1967 put_sync_write_buf(r1_bio, uptodate);
1968}
1969
1970static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1971 int sectors, struct page *page, blk_opf_t rw)
1972{
1973 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1974 /* success */
1975 return 1;
1976 if (rw == REQ_OP_WRITE) {
1977 set_bit(WriteErrorSeen, &rdev->flags);
1978 if (!test_and_set_bit(WantReplacement,
1979 &rdev->flags))
1980 set_bit(MD_RECOVERY_NEEDED, &
1981 rdev->mddev->recovery);
1982 }
1983 /* need to record an error - either for the block or the device */
1984 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1985 md_error(rdev->mddev, rdev);
1986 return 0;
1987}
1988
1989static int fix_sync_read_error(struct r1bio *r1_bio)
1990{
1991 /* Try some synchronous reads of other devices to get
1992 * good data, much like with normal read errors. Only
1993 * read into the pages we already have so we don't
1994 * need to re-issue the read request.
1995 * We don't need to freeze the array, because being in an
1996 * active sync request, there is no normal IO, and
1997 * no overlapping syncs.
1998 * We don't need to check is_badblock() again as we
1999 * made sure that anything with a bad block in range
2000 * will have bi_end_io clear.
2001 */
2002 struct mddev *mddev = r1_bio->mddev;
2003 struct r1conf *conf = mddev->private;
2004 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2005 struct page **pages = get_resync_pages(bio)->pages;
2006 sector_t sect = r1_bio->sector;
2007 int sectors = r1_bio->sectors;
2008 int idx = 0;
2009 struct md_rdev *rdev;
2010
2011 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2012 if (test_bit(FailFast, &rdev->flags)) {
2013 /* Don't try recovering from here - just fail it
2014 * ... unless it is the last working device of course */
2015 md_error(mddev, rdev);
2016 if (test_bit(Faulty, &rdev->flags))
2017 /* Don't try to read from here, but make sure
2018 * put_buf does it's thing
2019 */
2020 bio->bi_end_io = end_sync_write;
2021 }
2022
2023 while(sectors) {
2024 int s = sectors;
2025 int d = r1_bio->read_disk;
2026 int success = 0;
2027 int start;
2028
2029 if (s > (PAGE_SIZE>>9))
2030 s = PAGE_SIZE >> 9;
2031 do {
2032 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2033 /* No rcu protection needed here devices
2034 * can only be removed when no resync is
2035 * active, and resync is currently active
2036 */
2037 rdev = conf->mirrors[d].rdev;
2038 if (sync_page_io(rdev, sect, s<<9,
2039 pages[idx],
2040 REQ_OP_READ, false)) {
2041 success = 1;
2042 break;
2043 }
2044 }
2045 d++;
2046 if (d == conf->raid_disks * 2)
2047 d = 0;
2048 } while (!success && d != r1_bio->read_disk);
2049
2050 if (!success) {
2051 int abort = 0;
2052 /* Cannot read from anywhere, this block is lost.
2053 * Record a bad block on each device. If that doesn't
2054 * work just disable and interrupt the recovery.
2055 * Don't fail devices as that won't really help.
2056 */
2057 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2058 mdname(mddev), bio->bi_bdev,
2059 (unsigned long long)r1_bio->sector);
2060 for (d = 0; d < conf->raid_disks * 2; d++) {
2061 rdev = conf->mirrors[d].rdev;
2062 if (!rdev || test_bit(Faulty, &rdev->flags))
2063 continue;
2064 if (!rdev_set_badblocks(rdev, sect, s, 0))
2065 abort = 1;
2066 }
2067 if (abort) {
2068 conf->recovery_disabled =
2069 mddev->recovery_disabled;
2070 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2071 md_done_sync(mddev, r1_bio->sectors, 0);
2072 put_buf(r1_bio);
2073 return 0;
2074 }
2075 /* Try next page */
2076 sectors -= s;
2077 sect += s;
2078 idx++;
2079 continue;
2080 }
2081
2082 start = d;
2083 /* write it back and re-read */
2084 while (d != r1_bio->read_disk) {
2085 if (d == 0)
2086 d = conf->raid_disks * 2;
2087 d--;
2088 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2089 continue;
2090 rdev = conf->mirrors[d].rdev;
2091 if (r1_sync_page_io(rdev, sect, s,
2092 pages[idx],
2093 REQ_OP_WRITE) == 0) {
2094 r1_bio->bios[d]->bi_end_io = NULL;
2095 rdev_dec_pending(rdev, mddev);
2096 }
2097 }
2098 d = start;
2099 while (d != r1_bio->read_disk) {
2100 if (d == 0)
2101 d = conf->raid_disks * 2;
2102 d--;
2103 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2104 continue;
2105 rdev = conf->mirrors[d].rdev;
2106 if (r1_sync_page_io(rdev, sect, s,
2107 pages[idx],
2108 REQ_OP_READ) != 0)
2109 atomic_add(s, &rdev->corrected_errors);
2110 }
2111 sectors -= s;
2112 sect += s;
2113 idx ++;
2114 }
2115 set_bit(R1BIO_Uptodate, &r1_bio->state);
2116 bio->bi_status = 0;
2117 return 1;
2118}
2119
2120static void process_checks(struct r1bio *r1_bio)
2121{
2122 /* We have read all readable devices. If we haven't
2123 * got the block, then there is no hope left.
2124 * If we have, then we want to do a comparison
2125 * and skip the write if everything is the same.
2126 * If any blocks failed to read, then we need to
2127 * attempt an over-write
2128 */
2129 struct mddev *mddev = r1_bio->mddev;
2130 struct r1conf *conf = mddev->private;
2131 int primary;
2132 int i;
2133 int vcnt;
2134
2135 /* Fix variable parts of all bios */
2136 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2137 for (i = 0; i < conf->raid_disks * 2; i++) {
2138 blk_status_t status;
2139 struct bio *b = r1_bio->bios[i];
2140 struct resync_pages *rp = get_resync_pages(b);
2141 if (b->bi_end_io != end_sync_read)
2142 continue;
2143 /* fixup the bio for reuse, but preserve errno */
2144 status = b->bi_status;
2145 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2146 b->bi_status = status;
2147 b->bi_iter.bi_sector = r1_bio->sector +
2148 conf->mirrors[i].rdev->data_offset;
2149 b->bi_end_io = end_sync_read;
2150 rp->raid_bio = r1_bio;
2151 b->bi_private = rp;
2152
2153 /* initialize bvec table again */
2154 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2155 }
2156 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2157 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2158 !r1_bio->bios[primary]->bi_status) {
2159 r1_bio->bios[primary]->bi_end_io = NULL;
2160 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2161 break;
2162 }
2163 r1_bio->read_disk = primary;
2164 for (i = 0; i < conf->raid_disks * 2; i++) {
2165 int j = 0;
2166 struct bio *pbio = r1_bio->bios[primary];
2167 struct bio *sbio = r1_bio->bios[i];
2168 blk_status_t status = sbio->bi_status;
2169 struct page **ppages = get_resync_pages(pbio)->pages;
2170 struct page **spages = get_resync_pages(sbio)->pages;
2171 struct bio_vec *bi;
2172 int page_len[RESYNC_PAGES] = { 0 };
2173 struct bvec_iter_all iter_all;
2174
2175 if (sbio->bi_end_io != end_sync_read)
2176 continue;
2177 /* Now we can 'fixup' the error value */
2178 sbio->bi_status = 0;
2179
2180 bio_for_each_segment_all(bi, sbio, iter_all)
2181 page_len[j++] = bi->bv_len;
2182
2183 if (!status) {
2184 for (j = vcnt; j-- ; ) {
2185 if (memcmp(page_address(ppages[j]),
2186 page_address(spages[j]),
2187 page_len[j]))
2188 break;
2189 }
2190 } else
2191 j = 0;
2192 if (j >= 0)
2193 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2194 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2195 && !status)) {
2196 /* No need to write to this device. */
2197 sbio->bi_end_io = NULL;
2198 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2199 continue;
2200 }
2201
2202 bio_copy_data(sbio, pbio);
2203 }
2204}
2205
2206static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2207{
2208 struct r1conf *conf = mddev->private;
2209 int i;
2210 int disks = conf->raid_disks * 2;
2211 struct bio *wbio;
2212
2213 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2214 /* ouch - failed to read all of that. */
2215 if (!fix_sync_read_error(r1_bio))
2216 return;
2217
2218 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2219 process_checks(r1_bio);
2220
2221 /*
2222 * schedule writes
2223 */
2224 atomic_set(&r1_bio->remaining, 1);
2225 for (i = 0; i < disks ; i++) {
2226 wbio = r1_bio->bios[i];
2227 if (wbio->bi_end_io == NULL ||
2228 (wbio->bi_end_io == end_sync_read &&
2229 (i == r1_bio->read_disk ||
2230 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2231 continue;
2232 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2233 abort_sync_write(mddev, r1_bio);
2234 continue;
2235 }
2236
2237 wbio->bi_opf = REQ_OP_WRITE;
2238 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2239 wbio->bi_opf |= MD_FAILFAST;
2240
2241 wbio->bi_end_io = end_sync_write;
2242 atomic_inc(&r1_bio->remaining);
2243 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2244
2245 submit_bio_noacct(wbio);
2246 }
2247
2248 put_sync_write_buf(r1_bio, 1);
2249}
2250
2251/*
2252 * This is a kernel thread which:
2253 *
2254 * 1. Retries failed read operations on working mirrors.
2255 * 2. Updates the raid superblock when problems encounter.
2256 * 3. Performs writes following reads for array synchronising.
2257 */
2258
2259static void fix_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2260{
2261 sector_t sect = r1_bio->sector;
2262 int sectors = r1_bio->sectors;
2263 int read_disk = r1_bio->read_disk;
2264 struct mddev *mddev = conf->mddev;
2265 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2266
2267 if (exceed_read_errors(mddev, rdev)) {
2268 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2269 return;
2270 }
2271
2272 while(sectors) {
2273 int s = sectors;
2274 int d = read_disk;
2275 int success = 0;
2276 int start;
2277
2278 if (s > (PAGE_SIZE>>9))
2279 s = PAGE_SIZE >> 9;
2280
2281 do {
2282 sector_t first_bad;
2283 int bad_sectors;
2284
2285 rdev = conf->mirrors[d].rdev;
2286 if (rdev &&
2287 (test_bit(In_sync, &rdev->flags) ||
2288 (!test_bit(Faulty, &rdev->flags) &&
2289 rdev->recovery_offset >= sect + s)) &&
2290 is_badblock(rdev, sect, s,
2291 &first_bad, &bad_sectors) == 0) {
2292 atomic_inc(&rdev->nr_pending);
2293 if (sync_page_io(rdev, sect, s<<9,
2294 conf->tmppage, REQ_OP_READ, false))
2295 success = 1;
2296 rdev_dec_pending(rdev, mddev);
2297 if (success)
2298 break;
2299 }
2300
2301 d++;
2302 if (d == conf->raid_disks * 2)
2303 d = 0;
2304 } while (d != read_disk);
2305
2306 if (!success) {
2307 /* Cannot read from anywhere - mark it bad */
2308 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2309 if (!rdev_set_badblocks(rdev, sect, s, 0))
2310 md_error(mddev, rdev);
2311 break;
2312 }
2313 /* write it back and re-read */
2314 start = d;
2315 while (d != read_disk) {
2316 if (d==0)
2317 d = conf->raid_disks * 2;
2318 d--;
2319 rdev = conf->mirrors[d].rdev;
2320 if (rdev &&
2321 !test_bit(Faulty, &rdev->flags)) {
2322 atomic_inc(&rdev->nr_pending);
2323 r1_sync_page_io(rdev, sect, s,
2324 conf->tmppage, REQ_OP_WRITE);
2325 rdev_dec_pending(rdev, mddev);
2326 }
2327 }
2328 d = start;
2329 while (d != read_disk) {
2330 if (d==0)
2331 d = conf->raid_disks * 2;
2332 d--;
2333 rdev = conf->mirrors[d].rdev;
2334 if (rdev &&
2335 !test_bit(Faulty, &rdev->flags)) {
2336 atomic_inc(&rdev->nr_pending);
2337 if (r1_sync_page_io(rdev, sect, s,
2338 conf->tmppage, REQ_OP_READ)) {
2339 atomic_add(s, &rdev->corrected_errors);
2340 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2341 mdname(mddev), s,
2342 (unsigned long long)(sect +
2343 rdev->data_offset),
2344 rdev->bdev);
2345 }
2346 rdev_dec_pending(rdev, mddev);
2347 }
2348 }
2349 sectors -= s;
2350 sect += s;
2351 }
2352}
2353
2354static int narrow_write_error(struct r1bio *r1_bio, int i)
2355{
2356 struct mddev *mddev = r1_bio->mddev;
2357 struct r1conf *conf = mddev->private;
2358 struct md_rdev *rdev = conf->mirrors[i].rdev;
2359
2360 /* bio has the data to be written to device 'i' where
2361 * we just recently had a write error.
2362 * We repeatedly clone the bio and trim down to one block,
2363 * then try the write. Where the write fails we record
2364 * a bad block.
2365 * It is conceivable that the bio doesn't exactly align with
2366 * blocks. We must handle this somehow.
2367 *
2368 * We currently own a reference on the rdev.
2369 */
2370
2371 int block_sectors;
2372 sector_t sector;
2373 int sectors;
2374 int sect_to_write = r1_bio->sectors;
2375 int ok = 1;
2376
2377 if (rdev->badblocks.shift < 0)
2378 return 0;
2379
2380 block_sectors = roundup(1 << rdev->badblocks.shift,
2381 bdev_logical_block_size(rdev->bdev) >> 9);
2382 sector = r1_bio->sector;
2383 sectors = ((sector + block_sectors)
2384 & ~(sector_t)(block_sectors - 1))
2385 - sector;
2386
2387 while (sect_to_write) {
2388 struct bio *wbio;
2389 if (sectors > sect_to_write)
2390 sectors = sect_to_write;
2391 /* Write at 'sector' for 'sectors'*/
2392
2393 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2394 wbio = bio_alloc_clone(rdev->bdev,
2395 r1_bio->behind_master_bio,
2396 GFP_NOIO, &mddev->bio_set);
2397 } else {
2398 wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2399 GFP_NOIO, &mddev->bio_set);
2400 }
2401
2402 wbio->bi_opf = REQ_OP_WRITE;
2403 wbio->bi_iter.bi_sector = r1_bio->sector;
2404 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2405
2406 bio_trim(wbio, sector - r1_bio->sector, sectors);
2407 wbio->bi_iter.bi_sector += rdev->data_offset;
2408
2409 if (submit_bio_wait(wbio) < 0)
2410 /* failure! */
2411 ok = rdev_set_badblocks(rdev, sector,
2412 sectors, 0)
2413 && ok;
2414
2415 bio_put(wbio);
2416 sect_to_write -= sectors;
2417 sector += sectors;
2418 sectors = block_sectors;
2419 }
2420 return ok;
2421}
2422
2423static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2424{
2425 int m;
2426 int s = r1_bio->sectors;
2427 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2428 struct md_rdev *rdev = conf->mirrors[m].rdev;
2429 struct bio *bio = r1_bio->bios[m];
2430 if (bio->bi_end_io == NULL)
2431 continue;
2432 if (!bio->bi_status &&
2433 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2434 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2435 }
2436 if (bio->bi_status &&
2437 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2438 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2439 md_error(conf->mddev, rdev);
2440 }
2441 }
2442 put_buf(r1_bio);
2443 md_done_sync(conf->mddev, s, 1);
2444}
2445
2446static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2447{
2448 int m, idx;
2449 bool fail = false;
2450
2451 for (m = 0; m < conf->raid_disks * 2 ; m++)
2452 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2453 struct md_rdev *rdev = conf->mirrors[m].rdev;
2454 rdev_clear_badblocks(rdev,
2455 r1_bio->sector,
2456 r1_bio->sectors, 0);
2457 rdev_dec_pending(rdev, conf->mddev);
2458 } else if (r1_bio->bios[m] != NULL) {
2459 /* This drive got a write error. We need to
2460 * narrow down and record precise write
2461 * errors.
2462 */
2463 fail = true;
2464 if (!narrow_write_error(r1_bio, m)) {
2465 md_error(conf->mddev,
2466 conf->mirrors[m].rdev);
2467 /* an I/O failed, we can't clear the bitmap */
2468 set_bit(R1BIO_Degraded, &r1_bio->state);
2469 }
2470 rdev_dec_pending(conf->mirrors[m].rdev,
2471 conf->mddev);
2472 }
2473 if (fail) {
2474 spin_lock_irq(&conf->device_lock);
2475 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2476 idx = sector_to_idx(r1_bio->sector);
2477 atomic_inc(&conf->nr_queued[idx]);
2478 spin_unlock_irq(&conf->device_lock);
2479 /*
2480 * In case freeze_array() is waiting for condition
2481 * get_unqueued_pending() == extra to be true.
2482 */
2483 wake_up(&conf->wait_barrier);
2484 md_wakeup_thread(conf->mddev->thread);
2485 } else {
2486 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2487 close_write(r1_bio);
2488 raid_end_bio_io(r1_bio);
2489 }
2490}
2491
2492static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2493{
2494 struct mddev *mddev = conf->mddev;
2495 struct bio *bio;
2496 struct md_rdev *rdev;
2497 sector_t sector;
2498
2499 clear_bit(R1BIO_ReadError, &r1_bio->state);
2500 /* we got a read error. Maybe the drive is bad. Maybe just
2501 * the block and we can fix it.
2502 * We freeze all other IO, and try reading the block from
2503 * other devices. When we find one, we re-write
2504 * and check it that fixes the read error.
2505 * This is all done synchronously while the array is
2506 * frozen
2507 */
2508
2509 bio = r1_bio->bios[r1_bio->read_disk];
2510 bio_put(bio);
2511 r1_bio->bios[r1_bio->read_disk] = NULL;
2512
2513 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2514 if (mddev->ro == 0
2515 && !test_bit(FailFast, &rdev->flags)) {
2516 freeze_array(conf, 1);
2517 fix_read_error(conf, r1_bio);
2518 unfreeze_array(conf);
2519 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2520 md_error(mddev, rdev);
2521 } else {
2522 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2523 }
2524
2525 rdev_dec_pending(rdev, conf->mddev);
2526 sector = r1_bio->sector;
2527 bio = r1_bio->master_bio;
2528
2529 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2530 r1_bio->state = 0;
2531 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2532 allow_barrier(conf, sector);
2533}
2534
2535static void raid1d(struct md_thread *thread)
2536{
2537 struct mddev *mddev = thread->mddev;
2538 struct r1bio *r1_bio;
2539 unsigned long flags;
2540 struct r1conf *conf = mddev->private;
2541 struct list_head *head = &conf->retry_list;
2542 struct blk_plug plug;
2543 int idx;
2544
2545 md_check_recovery(mddev);
2546
2547 if (!list_empty_careful(&conf->bio_end_io_list) &&
2548 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2549 LIST_HEAD(tmp);
2550 spin_lock_irqsave(&conf->device_lock, flags);
2551 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2552 list_splice_init(&conf->bio_end_io_list, &tmp);
2553 spin_unlock_irqrestore(&conf->device_lock, flags);
2554 while (!list_empty(&tmp)) {
2555 r1_bio = list_first_entry(&tmp, struct r1bio,
2556 retry_list);
2557 list_del(&r1_bio->retry_list);
2558 idx = sector_to_idx(r1_bio->sector);
2559 atomic_dec(&conf->nr_queued[idx]);
2560 if (mddev->degraded)
2561 set_bit(R1BIO_Degraded, &r1_bio->state);
2562 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2563 close_write(r1_bio);
2564 raid_end_bio_io(r1_bio);
2565 }
2566 }
2567
2568 blk_start_plug(&plug);
2569 for (;;) {
2570
2571 flush_pending_writes(conf);
2572
2573 spin_lock_irqsave(&conf->device_lock, flags);
2574 if (list_empty(head)) {
2575 spin_unlock_irqrestore(&conf->device_lock, flags);
2576 break;
2577 }
2578 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2579 list_del(head->prev);
2580 idx = sector_to_idx(r1_bio->sector);
2581 atomic_dec(&conf->nr_queued[idx]);
2582 spin_unlock_irqrestore(&conf->device_lock, flags);
2583
2584 mddev = r1_bio->mddev;
2585 conf = mddev->private;
2586 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2587 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2588 test_bit(R1BIO_WriteError, &r1_bio->state))
2589 handle_sync_write_finished(conf, r1_bio);
2590 else
2591 sync_request_write(mddev, r1_bio);
2592 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2593 test_bit(R1BIO_WriteError, &r1_bio->state))
2594 handle_write_finished(conf, r1_bio);
2595 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2596 handle_read_error(conf, r1_bio);
2597 else
2598 WARN_ON_ONCE(1);
2599
2600 cond_resched();
2601 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2602 md_check_recovery(mddev);
2603 }
2604 blk_finish_plug(&plug);
2605}
2606
2607static int init_resync(struct r1conf *conf)
2608{
2609 int buffs;
2610
2611 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2612 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2613
2614 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2615 r1buf_pool_free, conf->poolinfo);
2616}
2617
2618static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2619{
2620 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2621 struct resync_pages *rps;
2622 struct bio *bio;
2623 int i;
2624
2625 for (i = conf->poolinfo->raid_disks; i--; ) {
2626 bio = r1bio->bios[i];
2627 rps = bio->bi_private;
2628 bio_reset(bio, NULL, 0);
2629 bio->bi_private = rps;
2630 }
2631 r1bio->master_bio = NULL;
2632 return r1bio;
2633}
2634
2635/*
2636 * perform a "sync" on one "block"
2637 *
2638 * We need to make sure that no normal I/O request - particularly write
2639 * requests - conflict with active sync requests.
2640 *
2641 * This is achieved by tracking pending requests and a 'barrier' concept
2642 * that can be installed to exclude normal IO requests.
2643 */
2644
2645static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2646 int *skipped)
2647{
2648 struct r1conf *conf = mddev->private;
2649 struct r1bio *r1_bio;
2650 struct bio *bio;
2651 sector_t max_sector, nr_sectors;
2652 int disk = -1;
2653 int i;
2654 int wonly = -1;
2655 int write_targets = 0, read_targets = 0;
2656 sector_t sync_blocks;
2657 int still_degraded = 0;
2658 int good_sectors = RESYNC_SECTORS;
2659 int min_bad = 0; /* number of sectors that are bad in all devices */
2660 int idx = sector_to_idx(sector_nr);
2661 int page_idx = 0;
2662
2663 if (!mempool_initialized(&conf->r1buf_pool))
2664 if (init_resync(conf))
2665 return 0;
2666
2667 max_sector = mddev->dev_sectors;
2668 if (sector_nr >= max_sector) {
2669 /* If we aborted, we need to abort the
2670 * sync on the 'current' bitmap chunk (there will
2671 * only be one in raid1 resync.
2672 * We can find the current addess in mddev->curr_resync
2673 */
2674 if (mddev->curr_resync < max_sector) /* aborted */
2675 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2676 &sync_blocks, 1);
2677 else /* completed sync */
2678 conf->fullsync = 0;
2679
2680 md_bitmap_close_sync(mddev->bitmap);
2681 close_sync(conf);
2682
2683 if (mddev_is_clustered(mddev)) {
2684 conf->cluster_sync_low = 0;
2685 conf->cluster_sync_high = 0;
2686 }
2687 return 0;
2688 }
2689
2690 if (mddev->bitmap == NULL &&
2691 mddev->recovery_cp == MaxSector &&
2692 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2693 conf->fullsync == 0) {
2694 *skipped = 1;
2695 return max_sector - sector_nr;
2696 }
2697 /* before building a request, check if we can skip these blocks..
2698 * This call the bitmap_start_sync doesn't actually record anything
2699 */
2700 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2701 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2702 /* We can skip this block, and probably several more */
2703 *skipped = 1;
2704 return sync_blocks;
2705 }
2706
2707 /*
2708 * If there is non-resync activity waiting for a turn, then let it
2709 * though before starting on this new sync request.
2710 */
2711 if (atomic_read(&conf->nr_waiting[idx]))
2712 schedule_timeout_uninterruptible(1);
2713
2714 /* we are incrementing sector_nr below. To be safe, we check against
2715 * sector_nr + two times RESYNC_SECTORS
2716 */
2717
2718 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2719 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2720
2721
2722 if (raise_barrier(conf, sector_nr))
2723 return 0;
2724
2725 r1_bio = raid1_alloc_init_r1buf(conf);
2726
2727 /*
2728 * If we get a correctably read error during resync or recovery,
2729 * we might want to read from a different device. So we
2730 * flag all drives that could conceivably be read from for READ,
2731 * and any others (which will be non-In_sync devices) for WRITE.
2732 * If a read fails, we try reading from something else for which READ
2733 * is OK.
2734 */
2735
2736 r1_bio->mddev = mddev;
2737 r1_bio->sector = sector_nr;
2738 r1_bio->state = 0;
2739 set_bit(R1BIO_IsSync, &r1_bio->state);
2740 /* make sure good_sectors won't go across barrier unit boundary */
2741 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2742
2743 for (i = 0; i < conf->raid_disks * 2; i++) {
2744 struct md_rdev *rdev;
2745 bio = r1_bio->bios[i];
2746
2747 rdev = conf->mirrors[i].rdev;
2748 if (rdev == NULL ||
2749 test_bit(Faulty, &rdev->flags)) {
2750 if (i < conf->raid_disks)
2751 still_degraded = 1;
2752 } else if (!test_bit(In_sync, &rdev->flags)) {
2753 bio->bi_opf = REQ_OP_WRITE;
2754 bio->bi_end_io = end_sync_write;
2755 write_targets ++;
2756 } else {
2757 /* may need to read from here */
2758 sector_t first_bad = MaxSector;
2759 int bad_sectors;
2760
2761 if (is_badblock(rdev, sector_nr, good_sectors,
2762 &first_bad, &bad_sectors)) {
2763 if (first_bad > sector_nr)
2764 good_sectors = first_bad - sector_nr;
2765 else {
2766 bad_sectors -= (sector_nr - first_bad);
2767 if (min_bad == 0 ||
2768 min_bad > bad_sectors)
2769 min_bad = bad_sectors;
2770 }
2771 }
2772 if (sector_nr < first_bad) {
2773 if (test_bit(WriteMostly, &rdev->flags)) {
2774 if (wonly < 0)
2775 wonly = i;
2776 } else {
2777 if (disk < 0)
2778 disk = i;
2779 }
2780 bio->bi_opf = REQ_OP_READ;
2781 bio->bi_end_io = end_sync_read;
2782 read_targets++;
2783 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2784 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2785 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2786 /*
2787 * The device is suitable for reading (InSync),
2788 * but has bad block(s) here. Let's try to correct them,
2789 * if we are doing resync or repair. Otherwise, leave
2790 * this device alone for this sync request.
2791 */
2792 bio->bi_opf = REQ_OP_WRITE;
2793 bio->bi_end_io = end_sync_write;
2794 write_targets++;
2795 }
2796 }
2797 if (rdev && bio->bi_end_io) {
2798 atomic_inc(&rdev->nr_pending);
2799 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2800 bio_set_dev(bio, rdev->bdev);
2801 if (test_bit(FailFast, &rdev->flags))
2802 bio->bi_opf |= MD_FAILFAST;
2803 }
2804 }
2805 if (disk < 0)
2806 disk = wonly;
2807 r1_bio->read_disk = disk;
2808
2809 if (read_targets == 0 && min_bad > 0) {
2810 /* These sectors are bad on all InSync devices, so we
2811 * need to mark them bad on all write targets
2812 */
2813 int ok = 1;
2814 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2815 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2816 struct md_rdev *rdev = conf->mirrors[i].rdev;
2817 ok = rdev_set_badblocks(rdev, sector_nr,
2818 min_bad, 0
2819 ) && ok;
2820 }
2821 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2822 *skipped = 1;
2823 put_buf(r1_bio);
2824
2825 if (!ok) {
2826 /* Cannot record the badblocks, so need to
2827 * abort the resync.
2828 * If there are multiple read targets, could just
2829 * fail the really bad ones ???
2830 */
2831 conf->recovery_disabled = mddev->recovery_disabled;
2832 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2833 return 0;
2834 } else
2835 return min_bad;
2836
2837 }
2838 if (min_bad > 0 && min_bad < good_sectors) {
2839 /* only resync enough to reach the next bad->good
2840 * transition */
2841 good_sectors = min_bad;
2842 }
2843
2844 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2845 /* extra read targets are also write targets */
2846 write_targets += read_targets-1;
2847
2848 if (write_targets == 0 || read_targets == 0) {
2849 /* There is nowhere to write, so all non-sync
2850 * drives must be failed - so we are finished
2851 */
2852 sector_t rv;
2853 if (min_bad > 0)
2854 max_sector = sector_nr + min_bad;
2855 rv = max_sector - sector_nr;
2856 *skipped = 1;
2857 put_buf(r1_bio);
2858 return rv;
2859 }
2860
2861 if (max_sector > mddev->resync_max)
2862 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2863 if (max_sector > sector_nr + good_sectors)
2864 max_sector = sector_nr + good_sectors;
2865 nr_sectors = 0;
2866 sync_blocks = 0;
2867 do {
2868 struct page *page;
2869 int len = PAGE_SIZE;
2870 if (sector_nr + (len>>9) > max_sector)
2871 len = (max_sector - sector_nr) << 9;
2872 if (len == 0)
2873 break;
2874 if (sync_blocks == 0) {
2875 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2876 &sync_blocks, still_degraded) &&
2877 !conf->fullsync &&
2878 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2879 break;
2880 if ((len >> 9) > sync_blocks)
2881 len = sync_blocks<<9;
2882 }
2883
2884 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2885 struct resync_pages *rp;
2886
2887 bio = r1_bio->bios[i];
2888 rp = get_resync_pages(bio);
2889 if (bio->bi_end_io) {
2890 page = resync_fetch_page(rp, page_idx);
2891
2892 /*
2893 * won't fail because the vec table is big
2894 * enough to hold all these pages
2895 */
2896 __bio_add_page(bio, page, len, 0);
2897 }
2898 }
2899 nr_sectors += len>>9;
2900 sector_nr += len>>9;
2901 sync_blocks -= (len>>9);
2902 } while (++page_idx < RESYNC_PAGES);
2903
2904 r1_bio->sectors = nr_sectors;
2905
2906 if (mddev_is_clustered(mddev) &&
2907 conf->cluster_sync_high < sector_nr + nr_sectors) {
2908 conf->cluster_sync_low = mddev->curr_resync_completed;
2909 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2910 /* Send resync message */
2911 md_cluster_ops->resync_info_update(mddev,
2912 conf->cluster_sync_low,
2913 conf->cluster_sync_high);
2914 }
2915
2916 /* For a user-requested sync, we read all readable devices and do a
2917 * compare
2918 */
2919 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2920 atomic_set(&r1_bio->remaining, read_targets);
2921 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2922 bio = r1_bio->bios[i];
2923 if (bio->bi_end_io == end_sync_read) {
2924 read_targets--;
2925 md_sync_acct_bio(bio, nr_sectors);
2926 if (read_targets == 1)
2927 bio->bi_opf &= ~MD_FAILFAST;
2928 submit_bio_noacct(bio);
2929 }
2930 }
2931 } else {
2932 atomic_set(&r1_bio->remaining, 1);
2933 bio = r1_bio->bios[r1_bio->read_disk];
2934 md_sync_acct_bio(bio, nr_sectors);
2935 if (read_targets == 1)
2936 bio->bi_opf &= ~MD_FAILFAST;
2937 submit_bio_noacct(bio);
2938 }
2939 return nr_sectors;
2940}
2941
2942static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2943{
2944 if (sectors)
2945 return sectors;
2946
2947 return mddev->dev_sectors;
2948}
2949
2950static struct r1conf *setup_conf(struct mddev *mddev)
2951{
2952 struct r1conf *conf;
2953 int i;
2954 struct raid1_info *disk;
2955 struct md_rdev *rdev;
2956 int err = -ENOMEM;
2957
2958 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2959 if (!conf)
2960 goto abort;
2961
2962 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2963 sizeof(atomic_t), GFP_KERNEL);
2964 if (!conf->nr_pending)
2965 goto abort;
2966
2967 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2968 sizeof(atomic_t), GFP_KERNEL);
2969 if (!conf->nr_waiting)
2970 goto abort;
2971
2972 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2973 sizeof(atomic_t), GFP_KERNEL);
2974 if (!conf->nr_queued)
2975 goto abort;
2976
2977 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2978 sizeof(atomic_t), GFP_KERNEL);
2979 if (!conf->barrier)
2980 goto abort;
2981
2982 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
2983 mddev->raid_disks, 2),
2984 GFP_KERNEL);
2985 if (!conf->mirrors)
2986 goto abort;
2987
2988 conf->tmppage = alloc_page(GFP_KERNEL);
2989 if (!conf->tmppage)
2990 goto abort;
2991
2992 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2993 if (!conf->poolinfo)
2994 goto abort;
2995 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2996 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
2997 rbio_pool_free, conf->poolinfo);
2998 if (err)
2999 goto abort;
3000
3001 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3002 if (err)
3003 goto abort;
3004
3005 conf->poolinfo->mddev = mddev;
3006
3007 err = -EINVAL;
3008 spin_lock_init(&conf->device_lock);
3009 rdev_for_each(rdev, mddev) {
3010 int disk_idx = rdev->raid_disk;
3011 if (disk_idx >= mddev->raid_disks
3012 || disk_idx < 0)
3013 continue;
3014 if (test_bit(Replacement, &rdev->flags))
3015 disk = conf->mirrors + mddev->raid_disks + disk_idx;
3016 else
3017 disk = conf->mirrors + disk_idx;
3018
3019 if (disk->rdev)
3020 goto abort;
3021 disk->rdev = rdev;
3022 disk->head_position = 0;
3023 disk->seq_start = MaxSector;
3024 }
3025 conf->raid_disks = mddev->raid_disks;
3026 conf->mddev = mddev;
3027 INIT_LIST_HEAD(&conf->retry_list);
3028 INIT_LIST_HEAD(&conf->bio_end_io_list);
3029
3030 spin_lock_init(&conf->resync_lock);
3031 init_waitqueue_head(&conf->wait_barrier);
3032
3033 bio_list_init(&conf->pending_bio_list);
3034 conf->recovery_disabled = mddev->recovery_disabled - 1;
3035
3036 err = -EIO;
3037 for (i = 0; i < conf->raid_disks * 2; i++) {
3038
3039 disk = conf->mirrors + i;
3040
3041 if (i < conf->raid_disks &&
3042 disk[conf->raid_disks].rdev) {
3043 /* This slot has a replacement. */
3044 if (!disk->rdev) {
3045 /* No original, just make the replacement
3046 * a recovering spare
3047 */
3048 disk->rdev =
3049 disk[conf->raid_disks].rdev;
3050 disk[conf->raid_disks].rdev = NULL;
3051 } else if (!test_bit(In_sync, &disk->rdev->flags))
3052 /* Original is not in_sync - bad */
3053 goto abort;
3054 }
3055
3056 if (!disk->rdev ||
3057 !test_bit(In_sync, &disk->rdev->flags)) {
3058 disk->head_position = 0;
3059 if (disk->rdev &&
3060 (disk->rdev->saved_raid_disk < 0))
3061 conf->fullsync = 1;
3062 }
3063 }
3064
3065 err = -ENOMEM;
3066 rcu_assign_pointer(conf->thread,
3067 md_register_thread(raid1d, mddev, "raid1"));
3068 if (!conf->thread)
3069 goto abort;
3070
3071 return conf;
3072
3073 abort:
3074 if (conf) {
3075 mempool_exit(&conf->r1bio_pool);
3076 kfree(conf->mirrors);
3077 safe_put_page(conf->tmppage);
3078 kfree(conf->poolinfo);
3079 kfree(conf->nr_pending);
3080 kfree(conf->nr_waiting);
3081 kfree(conf->nr_queued);
3082 kfree(conf->barrier);
3083 bioset_exit(&conf->bio_split);
3084 kfree(conf);
3085 }
3086 return ERR_PTR(err);
3087}
3088
3089static void raid1_free(struct mddev *mddev, void *priv);
3090static int raid1_run(struct mddev *mddev)
3091{
3092 struct r1conf *conf;
3093 int i;
3094 struct md_rdev *rdev;
3095 int ret;
3096
3097 if (mddev->level != 1) {
3098 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3099 mdname(mddev), mddev->level);
3100 return -EIO;
3101 }
3102 if (mddev->reshape_position != MaxSector) {
3103 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3104 mdname(mddev));
3105 return -EIO;
3106 }
3107
3108 /*
3109 * copy the already verified devices into our private RAID1
3110 * bookkeeping area. [whatever we allocate in run(),
3111 * should be freed in raid1_free()]
3112 */
3113 if (mddev->private == NULL)
3114 conf = setup_conf(mddev);
3115 else
3116 conf = mddev->private;
3117
3118 if (IS_ERR(conf))
3119 return PTR_ERR(conf);
3120
3121 if (mddev->queue)
3122 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3123
3124 rdev_for_each(rdev, mddev) {
3125 if (!mddev->gendisk)
3126 continue;
3127 disk_stack_limits(mddev->gendisk, rdev->bdev,
3128 rdev->data_offset << 9);
3129 }
3130
3131 mddev->degraded = 0;
3132 for (i = 0; i < conf->raid_disks; i++)
3133 if (conf->mirrors[i].rdev == NULL ||
3134 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3135 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3136 mddev->degraded++;
3137 /*
3138 * RAID1 needs at least one disk in active
3139 */
3140 if (conf->raid_disks - mddev->degraded < 1) {
3141 md_unregister_thread(mddev, &conf->thread);
3142 ret = -EINVAL;
3143 goto abort;
3144 }
3145
3146 if (conf->raid_disks - mddev->degraded == 1)
3147 mddev->recovery_cp = MaxSector;
3148
3149 if (mddev->recovery_cp != MaxSector)
3150 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3151 mdname(mddev));
3152 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3153 mdname(mddev), mddev->raid_disks - mddev->degraded,
3154 mddev->raid_disks);
3155
3156 /*
3157 * Ok, everything is just fine now
3158 */
3159 rcu_assign_pointer(mddev->thread, conf->thread);
3160 rcu_assign_pointer(conf->thread, NULL);
3161 mddev->private = conf;
3162 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3163
3164 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3165
3166 ret = md_integrity_register(mddev);
3167 if (ret) {
3168 md_unregister_thread(mddev, &mddev->thread);
3169 goto abort;
3170 }
3171 return 0;
3172
3173abort:
3174 raid1_free(mddev, conf);
3175 return ret;
3176}
3177
3178static void raid1_free(struct mddev *mddev, void *priv)
3179{
3180 struct r1conf *conf = priv;
3181
3182 mempool_exit(&conf->r1bio_pool);
3183 kfree(conf->mirrors);
3184 safe_put_page(conf->tmppage);
3185 kfree(conf->poolinfo);
3186 kfree(conf->nr_pending);
3187 kfree(conf->nr_waiting);
3188 kfree(conf->nr_queued);
3189 kfree(conf->barrier);
3190 bioset_exit(&conf->bio_split);
3191 kfree(conf);
3192}
3193
3194static int raid1_resize(struct mddev *mddev, sector_t sectors)
3195{
3196 /* no resync is happening, and there is enough space
3197 * on all devices, so we can resize.
3198 * We need to make sure resync covers any new space.
3199 * If the array is shrinking we should possibly wait until
3200 * any io in the removed space completes, but it hardly seems
3201 * worth it.
3202 */
3203 sector_t newsize = raid1_size(mddev, sectors, 0);
3204 if (mddev->external_size &&
3205 mddev->array_sectors > newsize)
3206 return -EINVAL;
3207 if (mddev->bitmap) {
3208 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3209 if (ret)
3210 return ret;
3211 }
3212 md_set_array_sectors(mddev, newsize);
3213 if (sectors > mddev->dev_sectors &&
3214 mddev->recovery_cp > mddev->dev_sectors) {
3215 mddev->recovery_cp = mddev->dev_sectors;
3216 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3217 }
3218 mddev->dev_sectors = sectors;
3219 mddev->resync_max_sectors = sectors;
3220 return 0;
3221}
3222
3223static int raid1_reshape(struct mddev *mddev)
3224{
3225 /* We need to:
3226 * 1/ resize the r1bio_pool
3227 * 2/ resize conf->mirrors
3228 *
3229 * We allocate a new r1bio_pool if we can.
3230 * Then raise a device barrier and wait until all IO stops.
3231 * Then resize conf->mirrors and swap in the new r1bio pool.
3232 *
3233 * At the same time, we "pack" the devices so that all the missing
3234 * devices have the higher raid_disk numbers.
3235 */
3236 mempool_t newpool, oldpool;
3237 struct pool_info *newpoolinfo;
3238 struct raid1_info *newmirrors;
3239 struct r1conf *conf = mddev->private;
3240 int cnt, raid_disks;
3241 unsigned long flags;
3242 int d, d2;
3243 int ret;
3244
3245 memset(&newpool, 0, sizeof(newpool));
3246 memset(&oldpool, 0, sizeof(oldpool));
3247
3248 /* Cannot change chunk_size, layout, or level */
3249 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3250 mddev->layout != mddev->new_layout ||
3251 mddev->level != mddev->new_level) {
3252 mddev->new_chunk_sectors = mddev->chunk_sectors;
3253 mddev->new_layout = mddev->layout;
3254 mddev->new_level = mddev->level;
3255 return -EINVAL;
3256 }
3257
3258 if (!mddev_is_clustered(mddev))
3259 md_allow_write(mddev);
3260
3261 raid_disks = mddev->raid_disks + mddev->delta_disks;
3262
3263 if (raid_disks < conf->raid_disks) {
3264 cnt=0;
3265 for (d= 0; d < conf->raid_disks; d++)
3266 if (conf->mirrors[d].rdev)
3267 cnt++;
3268 if (cnt > raid_disks)
3269 return -EBUSY;
3270 }
3271
3272 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3273 if (!newpoolinfo)
3274 return -ENOMEM;
3275 newpoolinfo->mddev = mddev;
3276 newpoolinfo->raid_disks = raid_disks * 2;
3277
3278 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3279 rbio_pool_free, newpoolinfo);
3280 if (ret) {
3281 kfree(newpoolinfo);
3282 return ret;
3283 }
3284 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3285 raid_disks, 2),
3286 GFP_KERNEL);
3287 if (!newmirrors) {
3288 kfree(newpoolinfo);
3289 mempool_exit(&newpool);
3290 return -ENOMEM;
3291 }
3292
3293 freeze_array(conf, 0);
3294
3295 /* ok, everything is stopped */
3296 oldpool = conf->r1bio_pool;
3297 conf->r1bio_pool = newpool;
3298
3299 for (d = d2 = 0; d < conf->raid_disks; d++) {
3300 struct md_rdev *rdev = conf->mirrors[d].rdev;
3301 if (rdev && rdev->raid_disk != d2) {
3302 sysfs_unlink_rdev(mddev, rdev);
3303 rdev->raid_disk = d2;
3304 sysfs_unlink_rdev(mddev, rdev);
3305 if (sysfs_link_rdev(mddev, rdev))
3306 pr_warn("md/raid1:%s: cannot register rd%d\n",
3307 mdname(mddev), rdev->raid_disk);
3308 }
3309 if (rdev)
3310 newmirrors[d2++].rdev = rdev;
3311 }
3312 kfree(conf->mirrors);
3313 conf->mirrors = newmirrors;
3314 kfree(conf->poolinfo);
3315 conf->poolinfo = newpoolinfo;
3316
3317 spin_lock_irqsave(&conf->device_lock, flags);
3318 mddev->degraded += (raid_disks - conf->raid_disks);
3319 spin_unlock_irqrestore(&conf->device_lock, flags);
3320 conf->raid_disks = mddev->raid_disks = raid_disks;
3321 mddev->delta_disks = 0;
3322
3323 unfreeze_array(conf);
3324
3325 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3326 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3327 md_wakeup_thread(mddev->thread);
3328
3329 mempool_exit(&oldpool);
3330 return 0;
3331}
3332
3333static void raid1_quiesce(struct mddev *mddev, int quiesce)
3334{
3335 struct r1conf *conf = mddev->private;
3336
3337 if (quiesce)
3338 freeze_array(conf, 0);
3339 else
3340 unfreeze_array(conf);
3341}
3342
3343static void *raid1_takeover(struct mddev *mddev)
3344{
3345 /* raid1 can take over:
3346 * raid5 with 2 devices, any layout or chunk size
3347 */
3348 if (mddev->level == 5 && mddev->raid_disks == 2) {
3349 struct r1conf *conf;
3350 mddev->new_level = 1;
3351 mddev->new_layout = 0;
3352 mddev->new_chunk_sectors = 0;
3353 conf = setup_conf(mddev);
3354 if (!IS_ERR(conf)) {
3355 /* Array must appear to be quiesced */
3356 conf->array_frozen = 1;
3357 mddev_clear_unsupported_flags(mddev,
3358 UNSUPPORTED_MDDEV_FLAGS);
3359 }
3360 return conf;
3361 }
3362 return ERR_PTR(-EINVAL);
3363}
3364
3365static struct md_personality raid1_personality =
3366{
3367 .name = "raid1",
3368 .level = 1,
3369 .owner = THIS_MODULE,
3370 .make_request = raid1_make_request,
3371 .run = raid1_run,
3372 .free = raid1_free,
3373 .status = raid1_status,
3374 .error_handler = raid1_error,
3375 .hot_add_disk = raid1_add_disk,
3376 .hot_remove_disk= raid1_remove_disk,
3377 .spare_active = raid1_spare_active,
3378 .sync_request = raid1_sync_request,
3379 .resize = raid1_resize,
3380 .size = raid1_size,
3381 .check_reshape = raid1_reshape,
3382 .quiesce = raid1_quiesce,
3383 .takeover = raid1_takeover,
3384};
3385
3386static int __init raid_init(void)
3387{
3388 return register_md_personality(&raid1_personality);
3389}
3390
3391static void raid_exit(void)
3392{
3393 unregister_md_personality(&raid1_personality);
3394}
3395
3396module_init(raid_init);
3397module_exit(raid_exit);
3398MODULE_LICENSE("GPL");
3399MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3400MODULE_ALIAS("md-personality-3"); /* RAID1 */
3401MODULE_ALIAS("md-raid1");
3402MODULE_ALIAS("md-level-1");
1/*
2 * raid1.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5 *
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7 *
8 * RAID-1 management functions.
9 *
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11 *
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
27 * any later version.
28 *
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32 */
33
34#include <linux/slab.h>
35#include <linux/delay.h>
36#include <linux/blkdev.h>
37#include <linux/module.h>
38#include <linux/seq_file.h>
39#include <linux/ratelimit.h>
40
41#include <trace/events/block.h>
42
43#include "md.h"
44#include "raid1.h"
45#include "md-bitmap.h"
46
47#define UNSUPPORTED_MDDEV_FLAGS \
48 ((1L << MD_HAS_JOURNAL) | \
49 (1L << MD_JOURNAL_CLEAN) | \
50 (1L << MD_HAS_PPL) | \
51 (1L << MD_HAS_MULTIPLE_PPLS))
52
53/*
54 * Number of guaranteed r1bios in case of extreme VM load:
55 */
56#define NR_RAID1_BIOS 256
57
58/* when we get a read error on a read-only array, we redirect to another
59 * device without failing the first device, or trying to over-write to
60 * correct the read error. To keep track of bad blocks on a per-bio
61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
62 */
63#define IO_BLOCKED ((struct bio *)1)
64/* When we successfully write to a known bad-block, we need to remove the
65 * bad-block marking which must be done from process context. So we record
66 * the success by setting devs[n].bio to IO_MADE_GOOD
67 */
68#define IO_MADE_GOOD ((struct bio *)2)
69
70#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
71
72/* When there are this many requests queue to be written by
73 * the raid1 thread, we become 'congested' to provide back-pressure
74 * for writeback.
75 */
76static int max_queued_requests = 1024;
77
78static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
79static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
80
81#define raid1_log(md, fmt, args...) \
82 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
83
84#include "raid1-10.c"
85
86/*
87 * for resync bio, r1bio pointer can be retrieved from the per-bio
88 * 'struct resync_pages'.
89 */
90static inline struct r1bio *get_resync_r1bio(struct bio *bio)
91{
92 return get_resync_pages(bio)->raid_bio;
93}
94
95static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
96{
97 struct pool_info *pi = data;
98 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
99
100 /* allocate a r1bio with room for raid_disks entries in the bios array */
101 return kzalloc(size, gfp_flags);
102}
103
104static void r1bio_pool_free(void *r1_bio, void *data)
105{
106 kfree(r1_bio);
107}
108
109#define RESYNC_DEPTH 32
110#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
111#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
112#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
113#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
114#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
115
116static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
117{
118 struct pool_info *pi = data;
119 struct r1bio *r1_bio;
120 struct bio *bio;
121 int need_pages;
122 int j;
123 struct resync_pages *rps;
124
125 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
126 if (!r1_bio)
127 return NULL;
128
129 rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
130 gfp_flags);
131 if (!rps)
132 goto out_free_r1bio;
133
134 /*
135 * Allocate bios : 1 for reading, n-1 for writing
136 */
137 for (j = pi->raid_disks ; j-- ; ) {
138 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
139 if (!bio)
140 goto out_free_bio;
141 r1_bio->bios[j] = bio;
142 }
143 /*
144 * Allocate RESYNC_PAGES data pages and attach them to
145 * the first bio.
146 * If this is a user-requested check/repair, allocate
147 * RESYNC_PAGES for each bio.
148 */
149 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
150 need_pages = pi->raid_disks;
151 else
152 need_pages = 1;
153 for (j = 0; j < pi->raid_disks; j++) {
154 struct resync_pages *rp = &rps[j];
155
156 bio = r1_bio->bios[j];
157
158 if (j < need_pages) {
159 if (resync_alloc_pages(rp, gfp_flags))
160 goto out_free_pages;
161 } else {
162 memcpy(rp, &rps[0], sizeof(*rp));
163 resync_get_all_pages(rp);
164 }
165
166 rp->raid_bio = r1_bio;
167 bio->bi_private = rp;
168 }
169
170 r1_bio->master_bio = NULL;
171
172 return r1_bio;
173
174out_free_pages:
175 while (--j >= 0)
176 resync_free_pages(&rps[j]);
177
178out_free_bio:
179 while (++j < pi->raid_disks)
180 bio_put(r1_bio->bios[j]);
181 kfree(rps);
182
183out_free_r1bio:
184 r1bio_pool_free(r1_bio, data);
185 return NULL;
186}
187
188static void r1buf_pool_free(void *__r1_bio, void *data)
189{
190 struct pool_info *pi = data;
191 int i;
192 struct r1bio *r1bio = __r1_bio;
193 struct resync_pages *rp = NULL;
194
195 for (i = pi->raid_disks; i--; ) {
196 rp = get_resync_pages(r1bio->bios[i]);
197 resync_free_pages(rp);
198 bio_put(r1bio->bios[i]);
199 }
200
201 /* resync pages array stored in the 1st bio's .bi_private */
202 kfree(rp);
203
204 r1bio_pool_free(r1bio, data);
205}
206
207static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
208{
209 int i;
210
211 for (i = 0; i < conf->raid_disks * 2; i++) {
212 struct bio **bio = r1_bio->bios + i;
213 if (!BIO_SPECIAL(*bio))
214 bio_put(*bio);
215 *bio = NULL;
216 }
217}
218
219static void free_r1bio(struct r1bio *r1_bio)
220{
221 struct r1conf *conf = r1_bio->mddev->private;
222
223 put_all_bios(conf, r1_bio);
224 mempool_free(r1_bio, conf->r1bio_pool);
225}
226
227static void put_buf(struct r1bio *r1_bio)
228{
229 struct r1conf *conf = r1_bio->mddev->private;
230 sector_t sect = r1_bio->sector;
231 int i;
232
233 for (i = 0; i < conf->raid_disks * 2; i++) {
234 struct bio *bio = r1_bio->bios[i];
235 if (bio->bi_end_io)
236 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
237 }
238
239 mempool_free(r1_bio, conf->r1buf_pool);
240
241 lower_barrier(conf, sect);
242}
243
244static void reschedule_retry(struct r1bio *r1_bio)
245{
246 unsigned long flags;
247 struct mddev *mddev = r1_bio->mddev;
248 struct r1conf *conf = mddev->private;
249 int idx;
250
251 idx = sector_to_idx(r1_bio->sector);
252 spin_lock_irqsave(&conf->device_lock, flags);
253 list_add(&r1_bio->retry_list, &conf->retry_list);
254 atomic_inc(&conf->nr_queued[idx]);
255 spin_unlock_irqrestore(&conf->device_lock, flags);
256
257 wake_up(&conf->wait_barrier);
258 md_wakeup_thread(mddev->thread);
259}
260
261/*
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
264 * cache layer.
265 */
266static void call_bio_endio(struct r1bio *r1_bio)
267{
268 struct bio *bio = r1_bio->master_bio;
269 struct r1conf *conf = r1_bio->mddev->private;
270
271 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
272 bio->bi_status = BLK_STS_IOERR;
273
274 bio_endio(bio);
275 /*
276 * Wake up any possible resync thread that waits for the device
277 * to go idle.
278 */
279 allow_barrier(conf, r1_bio->sector);
280}
281
282static void raid_end_bio_io(struct r1bio *r1_bio)
283{
284 struct bio *bio = r1_bio->master_bio;
285
286 /* if nobody has done the final endio yet, do it now */
287 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
288 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
289 (bio_data_dir(bio) == WRITE) ? "write" : "read",
290 (unsigned long long) bio->bi_iter.bi_sector,
291 (unsigned long long) bio_end_sector(bio) - 1);
292
293 call_bio_endio(r1_bio);
294 }
295 free_r1bio(r1_bio);
296}
297
298/*
299 * Update disk head position estimator based on IRQ completion info.
300 */
301static inline void update_head_pos(int disk, struct r1bio *r1_bio)
302{
303 struct r1conf *conf = r1_bio->mddev->private;
304
305 conf->mirrors[disk].head_position =
306 r1_bio->sector + (r1_bio->sectors);
307}
308
309/*
310 * Find the disk number which triggered given bio
311 */
312static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
313{
314 int mirror;
315 struct r1conf *conf = r1_bio->mddev->private;
316 int raid_disks = conf->raid_disks;
317
318 for (mirror = 0; mirror < raid_disks * 2; mirror++)
319 if (r1_bio->bios[mirror] == bio)
320 break;
321
322 BUG_ON(mirror == raid_disks * 2);
323 update_head_pos(mirror, r1_bio);
324
325 return mirror;
326}
327
328static void raid1_end_read_request(struct bio *bio)
329{
330 int uptodate = !bio->bi_status;
331 struct r1bio *r1_bio = bio->bi_private;
332 struct r1conf *conf = r1_bio->mddev->private;
333 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
334
335 /*
336 * this branch is our 'one mirror IO has finished' event handler:
337 */
338 update_head_pos(r1_bio->read_disk, r1_bio);
339
340 if (uptodate)
341 set_bit(R1BIO_Uptodate, &r1_bio->state);
342 else if (test_bit(FailFast, &rdev->flags) &&
343 test_bit(R1BIO_FailFast, &r1_bio->state))
344 /* This was a fail-fast read so we definitely
345 * want to retry */
346 ;
347 else {
348 /* If all other devices have failed, we want to return
349 * the error upwards rather than fail the last device.
350 * Here we redefine "uptodate" to mean "Don't want to retry"
351 */
352 unsigned long flags;
353 spin_lock_irqsave(&conf->device_lock, flags);
354 if (r1_bio->mddev->degraded == conf->raid_disks ||
355 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
356 test_bit(In_sync, &rdev->flags)))
357 uptodate = 1;
358 spin_unlock_irqrestore(&conf->device_lock, flags);
359 }
360
361 if (uptodate) {
362 raid_end_bio_io(r1_bio);
363 rdev_dec_pending(rdev, conf->mddev);
364 } else {
365 /*
366 * oops, read error:
367 */
368 char b[BDEVNAME_SIZE];
369 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
370 mdname(conf->mddev),
371 bdevname(rdev->bdev, b),
372 (unsigned long long)r1_bio->sector);
373 set_bit(R1BIO_ReadError, &r1_bio->state);
374 reschedule_retry(r1_bio);
375 /* don't drop the reference on read_disk yet */
376 }
377}
378
379static void close_write(struct r1bio *r1_bio)
380{
381 /* it really is the end of this request */
382 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383 bio_free_pages(r1_bio->behind_master_bio);
384 bio_put(r1_bio->behind_master_bio);
385 r1_bio->behind_master_bio = NULL;
386 }
387 /* clear the bitmap if all writes complete successfully */
388 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
389 r1_bio->sectors,
390 !test_bit(R1BIO_Degraded, &r1_bio->state),
391 test_bit(R1BIO_BehindIO, &r1_bio->state));
392 md_write_end(r1_bio->mddev);
393}
394
395static void r1_bio_write_done(struct r1bio *r1_bio)
396{
397 if (!atomic_dec_and_test(&r1_bio->remaining))
398 return;
399
400 if (test_bit(R1BIO_WriteError, &r1_bio->state))
401 reschedule_retry(r1_bio);
402 else {
403 close_write(r1_bio);
404 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
405 reschedule_retry(r1_bio);
406 else
407 raid_end_bio_io(r1_bio);
408 }
409}
410
411static void raid1_end_write_request(struct bio *bio)
412{
413 struct r1bio *r1_bio = bio->bi_private;
414 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
415 struct r1conf *conf = r1_bio->mddev->private;
416 struct bio *to_put = NULL;
417 int mirror = find_bio_disk(r1_bio, bio);
418 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
419 bool discard_error;
420
421 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
422
423 /*
424 * 'one mirror IO has finished' event handler:
425 */
426 if (bio->bi_status && !discard_error) {
427 set_bit(WriteErrorSeen, &rdev->flags);
428 if (!test_and_set_bit(WantReplacement, &rdev->flags))
429 set_bit(MD_RECOVERY_NEEDED, &
430 conf->mddev->recovery);
431
432 if (test_bit(FailFast, &rdev->flags) &&
433 (bio->bi_opf & MD_FAILFAST) &&
434 /* We never try FailFast to WriteMostly devices */
435 !test_bit(WriteMostly, &rdev->flags)) {
436 md_error(r1_bio->mddev, rdev);
437 if (!test_bit(Faulty, &rdev->flags))
438 /* This is the only remaining device,
439 * We need to retry the write without
440 * FailFast
441 */
442 set_bit(R1BIO_WriteError, &r1_bio->state);
443 else {
444 /* Finished with this branch */
445 r1_bio->bios[mirror] = NULL;
446 to_put = bio;
447 }
448 } else
449 set_bit(R1BIO_WriteError, &r1_bio->state);
450 } else {
451 /*
452 * Set R1BIO_Uptodate in our master bio, so that we
453 * will return a good error code for to the higher
454 * levels even if IO on some other mirrored buffer
455 * fails.
456 *
457 * The 'master' represents the composite IO operation
458 * to user-side. So if something waits for IO, then it
459 * will wait for the 'master' bio.
460 */
461 sector_t first_bad;
462 int bad_sectors;
463
464 r1_bio->bios[mirror] = NULL;
465 to_put = bio;
466 /*
467 * Do not set R1BIO_Uptodate if the current device is
468 * rebuilding or Faulty. This is because we cannot use
469 * such device for properly reading the data back (we could
470 * potentially use it, if the current write would have felt
471 * before rdev->recovery_offset, but for simplicity we don't
472 * check this here.
473 */
474 if (test_bit(In_sync, &rdev->flags) &&
475 !test_bit(Faulty, &rdev->flags))
476 set_bit(R1BIO_Uptodate, &r1_bio->state);
477
478 /* Maybe we can clear some bad blocks. */
479 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
480 &first_bad, &bad_sectors) && !discard_error) {
481 r1_bio->bios[mirror] = IO_MADE_GOOD;
482 set_bit(R1BIO_MadeGood, &r1_bio->state);
483 }
484 }
485
486 if (behind) {
487 if (test_bit(WriteMostly, &rdev->flags))
488 atomic_dec(&r1_bio->behind_remaining);
489
490 /*
491 * In behind mode, we ACK the master bio once the I/O
492 * has safely reached all non-writemostly
493 * disks. Setting the Returned bit ensures that this
494 * gets done only once -- we don't ever want to return
495 * -EIO here, instead we'll wait
496 */
497 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
498 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
499 /* Maybe we can return now */
500 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
501 struct bio *mbio = r1_bio->master_bio;
502 pr_debug("raid1: behind end write sectors"
503 " %llu-%llu\n",
504 (unsigned long long) mbio->bi_iter.bi_sector,
505 (unsigned long long) bio_end_sector(mbio) - 1);
506 call_bio_endio(r1_bio);
507 }
508 }
509 }
510 if (r1_bio->bios[mirror] == NULL)
511 rdev_dec_pending(rdev, conf->mddev);
512
513 /*
514 * Let's see if all mirrored write operations have finished
515 * already.
516 */
517 r1_bio_write_done(r1_bio);
518
519 if (to_put)
520 bio_put(to_put);
521}
522
523static sector_t align_to_barrier_unit_end(sector_t start_sector,
524 sector_t sectors)
525{
526 sector_t len;
527
528 WARN_ON(sectors == 0);
529 /*
530 * len is the number of sectors from start_sector to end of the
531 * barrier unit which start_sector belongs to.
532 */
533 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
534 start_sector;
535
536 if (len > sectors)
537 len = sectors;
538
539 return len;
540}
541
542/*
543 * This routine returns the disk from which the requested read should
544 * be done. There is a per-array 'next expected sequential IO' sector
545 * number - if this matches on the next IO then we use the last disk.
546 * There is also a per-disk 'last know head position' sector that is
547 * maintained from IRQ contexts, both the normal and the resync IO
548 * completion handlers update this position correctly. If there is no
549 * perfect sequential match then we pick the disk whose head is closest.
550 *
551 * If there are 2 mirrors in the same 2 devices, performance degrades
552 * because position is mirror, not device based.
553 *
554 * The rdev for the device selected will have nr_pending incremented.
555 */
556static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
557{
558 const sector_t this_sector = r1_bio->sector;
559 int sectors;
560 int best_good_sectors;
561 int best_disk, best_dist_disk, best_pending_disk;
562 int has_nonrot_disk;
563 int disk;
564 sector_t best_dist;
565 unsigned int min_pending;
566 struct md_rdev *rdev;
567 int choose_first;
568 int choose_next_idle;
569
570 rcu_read_lock();
571 /*
572 * Check if we can balance. We can balance on the whole
573 * device if no resync is going on, or below the resync window.
574 * We take the first readable disk when above the resync window.
575 */
576 retry:
577 sectors = r1_bio->sectors;
578 best_disk = -1;
579 best_dist_disk = -1;
580 best_dist = MaxSector;
581 best_pending_disk = -1;
582 min_pending = UINT_MAX;
583 best_good_sectors = 0;
584 has_nonrot_disk = 0;
585 choose_next_idle = 0;
586 clear_bit(R1BIO_FailFast, &r1_bio->state);
587
588 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
589 (mddev_is_clustered(conf->mddev) &&
590 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
591 this_sector + sectors)))
592 choose_first = 1;
593 else
594 choose_first = 0;
595
596 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
597 sector_t dist;
598 sector_t first_bad;
599 int bad_sectors;
600 unsigned int pending;
601 bool nonrot;
602
603 rdev = rcu_dereference(conf->mirrors[disk].rdev);
604 if (r1_bio->bios[disk] == IO_BLOCKED
605 || rdev == NULL
606 || test_bit(Faulty, &rdev->flags))
607 continue;
608 if (!test_bit(In_sync, &rdev->flags) &&
609 rdev->recovery_offset < this_sector + sectors)
610 continue;
611 if (test_bit(WriteMostly, &rdev->flags)) {
612 /* Don't balance among write-mostly, just
613 * use the first as a last resort */
614 if (best_dist_disk < 0) {
615 if (is_badblock(rdev, this_sector, sectors,
616 &first_bad, &bad_sectors)) {
617 if (first_bad <= this_sector)
618 /* Cannot use this */
619 continue;
620 best_good_sectors = first_bad - this_sector;
621 } else
622 best_good_sectors = sectors;
623 best_dist_disk = disk;
624 best_pending_disk = disk;
625 }
626 continue;
627 }
628 /* This is a reasonable device to use. It might
629 * even be best.
630 */
631 if (is_badblock(rdev, this_sector, sectors,
632 &first_bad, &bad_sectors)) {
633 if (best_dist < MaxSector)
634 /* already have a better device */
635 continue;
636 if (first_bad <= this_sector) {
637 /* cannot read here. If this is the 'primary'
638 * device, then we must not read beyond
639 * bad_sectors from another device..
640 */
641 bad_sectors -= (this_sector - first_bad);
642 if (choose_first && sectors > bad_sectors)
643 sectors = bad_sectors;
644 if (best_good_sectors > sectors)
645 best_good_sectors = sectors;
646
647 } else {
648 sector_t good_sectors = first_bad - this_sector;
649 if (good_sectors > best_good_sectors) {
650 best_good_sectors = good_sectors;
651 best_disk = disk;
652 }
653 if (choose_first)
654 break;
655 }
656 continue;
657 } else {
658 if ((sectors > best_good_sectors) && (best_disk >= 0))
659 best_disk = -1;
660 best_good_sectors = sectors;
661 }
662
663 if (best_disk >= 0)
664 /* At least two disks to choose from so failfast is OK */
665 set_bit(R1BIO_FailFast, &r1_bio->state);
666
667 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
668 has_nonrot_disk |= nonrot;
669 pending = atomic_read(&rdev->nr_pending);
670 dist = abs(this_sector - conf->mirrors[disk].head_position);
671 if (choose_first) {
672 best_disk = disk;
673 break;
674 }
675 /* Don't change to another disk for sequential reads */
676 if (conf->mirrors[disk].next_seq_sect == this_sector
677 || dist == 0) {
678 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
679 struct raid1_info *mirror = &conf->mirrors[disk];
680
681 best_disk = disk;
682 /*
683 * If buffered sequential IO size exceeds optimal
684 * iosize, check if there is idle disk. If yes, choose
685 * the idle disk. read_balance could already choose an
686 * idle disk before noticing it's a sequential IO in
687 * this disk. This doesn't matter because this disk
688 * will idle, next time it will be utilized after the
689 * first disk has IO size exceeds optimal iosize. In
690 * this way, iosize of the first disk will be optimal
691 * iosize at least. iosize of the second disk might be
692 * small, but not a big deal since when the second disk
693 * starts IO, the first disk is likely still busy.
694 */
695 if (nonrot && opt_iosize > 0 &&
696 mirror->seq_start != MaxSector &&
697 mirror->next_seq_sect > opt_iosize &&
698 mirror->next_seq_sect - opt_iosize >=
699 mirror->seq_start) {
700 choose_next_idle = 1;
701 continue;
702 }
703 break;
704 }
705
706 if (choose_next_idle)
707 continue;
708
709 if (min_pending > pending) {
710 min_pending = pending;
711 best_pending_disk = disk;
712 }
713
714 if (dist < best_dist) {
715 best_dist = dist;
716 best_dist_disk = disk;
717 }
718 }
719
720 /*
721 * If all disks are rotational, choose the closest disk. If any disk is
722 * non-rotational, choose the disk with less pending request even the
723 * disk is rotational, which might/might not be optimal for raids with
724 * mixed ratation/non-rotational disks depending on workload.
725 */
726 if (best_disk == -1) {
727 if (has_nonrot_disk || min_pending == 0)
728 best_disk = best_pending_disk;
729 else
730 best_disk = best_dist_disk;
731 }
732
733 if (best_disk >= 0) {
734 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
735 if (!rdev)
736 goto retry;
737 atomic_inc(&rdev->nr_pending);
738 sectors = best_good_sectors;
739
740 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
741 conf->mirrors[best_disk].seq_start = this_sector;
742
743 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
744 }
745 rcu_read_unlock();
746 *max_sectors = sectors;
747
748 return best_disk;
749}
750
751static int raid1_congested(struct mddev *mddev, int bits)
752{
753 struct r1conf *conf = mddev->private;
754 int i, ret = 0;
755
756 if ((bits & (1 << WB_async_congested)) &&
757 conf->pending_count >= max_queued_requests)
758 return 1;
759
760 rcu_read_lock();
761 for (i = 0; i < conf->raid_disks * 2; i++) {
762 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
763 if (rdev && !test_bit(Faulty, &rdev->flags)) {
764 struct request_queue *q = bdev_get_queue(rdev->bdev);
765
766 BUG_ON(!q);
767
768 /* Note the '|| 1' - when read_balance prefers
769 * non-congested targets, it can be removed
770 */
771 if ((bits & (1 << WB_async_congested)) || 1)
772 ret |= bdi_congested(q->backing_dev_info, bits);
773 else
774 ret &= bdi_congested(q->backing_dev_info, bits);
775 }
776 }
777 rcu_read_unlock();
778 return ret;
779}
780
781static void flush_bio_list(struct r1conf *conf, struct bio *bio)
782{
783 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
784 bitmap_unplug(conf->mddev->bitmap);
785 wake_up(&conf->wait_barrier);
786
787 while (bio) { /* submit pending writes */
788 struct bio *next = bio->bi_next;
789 struct md_rdev *rdev = (void *)bio->bi_disk;
790 bio->bi_next = NULL;
791 bio_set_dev(bio, rdev->bdev);
792 if (test_bit(Faulty, &rdev->flags)) {
793 bio_io_error(bio);
794 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
795 !blk_queue_discard(bio->bi_disk->queue)))
796 /* Just ignore it */
797 bio_endio(bio);
798 else
799 generic_make_request(bio);
800 bio = next;
801 }
802}
803
804static void flush_pending_writes(struct r1conf *conf)
805{
806 /* Any writes that have been queued but are awaiting
807 * bitmap updates get flushed here.
808 */
809 spin_lock_irq(&conf->device_lock);
810
811 if (conf->pending_bio_list.head) {
812 struct blk_plug plug;
813 struct bio *bio;
814
815 bio = bio_list_get(&conf->pending_bio_list);
816 conf->pending_count = 0;
817 spin_unlock_irq(&conf->device_lock);
818
819 /*
820 * As this is called in a wait_event() loop (see freeze_array),
821 * current->state might be TASK_UNINTERRUPTIBLE which will
822 * cause a warning when we prepare to wait again. As it is
823 * rare that this path is taken, it is perfectly safe to force
824 * us to go around the wait_event() loop again, so the warning
825 * is a false-positive. Silence the warning by resetting
826 * thread state
827 */
828 __set_current_state(TASK_RUNNING);
829 blk_start_plug(&plug);
830 flush_bio_list(conf, bio);
831 blk_finish_plug(&plug);
832 } else
833 spin_unlock_irq(&conf->device_lock);
834}
835
836/* Barriers....
837 * Sometimes we need to suspend IO while we do something else,
838 * either some resync/recovery, or reconfigure the array.
839 * To do this we raise a 'barrier'.
840 * The 'barrier' is a counter that can be raised multiple times
841 * to count how many activities are happening which preclude
842 * normal IO.
843 * We can only raise the barrier if there is no pending IO.
844 * i.e. if nr_pending == 0.
845 * We choose only to raise the barrier if no-one is waiting for the
846 * barrier to go down. This means that as soon as an IO request
847 * is ready, no other operations which require a barrier will start
848 * until the IO request has had a chance.
849 *
850 * So: regular IO calls 'wait_barrier'. When that returns there
851 * is no backgroup IO happening, It must arrange to call
852 * allow_barrier when it has finished its IO.
853 * backgroup IO calls must call raise_barrier. Once that returns
854 * there is no normal IO happeing. It must arrange to call
855 * lower_barrier when the particular background IO completes.
856 */
857static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
858{
859 int idx = sector_to_idx(sector_nr);
860
861 spin_lock_irq(&conf->resync_lock);
862
863 /* Wait until no block IO is waiting */
864 wait_event_lock_irq(conf->wait_barrier,
865 !atomic_read(&conf->nr_waiting[idx]),
866 conf->resync_lock);
867
868 /* block any new IO from starting */
869 atomic_inc(&conf->barrier[idx]);
870 /*
871 * In raise_barrier() we firstly increase conf->barrier[idx] then
872 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
873 * increase conf->nr_pending[idx] then check conf->barrier[idx].
874 * A memory barrier here to make sure conf->nr_pending[idx] won't
875 * be fetched before conf->barrier[idx] is increased. Otherwise
876 * there will be a race between raise_barrier() and _wait_barrier().
877 */
878 smp_mb__after_atomic();
879
880 /* For these conditions we must wait:
881 * A: while the array is in frozen state
882 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
883 * existing in corresponding I/O barrier bucket.
884 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
885 * max resync count which allowed on current I/O barrier bucket.
886 */
887 wait_event_lock_irq(conf->wait_barrier,
888 (!conf->array_frozen &&
889 !atomic_read(&conf->nr_pending[idx]) &&
890 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
891 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
892 conf->resync_lock);
893
894 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
895 atomic_dec(&conf->barrier[idx]);
896 spin_unlock_irq(&conf->resync_lock);
897 wake_up(&conf->wait_barrier);
898 return -EINTR;
899 }
900
901 atomic_inc(&conf->nr_sync_pending);
902 spin_unlock_irq(&conf->resync_lock);
903
904 return 0;
905}
906
907static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
908{
909 int idx = sector_to_idx(sector_nr);
910
911 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
912
913 atomic_dec(&conf->barrier[idx]);
914 atomic_dec(&conf->nr_sync_pending);
915 wake_up(&conf->wait_barrier);
916}
917
918static void _wait_barrier(struct r1conf *conf, int idx)
919{
920 /*
921 * We need to increase conf->nr_pending[idx] very early here,
922 * then raise_barrier() can be blocked when it waits for
923 * conf->nr_pending[idx] to be 0. Then we can avoid holding
924 * conf->resync_lock when there is no barrier raised in same
925 * barrier unit bucket. Also if the array is frozen, I/O
926 * should be blocked until array is unfrozen.
927 */
928 atomic_inc(&conf->nr_pending[idx]);
929 /*
930 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
931 * check conf->barrier[idx]. In raise_barrier() we firstly increase
932 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
933 * barrier is necessary here to make sure conf->barrier[idx] won't be
934 * fetched before conf->nr_pending[idx] is increased. Otherwise there
935 * will be a race between _wait_barrier() and raise_barrier().
936 */
937 smp_mb__after_atomic();
938
939 /*
940 * Don't worry about checking two atomic_t variables at same time
941 * here. If during we check conf->barrier[idx], the array is
942 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
943 * 0, it is safe to return and make the I/O continue. Because the
944 * array is frozen, all I/O returned here will eventually complete
945 * or be queued, no race will happen. See code comment in
946 * frozen_array().
947 */
948 if (!READ_ONCE(conf->array_frozen) &&
949 !atomic_read(&conf->barrier[idx]))
950 return;
951
952 /*
953 * After holding conf->resync_lock, conf->nr_pending[idx]
954 * should be decreased before waiting for barrier to drop.
955 * Otherwise, we may encounter a race condition because
956 * raise_barrer() might be waiting for conf->nr_pending[idx]
957 * to be 0 at same time.
958 */
959 spin_lock_irq(&conf->resync_lock);
960 atomic_inc(&conf->nr_waiting[idx]);
961 atomic_dec(&conf->nr_pending[idx]);
962 /*
963 * In case freeze_array() is waiting for
964 * get_unqueued_pending() == extra
965 */
966 wake_up(&conf->wait_barrier);
967 /* Wait for the barrier in same barrier unit bucket to drop. */
968 wait_event_lock_irq(conf->wait_barrier,
969 !conf->array_frozen &&
970 !atomic_read(&conf->barrier[idx]),
971 conf->resync_lock);
972 atomic_inc(&conf->nr_pending[idx]);
973 atomic_dec(&conf->nr_waiting[idx]);
974 spin_unlock_irq(&conf->resync_lock);
975}
976
977static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
978{
979 int idx = sector_to_idx(sector_nr);
980
981 /*
982 * Very similar to _wait_barrier(). The difference is, for read
983 * I/O we don't need wait for sync I/O, but if the whole array
984 * is frozen, the read I/O still has to wait until the array is
985 * unfrozen. Since there is no ordering requirement with
986 * conf->barrier[idx] here, memory barrier is unnecessary as well.
987 */
988 atomic_inc(&conf->nr_pending[idx]);
989
990 if (!READ_ONCE(conf->array_frozen))
991 return;
992
993 spin_lock_irq(&conf->resync_lock);
994 atomic_inc(&conf->nr_waiting[idx]);
995 atomic_dec(&conf->nr_pending[idx]);
996 /*
997 * In case freeze_array() is waiting for
998 * get_unqueued_pending() == extra
999 */
1000 wake_up(&conf->wait_barrier);
1001 /* Wait for array to be unfrozen */
1002 wait_event_lock_irq(conf->wait_barrier,
1003 !conf->array_frozen,
1004 conf->resync_lock);
1005 atomic_inc(&conf->nr_pending[idx]);
1006 atomic_dec(&conf->nr_waiting[idx]);
1007 spin_unlock_irq(&conf->resync_lock);
1008}
1009
1010static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1011{
1012 int idx = sector_to_idx(sector_nr);
1013
1014 _wait_barrier(conf, idx);
1015}
1016
1017static void _allow_barrier(struct r1conf *conf, int idx)
1018{
1019 atomic_dec(&conf->nr_pending[idx]);
1020 wake_up(&conf->wait_barrier);
1021}
1022
1023static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1024{
1025 int idx = sector_to_idx(sector_nr);
1026
1027 _allow_barrier(conf, idx);
1028}
1029
1030/* conf->resync_lock should be held */
1031static int get_unqueued_pending(struct r1conf *conf)
1032{
1033 int idx, ret;
1034
1035 ret = atomic_read(&conf->nr_sync_pending);
1036 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1037 ret += atomic_read(&conf->nr_pending[idx]) -
1038 atomic_read(&conf->nr_queued[idx]);
1039
1040 return ret;
1041}
1042
1043static void freeze_array(struct r1conf *conf, int extra)
1044{
1045 /* Stop sync I/O and normal I/O and wait for everything to
1046 * go quiet.
1047 * This is called in two situations:
1048 * 1) management command handlers (reshape, remove disk, quiesce).
1049 * 2) one normal I/O request failed.
1050
1051 * After array_frozen is set to 1, new sync IO will be blocked at
1052 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1053 * or wait_read_barrier(). The flying I/Os will either complete or be
1054 * queued. When everything goes quite, there are only queued I/Os left.
1055
1056 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1057 * barrier bucket index which this I/O request hits. When all sync and
1058 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1059 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1060 * in handle_read_error(), we may call freeze_array() before trying to
1061 * fix the read error. In this case, the error read I/O is not queued,
1062 * so get_unqueued_pending() == 1.
1063 *
1064 * Therefore before this function returns, we need to wait until
1065 * get_unqueued_pendings(conf) gets equal to extra. For
1066 * normal I/O context, extra is 1, in rested situations extra is 0.
1067 */
1068 spin_lock_irq(&conf->resync_lock);
1069 conf->array_frozen = 1;
1070 raid1_log(conf->mddev, "wait freeze");
1071 wait_event_lock_irq_cmd(
1072 conf->wait_barrier,
1073 get_unqueued_pending(conf) == extra,
1074 conf->resync_lock,
1075 flush_pending_writes(conf));
1076 spin_unlock_irq(&conf->resync_lock);
1077}
1078static void unfreeze_array(struct r1conf *conf)
1079{
1080 /* reverse the effect of the freeze */
1081 spin_lock_irq(&conf->resync_lock);
1082 conf->array_frozen = 0;
1083 spin_unlock_irq(&conf->resync_lock);
1084 wake_up(&conf->wait_barrier);
1085}
1086
1087static void alloc_behind_master_bio(struct r1bio *r1_bio,
1088 struct bio *bio)
1089{
1090 int size = bio->bi_iter.bi_size;
1091 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1092 int i = 0;
1093 struct bio *behind_bio = NULL;
1094
1095 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1096 if (!behind_bio)
1097 return;
1098
1099 /* discard op, we don't support writezero/writesame yet */
1100 if (!bio_has_data(bio)) {
1101 behind_bio->bi_iter.bi_size = size;
1102 goto skip_copy;
1103 }
1104
1105 behind_bio->bi_write_hint = bio->bi_write_hint;
1106
1107 while (i < vcnt && size) {
1108 struct page *page;
1109 int len = min_t(int, PAGE_SIZE, size);
1110
1111 page = alloc_page(GFP_NOIO);
1112 if (unlikely(!page))
1113 goto free_pages;
1114
1115 bio_add_page(behind_bio, page, len, 0);
1116
1117 size -= len;
1118 i++;
1119 }
1120
1121 bio_copy_data(behind_bio, bio);
1122skip_copy:
1123 r1_bio->behind_master_bio = behind_bio;
1124 set_bit(R1BIO_BehindIO, &r1_bio->state);
1125
1126 return;
1127
1128free_pages:
1129 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1130 bio->bi_iter.bi_size);
1131 bio_free_pages(behind_bio);
1132 bio_put(behind_bio);
1133}
1134
1135struct raid1_plug_cb {
1136 struct blk_plug_cb cb;
1137 struct bio_list pending;
1138 int pending_cnt;
1139};
1140
1141static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1142{
1143 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1144 cb);
1145 struct mddev *mddev = plug->cb.data;
1146 struct r1conf *conf = mddev->private;
1147 struct bio *bio;
1148
1149 if (from_schedule || current->bio_list) {
1150 spin_lock_irq(&conf->device_lock);
1151 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1152 conf->pending_count += plug->pending_cnt;
1153 spin_unlock_irq(&conf->device_lock);
1154 wake_up(&conf->wait_barrier);
1155 md_wakeup_thread(mddev->thread);
1156 kfree(plug);
1157 return;
1158 }
1159
1160 /* we aren't scheduling, so we can do the write-out directly. */
1161 bio = bio_list_get(&plug->pending);
1162 flush_bio_list(conf, bio);
1163 kfree(plug);
1164}
1165
1166static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1167{
1168 r1_bio->master_bio = bio;
1169 r1_bio->sectors = bio_sectors(bio);
1170 r1_bio->state = 0;
1171 r1_bio->mddev = mddev;
1172 r1_bio->sector = bio->bi_iter.bi_sector;
1173}
1174
1175static inline struct r1bio *
1176alloc_r1bio(struct mddev *mddev, struct bio *bio)
1177{
1178 struct r1conf *conf = mddev->private;
1179 struct r1bio *r1_bio;
1180
1181 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1182 /* Ensure no bio records IO_BLOCKED */
1183 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1184 init_r1bio(r1_bio, mddev, bio);
1185 return r1_bio;
1186}
1187
1188static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1189 int max_read_sectors, struct r1bio *r1_bio)
1190{
1191 struct r1conf *conf = mddev->private;
1192 struct raid1_info *mirror;
1193 struct bio *read_bio;
1194 struct bitmap *bitmap = mddev->bitmap;
1195 const int op = bio_op(bio);
1196 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1197 int max_sectors;
1198 int rdisk;
1199 bool print_msg = !!r1_bio;
1200 char b[BDEVNAME_SIZE];
1201
1202 /*
1203 * If r1_bio is set, we are blocking the raid1d thread
1204 * so there is a tiny risk of deadlock. So ask for
1205 * emergency memory if needed.
1206 */
1207 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1208
1209 if (print_msg) {
1210 /* Need to get the block device name carefully */
1211 struct md_rdev *rdev;
1212 rcu_read_lock();
1213 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1214 if (rdev)
1215 bdevname(rdev->bdev, b);
1216 else
1217 strcpy(b, "???");
1218 rcu_read_unlock();
1219 }
1220
1221 /*
1222 * Still need barrier for READ in case that whole
1223 * array is frozen.
1224 */
1225 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1226
1227 if (!r1_bio)
1228 r1_bio = alloc_r1bio(mddev, bio);
1229 else
1230 init_r1bio(r1_bio, mddev, bio);
1231 r1_bio->sectors = max_read_sectors;
1232
1233 /*
1234 * make_request() can abort the operation when read-ahead is being
1235 * used and no empty request is available.
1236 */
1237 rdisk = read_balance(conf, r1_bio, &max_sectors);
1238
1239 if (rdisk < 0) {
1240 /* couldn't find anywhere to read from */
1241 if (print_msg) {
1242 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1243 mdname(mddev),
1244 b,
1245 (unsigned long long)r1_bio->sector);
1246 }
1247 raid_end_bio_io(r1_bio);
1248 return;
1249 }
1250 mirror = conf->mirrors + rdisk;
1251
1252 if (print_msg)
1253 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1254 mdname(mddev),
1255 (unsigned long long)r1_bio->sector,
1256 bdevname(mirror->rdev->bdev, b));
1257
1258 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1259 bitmap) {
1260 /*
1261 * Reading from a write-mostly device must take care not to
1262 * over-take any writes that are 'behind'
1263 */
1264 raid1_log(mddev, "wait behind writes");
1265 wait_event(bitmap->behind_wait,
1266 atomic_read(&bitmap->behind_writes) == 0);
1267 }
1268
1269 if (max_sectors < bio_sectors(bio)) {
1270 struct bio *split = bio_split(bio, max_sectors,
1271 gfp, conf->bio_split);
1272 bio_chain(split, bio);
1273 generic_make_request(bio);
1274 bio = split;
1275 r1_bio->master_bio = bio;
1276 r1_bio->sectors = max_sectors;
1277 }
1278
1279 r1_bio->read_disk = rdisk;
1280
1281 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1282
1283 r1_bio->bios[rdisk] = read_bio;
1284
1285 read_bio->bi_iter.bi_sector = r1_bio->sector +
1286 mirror->rdev->data_offset;
1287 bio_set_dev(read_bio, mirror->rdev->bdev);
1288 read_bio->bi_end_io = raid1_end_read_request;
1289 bio_set_op_attrs(read_bio, op, do_sync);
1290 if (test_bit(FailFast, &mirror->rdev->flags) &&
1291 test_bit(R1BIO_FailFast, &r1_bio->state))
1292 read_bio->bi_opf |= MD_FAILFAST;
1293 read_bio->bi_private = r1_bio;
1294
1295 if (mddev->gendisk)
1296 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1297 disk_devt(mddev->gendisk), r1_bio->sector);
1298
1299 generic_make_request(read_bio);
1300}
1301
1302static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1303 int max_write_sectors)
1304{
1305 struct r1conf *conf = mddev->private;
1306 struct r1bio *r1_bio;
1307 int i, disks;
1308 struct bitmap *bitmap = mddev->bitmap;
1309 unsigned long flags;
1310 struct md_rdev *blocked_rdev;
1311 struct blk_plug_cb *cb;
1312 struct raid1_plug_cb *plug = NULL;
1313 int first_clone;
1314 int max_sectors;
1315
1316 if (mddev_is_clustered(mddev) &&
1317 md_cluster_ops->area_resyncing(mddev, WRITE,
1318 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1319
1320 DEFINE_WAIT(w);
1321 for (;;) {
1322 prepare_to_wait(&conf->wait_barrier,
1323 &w, TASK_IDLE);
1324 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325 bio->bi_iter.bi_sector,
1326 bio_end_sector(bio)))
1327 break;
1328 schedule();
1329 }
1330 finish_wait(&conf->wait_barrier, &w);
1331 }
1332
1333 /*
1334 * Register the new request and wait if the reconstruction
1335 * thread has put up a bar for new requests.
1336 * Continue immediately if no resync is active currently.
1337 */
1338 wait_barrier(conf, bio->bi_iter.bi_sector);
1339
1340 r1_bio = alloc_r1bio(mddev, bio);
1341 r1_bio->sectors = max_write_sectors;
1342
1343 if (conf->pending_count >= max_queued_requests) {
1344 md_wakeup_thread(mddev->thread);
1345 raid1_log(mddev, "wait queued");
1346 wait_event(conf->wait_barrier,
1347 conf->pending_count < max_queued_requests);
1348 }
1349 /* first select target devices under rcu_lock and
1350 * inc refcount on their rdev. Record them by setting
1351 * bios[x] to bio
1352 * If there are known/acknowledged bad blocks on any device on
1353 * which we have seen a write error, we want to avoid writing those
1354 * blocks.
1355 * This potentially requires several writes to write around
1356 * the bad blocks. Each set of writes gets it's own r1bio
1357 * with a set of bios attached.
1358 */
1359
1360 disks = conf->raid_disks * 2;
1361 retry_write:
1362 blocked_rdev = NULL;
1363 rcu_read_lock();
1364 max_sectors = r1_bio->sectors;
1365 for (i = 0; i < disks; i++) {
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1367 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1368 atomic_inc(&rdev->nr_pending);
1369 blocked_rdev = rdev;
1370 break;
1371 }
1372 r1_bio->bios[i] = NULL;
1373 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1374 if (i < conf->raid_disks)
1375 set_bit(R1BIO_Degraded, &r1_bio->state);
1376 continue;
1377 }
1378
1379 atomic_inc(&rdev->nr_pending);
1380 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1381 sector_t first_bad;
1382 int bad_sectors;
1383 int is_bad;
1384
1385 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1386 &first_bad, &bad_sectors);
1387 if (is_bad < 0) {
1388 /* mustn't write here until the bad block is
1389 * acknowledged*/
1390 set_bit(BlockedBadBlocks, &rdev->flags);
1391 blocked_rdev = rdev;
1392 break;
1393 }
1394 if (is_bad && first_bad <= r1_bio->sector) {
1395 /* Cannot write here at all */
1396 bad_sectors -= (r1_bio->sector - first_bad);
1397 if (bad_sectors < max_sectors)
1398 /* mustn't write more than bad_sectors
1399 * to other devices yet
1400 */
1401 max_sectors = bad_sectors;
1402 rdev_dec_pending(rdev, mddev);
1403 /* We don't set R1BIO_Degraded as that
1404 * only applies if the disk is
1405 * missing, so it might be re-added,
1406 * and we want to know to recover this
1407 * chunk.
1408 * In this case the device is here,
1409 * and the fact that this chunk is not
1410 * in-sync is recorded in the bad
1411 * block log
1412 */
1413 continue;
1414 }
1415 if (is_bad) {
1416 int good_sectors = first_bad - r1_bio->sector;
1417 if (good_sectors < max_sectors)
1418 max_sectors = good_sectors;
1419 }
1420 }
1421 r1_bio->bios[i] = bio;
1422 }
1423 rcu_read_unlock();
1424
1425 if (unlikely(blocked_rdev)) {
1426 /* Wait for this device to become unblocked */
1427 int j;
1428
1429 for (j = 0; j < i; j++)
1430 if (r1_bio->bios[j])
1431 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1432 r1_bio->state = 0;
1433 allow_barrier(conf, bio->bi_iter.bi_sector);
1434 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1435 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1436 wait_barrier(conf, bio->bi_iter.bi_sector);
1437 goto retry_write;
1438 }
1439
1440 if (max_sectors < bio_sectors(bio)) {
1441 struct bio *split = bio_split(bio, max_sectors,
1442 GFP_NOIO, conf->bio_split);
1443 bio_chain(split, bio);
1444 generic_make_request(bio);
1445 bio = split;
1446 r1_bio->master_bio = bio;
1447 r1_bio->sectors = max_sectors;
1448 }
1449
1450 atomic_set(&r1_bio->remaining, 1);
1451 atomic_set(&r1_bio->behind_remaining, 0);
1452
1453 first_clone = 1;
1454
1455 for (i = 0; i < disks; i++) {
1456 struct bio *mbio = NULL;
1457 if (!r1_bio->bios[i])
1458 continue;
1459
1460
1461 if (first_clone) {
1462 /* do behind I/O ?
1463 * Not if there are too many, or cannot
1464 * allocate memory, or a reader on WriteMostly
1465 * is waiting for behind writes to flush */
1466 if (bitmap &&
1467 (atomic_read(&bitmap->behind_writes)
1468 < mddev->bitmap_info.max_write_behind) &&
1469 !waitqueue_active(&bitmap->behind_wait)) {
1470 alloc_behind_master_bio(r1_bio, bio);
1471 }
1472
1473 bitmap_startwrite(bitmap, r1_bio->sector,
1474 r1_bio->sectors,
1475 test_bit(R1BIO_BehindIO,
1476 &r1_bio->state));
1477 first_clone = 0;
1478 }
1479
1480 if (r1_bio->behind_master_bio)
1481 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1482 GFP_NOIO, mddev->bio_set);
1483 else
1484 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1485
1486 if (r1_bio->behind_master_bio) {
1487 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1488 atomic_inc(&r1_bio->behind_remaining);
1489 }
1490
1491 r1_bio->bios[i] = mbio;
1492
1493 mbio->bi_iter.bi_sector = (r1_bio->sector +
1494 conf->mirrors[i].rdev->data_offset);
1495 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1496 mbio->bi_end_io = raid1_end_write_request;
1497 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1498 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1499 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1500 conf->raid_disks - mddev->degraded > 1)
1501 mbio->bi_opf |= MD_FAILFAST;
1502 mbio->bi_private = r1_bio;
1503
1504 atomic_inc(&r1_bio->remaining);
1505
1506 if (mddev->gendisk)
1507 trace_block_bio_remap(mbio->bi_disk->queue,
1508 mbio, disk_devt(mddev->gendisk),
1509 r1_bio->sector);
1510 /* flush_pending_writes() needs access to the rdev so...*/
1511 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1512
1513 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1514 if (cb)
1515 plug = container_of(cb, struct raid1_plug_cb, cb);
1516 else
1517 plug = NULL;
1518 if (plug) {
1519 bio_list_add(&plug->pending, mbio);
1520 plug->pending_cnt++;
1521 } else {
1522 spin_lock_irqsave(&conf->device_lock, flags);
1523 bio_list_add(&conf->pending_bio_list, mbio);
1524 conf->pending_count++;
1525 spin_unlock_irqrestore(&conf->device_lock, flags);
1526 md_wakeup_thread(mddev->thread);
1527 }
1528 }
1529
1530 r1_bio_write_done(r1_bio);
1531
1532 /* In case raid1d snuck in to freeze_array */
1533 wake_up(&conf->wait_barrier);
1534}
1535
1536static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1537{
1538 sector_t sectors;
1539
1540 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1541 md_flush_request(mddev, bio);
1542 return true;
1543 }
1544
1545 /*
1546 * There is a limit to the maximum size, but
1547 * the read/write handler might find a lower limit
1548 * due to bad blocks. To avoid multiple splits,
1549 * we pass the maximum number of sectors down
1550 * and let the lower level perform the split.
1551 */
1552 sectors = align_to_barrier_unit_end(
1553 bio->bi_iter.bi_sector, bio_sectors(bio));
1554
1555 if (bio_data_dir(bio) == READ)
1556 raid1_read_request(mddev, bio, sectors, NULL);
1557 else {
1558 if (!md_write_start(mddev,bio))
1559 return false;
1560 raid1_write_request(mddev, bio, sectors);
1561 }
1562 return true;
1563}
1564
1565static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1566{
1567 struct r1conf *conf = mddev->private;
1568 int i;
1569
1570 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1571 conf->raid_disks - mddev->degraded);
1572 rcu_read_lock();
1573 for (i = 0; i < conf->raid_disks; i++) {
1574 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1575 seq_printf(seq, "%s",
1576 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1577 }
1578 rcu_read_unlock();
1579 seq_printf(seq, "]");
1580}
1581
1582static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1583{
1584 char b[BDEVNAME_SIZE];
1585 struct r1conf *conf = mddev->private;
1586 unsigned long flags;
1587
1588 /*
1589 * If it is not operational, then we have already marked it as dead
1590 * else if it is the last working disks, ignore the error, let the
1591 * next level up know.
1592 * else mark the drive as failed
1593 */
1594 spin_lock_irqsave(&conf->device_lock, flags);
1595 if (test_bit(In_sync, &rdev->flags)
1596 && (conf->raid_disks - mddev->degraded) == 1) {
1597 /*
1598 * Don't fail the drive, act as though we were just a
1599 * normal single drive.
1600 * However don't try a recovery from this drive as
1601 * it is very likely to fail.
1602 */
1603 conf->recovery_disabled = mddev->recovery_disabled;
1604 spin_unlock_irqrestore(&conf->device_lock, flags);
1605 return;
1606 }
1607 set_bit(Blocked, &rdev->flags);
1608 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1609 mddev->degraded++;
1610 set_bit(Faulty, &rdev->flags);
1611 } else
1612 set_bit(Faulty, &rdev->flags);
1613 spin_unlock_irqrestore(&conf->device_lock, flags);
1614 /*
1615 * if recovery is running, make sure it aborts.
1616 */
1617 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1618 set_mask_bits(&mddev->sb_flags, 0,
1619 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1620 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1621 "md/raid1:%s: Operation continuing on %d devices.\n",
1622 mdname(mddev), bdevname(rdev->bdev, b),
1623 mdname(mddev), conf->raid_disks - mddev->degraded);
1624}
1625
1626static void print_conf(struct r1conf *conf)
1627{
1628 int i;
1629
1630 pr_debug("RAID1 conf printout:\n");
1631 if (!conf) {
1632 pr_debug("(!conf)\n");
1633 return;
1634 }
1635 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1636 conf->raid_disks);
1637
1638 rcu_read_lock();
1639 for (i = 0; i < conf->raid_disks; i++) {
1640 char b[BDEVNAME_SIZE];
1641 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1642 if (rdev)
1643 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1644 i, !test_bit(In_sync, &rdev->flags),
1645 !test_bit(Faulty, &rdev->flags),
1646 bdevname(rdev->bdev,b));
1647 }
1648 rcu_read_unlock();
1649}
1650
1651static void close_sync(struct r1conf *conf)
1652{
1653 int idx;
1654
1655 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1656 _wait_barrier(conf, idx);
1657 _allow_barrier(conf, idx);
1658 }
1659
1660 mempool_destroy(conf->r1buf_pool);
1661 conf->r1buf_pool = NULL;
1662}
1663
1664static int raid1_spare_active(struct mddev *mddev)
1665{
1666 int i;
1667 struct r1conf *conf = mddev->private;
1668 int count = 0;
1669 unsigned long flags;
1670
1671 /*
1672 * Find all failed disks within the RAID1 configuration
1673 * and mark them readable.
1674 * Called under mddev lock, so rcu protection not needed.
1675 * device_lock used to avoid races with raid1_end_read_request
1676 * which expects 'In_sync' flags and ->degraded to be consistent.
1677 */
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 for (i = 0; i < conf->raid_disks; i++) {
1680 struct md_rdev *rdev = conf->mirrors[i].rdev;
1681 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1682 if (repl
1683 && !test_bit(Candidate, &repl->flags)
1684 && repl->recovery_offset == MaxSector
1685 && !test_bit(Faulty, &repl->flags)
1686 && !test_and_set_bit(In_sync, &repl->flags)) {
1687 /* replacement has just become active */
1688 if (!rdev ||
1689 !test_and_clear_bit(In_sync, &rdev->flags))
1690 count++;
1691 if (rdev) {
1692 /* Replaced device not technically
1693 * faulty, but we need to be sure
1694 * it gets removed and never re-added
1695 */
1696 set_bit(Faulty, &rdev->flags);
1697 sysfs_notify_dirent_safe(
1698 rdev->sysfs_state);
1699 }
1700 }
1701 if (rdev
1702 && rdev->recovery_offset == MaxSector
1703 && !test_bit(Faulty, &rdev->flags)
1704 && !test_and_set_bit(In_sync, &rdev->flags)) {
1705 count++;
1706 sysfs_notify_dirent_safe(rdev->sysfs_state);
1707 }
1708 }
1709 mddev->degraded -= count;
1710 spin_unlock_irqrestore(&conf->device_lock, flags);
1711
1712 print_conf(conf);
1713 return count;
1714}
1715
1716static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1717{
1718 struct r1conf *conf = mddev->private;
1719 int err = -EEXIST;
1720 int mirror = 0;
1721 struct raid1_info *p;
1722 int first = 0;
1723 int last = conf->raid_disks - 1;
1724
1725 if (mddev->recovery_disabled == conf->recovery_disabled)
1726 return -EBUSY;
1727
1728 if (md_integrity_add_rdev(rdev, mddev))
1729 return -ENXIO;
1730
1731 if (rdev->raid_disk >= 0)
1732 first = last = rdev->raid_disk;
1733
1734 /*
1735 * find the disk ... but prefer rdev->saved_raid_disk
1736 * if possible.
1737 */
1738 if (rdev->saved_raid_disk >= 0 &&
1739 rdev->saved_raid_disk >= first &&
1740 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1741 first = last = rdev->saved_raid_disk;
1742
1743 for (mirror = first; mirror <= last; mirror++) {
1744 p = conf->mirrors+mirror;
1745 if (!p->rdev) {
1746
1747 if (mddev->gendisk)
1748 disk_stack_limits(mddev->gendisk, rdev->bdev,
1749 rdev->data_offset << 9);
1750
1751 p->head_position = 0;
1752 rdev->raid_disk = mirror;
1753 err = 0;
1754 /* As all devices are equivalent, we don't need a full recovery
1755 * if this was recently any drive of the array
1756 */
1757 if (rdev->saved_raid_disk < 0)
1758 conf->fullsync = 1;
1759 rcu_assign_pointer(p->rdev, rdev);
1760 break;
1761 }
1762 if (test_bit(WantReplacement, &p->rdev->flags) &&
1763 p[conf->raid_disks].rdev == NULL) {
1764 /* Add this device as a replacement */
1765 clear_bit(In_sync, &rdev->flags);
1766 set_bit(Replacement, &rdev->flags);
1767 rdev->raid_disk = mirror;
1768 err = 0;
1769 conf->fullsync = 1;
1770 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1771 break;
1772 }
1773 }
1774 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1775 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1776 print_conf(conf);
1777 return err;
1778}
1779
1780static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1781{
1782 struct r1conf *conf = mddev->private;
1783 int err = 0;
1784 int number = rdev->raid_disk;
1785 struct raid1_info *p = conf->mirrors + number;
1786
1787 if (rdev != p->rdev)
1788 p = conf->mirrors + conf->raid_disks + number;
1789
1790 print_conf(conf);
1791 if (rdev == p->rdev) {
1792 if (test_bit(In_sync, &rdev->flags) ||
1793 atomic_read(&rdev->nr_pending)) {
1794 err = -EBUSY;
1795 goto abort;
1796 }
1797 /* Only remove non-faulty devices if recovery
1798 * is not possible.
1799 */
1800 if (!test_bit(Faulty, &rdev->flags) &&
1801 mddev->recovery_disabled != conf->recovery_disabled &&
1802 mddev->degraded < conf->raid_disks) {
1803 err = -EBUSY;
1804 goto abort;
1805 }
1806 p->rdev = NULL;
1807 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1808 synchronize_rcu();
1809 if (atomic_read(&rdev->nr_pending)) {
1810 /* lost the race, try later */
1811 err = -EBUSY;
1812 p->rdev = rdev;
1813 goto abort;
1814 }
1815 }
1816 if (conf->mirrors[conf->raid_disks + number].rdev) {
1817 /* We just removed a device that is being replaced.
1818 * Move down the replacement. We drain all IO before
1819 * doing this to avoid confusion.
1820 */
1821 struct md_rdev *repl =
1822 conf->mirrors[conf->raid_disks + number].rdev;
1823 freeze_array(conf, 0);
1824 if (atomic_read(&repl->nr_pending)) {
1825 /* It means that some queued IO of retry_list
1826 * hold repl. Thus, we cannot set replacement
1827 * as NULL, avoiding rdev NULL pointer
1828 * dereference in sync_request_write and
1829 * handle_write_finished.
1830 */
1831 err = -EBUSY;
1832 unfreeze_array(conf);
1833 goto abort;
1834 }
1835 clear_bit(Replacement, &repl->flags);
1836 p->rdev = repl;
1837 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1838 unfreeze_array(conf);
1839 }
1840
1841 clear_bit(WantReplacement, &rdev->flags);
1842 err = md_integrity_register(mddev);
1843 }
1844abort:
1845
1846 print_conf(conf);
1847 return err;
1848}
1849
1850static void end_sync_read(struct bio *bio)
1851{
1852 struct r1bio *r1_bio = get_resync_r1bio(bio);
1853
1854 update_head_pos(r1_bio->read_disk, r1_bio);
1855
1856 /*
1857 * we have read a block, now it needs to be re-written,
1858 * or re-read if the read failed.
1859 * We don't do much here, just schedule handling by raid1d
1860 */
1861 if (!bio->bi_status)
1862 set_bit(R1BIO_Uptodate, &r1_bio->state);
1863
1864 if (atomic_dec_and_test(&r1_bio->remaining))
1865 reschedule_retry(r1_bio);
1866}
1867
1868static void end_sync_write(struct bio *bio)
1869{
1870 int uptodate = !bio->bi_status;
1871 struct r1bio *r1_bio = get_resync_r1bio(bio);
1872 struct mddev *mddev = r1_bio->mddev;
1873 struct r1conf *conf = mddev->private;
1874 sector_t first_bad;
1875 int bad_sectors;
1876 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1877
1878 if (!uptodate) {
1879 sector_t sync_blocks = 0;
1880 sector_t s = r1_bio->sector;
1881 long sectors_to_go = r1_bio->sectors;
1882 /* make sure these bits doesn't get cleared. */
1883 do {
1884 bitmap_end_sync(mddev->bitmap, s,
1885 &sync_blocks, 1);
1886 s += sync_blocks;
1887 sectors_to_go -= sync_blocks;
1888 } while (sectors_to_go > 0);
1889 set_bit(WriteErrorSeen, &rdev->flags);
1890 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1891 set_bit(MD_RECOVERY_NEEDED, &
1892 mddev->recovery);
1893 set_bit(R1BIO_WriteError, &r1_bio->state);
1894 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1895 &first_bad, &bad_sectors) &&
1896 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1897 r1_bio->sector,
1898 r1_bio->sectors,
1899 &first_bad, &bad_sectors)
1900 )
1901 set_bit(R1BIO_MadeGood, &r1_bio->state);
1902
1903 if (atomic_dec_and_test(&r1_bio->remaining)) {
1904 int s = r1_bio->sectors;
1905 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1906 test_bit(R1BIO_WriteError, &r1_bio->state))
1907 reschedule_retry(r1_bio);
1908 else {
1909 put_buf(r1_bio);
1910 md_done_sync(mddev, s, uptodate);
1911 }
1912 }
1913}
1914
1915static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1916 int sectors, struct page *page, int rw)
1917{
1918 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1919 /* success */
1920 return 1;
1921 if (rw == WRITE) {
1922 set_bit(WriteErrorSeen, &rdev->flags);
1923 if (!test_and_set_bit(WantReplacement,
1924 &rdev->flags))
1925 set_bit(MD_RECOVERY_NEEDED, &
1926 rdev->mddev->recovery);
1927 }
1928 /* need to record an error - either for the block or the device */
1929 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1930 md_error(rdev->mddev, rdev);
1931 return 0;
1932}
1933
1934static int fix_sync_read_error(struct r1bio *r1_bio)
1935{
1936 /* Try some synchronous reads of other devices to get
1937 * good data, much like with normal read errors. Only
1938 * read into the pages we already have so we don't
1939 * need to re-issue the read request.
1940 * We don't need to freeze the array, because being in an
1941 * active sync request, there is no normal IO, and
1942 * no overlapping syncs.
1943 * We don't need to check is_badblock() again as we
1944 * made sure that anything with a bad block in range
1945 * will have bi_end_io clear.
1946 */
1947 struct mddev *mddev = r1_bio->mddev;
1948 struct r1conf *conf = mddev->private;
1949 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1950 struct page **pages = get_resync_pages(bio)->pages;
1951 sector_t sect = r1_bio->sector;
1952 int sectors = r1_bio->sectors;
1953 int idx = 0;
1954 struct md_rdev *rdev;
1955
1956 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1957 if (test_bit(FailFast, &rdev->flags)) {
1958 /* Don't try recovering from here - just fail it
1959 * ... unless it is the last working device of course */
1960 md_error(mddev, rdev);
1961 if (test_bit(Faulty, &rdev->flags))
1962 /* Don't try to read from here, but make sure
1963 * put_buf does it's thing
1964 */
1965 bio->bi_end_io = end_sync_write;
1966 }
1967
1968 while(sectors) {
1969 int s = sectors;
1970 int d = r1_bio->read_disk;
1971 int success = 0;
1972 int start;
1973
1974 if (s > (PAGE_SIZE>>9))
1975 s = PAGE_SIZE >> 9;
1976 do {
1977 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1978 /* No rcu protection needed here devices
1979 * can only be removed when no resync is
1980 * active, and resync is currently active
1981 */
1982 rdev = conf->mirrors[d].rdev;
1983 if (sync_page_io(rdev, sect, s<<9,
1984 pages[idx],
1985 REQ_OP_READ, 0, false)) {
1986 success = 1;
1987 break;
1988 }
1989 }
1990 d++;
1991 if (d == conf->raid_disks * 2)
1992 d = 0;
1993 } while (!success && d != r1_bio->read_disk);
1994
1995 if (!success) {
1996 char b[BDEVNAME_SIZE];
1997 int abort = 0;
1998 /* Cannot read from anywhere, this block is lost.
1999 * Record a bad block on each device. If that doesn't
2000 * work just disable and interrupt the recovery.
2001 * Don't fail devices as that won't really help.
2002 */
2003 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2004 mdname(mddev), bio_devname(bio, b),
2005 (unsigned long long)r1_bio->sector);
2006 for (d = 0; d < conf->raid_disks * 2; d++) {
2007 rdev = conf->mirrors[d].rdev;
2008 if (!rdev || test_bit(Faulty, &rdev->flags))
2009 continue;
2010 if (!rdev_set_badblocks(rdev, sect, s, 0))
2011 abort = 1;
2012 }
2013 if (abort) {
2014 conf->recovery_disabled =
2015 mddev->recovery_disabled;
2016 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2017 md_done_sync(mddev, r1_bio->sectors, 0);
2018 put_buf(r1_bio);
2019 return 0;
2020 }
2021 /* Try next page */
2022 sectors -= s;
2023 sect += s;
2024 idx++;
2025 continue;
2026 }
2027
2028 start = d;
2029 /* write it back and re-read */
2030 while (d != r1_bio->read_disk) {
2031 if (d == 0)
2032 d = conf->raid_disks * 2;
2033 d--;
2034 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2035 continue;
2036 rdev = conf->mirrors[d].rdev;
2037 if (r1_sync_page_io(rdev, sect, s,
2038 pages[idx],
2039 WRITE) == 0) {
2040 r1_bio->bios[d]->bi_end_io = NULL;
2041 rdev_dec_pending(rdev, mddev);
2042 }
2043 }
2044 d = start;
2045 while (d != r1_bio->read_disk) {
2046 if (d == 0)
2047 d = conf->raid_disks * 2;
2048 d--;
2049 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2050 continue;
2051 rdev = conf->mirrors[d].rdev;
2052 if (r1_sync_page_io(rdev, sect, s,
2053 pages[idx],
2054 READ) != 0)
2055 atomic_add(s, &rdev->corrected_errors);
2056 }
2057 sectors -= s;
2058 sect += s;
2059 idx ++;
2060 }
2061 set_bit(R1BIO_Uptodate, &r1_bio->state);
2062 bio->bi_status = 0;
2063 return 1;
2064}
2065
2066static void process_checks(struct r1bio *r1_bio)
2067{
2068 /* We have read all readable devices. If we haven't
2069 * got the block, then there is no hope left.
2070 * If we have, then we want to do a comparison
2071 * and skip the write if everything is the same.
2072 * If any blocks failed to read, then we need to
2073 * attempt an over-write
2074 */
2075 struct mddev *mddev = r1_bio->mddev;
2076 struct r1conf *conf = mddev->private;
2077 int primary;
2078 int i;
2079 int vcnt;
2080
2081 /* Fix variable parts of all bios */
2082 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2083 for (i = 0; i < conf->raid_disks * 2; i++) {
2084 blk_status_t status;
2085 struct bio *b = r1_bio->bios[i];
2086 struct resync_pages *rp = get_resync_pages(b);
2087 if (b->bi_end_io != end_sync_read)
2088 continue;
2089 /* fixup the bio for reuse, but preserve errno */
2090 status = b->bi_status;
2091 bio_reset(b);
2092 b->bi_status = status;
2093 b->bi_iter.bi_sector = r1_bio->sector +
2094 conf->mirrors[i].rdev->data_offset;
2095 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2096 b->bi_end_io = end_sync_read;
2097 rp->raid_bio = r1_bio;
2098 b->bi_private = rp;
2099
2100 /* initialize bvec table again */
2101 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2102 }
2103 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2104 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2105 !r1_bio->bios[primary]->bi_status) {
2106 r1_bio->bios[primary]->bi_end_io = NULL;
2107 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2108 break;
2109 }
2110 r1_bio->read_disk = primary;
2111 for (i = 0; i < conf->raid_disks * 2; i++) {
2112 int j;
2113 struct bio *pbio = r1_bio->bios[primary];
2114 struct bio *sbio = r1_bio->bios[i];
2115 blk_status_t status = sbio->bi_status;
2116 struct page **ppages = get_resync_pages(pbio)->pages;
2117 struct page **spages = get_resync_pages(sbio)->pages;
2118 struct bio_vec *bi;
2119 int page_len[RESYNC_PAGES] = { 0 };
2120
2121 if (sbio->bi_end_io != end_sync_read)
2122 continue;
2123 /* Now we can 'fixup' the error value */
2124 sbio->bi_status = 0;
2125
2126 bio_for_each_segment_all(bi, sbio, j)
2127 page_len[j] = bi->bv_len;
2128
2129 if (!status) {
2130 for (j = vcnt; j-- ; ) {
2131 if (memcmp(page_address(ppages[j]),
2132 page_address(spages[j]),
2133 page_len[j]))
2134 break;
2135 }
2136 } else
2137 j = 0;
2138 if (j >= 0)
2139 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2140 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2141 && !status)) {
2142 /* No need to write to this device. */
2143 sbio->bi_end_io = NULL;
2144 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2145 continue;
2146 }
2147
2148 bio_copy_data(sbio, pbio);
2149 }
2150}
2151
2152static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2153{
2154 struct r1conf *conf = mddev->private;
2155 int i;
2156 int disks = conf->raid_disks * 2;
2157 struct bio *wbio;
2158
2159 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2160 /* ouch - failed to read all of that. */
2161 if (!fix_sync_read_error(r1_bio))
2162 return;
2163
2164 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2165 process_checks(r1_bio);
2166
2167 /*
2168 * schedule writes
2169 */
2170 atomic_set(&r1_bio->remaining, 1);
2171 for (i = 0; i < disks ; i++) {
2172 wbio = r1_bio->bios[i];
2173 if (wbio->bi_end_io == NULL ||
2174 (wbio->bi_end_io == end_sync_read &&
2175 (i == r1_bio->read_disk ||
2176 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2177 continue;
2178 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2179 continue;
2180
2181 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2182 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2183 wbio->bi_opf |= MD_FAILFAST;
2184
2185 wbio->bi_end_io = end_sync_write;
2186 atomic_inc(&r1_bio->remaining);
2187 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2188
2189 generic_make_request(wbio);
2190 }
2191
2192 if (atomic_dec_and_test(&r1_bio->remaining)) {
2193 /* if we're here, all write(s) have completed, so clean up */
2194 int s = r1_bio->sectors;
2195 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2196 test_bit(R1BIO_WriteError, &r1_bio->state))
2197 reschedule_retry(r1_bio);
2198 else {
2199 put_buf(r1_bio);
2200 md_done_sync(mddev, s, 1);
2201 }
2202 }
2203}
2204
2205/*
2206 * This is a kernel thread which:
2207 *
2208 * 1. Retries failed read operations on working mirrors.
2209 * 2. Updates the raid superblock when problems encounter.
2210 * 3. Performs writes following reads for array synchronising.
2211 */
2212
2213static void fix_read_error(struct r1conf *conf, int read_disk,
2214 sector_t sect, int sectors)
2215{
2216 struct mddev *mddev = conf->mddev;
2217 while(sectors) {
2218 int s = sectors;
2219 int d = read_disk;
2220 int success = 0;
2221 int start;
2222 struct md_rdev *rdev;
2223
2224 if (s > (PAGE_SIZE>>9))
2225 s = PAGE_SIZE >> 9;
2226
2227 do {
2228 sector_t first_bad;
2229 int bad_sectors;
2230
2231 rcu_read_lock();
2232 rdev = rcu_dereference(conf->mirrors[d].rdev);
2233 if (rdev &&
2234 (test_bit(In_sync, &rdev->flags) ||
2235 (!test_bit(Faulty, &rdev->flags) &&
2236 rdev->recovery_offset >= sect + s)) &&
2237 is_badblock(rdev, sect, s,
2238 &first_bad, &bad_sectors) == 0) {
2239 atomic_inc(&rdev->nr_pending);
2240 rcu_read_unlock();
2241 if (sync_page_io(rdev, sect, s<<9,
2242 conf->tmppage, REQ_OP_READ, 0, false))
2243 success = 1;
2244 rdev_dec_pending(rdev, mddev);
2245 if (success)
2246 break;
2247 } else
2248 rcu_read_unlock();
2249 d++;
2250 if (d == conf->raid_disks * 2)
2251 d = 0;
2252 } while (!success && d != read_disk);
2253
2254 if (!success) {
2255 /* Cannot read from anywhere - mark it bad */
2256 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2257 if (!rdev_set_badblocks(rdev, sect, s, 0))
2258 md_error(mddev, rdev);
2259 break;
2260 }
2261 /* write it back and re-read */
2262 start = d;
2263 while (d != read_disk) {
2264 if (d==0)
2265 d = conf->raid_disks * 2;
2266 d--;
2267 rcu_read_lock();
2268 rdev = rcu_dereference(conf->mirrors[d].rdev);
2269 if (rdev &&
2270 !test_bit(Faulty, &rdev->flags)) {
2271 atomic_inc(&rdev->nr_pending);
2272 rcu_read_unlock();
2273 r1_sync_page_io(rdev, sect, s,
2274 conf->tmppage, WRITE);
2275 rdev_dec_pending(rdev, mddev);
2276 } else
2277 rcu_read_unlock();
2278 }
2279 d = start;
2280 while (d != read_disk) {
2281 char b[BDEVNAME_SIZE];
2282 if (d==0)
2283 d = conf->raid_disks * 2;
2284 d--;
2285 rcu_read_lock();
2286 rdev = rcu_dereference(conf->mirrors[d].rdev);
2287 if (rdev &&
2288 !test_bit(Faulty, &rdev->flags)) {
2289 atomic_inc(&rdev->nr_pending);
2290 rcu_read_unlock();
2291 if (r1_sync_page_io(rdev, sect, s,
2292 conf->tmppage, READ)) {
2293 atomic_add(s, &rdev->corrected_errors);
2294 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2295 mdname(mddev), s,
2296 (unsigned long long)(sect +
2297 rdev->data_offset),
2298 bdevname(rdev->bdev, b));
2299 }
2300 rdev_dec_pending(rdev, mddev);
2301 } else
2302 rcu_read_unlock();
2303 }
2304 sectors -= s;
2305 sect += s;
2306 }
2307}
2308
2309static int narrow_write_error(struct r1bio *r1_bio, int i)
2310{
2311 struct mddev *mddev = r1_bio->mddev;
2312 struct r1conf *conf = mddev->private;
2313 struct md_rdev *rdev = conf->mirrors[i].rdev;
2314
2315 /* bio has the data to be written to device 'i' where
2316 * we just recently had a write error.
2317 * We repeatedly clone the bio and trim down to one block,
2318 * then try the write. Where the write fails we record
2319 * a bad block.
2320 * It is conceivable that the bio doesn't exactly align with
2321 * blocks. We must handle this somehow.
2322 *
2323 * We currently own a reference on the rdev.
2324 */
2325
2326 int block_sectors;
2327 sector_t sector;
2328 int sectors;
2329 int sect_to_write = r1_bio->sectors;
2330 int ok = 1;
2331
2332 if (rdev->badblocks.shift < 0)
2333 return 0;
2334
2335 block_sectors = roundup(1 << rdev->badblocks.shift,
2336 bdev_logical_block_size(rdev->bdev) >> 9);
2337 sector = r1_bio->sector;
2338 sectors = ((sector + block_sectors)
2339 & ~(sector_t)(block_sectors - 1))
2340 - sector;
2341
2342 while (sect_to_write) {
2343 struct bio *wbio;
2344 if (sectors > sect_to_write)
2345 sectors = sect_to_write;
2346 /* Write at 'sector' for 'sectors'*/
2347
2348 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2349 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2350 GFP_NOIO,
2351 mddev->bio_set);
2352 } else {
2353 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2354 mddev->bio_set);
2355 }
2356
2357 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2358 wbio->bi_iter.bi_sector = r1_bio->sector;
2359 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2360
2361 bio_trim(wbio, sector - r1_bio->sector, sectors);
2362 wbio->bi_iter.bi_sector += rdev->data_offset;
2363 bio_set_dev(wbio, rdev->bdev);
2364
2365 if (submit_bio_wait(wbio) < 0)
2366 /* failure! */
2367 ok = rdev_set_badblocks(rdev, sector,
2368 sectors, 0)
2369 && ok;
2370
2371 bio_put(wbio);
2372 sect_to_write -= sectors;
2373 sector += sectors;
2374 sectors = block_sectors;
2375 }
2376 return ok;
2377}
2378
2379static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2380{
2381 int m;
2382 int s = r1_bio->sectors;
2383 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2384 struct md_rdev *rdev = conf->mirrors[m].rdev;
2385 struct bio *bio = r1_bio->bios[m];
2386 if (bio->bi_end_io == NULL)
2387 continue;
2388 if (!bio->bi_status &&
2389 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2390 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2391 }
2392 if (bio->bi_status &&
2393 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2394 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2395 md_error(conf->mddev, rdev);
2396 }
2397 }
2398 put_buf(r1_bio);
2399 md_done_sync(conf->mddev, s, 1);
2400}
2401
2402static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2403{
2404 int m, idx;
2405 bool fail = false;
2406
2407 for (m = 0; m < conf->raid_disks * 2 ; m++)
2408 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2409 struct md_rdev *rdev = conf->mirrors[m].rdev;
2410 rdev_clear_badblocks(rdev,
2411 r1_bio->sector,
2412 r1_bio->sectors, 0);
2413 rdev_dec_pending(rdev, conf->mddev);
2414 } else if (r1_bio->bios[m] != NULL) {
2415 /* This drive got a write error. We need to
2416 * narrow down and record precise write
2417 * errors.
2418 */
2419 fail = true;
2420 if (!narrow_write_error(r1_bio, m)) {
2421 md_error(conf->mddev,
2422 conf->mirrors[m].rdev);
2423 /* an I/O failed, we can't clear the bitmap */
2424 set_bit(R1BIO_Degraded, &r1_bio->state);
2425 }
2426 rdev_dec_pending(conf->mirrors[m].rdev,
2427 conf->mddev);
2428 }
2429 if (fail) {
2430 spin_lock_irq(&conf->device_lock);
2431 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2432 idx = sector_to_idx(r1_bio->sector);
2433 atomic_inc(&conf->nr_queued[idx]);
2434 spin_unlock_irq(&conf->device_lock);
2435 /*
2436 * In case freeze_array() is waiting for condition
2437 * get_unqueued_pending() == extra to be true.
2438 */
2439 wake_up(&conf->wait_barrier);
2440 md_wakeup_thread(conf->mddev->thread);
2441 } else {
2442 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2443 close_write(r1_bio);
2444 raid_end_bio_io(r1_bio);
2445 }
2446}
2447
2448static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2449{
2450 struct mddev *mddev = conf->mddev;
2451 struct bio *bio;
2452 struct md_rdev *rdev;
2453 sector_t bio_sector;
2454
2455 clear_bit(R1BIO_ReadError, &r1_bio->state);
2456 /* we got a read error. Maybe the drive is bad. Maybe just
2457 * the block and we can fix it.
2458 * We freeze all other IO, and try reading the block from
2459 * other devices. When we find one, we re-write
2460 * and check it that fixes the read error.
2461 * This is all done synchronously while the array is
2462 * frozen
2463 */
2464
2465 bio = r1_bio->bios[r1_bio->read_disk];
2466 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2467 bio_put(bio);
2468 r1_bio->bios[r1_bio->read_disk] = NULL;
2469
2470 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2471 if (mddev->ro == 0
2472 && !test_bit(FailFast, &rdev->flags)) {
2473 freeze_array(conf, 1);
2474 fix_read_error(conf, r1_bio->read_disk,
2475 r1_bio->sector, r1_bio->sectors);
2476 unfreeze_array(conf);
2477 } else {
2478 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2479 }
2480
2481 rdev_dec_pending(rdev, conf->mddev);
2482 allow_barrier(conf, r1_bio->sector);
2483 bio = r1_bio->master_bio;
2484
2485 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2486 r1_bio->state = 0;
2487 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2488}
2489
2490static void raid1d(struct md_thread *thread)
2491{
2492 struct mddev *mddev = thread->mddev;
2493 struct r1bio *r1_bio;
2494 unsigned long flags;
2495 struct r1conf *conf = mddev->private;
2496 struct list_head *head = &conf->retry_list;
2497 struct blk_plug plug;
2498 int idx;
2499
2500 md_check_recovery(mddev);
2501
2502 if (!list_empty_careful(&conf->bio_end_io_list) &&
2503 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2504 LIST_HEAD(tmp);
2505 spin_lock_irqsave(&conf->device_lock, flags);
2506 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2507 list_splice_init(&conf->bio_end_io_list, &tmp);
2508 spin_unlock_irqrestore(&conf->device_lock, flags);
2509 while (!list_empty(&tmp)) {
2510 r1_bio = list_first_entry(&tmp, struct r1bio,
2511 retry_list);
2512 list_del(&r1_bio->retry_list);
2513 idx = sector_to_idx(r1_bio->sector);
2514 atomic_dec(&conf->nr_queued[idx]);
2515 if (mddev->degraded)
2516 set_bit(R1BIO_Degraded, &r1_bio->state);
2517 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2518 close_write(r1_bio);
2519 raid_end_bio_io(r1_bio);
2520 }
2521 }
2522
2523 blk_start_plug(&plug);
2524 for (;;) {
2525
2526 flush_pending_writes(conf);
2527
2528 spin_lock_irqsave(&conf->device_lock, flags);
2529 if (list_empty(head)) {
2530 spin_unlock_irqrestore(&conf->device_lock, flags);
2531 break;
2532 }
2533 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2534 list_del(head->prev);
2535 idx = sector_to_idx(r1_bio->sector);
2536 atomic_dec(&conf->nr_queued[idx]);
2537 spin_unlock_irqrestore(&conf->device_lock, flags);
2538
2539 mddev = r1_bio->mddev;
2540 conf = mddev->private;
2541 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2542 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2543 test_bit(R1BIO_WriteError, &r1_bio->state))
2544 handle_sync_write_finished(conf, r1_bio);
2545 else
2546 sync_request_write(mddev, r1_bio);
2547 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2548 test_bit(R1BIO_WriteError, &r1_bio->state))
2549 handle_write_finished(conf, r1_bio);
2550 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2551 handle_read_error(conf, r1_bio);
2552 else
2553 WARN_ON_ONCE(1);
2554
2555 cond_resched();
2556 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2557 md_check_recovery(mddev);
2558 }
2559 blk_finish_plug(&plug);
2560}
2561
2562static int init_resync(struct r1conf *conf)
2563{
2564 int buffs;
2565
2566 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2567 BUG_ON(conf->r1buf_pool);
2568 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2569 conf->poolinfo);
2570 if (!conf->r1buf_pool)
2571 return -ENOMEM;
2572 return 0;
2573}
2574
2575static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2576{
2577 struct r1bio *r1bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2578 struct resync_pages *rps;
2579 struct bio *bio;
2580 int i;
2581
2582 for (i = conf->poolinfo->raid_disks; i--; ) {
2583 bio = r1bio->bios[i];
2584 rps = bio->bi_private;
2585 bio_reset(bio);
2586 bio->bi_private = rps;
2587 }
2588 r1bio->master_bio = NULL;
2589 return r1bio;
2590}
2591
2592/*
2593 * perform a "sync" on one "block"
2594 *
2595 * We need to make sure that no normal I/O request - particularly write
2596 * requests - conflict with active sync requests.
2597 *
2598 * This is achieved by tracking pending requests and a 'barrier' concept
2599 * that can be installed to exclude normal IO requests.
2600 */
2601
2602static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2603 int *skipped)
2604{
2605 struct r1conf *conf = mddev->private;
2606 struct r1bio *r1_bio;
2607 struct bio *bio;
2608 sector_t max_sector, nr_sectors;
2609 int disk = -1;
2610 int i;
2611 int wonly = -1;
2612 int write_targets = 0, read_targets = 0;
2613 sector_t sync_blocks;
2614 int still_degraded = 0;
2615 int good_sectors = RESYNC_SECTORS;
2616 int min_bad = 0; /* number of sectors that are bad in all devices */
2617 int idx = sector_to_idx(sector_nr);
2618 int page_idx = 0;
2619
2620 if (!conf->r1buf_pool)
2621 if (init_resync(conf))
2622 return 0;
2623
2624 max_sector = mddev->dev_sectors;
2625 if (sector_nr >= max_sector) {
2626 /* If we aborted, we need to abort the
2627 * sync on the 'current' bitmap chunk (there will
2628 * only be one in raid1 resync.
2629 * We can find the current addess in mddev->curr_resync
2630 */
2631 if (mddev->curr_resync < max_sector) /* aborted */
2632 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2633 &sync_blocks, 1);
2634 else /* completed sync */
2635 conf->fullsync = 0;
2636
2637 bitmap_close_sync(mddev->bitmap);
2638 close_sync(conf);
2639
2640 if (mddev_is_clustered(mddev)) {
2641 conf->cluster_sync_low = 0;
2642 conf->cluster_sync_high = 0;
2643 }
2644 return 0;
2645 }
2646
2647 if (mddev->bitmap == NULL &&
2648 mddev->recovery_cp == MaxSector &&
2649 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2650 conf->fullsync == 0) {
2651 *skipped = 1;
2652 return max_sector - sector_nr;
2653 }
2654 /* before building a request, check if we can skip these blocks..
2655 * This call the bitmap_start_sync doesn't actually record anything
2656 */
2657 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2658 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2659 /* We can skip this block, and probably several more */
2660 *skipped = 1;
2661 return sync_blocks;
2662 }
2663
2664 /*
2665 * If there is non-resync activity waiting for a turn, then let it
2666 * though before starting on this new sync request.
2667 */
2668 if (atomic_read(&conf->nr_waiting[idx]))
2669 schedule_timeout_uninterruptible(1);
2670
2671 /* we are incrementing sector_nr below. To be safe, we check against
2672 * sector_nr + two times RESYNC_SECTORS
2673 */
2674
2675 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2676 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2677
2678
2679 if (raise_barrier(conf, sector_nr))
2680 return 0;
2681
2682 r1_bio = raid1_alloc_init_r1buf(conf);
2683
2684 rcu_read_lock();
2685 /*
2686 * If we get a correctably read error during resync or recovery,
2687 * we might want to read from a different device. So we
2688 * flag all drives that could conceivably be read from for READ,
2689 * and any others (which will be non-In_sync devices) for WRITE.
2690 * If a read fails, we try reading from something else for which READ
2691 * is OK.
2692 */
2693
2694 r1_bio->mddev = mddev;
2695 r1_bio->sector = sector_nr;
2696 r1_bio->state = 0;
2697 set_bit(R1BIO_IsSync, &r1_bio->state);
2698 /* make sure good_sectors won't go across barrier unit boundary */
2699 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2700
2701 for (i = 0; i < conf->raid_disks * 2; i++) {
2702 struct md_rdev *rdev;
2703 bio = r1_bio->bios[i];
2704
2705 rdev = rcu_dereference(conf->mirrors[i].rdev);
2706 if (rdev == NULL ||
2707 test_bit(Faulty, &rdev->flags)) {
2708 if (i < conf->raid_disks)
2709 still_degraded = 1;
2710 } else if (!test_bit(In_sync, &rdev->flags)) {
2711 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2712 bio->bi_end_io = end_sync_write;
2713 write_targets ++;
2714 } else {
2715 /* may need to read from here */
2716 sector_t first_bad = MaxSector;
2717 int bad_sectors;
2718
2719 if (is_badblock(rdev, sector_nr, good_sectors,
2720 &first_bad, &bad_sectors)) {
2721 if (first_bad > sector_nr)
2722 good_sectors = first_bad - sector_nr;
2723 else {
2724 bad_sectors -= (sector_nr - first_bad);
2725 if (min_bad == 0 ||
2726 min_bad > bad_sectors)
2727 min_bad = bad_sectors;
2728 }
2729 }
2730 if (sector_nr < first_bad) {
2731 if (test_bit(WriteMostly, &rdev->flags)) {
2732 if (wonly < 0)
2733 wonly = i;
2734 } else {
2735 if (disk < 0)
2736 disk = i;
2737 }
2738 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2739 bio->bi_end_io = end_sync_read;
2740 read_targets++;
2741 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2742 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2743 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2744 /*
2745 * The device is suitable for reading (InSync),
2746 * but has bad block(s) here. Let's try to correct them,
2747 * if we are doing resync or repair. Otherwise, leave
2748 * this device alone for this sync request.
2749 */
2750 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2751 bio->bi_end_io = end_sync_write;
2752 write_targets++;
2753 }
2754 }
2755 if (bio->bi_end_io) {
2756 atomic_inc(&rdev->nr_pending);
2757 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2758 bio_set_dev(bio, rdev->bdev);
2759 if (test_bit(FailFast, &rdev->flags))
2760 bio->bi_opf |= MD_FAILFAST;
2761 }
2762 }
2763 rcu_read_unlock();
2764 if (disk < 0)
2765 disk = wonly;
2766 r1_bio->read_disk = disk;
2767
2768 if (read_targets == 0 && min_bad > 0) {
2769 /* These sectors are bad on all InSync devices, so we
2770 * need to mark them bad on all write targets
2771 */
2772 int ok = 1;
2773 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2774 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2775 struct md_rdev *rdev = conf->mirrors[i].rdev;
2776 ok = rdev_set_badblocks(rdev, sector_nr,
2777 min_bad, 0
2778 ) && ok;
2779 }
2780 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2781 *skipped = 1;
2782 put_buf(r1_bio);
2783
2784 if (!ok) {
2785 /* Cannot record the badblocks, so need to
2786 * abort the resync.
2787 * If there are multiple read targets, could just
2788 * fail the really bad ones ???
2789 */
2790 conf->recovery_disabled = mddev->recovery_disabled;
2791 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2792 return 0;
2793 } else
2794 return min_bad;
2795
2796 }
2797 if (min_bad > 0 && min_bad < good_sectors) {
2798 /* only resync enough to reach the next bad->good
2799 * transition */
2800 good_sectors = min_bad;
2801 }
2802
2803 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2804 /* extra read targets are also write targets */
2805 write_targets += read_targets-1;
2806
2807 if (write_targets == 0 || read_targets == 0) {
2808 /* There is nowhere to write, so all non-sync
2809 * drives must be failed - so we are finished
2810 */
2811 sector_t rv;
2812 if (min_bad > 0)
2813 max_sector = sector_nr + min_bad;
2814 rv = max_sector - sector_nr;
2815 *skipped = 1;
2816 put_buf(r1_bio);
2817 return rv;
2818 }
2819
2820 if (max_sector > mddev->resync_max)
2821 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2822 if (max_sector > sector_nr + good_sectors)
2823 max_sector = sector_nr + good_sectors;
2824 nr_sectors = 0;
2825 sync_blocks = 0;
2826 do {
2827 struct page *page;
2828 int len = PAGE_SIZE;
2829 if (sector_nr + (len>>9) > max_sector)
2830 len = (max_sector - sector_nr) << 9;
2831 if (len == 0)
2832 break;
2833 if (sync_blocks == 0) {
2834 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2835 &sync_blocks, still_degraded) &&
2836 !conf->fullsync &&
2837 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2838 break;
2839 if ((len >> 9) > sync_blocks)
2840 len = sync_blocks<<9;
2841 }
2842
2843 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2844 struct resync_pages *rp;
2845
2846 bio = r1_bio->bios[i];
2847 rp = get_resync_pages(bio);
2848 if (bio->bi_end_io) {
2849 page = resync_fetch_page(rp, page_idx);
2850
2851 /*
2852 * won't fail because the vec table is big
2853 * enough to hold all these pages
2854 */
2855 bio_add_page(bio, page, len, 0);
2856 }
2857 }
2858 nr_sectors += len>>9;
2859 sector_nr += len>>9;
2860 sync_blocks -= (len>>9);
2861 } while (++page_idx < RESYNC_PAGES);
2862
2863 r1_bio->sectors = nr_sectors;
2864
2865 if (mddev_is_clustered(mddev) &&
2866 conf->cluster_sync_high < sector_nr + nr_sectors) {
2867 conf->cluster_sync_low = mddev->curr_resync_completed;
2868 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2869 /* Send resync message */
2870 md_cluster_ops->resync_info_update(mddev,
2871 conf->cluster_sync_low,
2872 conf->cluster_sync_high);
2873 }
2874
2875 /* For a user-requested sync, we read all readable devices and do a
2876 * compare
2877 */
2878 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2879 atomic_set(&r1_bio->remaining, read_targets);
2880 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2881 bio = r1_bio->bios[i];
2882 if (bio->bi_end_io == end_sync_read) {
2883 read_targets--;
2884 md_sync_acct_bio(bio, nr_sectors);
2885 if (read_targets == 1)
2886 bio->bi_opf &= ~MD_FAILFAST;
2887 generic_make_request(bio);
2888 }
2889 }
2890 } else {
2891 atomic_set(&r1_bio->remaining, 1);
2892 bio = r1_bio->bios[r1_bio->read_disk];
2893 md_sync_acct_bio(bio, nr_sectors);
2894 if (read_targets == 1)
2895 bio->bi_opf &= ~MD_FAILFAST;
2896 generic_make_request(bio);
2897
2898 }
2899 return nr_sectors;
2900}
2901
2902static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2903{
2904 if (sectors)
2905 return sectors;
2906
2907 return mddev->dev_sectors;
2908}
2909
2910static struct r1conf *setup_conf(struct mddev *mddev)
2911{
2912 struct r1conf *conf;
2913 int i;
2914 struct raid1_info *disk;
2915 struct md_rdev *rdev;
2916 int err = -ENOMEM;
2917
2918 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2919 if (!conf)
2920 goto abort;
2921
2922 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2923 sizeof(atomic_t), GFP_KERNEL);
2924 if (!conf->nr_pending)
2925 goto abort;
2926
2927 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2928 sizeof(atomic_t), GFP_KERNEL);
2929 if (!conf->nr_waiting)
2930 goto abort;
2931
2932 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2933 sizeof(atomic_t), GFP_KERNEL);
2934 if (!conf->nr_queued)
2935 goto abort;
2936
2937 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2938 sizeof(atomic_t), GFP_KERNEL);
2939 if (!conf->barrier)
2940 goto abort;
2941
2942 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2943 * mddev->raid_disks * 2,
2944 GFP_KERNEL);
2945 if (!conf->mirrors)
2946 goto abort;
2947
2948 conf->tmppage = alloc_page(GFP_KERNEL);
2949 if (!conf->tmppage)
2950 goto abort;
2951
2952 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2953 if (!conf->poolinfo)
2954 goto abort;
2955 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2956 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2957 r1bio_pool_free,
2958 conf->poolinfo);
2959 if (!conf->r1bio_pool)
2960 goto abort;
2961
2962 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2963 if (!conf->bio_split)
2964 goto abort;
2965
2966 conf->poolinfo->mddev = mddev;
2967
2968 err = -EINVAL;
2969 spin_lock_init(&conf->device_lock);
2970 rdev_for_each(rdev, mddev) {
2971 int disk_idx = rdev->raid_disk;
2972 if (disk_idx >= mddev->raid_disks
2973 || disk_idx < 0)
2974 continue;
2975 if (test_bit(Replacement, &rdev->flags))
2976 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2977 else
2978 disk = conf->mirrors + disk_idx;
2979
2980 if (disk->rdev)
2981 goto abort;
2982 disk->rdev = rdev;
2983 disk->head_position = 0;
2984 disk->seq_start = MaxSector;
2985 }
2986 conf->raid_disks = mddev->raid_disks;
2987 conf->mddev = mddev;
2988 INIT_LIST_HEAD(&conf->retry_list);
2989 INIT_LIST_HEAD(&conf->bio_end_io_list);
2990
2991 spin_lock_init(&conf->resync_lock);
2992 init_waitqueue_head(&conf->wait_barrier);
2993
2994 bio_list_init(&conf->pending_bio_list);
2995 conf->pending_count = 0;
2996 conf->recovery_disabled = mddev->recovery_disabled - 1;
2997
2998 err = -EIO;
2999 for (i = 0; i < conf->raid_disks * 2; i++) {
3000
3001 disk = conf->mirrors + i;
3002
3003 if (i < conf->raid_disks &&
3004 disk[conf->raid_disks].rdev) {
3005 /* This slot has a replacement. */
3006 if (!disk->rdev) {
3007 /* No original, just make the replacement
3008 * a recovering spare
3009 */
3010 disk->rdev =
3011 disk[conf->raid_disks].rdev;
3012 disk[conf->raid_disks].rdev = NULL;
3013 } else if (!test_bit(In_sync, &disk->rdev->flags))
3014 /* Original is not in_sync - bad */
3015 goto abort;
3016 }
3017
3018 if (!disk->rdev ||
3019 !test_bit(In_sync, &disk->rdev->flags)) {
3020 disk->head_position = 0;
3021 if (disk->rdev &&
3022 (disk->rdev->saved_raid_disk < 0))
3023 conf->fullsync = 1;
3024 }
3025 }
3026
3027 err = -ENOMEM;
3028 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3029 if (!conf->thread)
3030 goto abort;
3031
3032 return conf;
3033
3034 abort:
3035 if (conf) {
3036 mempool_destroy(conf->r1bio_pool);
3037 kfree(conf->mirrors);
3038 safe_put_page(conf->tmppage);
3039 kfree(conf->poolinfo);
3040 kfree(conf->nr_pending);
3041 kfree(conf->nr_waiting);
3042 kfree(conf->nr_queued);
3043 kfree(conf->barrier);
3044 if (conf->bio_split)
3045 bioset_free(conf->bio_split);
3046 kfree(conf);
3047 }
3048 return ERR_PTR(err);
3049}
3050
3051static void raid1_free(struct mddev *mddev, void *priv);
3052static int raid1_run(struct mddev *mddev)
3053{
3054 struct r1conf *conf;
3055 int i;
3056 struct md_rdev *rdev;
3057 int ret;
3058 bool discard_supported = false;
3059
3060 if (mddev->level != 1) {
3061 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3062 mdname(mddev), mddev->level);
3063 return -EIO;
3064 }
3065 if (mddev->reshape_position != MaxSector) {
3066 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3067 mdname(mddev));
3068 return -EIO;
3069 }
3070 if (mddev_init_writes_pending(mddev) < 0)
3071 return -ENOMEM;
3072 /*
3073 * copy the already verified devices into our private RAID1
3074 * bookkeeping area. [whatever we allocate in run(),
3075 * should be freed in raid1_free()]
3076 */
3077 if (mddev->private == NULL)
3078 conf = setup_conf(mddev);
3079 else
3080 conf = mddev->private;
3081
3082 if (IS_ERR(conf))
3083 return PTR_ERR(conf);
3084
3085 if (mddev->queue) {
3086 blk_queue_max_write_same_sectors(mddev->queue, 0);
3087 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3088 }
3089
3090 rdev_for_each(rdev, mddev) {
3091 if (!mddev->gendisk)
3092 continue;
3093 disk_stack_limits(mddev->gendisk, rdev->bdev,
3094 rdev->data_offset << 9);
3095 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3096 discard_supported = true;
3097 }
3098
3099 mddev->degraded = 0;
3100 for (i=0; i < conf->raid_disks; i++)
3101 if (conf->mirrors[i].rdev == NULL ||
3102 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3103 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3104 mddev->degraded++;
3105
3106 if (conf->raid_disks - mddev->degraded == 1)
3107 mddev->recovery_cp = MaxSector;
3108
3109 if (mddev->recovery_cp != MaxSector)
3110 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3111 mdname(mddev));
3112 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3113 mdname(mddev), mddev->raid_disks - mddev->degraded,
3114 mddev->raid_disks);
3115
3116 /*
3117 * Ok, everything is just fine now
3118 */
3119 mddev->thread = conf->thread;
3120 conf->thread = NULL;
3121 mddev->private = conf;
3122 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3123
3124 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3125
3126 if (mddev->queue) {
3127 if (discard_supported)
3128 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3129 mddev->queue);
3130 else
3131 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3132 mddev->queue);
3133 }
3134
3135 ret = md_integrity_register(mddev);
3136 if (ret) {
3137 md_unregister_thread(&mddev->thread);
3138 raid1_free(mddev, conf);
3139 }
3140 return ret;
3141}
3142
3143static void raid1_free(struct mddev *mddev, void *priv)
3144{
3145 struct r1conf *conf = priv;
3146
3147 mempool_destroy(conf->r1bio_pool);
3148 kfree(conf->mirrors);
3149 safe_put_page(conf->tmppage);
3150 kfree(conf->poolinfo);
3151 kfree(conf->nr_pending);
3152 kfree(conf->nr_waiting);
3153 kfree(conf->nr_queued);
3154 kfree(conf->barrier);
3155 if (conf->bio_split)
3156 bioset_free(conf->bio_split);
3157 kfree(conf);
3158}
3159
3160static int raid1_resize(struct mddev *mddev, sector_t sectors)
3161{
3162 /* no resync is happening, and there is enough space
3163 * on all devices, so we can resize.
3164 * We need to make sure resync covers any new space.
3165 * If the array is shrinking we should possibly wait until
3166 * any io in the removed space completes, but it hardly seems
3167 * worth it.
3168 */
3169 sector_t newsize = raid1_size(mddev, sectors, 0);
3170 if (mddev->external_size &&
3171 mddev->array_sectors > newsize)
3172 return -EINVAL;
3173 if (mddev->bitmap) {
3174 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3175 if (ret)
3176 return ret;
3177 }
3178 md_set_array_sectors(mddev, newsize);
3179 if (sectors > mddev->dev_sectors &&
3180 mddev->recovery_cp > mddev->dev_sectors) {
3181 mddev->recovery_cp = mddev->dev_sectors;
3182 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3183 }
3184 mddev->dev_sectors = sectors;
3185 mddev->resync_max_sectors = sectors;
3186 return 0;
3187}
3188
3189static int raid1_reshape(struct mddev *mddev)
3190{
3191 /* We need to:
3192 * 1/ resize the r1bio_pool
3193 * 2/ resize conf->mirrors
3194 *
3195 * We allocate a new r1bio_pool if we can.
3196 * Then raise a device barrier and wait until all IO stops.
3197 * Then resize conf->mirrors and swap in the new r1bio pool.
3198 *
3199 * At the same time, we "pack" the devices so that all the missing
3200 * devices have the higher raid_disk numbers.
3201 */
3202 mempool_t *newpool, *oldpool;
3203 struct pool_info *newpoolinfo;
3204 struct raid1_info *newmirrors;
3205 struct r1conf *conf = mddev->private;
3206 int cnt, raid_disks;
3207 unsigned long flags;
3208 int d, d2;
3209
3210 /* Cannot change chunk_size, layout, or level */
3211 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3212 mddev->layout != mddev->new_layout ||
3213 mddev->level != mddev->new_level) {
3214 mddev->new_chunk_sectors = mddev->chunk_sectors;
3215 mddev->new_layout = mddev->layout;
3216 mddev->new_level = mddev->level;
3217 return -EINVAL;
3218 }
3219
3220 if (!mddev_is_clustered(mddev))
3221 md_allow_write(mddev);
3222
3223 raid_disks = mddev->raid_disks + mddev->delta_disks;
3224
3225 if (raid_disks < conf->raid_disks) {
3226 cnt=0;
3227 for (d= 0; d < conf->raid_disks; d++)
3228 if (conf->mirrors[d].rdev)
3229 cnt++;
3230 if (cnt > raid_disks)
3231 return -EBUSY;
3232 }
3233
3234 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3235 if (!newpoolinfo)
3236 return -ENOMEM;
3237 newpoolinfo->mddev = mddev;
3238 newpoolinfo->raid_disks = raid_disks * 2;
3239
3240 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3241 r1bio_pool_free, newpoolinfo);
3242 if (!newpool) {
3243 kfree(newpoolinfo);
3244 return -ENOMEM;
3245 }
3246 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3247 GFP_KERNEL);
3248 if (!newmirrors) {
3249 kfree(newpoolinfo);
3250 mempool_destroy(newpool);
3251 return -ENOMEM;
3252 }
3253
3254 freeze_array(conf, 0);
3255
3256 /* ok, everything is stopped */
3257 oldpool = conf->r1bio_pool;
3258 conf->r1bio_pool = newpool;
3259
3260 for (d = d2 = 0; d < conf->raid_disks; d++) {
3261 struct md_rdev *rdev = conf->mirrors[d].rdev;
3262 if (rdev && rdev->raid_disk != d2) {
3263 sysfs_unlink_rdev(mddev, rdev);
3264 rdev->raid_disk = d2;
3265 sysfs_unlink_rdev(mddev, rdev);
3266 if (sysfs_link_rdev(mddev, rdev))
3267 pr_warn("md/raid1:%s: cannot register rd%d\n",
3268 mdname(mddev), rdev->raid_disk);
3269 }
3270 if (rdev)
3271 newmirrors[d2++].rdev = rdev;
3272 }
3273 kfree(conf->mirrors);
3274 conf->mirrors = newmirrors;
3275 kfree(conf->poolinfo);
3276 conf->poolinfo = newpoolinfo;
3277
3278 spin_lock_irqsave(&conf->device_lock, flags);
3279 mddev->degraded += (raid_disks - conf->raid_disks);
3280 spin_unlock_irqrestore(&conf->device_lock, flags);
3281 conf->raid_disks = mddev->raid_disks = raid_disks;
3282 mddev->delta_disks = 0;
3283
3284 unfreeze_array(conf);
3285
3286 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3287 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3288 md_wakeup_thread(mddev->thread);
3289
3290 mempool_destroy(oldpool);
3291 return 0;
3292}
3293
3294static void raid1_quiesce(struct mddev *mddev, int quiesce)
3295{
3296 struct r1conf *conf = mddev->private;
3297
3298 if (quiesce)
3299 freeze_array(conf, 0);
3300 else
3301 unfreeze_array(conf);
3302}
3303
3304static void *raid1_takeover(struct mddev *mddev)
3305{
3306 /* raid1 can take over:
3307 * raid5 with 2 devices, any layout or chunk size
3308 */
3309 if (mddev->level == 5 && mddev->raid_disks == 2) {
3310 struct r1conf *conf;
3311 mddev->new_level = 1;
3312 mddev->new_layout = 0;
3313 mddev->new_chunk_sectors = 0;
3314 conf = setup_conf(mddev);
3315 if (!IS_ERR(conf)) {
3316 /* Array must appear to be quiesced */
3317 conf->array_frozen = 1;
3318 mddev_clear_unsupported_flags(mddev,
3319 UNSUPPORTED_MDDEV_FLAGS);
3320 }
3321 return conf;
3322 }
3323 return ERR_PTR(-EINVAL);
3324}
3325
3326static struct md_personality raid1_personality =
3327{
3328 .name = "raid1",
3329 .level = 1,
3330 .owner = THIS_MODULE,
3331 .make_request = raid1_make_request,
3332 .run = raid1_run,
3333 .free = raid1_free,
3334 .status = raid1_status,
3335 .error_handler = raid1_error,
3336 .hot_add_disk = raid1_add_disk,
3337 .hot_remove_disk= raid1_remove_disk,
3338 .spare_active = raid1_spare_active,
3339 .sync_request = raid1_sync_request,
3340 .resize = raid1_resize,
3341 .size = raid1_size,
3342 .check_reshape = raid1_reshape,
3343 .quiesce = raid1_quiesce,
3344 .takeover = raid1_takeover,
3345 .congested = raid1_congested,
3346};
3347
3348static int __init raid_init(void)
3349{
3350 return register_md_personality(&raid1_personality);
3351}
3352
3353static void raid_exit(void)
3354{
3355 unregister_md_personality(&raid1_personality);
3356}
3357
3358module_init(raid_init);
3359module_exit(raid_exit);
3360MODULE_LICENSE("GPL");
3361MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3362MODULE_ALIAS("md-personality-3"); /* RAID1 */
3363MODULE_ALIAS("md-raid1");
3364MODULE_ALIAS("md-level-1");
3365
3366module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);