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