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