]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - drivers/md/raid1.c
md: remove 'idx' from 'struct resync_pages'
[karo-tx-linux.git] / drivers / md / raid1.c
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 Ã˜stergaard" <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 <linux/sched/signal.h>
41
42 #include <trace/events/block.h>
43
44 #include "md.h"
45 #include "raid1.h"
46 #include "bitmap.h"
47
48 #define UNSUPPORTED_MDDEV_FLAGS         \
49         ((1L << MD_HAS_JOURNAL) |       \
50          (1L << MD_JOURNAL_CLEAN) |     \
51          (1L << MD_HAS_PPL))
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  */
76 static int max_queued_requests = 1024;
77
78 static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
79 static 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 /*
85  * 'strct resync_pages' stores actual pages used for doing the resync
86  *  IO, and it is per-bio, so make .bi_private points to it.
87  */
88 static inline struct resync_pages *get_resync_pages(struct bio *bio)
89 {
90         return bio->bi_private;
91 }
92
93 /*
94  * for resync bio, r1bio pointer can be retrieved from the per-bio
95  * 'struct resync_pages'.
96  */
97 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
98 {
99         return get_resync_pages(bio)->raid_bio;
100 }
101
102 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
103 {
104         struct pool_info *pi = data;
105         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
106
107         /* allocate a r1bio with room for raid_disks entries in the bios array */
108         return kzalloc(size, gfp_flags);
109 }
110
111 static void r1bio_pool_free(void *r1_bio, void *data)
112 {
113         kfree(r1_bio);
114 }
115
116 #define RESYNC_DEPTH 32
117 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
118 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
119 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
120 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
121 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
122
123 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
124 {
125         struct pool_info *pi = data;
126         struct r1bio *r1_bio;
127         struct bio *bio;
128         int need_pages;
129         int j;
130         struct resync_pages *rps;
131
132         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
133         if (!r1_bio)
134                 return NULL;
135
136         rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
137                       gfp_flags);
138         if (!rps)
139                 goto out_free_r1bio;
140
141         /*
142          * Allocate bios : 1 for reading, n-1 for writing
143          */
144         for (j = pi->raid_disks ; j-- ; ) {
145                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
146                 if (!bio)
147                         goto out_free_bio;
148                 r1_bio->bios[j] = bio;
149         }
150         /*
151          * Allocate RESYNC_PAGES data pages and attach them to
152          * the first bio.
153          * If this is a user-requested check/repair, allocate
154          * RESYNC_PAGES for each bio.
155          */
156         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
157                 need_pages = pi->raid_disks;
158         else
159                 need_pages = 1;
160         for (j = 0; j < pi->raid_disks; j++) {
161                 struct resync_pages *rp = &rps[j];
162
163                 bio = r1_bio->bios[j];
164
165                 if (j < need_pages) {
166                         if (resync_alloc_pages(rp, gfp_flags))
167                                 goto out_free_pages;
168                 } else {
169                         memcpy(rp, &rps[0], sizeof(*rp));
170                         resync_get_all_pages(rp);
171                 }
172
173                 rp->raid_bio = r1_bio;
174                 bio->bi_private = rp;
175         }
176
177         r1_bio->master_bio = NULL;
178
179         return r1_bio;
180
181 out_free_pages:
182         while (--j >= 0)
183                 resync_free_pages(&rps[j]);
184
185 out_free_bio:
186         while (++j < pi->raid_disks)
187                 bio_put(r1_bio->bios[j]);
188         kfree(rps);
189
190 out_free_r1bio:
191         r1bio_pool_free(r1_bio, data);
192         return NULL;
193 }
194
195 static void r1buf_pool_free(void *__r1_bio, void *data)
196 {
197         struct pool_info *pi = data;
198         int i;
199         struct r1bio *r1bio = __r1_bio;
200         struct resync_pages *rp = NULL;
201
202         for (i = pi->raid_disks; i--; ) {
203                 rp = get_resync_pages(r1bio->bios[i]);
204                 resync_free_pages(rp);
205                 bio_put(r1bio->bios[i]);
206         }
207
208         /* resync pages array stored in the 1st bio's .bi_private */
209         kfree(rp);
210
211         r1bio_pool_free(r1bio, data);
212 }
213
214 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
215 {
216         int i;
217
218         for (i = 0; i < conf->raid_disks * 2; i++) {
219                 struct bio **bio = r1_bio->bios + i;
220                 if (!BIO_SPECIAL(*bio))
221                         bio_put(*bio);
222                 *bio = NULL;
223         }
224 }
225
226 static void free_r1bio(struct r1bio *r1_bio)
227 {
228         struct r1conf *conf = r1_bio->mddev->private;
229
230         put_all_bios(conf, r1_bio);
231         mempool_free(r1_bio, conf->r1bio_pool);
232 }
233
234 static void put_buf(struct r1bio *r1_bio)
235 {
236         struct r1conf *conf = r1_bio->mddev->private;
237         sector_t sect = r1_bio->sector;
238         int i;
239
240         for (i = 0; i < conf->raid_disks * 2; i++) {
241                 struct bio *bio = r1_bio->bios[i];
242                 if (bio->bi_end_io)
243                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
244         }
245
246         mempool_free(r1_bio, conf->r1buf_pool);
247
248         lower_barrier(conf, sect);
249 }
250
251 static void reschedule_retry(struct r1bio *r1_bio)
252 {
253         unsigned long flags;
254         struct mddev *mddev = r1_bio->mddev;
255         struct r1conf *conf = mddev->private;
256         int idx;
257
258         idx = sector_to_idx(r1_bio->sector);
259         spin_lock_irqsave(&conf->device_lock, flags);
260         list_add(&r1_bio->retry_list, &conf->retry_list);
261         atomic_inc(&conf->nr_queued[idx]);
262         spin_unlock_irqrestore(&conf->device_lock, flags);
263
264         wake_up(&conf->wait_barrier);
265         md_wakeup_thread(mddev->thread);
266 }
267
268 /*
269  * raid_end_bio_io() is called when we have finished servicing a mirrored
270  * operation and are ready to return a success/failure code to the buffer
271  * cache layer.
272  */
273 static void call_bio_endio(struct r1bio *r1_bio)
274 {
275         struct bio *bio = r1_bio->master_bio;
276         struct r1conf *conf = r1_bio->mddev->private;
277
278         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
279                 bio->bi_status = BLK_STS_IOERR;
280
281         bio_endio(bio);
282         /*
283          * Wake up any possible resync thread that waits for the device
284          * to go idle.
285          */
286         allow_barrier(conf, r1_bio->sector);
287 }
288
289 static void raid_end_bio_io(struct r1bio *r1_bio)
290 {
291         struct bio *bio = r1_bio->master_bio;
292
293         /* if nobody has done the final endio yet, do it now */
294         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
295                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
296                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
297                          (unsigned long long) bio->bi_iter.bi_sector,
298                          (unsigned long long) bio_end_sector(bio) - 1);
299
300                 call_bio_endio(r1_bio);
301         }
302         free_r1bio(r1_bio);
303 }
304
305 /*
306  * Update disk head position estimator based on IRQ completion info.
307  */
308 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
309 {
310         struct r1conf *conf = r1_bio->mddev->private;
311
312         conf->mirrors[disk].head_position =
313                 r1_bio->sector + (r1_bio->sectors);
314 }
315
316 /*
317  * Find the disk number which triggered given bio
318  */
319 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
320 {
321         int mirror;
322         struct r1conf *conf = r1_bio->mddev->private;
323         int raid_disks = conf->raid_disks;
324
325         for (mirror = 0; mirror < raid_disks * 2; mirror++)
326                 if (r1_bio->bios[mirror] == bio)
327                         break;
328
329         BUG_ON(mirror == raid_disks * 2);
330         update_head_pos(mirror, r1_bio);
331
332         return mirror;
333 }
334
335 static void raid1_end_read_request(struct bio *bio)
336 {
337         int uptodate = !bio->bi_status;
338         struct r1bio *r1_bio = bio->bi_private;
339         struct r1conf *conf = r1_bio->mddev->private;
340         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
341
342         /*
343          * this branch is our 'one mirror IO has finished' event handler:
344          */
345         update_head_pos(r1_bio->read_disk, r1_bio);
346
347         if (uptodate)
348                 set_bit(R1BIO_Uptodate, &r1_bio->state);
349         else if (test_bit(FailFast, &rdev->flags) &&
350                  test_bit(R1BIO_FailFast, &r1_bio->state))
351                 /* This was a fail-fast read so we definitely
352                  * want to retry */
353                 ;
354         else {
355                 /* If all other devices have failed, we want to return
356                  * the error upwards rather than fail the last device.
357                  * Here we redefine "uptodate" to mean "Don't want to retry"
358                  */
359                 unsigned long flags;
360                 spin_lock_irqsave(&conf->device_lock, flags);
361                 if (r1_bio->mddev->degraded == conf->raid_disks ||
362                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
363                      test_bit(In_sync, &rdev->flags)))
364                         uptodate = 1;
365                 spin_unlock_irqrestore(&conf->device_lock, flags);
366         }
367
368         if (uptodate) {
369                 raid_end_bio_io(r1_bio);
370                 rdev_dec_pending(rdev, conf->mddev);
371         } else {
372                 /*
373                  * oops, read error:
374                  */
375                 char b[BDEVNAME_SIZE];
376                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
377                                    mdname(conf->mddev),
378                                    bdevname(rdev->bdev, b),
379                                    (unsigned long long)r1_bio->sector);
380                 set_bit(R1BIO_ReadError, &r1_bio->state);
381                 reschedule_retry(r1_bio);
382                 /* don't drop the reference on read_disk yet */
383         }
384 }
385
386 static void close_write(struct r1bio *r1_bio)
387 {
388         /* it really is the end of this request */
389         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
390                 bio_free_pages(r1_bio->behind_master_bio);
391                 bio_put(r1_bio->behind_master_bio);
392                 r1_bio->behind_master_bio = NULL;
393         }
394         /* clear the bitmap if all writes complete successfully */
395         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
396                         r1_bio->sectors,
397                         !test_bit(R1BIO_Degraded, &r1_bio->state),
398                         test_bit(R1BIO_BehindIO, &r1_bio->state));
399         md_write_end(r1_bio->mddev);
400 }
401
402 static void r1_bio_write_done(struct r1bio *r1_bio)
403 {
404         if (!atomic_dec_and_test(&r1_bio->remaining))
405                 return;
406
407         if (test_bit(R1BIO_WriteError, &r1_bio->state))
408                 reschedule_retry(r1_bio);
409         else {
410                 close_write(r1_bio);
411                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
412                         reschedule_retry(r1_bio);
413                 else
414                         raid_end_bio_io(r1_bio);
415         }
416 }
417
418 static void raid1_end_write_request(struct bio *bio)
419 {
420         struct r1bio *r1_bio = bio->bi_private;
421         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
422         struct r1conf *conf = r1_bio->mddev->private;
423         struct bio *to_put = NULL;
424         int mirror = find_bio_disk(r1_bio, bio);
425         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
426         bool discard_error;
427
428         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
429
430         /*
431          * 'one mirror IO has finished' event handler:
432          */
433         if (bio->bi_status && !discard_error) {
434                 set_bit(WriteErrorSeen, &rdev->flags);
435                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
436                         set_bit(MD_RECOVERY_NEEDED, &
437                                 conf->mddev->recovery);
438
439                 if (test_bit(FailFast, &rdev->flags) &&
440                     (bio->bi_opf & MD_FAILFAST) &&
441                     /* We never try FailFast to WriteMostly devices */
442                     !test_bit(WriteMostly, &rdev->flags)) {
443                         md_error(r1_bio->mddev, rdev);
444                         if (!test_bit(Faulty, &rdev->flags))
445                                 /* This is the only remaining device,
446                                  * We need to retry the write without
447                                  * FailFast
448                                  */
449                                 set_bit(R1BIO_WriteError, &r1_bio->state);
450                         else {
451                                 /* Finished with this branch */
452                                 r1_bio->bios[mirror] = NULL;
453                                 to_put = bio;
454                         }
455                 } else
456                         set_bit(R1BIO_WriteError, &r1_bio->state);
457         } else {
458                 /*
459                  * Set R1BIO_Uptodate in our master bio, so that we
460                  * will return a good error code for to the higher
461                  * levels even if IO on some other mirrored buffer
462                  * fails.
463                  *
464                  * The 'master' represents the composite IO operation
465                  * to user-side. So if something waits for IO, then it
466                  * will wait for the 'master' bio.
467                  */
468                 sector_t first_bad;
469                 int bad_sectors;
470
471                 r1_bio->bios[mirror] = NULL;
472                 to_put = bio;
473                 /*
474                  * Do not set R1BIO_Uptodate if the current device is
475                  * rebuilding or Faulty. This is because we cannot use
476                  * such device for properly reading the data back (we could
477                  * potentially use it, if the current write would have felt
478                  * before rdev->recovery_offset, but for simplicity we don't
479                  * check this here.
480                  */
481                 if (test_bit(In_sync, &rdev->flags) &&
482                     !test_bit(Faulty, &rdev->flags))
483                         set_bit(R1BIO_Uptodate, &r1_bio->state);
484
485                 /* Maybe we can clear some bad blocks. */
486                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
487                                 &first_bad, &bad_sectors) && !discard_error) {
488                         r1_bio->bios[mirror] = IO_MADE_GOOD;
489                         set_bit(R1BIO_MadeGood, &r1_bio->state);
490                 }
491         }
492
493         if (behind) {
494                 /* we release behind master bio when all write are done */
495                 if (r1_bio->behind_master_bio == bio)
496                         to_put = NULL;
497
498                 if (test_bit(WriteMostly, &rdev->flags))
499                         atomic_dec(&r1_bio->behind_remaining);
500
501                 /*
502                  * In behind mode, we ACK the master bio once the I/O
503                  * has safely reached all non-writemostly
504                  * disks. Setting the Returned bit ensures that this
505                  * gets done only once -- we don't ever want to return
506                  * -EIO here, instead we'll wait
507                  */
508                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
509                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
510                         /* Maybe we can return now */
511                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
512                                 struct bio *mbio = r1_bio->master_bio;
513                                 pr_debug("raid1: behind end write sectors"
514                                          " %llu-%llu\n",
515                                          (unsigned long long) mbio->bi_iter.bi_sector,
516                                          (unsigned long long) bio_end_sector(mbio) - 1);
517                                 call_bio_endio(r1_bio);
518                         }
519                 }
520         }
521         if (r1_bio->bios[mirror] == NULL)
522                 rdev_dec_pending(rdev, conf->mddev);
523
524         /*
525          * Let's see if all mirrored write operations have finished
526          * already.
527          */
528         r1_bio_write_done(r1_bio);
529
530         if (to_put)
531                 bio_put(to_put);
532 }
533
534 static sector_t align_to_barrier_unit_end(sector_t start_sector,
535                                           sector_t sectors)
536 {
537         sector_t len;
538
539         WARN_ON(sectors == 0);
540         /*
541          * len is the number of sectors from start_sector to end of the
542          * barrier unit which start_sector belongs to.
543          */
544         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
545               start_sector;
546
547         if (len > sectors)
548                 len = sectors;
549
550         return len;
551 }
552
553 /*
554  * This routine returns the disk from which the requested read should
555  * be done. There is a per-array 'next expected sequential IO' sector
556  * number - if this matches on the next IO then we use the last disk.
557  * There is also a per-disk 'last know head position' sector that is
558  * maintained from IRQ contexts, both the normal and the resync IO
559  * completion handlers update this position correctly. If there is no
560  * perfect sequential match then we pick the disk whose head is closest.
561  *
562  * If there are 2 mirrors in the same 2 devices, performance degrades
563  * because position is mirror, not device based.
564  *
565  * The rdev for the device selected will have nr_pending incremented.
566  */
567 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
568 {
569         const sector_t this_sector = r1_bio->sector;
570         int sectors;
571         int best_good_sectors;
572         int best_disk, best_dist_disk, best_pending_disk;
573         int has_nonrot_disk;
574         int disk;
575         sector_t best_dist;
576         unsigned int min_pending;
577         struct md_rdev *rdev;
578         int choose_first;
579         int choose_next_idle;
580
581         rcu_read_lock();
582         /*
583          * Check if we can balance. We can balance on the whole
584          * device if no resync is going on, or below the resync window.
585          * We take the first readable disk when above the resync window.
586          */
587  retry:
588         sectors = r1_bio->sectors;
589         best_disk = -1;
590         best_dist_disk = -1;
591         best_dist = MaxSector;
592         best_pending_disk = -1;
593         min_pending = UINT_MAX;
594         best_good_sectors = 0;
595         has_nonrot_disk = 0;
596         choose_next_idle = 0;
597         clear_bit(R1BIO_FailFast, &r1_bio->state);
598
599         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
600             (mddev_is_clustered(conf->mddev) &&
601             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
602                     this_sector + sectors)))
603                 choose_first = 1;
604         else
605                 choose_first = 0;
606
607         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
608                 sector_t dist;
609                 sector_t first_bad;
610                 int bad_sectors;
611                 unsigned int pending;
612                 bool nonrot;
613
614                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
615                 if (r1_bio->bios[disk] == IO_BLOCKED
616                     || rdev == NULL
617                     || test_bit(Faulty, &rdev->flags))
618                         continue;
619                 if (!test_bit(In_sync, &rdev->flags) &&
620                     rdev->recovery_offset < this_sector + sectors)
621                         continue;
622                 if (test_bit(WriteMostly, &rdev->flags)) {
623                         /* Don't balance among write-mostly, just
624                          * use the first as a last resort */
625                         if (best_dist_disk < 0) {
626                                 if (is_badblock(rdev, this_sector, sectors,
627                                                 &first_bad, &bad_sectors)) {
628                                         if (first_bad <= this_sector)
629                                                 /* Cannot use this */
630                                                 continue;
631                                         best_good_sectors = first_bad - this_sector;
632                                 } else
633                                         best_good_sectors = sectors;
634                                 best_dist_disk = disk;
635                                 best_pending_disk = disk;
636                         }
637                         continue;
638                 }
639                 /* This is a reasonable device to use.  It might
640                  * even be best.
641                  */
642                 if (is_badblock(rdev, this_sector, sectors,
643                                 &first_bad, &bad_sectors)) {
644                         if (best_dist < MaxSector)
645                                 /* already have a better device */
646                                 continue;
647                         if (first_bad <= this_sector) {
648                                 /* cannot read here. If this is the 'primary'
649                                  * device, then we must not read beyond
650                                  * bad_sectors from another device..
651                                  */
652                                 bad_sectors -= (this_sector - first_bad);
653                                 if (choose_first && sectors > bad_sectors)
654                                         sectors = bad_sectors;
655                                 if (best_good_sectors > sectors)
656                                         best_good_sectors = sectors;
657
658                         } else {
659                                 sector_t good_sectors = first_bad - this_sector;
660                                 if (good_sectors > best_good_sectors) {
661                                         best_good_sectors = good_sectors;
662                                         best_disk = disk;
663                                 }
664                                 if (choose_first)
665                                         break;
666                         }
667                         continue;
668                 } else {
669                         if ((sectors > best_good_sectors) && (best_disk >= 0))
670                                 best_disk = -1;
671                         best_good_sectors = sectors;
672                 }
673
674                 if (best_disk >= 0)
675                         /* At least two disks to choose from so failfast is OK */
676                         set_bit(R1BIO_FailFast, &r1_bio->state);
677
678                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
679                 has_nonrot_disk |= nonrot;
680                 pending = atomic_read(&rdev->nr_pending);
681                 dist = abs(this_sector - conf->mirrors[disk].head_position);
682                 if (choose_first) {
683                         best_disk = disk;
684                         break;
685                 }
686                 /* Don't change to another disk for sequential reads */
687                 if (conf->mirrors[disk].next_seq_sect == this_sector
688                     || dist == 0) {
689                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
690                         struct raid1_info *mirror = &conf->mirrors[disk];
691
692                         best_disk = disk;
693                         /*
694                          * If buffered sequential IO size exceeds optimal
695                          * iosize, check if there is idle disk. If yes, choose
696                          * the idle disk. read_balance could already choose an
697                          * idle disk before noticing it's a sequential IO in
698                          * this disk. This doesn't matter because this disk
699                          * will idle, next time it will be utilized after the
700                          * first disk has IO size exceeds optimal iosize. In
701                          * this way, iosize of the first disk will be optimal
702                          * iosize at least. iosize of the second disk might be
703                          * small, but not a big deal since when the second disk
704                          * starts IO, the first disk is likely still busy.
705                          */
706                         if (nonrot && opt_iosize > 0 &&
707                             mirror->seq_start != MaxSector &&
708                             mirror->next_seq_sect > opt_iosize &&
709                             mirror->next_seq_sect - opt_iosize >=
710                             mirror->seq_start) {
711                                 choose_next_idle = 1;
712                                 continue;
713                         }
714                         break;
715                 }
716
717                 if (choose_next_idle)
718                         continue;
719
720                 if (min_pending > pending) {
721                         min_pending = pending;
722                         best_pending_disk = disk;
723                 }
724
725                 if (dist < best_dist) {
726                         best_dist = dist;
727                         best_dist_disk = disk;
728                 }
729         }
730
731         /*
732          * If all disks are rotational, choose the closest disk. If any disk is
733          * non-rotational, choose the disk with less pending request even the
734          * disk is rotational, which might/might not be optimal for raids with
735          * mixed ratation/non-rotational disks depending on workload.
736          */
737         if (best_disk == -1) {
738                 if (has_nonrot_disk || min_pending == 0)
739                         best_disk = best_pending_disk;
740                 else
741                         best_disk = best_dist_disk;
742         }
743
744         if (best_disk >= 0) {
745                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
746                 if (!rdev)
747                         goto retry;
748                 atomic_inc(&rdev->nr_pending);
749                 sectors = best_good_sectors;
750
751                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
752                         conf->mirrors[best_disk].seq_start = this_sector;
753
754                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
755         }
756         rcu_read_unlock();
757         *max_sectors = sectors;
758
759         return best_disk;
760 }
761
762 static int raid1_congested(struct mddev *mddev, int bits)
763 {
764         struct r1conf *conf = mddev->private;
765         int i, ret = 0;
766
767         if ((bits & (1 << WB_async_congested)) &&
768             conf->pending_count >= max_queued_requests)
769                 return 1;
770
771         rcu_read_lock();
772         for (i = 0; i < conf->raid_disks * 2; i++) {
773                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
774                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
775                         struct request_queue *q = bdev_get_queue(rdev->bdev);
776
777                         BUG_ON(!q);
778
779                         /* Note the '|| 1' - when read_balance prefers
780                          * non-congested targets, it can be removed
781                          */
782                         if ((bits & (1 << WB_async_congested)) || 1)
783                                 ret |= bdi_congested(q->backing_dev_info, bits);
784                         else
785                                 ret &= bdi_congested(q->backing_dev_info, bits);
786                 }
787         }
788         rcu_read_unlock();
789         return ret;
790 }
791
792 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
793 {
794         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
795         bitmap_unplug(conf->mddev->bitmap);
796         wake_up(&conf->wait_barrier);
797
798         while (bio) { /* submit pending writes */
799                 struct bio *next = bio->bi_next;
800                 struct md_rdev *rdev = (void*)bio->bi_bdev;
801                 bio->bi_next = NULL;
802                 bio->bi_bdev = rdev->bdev;
803                 if (test_bit(Faulty, &rdev->flags)) {
804                         bio->bi_status = BLK_STS_IOERR;
805                         bio_endio(bio);
806                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
807                                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
808                         /* Just ignore it */
809                         bio_endio(bio);
810                 else
811                         generic_make_request(bio);
812                 bio = next;
813         }
814 }
815
816 static void flush_pending_writes(struct r1conf *conf)
817 {
818         /* Any writes that have been queued but are awaiting
819          * bitmap updates get flushed here.
820          */
821         spin_lock_irq(&conf->device_lock);
822
823         if (conf->pending_bio_list.head) {
824                 struct bio *bio;
825                 bio = bio_list_get(&conf->pending_bio_list);
826                 conf->pending_count = 0;
827                 spin_unlock_irq(&conf->device_lock);
828                 flush_bio_list(conf, bio);
829         } else
830                 spin_unlock_irq(&conf->device_lock);
831 }
832
833 /* Barriers....
834  * Sometimes we need to suspend IO while we do something else,
835  * either some resync/recovery, or reconfigure the array.
836  * To do this we raise a 'barrier'.
837  * The 'barrier' is a counter that can be raised multiple times
838  * to count how many activities are happening which preclude
839  * normal IO.
840  * We can only raise the barrier if there is no pending IO.
841  * i.e. if nr_pending == 0.
842  * We choose only to raise the barrier if no-one is waiting for the
843  * barrier to go down.  This means that as soon as an IO request
844  * is ready, no other operations which require a barrier will start
845  * until the IO request has had a chance.
846  *
847  * So: regular IO calls 'wait_barrier'.  When that returns there
848  *    is no backgroup IO happening,  It must arrange to call
849  *    allow_barrier when it has finished its IO.
850  * backgroup IO calls must call raise_barrier.  Once that returns
851  *    there is no normal IO happeing.  It must arrange to call
852  *    lower_barrier when the particular background IO completes.
853  */
854 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
855 {
856         int idx = sector_to_idx(sector_nr);
857
858         spin_lock_irq(&conf->resync_lock);
859
860         /* Wait until no block IO is waiting */
861         wait_event_lock_irq(conf->wait_barrier,
862                             !atomic_read(&conf->nr_waiting[idx]),
863                             conf->resync_lock);
864
865         /* block any new IO from starting */
866         atomic_inc(&conf->barrier[idx]);
867         /*
868          * In raise_barrier() we firstly increase conf->barrier[idx] then
869          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
870          * increase conf->nr_pending[idx] then check conf->barrier[idx].
871          * A memory barrier here to make sure conf->nr_pending[idx] won't
872          * be fetched before conf->barrier[idx] is increased. Otherwise
873          * there will be a race between raise_barrier() and _wait_barrier().
874          */
875         smp_mb__after_atomic();
876
877         /* For these conditions we must wait:
878          * A: while the array is in frozen state
879          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
880          *    existing in corresponding I/O barrier bucket.
881          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
882          *    max resync count which allowed on current I/O barrier bucket.
883          */
884         wait_event_lock_irq(conf->wait_barrier,
885                             !conf->array_frozen &&
886                              !atomic_read(&conf->nr_pending[idx]) &&
887                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
888                             conf->resync_lock);
889
890         atomic_inc(&conf->nr_sync_pending);
891         spin_unlock_irq(&conf->resync_lock);
892 }
893
894 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
895 {
896         int idx = sector_to_idx(sector_nr);
897
898         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
899
900         atomic_dec(&conf->barrier[idx]);
901         atomic_dec(&conf->nr_sync_pending);
902         wake_up(&conf->wait_barrier);
903 }
904
905 static void _wait_barrier(struct r1conf *conf, int idx)
906 {
907         /*
908          * We need to increase conf->nr_pending[idx] very early here,
909          * then raise_barrier() can be blocked when it waits for
910          * conf->nr_pending[idx] to be 0. Then we can avoid holding
911          * conf->resync_lock when there is no barrier raised in same
912          * barrier unit bucket. Also if the array is frozen, I/O
913          * should be blocked until array is unfrozen.
914          */
915         atomic_inc(&conf->nr_pending[idx]);
916         /*
917          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
918          * check conf->barrier[idx]. In raise_barrier() we firstly increase
919          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
920          * barrier is necessary here to make sure conf->barrier[idx] won't be
921          * fetched before conf->nr_pending[idx] is increased. Otherwise there
922          * will be a race between _wait_barrier() and raise_barrier().
923          */
924         smp_mb__after_atomic();
925
926         /*
927          * Don't worry about checking two atomic_t variables at same time
928          * here. If during we check conf->barrier[idx], the array is
929          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
930          * 0, it is safe to return and make the I/O continue. Because the
931          * array is frozen, all I/O returned here will eventually complete
932          * or be queued, no race will happen. See code comment in
933          * frozen_array().
934          */
935         if (!READ_ONCE(conf->array_frozen) &&
936             !atomic_read(&conf->barrier[idx]))
937                 return;
938
939         /*
940          * After holding conf->resync_lock, conf->nr_pending[idx]
941          * should be decreased before waiting for barrier to drop.
942          * Otherwise, we may encounter a race condition because
943          * raise_barrer() might be waiting for conf->nr_pending[idx]
944          * to be 0 at same time.
945          */
946         spin_lock_irq(&conf->resync_lock);
947         atomic_inc(&conf->nr_waiting[idx]);
948         atomic_dec(&conf->nr_pending[idx]);
949         /*
950          * In case freeze_array() is waiting for
951          * get_unqueued_pending() == extra
952          */
953         wake_up(&conf->wait_barrier);
954         /* Wait for the barrier in same barrier unit bucket to drop. */
955         wait_event_lock_irq(conf->wait_barrier,
956                             !conf->array_frozen &&
957                              !atomic_read(&conf->barrier[idx]),
958                             conf->resync_lock);
959         atomic_inc(&conf->nr_pending[idx]);
960         atomic_dec(&conf->nr_waiting[idx]);
961         spin_unlock_irq(&conf->resync_lock);
962 }
963
964 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
965 {
966         int idx = sector_to_idx(sector_nr);
967
968         /*
969          * Very similar to _wait_barrier(). The difference is, for read
970          * I/O we don't need wait for sync I/O, but if the whole array
971          * is frozen, the read I/O still has to wait until the array is
972          * unfrozen. Since there is no ordering requirement with
973          * conf->barrier[idx] here, memory barrier is unnecessary as well.
974          */
975         atomic_inc(&conf->nr_pending[idx]);
976
977         if (!READ_ONCE(conf->array_frozen))
978                 return;
979
980         spin_lock_irq(&conf->resync_lock);
981         atomic_inc(&conf->nr_waiting[idx]);
982         atomic_dec(&conf->nr_pending[idx]);
983         /*
984          * In case freeze_array() is waiting for
985          * get_unqueued_pending() == extra
986          */
987         wake_up(&conf->wait_barrier);
988         /* Wait for array to be unfrozen */
989         wait_event_lock_irq(conf->wait_barrier,
990                             !conf->array_frozen,
991                             conf->resync_lock);
992         atomic_inc(&conf->nr_pending[idx]);
993         atomic_dec(&conf->nr_waiting[idx]);
994         spin_unlock_irq(&conf->resync_lock);
995 }
996
997 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
998 {
999         int idx = sector_to_idx(sector_nr);
1000
1001         _wait_barrier(conf, idx);
1002 }
1003
1004 static void wait_all_barriers(struct r1conf *conf)
1005 {
1006         int idx;
1007
1008         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1009                 _wait_barrier(conf, idx);
1010 }
1011
1012 static void _allow_barrier(struct r1conf *conf, int idx)
1013 {
1014         atomic_dec(&conf->nr_pending[idx]);
1015         wake_up(&conf->wait_barrier);
1016 }
1017
1018 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1019 {
1020         int idx = sector_to_idx(sector_nr);
1021
1022         _allow_barrier(conf, idx);
1023 }
1024
1025 static void allow_all_barriers(struct r1conf *conf)
1026 {
1027         int idx;
1028
1029         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1030                 _allow_barrier(conf, idx);
1031 }
1032
1033 /* conf->resync_lock should be held */
1034 static int get_unqueued_pending(struct r1conf *conf)
1035 {
1036         int idx, ret;
1037
1038         ret = atomic_read(&conf->nr_sync_pending);
1039         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1040                 ret += atomic_read(&conf->nr_pending[idx]) -
1041                         atomic_read(&conf->nr_queued[idx]);
1042
1043         return ret;
1044 }
1045
1046 static void freeze_array(struct r1conf *conf, int extra)
1047 {
1048         /* Stop sync I/O and normal I/O and wait for everything to
1049          * go quiet.
1050          * This is called in two situations:
1051          * 1) management command handlers (reshape, remove disk, quiesce).
1052          * 2) one normal I/O request failed.
1053
1054          * After array_frozen is set to 1, new sync IO will be blocked at
1055          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1056          * or wait_read_barrier(). The flying I/Os will either complete or be
1057          * queued. When everything goes quite, there are only queued I/Os left.
1058
1059          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1060          * barrier bucket index which this I/O request hits. When all sync and
1061          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1062          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1063          * in handle_read_error(), we may call freeze_array() before trying to
1064          * fix the read error. In this case, the error read I/O is not queued,
1065          * so get_unqueued_pending() == 1.
1066          *
1067          * Therefore before this function returns, we need to wait until
1068          * get_unqueued_pendings(conf) gets equal to extra. For
1069          * normal I/O context, extra is 1, in rested situations extra is 0.
1070          */
1071         spin_lock_irq(&conf->resync_lock);
1072         conf->array_frozen = 1;
1073         raid1_log(conf->mddev, "wait freeze");
1074         wait_event_lock_irq_cmd(
1075                 conf->wait_barrier,
1076                 get_unqueued_pending(conf) == extra,
1077                 conf->resync_lock,
1078                 flush_pending_writes(conf));
1079         spin_unlock_irq(&conf->resync_lock);
1080 }
1081 static void unfreeze_array(struct r1conf *conf)
1082 {
1083         /* reverse the effect of the freeze */
1084         spin_lock_irq(&conf->resync_lock);
1085         conf->array_frozen = 0;
1086         spin_unlock_irq(&conf->resync_lock);
1087         wake_up(&conf->wait_barrier);
1088 }
1089
1090 static struct bio *alloc_behind_master_bio(struct r1bio *r1_bio,
1091                                            struct bio *bio)
1092 {
1093         int size = bio->bi_iter.bi_size;
1094         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1095         int i = 0;
1096         struct bio *behind_bio = NULL;
1097
1098         behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1099         if (!behind_bio)
1100                 goto fail;
1101
1102         /* discard op, we don't support writezero/writesame yet */
1103         if (!bio_has_data(bio))
1104                 goto skip_copy;
1105
1106         while (i < vcnt && size) {
1107                 struct page *page;
1108                 int len = min_t(int, PAGE_SIZE, size);
1109
1110                 page = alloc_page(GFP_NOIO);
1111                 if (unlikely(!page))
1112                         goto free_pages;
1113
1114                 bio_add_page(behind_bio, page, len, 0);
1115
1116                 size -= len;
1117                 i++;
1118         }
1119
1120         bio_copy_data(behind_bio, bio);
1121 skip_copy:
1122         r1_bio->behind_master_bio = behind_bio;;
1123         set_bit(R1BIO_BehindIO, &r1_bio->state);
1124
1125         return behind_bio;
1126
1127 free_pages:
1128         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1129                  bio->bi_iter.bi_size);
1130         bio_free_pages(behind_bio);
1131 fail:
1132         return behind_bio;
1133 }
1134
1135 struct raid1_plug_cb {
1136         struct blk_plug_cb      cb;
1137         struct bio_list         pending;
1138         int                     pending_cnt;
1139 };
1140
1141 static 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
1166 static 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
1175 static inline struct r1bio *
1176 alloc_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
1188 static 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         read_bio->bi_bdev = 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(bdev_get_queue(read_bio->bi_bdev),
1297                                       read_bio, disk_devt(mddev->gendisk),
1298                                       r1_bio->sector);
1299
1300         generic_make_request(read_bio);
1301 }
1302
1303 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1304                                 int max_write_sectors)
1305 {
1306         struct r1conf *conf = mddev->private;
1307         struct r1bio *r1_bio;
1308         int i, disks;
1309         struct bitmap *bitmap = mddev->bitmap;
1310         unsigned long flags;
1311         struct md_rdev *blocked_rdev;
1312         struct blk_plug_cb *cb;
1313         struct raid1_plug_cb *plug = NULL;
1314         int first_clone;
1315         int max_sectors;
1316
1317         /*
1318          * Register the new request and wait if the reconstruction
1319          * thread has put up a bar for new requests.
1320          * Continue immediately if no resync is active currently.
1321          */
1322
1323
1324         if ((bio_end_sector(bio) > mddev->suspend_lo &&
1325             bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1326             (mddev_is_clustered(mddev) &&
1327              md_cluster_ops->area_resyncing(mddev, WRITE,
1328                      bio->bi_iter.bi_sector, bio_end_sector(bio)))) {
1329
1330                 /*
1331                  * As the suspend_* range is controlled by userspace, we want
1332                  * an interruptible wait.
1333                  */
1334                 DEFINE_WAIT(w);
1335                 for (;;) {
1336                         sigset_t full, old;
1337                         prepare_to_wait(&conf->wait_barrier,
1338                                         &w, TASK_INTERRUPTIBLE);
1339                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1340                             bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1341                             (mddev_is_clustered(mddev) &&
1342                              !md_cluster_ops->area_resyncing(mddev, WRITE,
1343                                      bio->bi_iter.bi_sector,
1344                                      bio_end_sector(bio))))
1345                                 break;
1346                         sigfillset(&full);
1347                         sigprocmask(SIG_BLOCK, &full, &old);
1348                         schedule();
1349                         sigprocmask(SIG_SETMASK, &old, NULL);
1350                 }
1351                 finish_wait(&conf->wait_barrier, &w);
1352         }
1353         wait_barrier(conf, bio->bi_iter.bi_sector);
1354
1355         r1_bio = alloc_r1bio(mddev, bio);
1356         r1_bio->sectors = max_write_sectors;
1357
1358         if (conf->pending_count >= max_queued_requests) {
1359                 md_wakeup_thread(mddev->thread);
1360                 raid1_log(mddev, "wait queued");
1361                 wait_event(conf->wait_barrier,
1362                            conf->pending_count < max_queued_requests);
1363         }
1364         /* first select target devices under rcu_lock and
1365          * inc refcount on their rdev.  Record them by setting
1366          * bios[x] to bio
1367          * If there are known/acknowledged bad blocks on any device on
1368          * which we have seen a write error, we want to avoid writing those
1369          * blocks.
1370          * This potentially requires several writes to write around
1371          * the bad blocks.  Each set of writes gets it's own r1bio
1372          * with a set of bios attached.
1373          */
1374
1375         disks = conf->raid_disks * 2;
1376  retry_write:
1377         blocked_rdev = NULL;
1378         rcu_read_lock();
1379         max_sectors = r1_bio->sectors;
1380         for (i = 0;  i < disks; i++) {
1381                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1382                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1383                         atomic_inc(&rdev->nr_pending);
1384                         blocked_rdev = rdev;
1385                         break;
1386                 }
1387                 r1_bio->bios[i] = NULL;
1388                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1389                         if (i < conf->raid_disks)
1390                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1391                         continue;
1392                 }
1393
1394                 atomic_inc(&rdev->nr_pending);
1395                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1396                         sector_t first_bad;
1397                         int bad_sectors;
1398                         int is_bad;
1399
1400                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1401                                              &first_bad, &bad_sectors);
1402                         if (is_bad < 0) {
1403                                 /* mustn't write here until the bad block is
1404                                  * acknowledged*/
1405                                 set_bit(BlockedBadBlocks, &rdev->flags);
1406                                 blocked_rdev = rdev;
1407                                 break;
1408                         }
1409                         if (is_bad && first_bad <= r1_bio->sector) {
1410                                 /* Cannot write here at all */
1411                                 bad_sectors -= (r1_bio->sector - first_bad);
1412                                 if (bad_sectors < max_sectors)
1413                                         /* mustn't write more than bad_sectors
1414                                          * to other devices yet
1415                                          */
1416                                         max_sectors = bad_sectors;
1417                                 rdev_dec_pending(rdev, mddev);
1418                                 /* We don't set R1BIO_Degraded as that
1419                                  * only applies if the disk is
1420                                  * missing, so it might be re-added,
1421                                  * and we want to know to recover this
1422                                  * chunk.
1423                                  * In this case the device is here,
1424                                  * and the fact that this chunk is not
1425                                  * in-sync is recorded in the bad
1426                                  * block log
1427                                  */
1428                                 continue;
1429                         }
1430                         if (is_bad) {
1431                                 int good_sectors = first_bad - r1_bio->sector;
1432                                 if (good_sectors < max_sectors)
1433                                         max_sectors = good_sectors;
1434                         }
1435                 }
1436                 r1_bio->bios[i] = bio;
1437         }
1438         rcu_read_unlock();
1439
1440         if (unlikely(blocked_rdev)) {
1441                 /* Wait for this device to become unblocked */
1442                 int j;
1443
1444                 for (j = 0; j < i; j++)
1445                         if (r1_bio->bios[j])
1446                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1447                 r1_bio->state = 0;
1448                 allow_barrier(conf, bio->bi_iter.bi_sector);
1449                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1450                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1451                 wait_barrier(conf, bio->bi_iter.bi_sector);
1452                 goto retry_write;
1453         }
1454
1455         if (max_sectors < bio_sectors(bio)) {
1456                 struct bio *split = bio_split(bio, max_sectors,
1457                                               GFP_NOIO, conf->bio_split);
1458                 bio_chain(split, bio);
1459                 generic_make_request(bio);
1460                 bio = split;
1461                 r1_bio->master_bio = bio;
1462                 r1_bio->sectors = max_sectors;
1463         }
1464
1465         atomic_set(&r1_bio->remaining, 1);
1466         atomic_set(&r1_bio->behind_remaining, 0);
1467
1468         first_clone = 1;
1469
1470         for (i = 0; i < disks; i++) {
1471                 struct bio *mbio = NULL;
1472                 if (!r1_bio->bios[i])
1473                         continue;
1474
1475
1476                 if (first_clone) {
1477                         /* do behind I/O ?
1478                          * Not if there are too many, or cannot
1479                          * allocate memory, or a reader on WriteMostly
1480                          * is waiting for behind writes to flush */
1481                         if (bitmap &&
1482                             (atomic_read(&bitmap->behind_writes)
1483                              < mddev->bitmap_info.max_write_behind) &&
1484                             !waitqueue_active(&bitmap->behind_wait)) {
1485                                 mbio = alloc_behind_master_bio(r1_bio, bio);
1486                         }
1487
1488                         bitmap_startwrite(bitmap, r1_bio->sector,
1489                                           r1_bio->sectors,
1490                                           test_bit(R1BIO_BehindIO,
1491                                                    &r1_bio->state));
1492                         first_clone = 0;
1493                 }
1494
1495                 if (!mbio) {
1496                         if (r1_bio->behind_master_bio)
1497                                 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1498                                                       GFP_NOIO,
1499                                                       mddev->bio_set);
1500                         else
1501                                 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1502                 }
1503
1504                 if (r1_bio->behind_master_bio) {
1505                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1506                                 atomic_inc(&r1_bio->behind_remaining);
1507                 }
1508
1509                 r1_bio->bios[i] = mbio;
1510
1511                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1512                                    conf->mirrors[i].rdev->data_offset);
1513                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1514                 mbio->bi_end_io = raid1_end_write_request;
1515                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1516                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1517                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1518                     conf->raid_disks - mddev->degraded > 1)
1519                         mbio->bi_opf |= MD_FAILFAST;
1520                 mbio->bi_private = r1_bio;
1521
1522                 atomic_inc(&r1_bio->remaining);
1523
1524                 if (mddev->gendisk)
1525                         trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1526                                               mbio, disk_devt(mddev->gendisk),
1527                                               r1_bio->sector);
1528                 /* flush_pending_writes() needs access to the rdev so...*/
1529                 mbio->bi_bdev = (void*)conf->mirrors[i].rdev;
1530
1531                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1532                 if (cb)
1533                         plug = container_of(cb, struct raid1_plug_cb, cb);
1534                 else
1535                         plug = NULL;
1536                 if (plug) {
1537                         bio_list_add(&plug->pending, mbio);
1538                         plug->pending_cnt++;
1539                 } else {
1540                         spin_lock_irqsave(&conf->device_lock, flags);
1541                         bio_list_add(&conf->pending_bio_list, mbio);
1542                         conf->pending_count++;
1543                         spin_unlock_irqrestore(&conf->device_lock, flags);
1544                         md_wakeup_thread(mddev->thread);
1545                 }
1546         }
1547
1548         r1_bio_write_done(r1_bio);
1549
1550         /* In case raid1d snuck in to freeze_array */
1551         wake_up(&conf->wait_barrier);
1552 }
1553
1554 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1555 {
1556         sector_t sectors;
1557
1558         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1559                 md_flush_request(mddev, bio);
1560                 return true;
1561         }
1562
1563         /*
1564          * There is a limit to the maximum size, but
1565          * the read/write handler might find a lower limit
1566          * due to bad blocks.  To avoid multiple splits,
1567          * we pass the maximum number of sectors down
1568          * and let the lower level perform the split.
1569          */
1570         sectors = align_to_barrier_unit_end(
1571                 bio->bi_iter.bi_sector, bio_sectors(bio));
1572
1573         if (bio_data_dir(bio) == READ)
1574                 raid1_read_request(mddev, bio, sectors, NULL);
1575         else {
1576                 if (!md_write_start(mddev,bio))
1577                         return false;
1578                 raid1_write_request(mddev, bio, sectors);
1579         }
1580         return true;
1581 }
1582
1583 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1584 {
1585         struct r1conf *conf = mddev->private;
1586         int i;
1587
1588         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1589                    conf->raid_disks - mddev->degraded);
1590         rcu_read_lock();
1591         for (i = 0; i < conf->raid_disks; i++) {
1592                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1593                 seq_printf(seq, "%s",
1594                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1595         }
1596         rcu_read_unlock();
1597         seq_printf(seq, "]");
1598 }
1599
1600 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1601 {
1602         char b[BDEVNAME_SIZE];
1603         struct r1conf *conf = mddev->private;
1604         unsigned long flags;
1605
1606         /*
1607          * If it is not operational, then we have already marked it as dead
1608          * else if it is the last working disks, ignore the error, let the
1609          * next level up know.
1610          * else mark the drive as failed
1611          */
1612         spin_lock_irqsave(&conf->device_lock, flags);
1613         if (test_bit(In_sync, &rdev->flags)
1614             && (conf->raid_disks - mddev->degraded) == 1) {
1615                 /*
1616                  * Don't fail the drive, act as though we were just a
1617                  * normal single drive.
1618                  * However don't try a recovery from this drive as
1619                  * it is very likely to fail.
1620                  */
1621                 conf->recovery_disabled = mddev->recovery_disabled;
1622                 spin_unlock_irqrestore(&conf->device_lock, flags);
1623                 return;
1624         }
1625         set_bit(Blocked, &rdev->flags);
1626         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1627                 mddev->degraded++;
1628                 set_bit(Faulty, &rdev->flags);
1629         } else
1630                 set_bit(Faulty, &rdev->flags);
1631         spin_unlock_irqrestore(&conf->device_lock, flags);
1632         /*
1633          * if recovery is running, make sure it aborts.
1634          */
1635         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1636         set_mask_bits(&mddev->sb_flags, 0,
1637                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1638         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1639                 "md/raid1:%s: Operation continuing on %d devices.\n",
1640                 mdname(mddev), bdevname(rdev->bdev, b),
1641                 mdname(mddev), conf->raid_disks - mddev->degraded);
1642 }
1643
1644 static void print_conf(struct r1conf *conf)
1645 {
1646         int i;
1647
1648         pr_debug("RAID1 conf printout:\n");
1649         if (!conf) {
1650                 pr_debug("(!conf)\n");
1651                 return;
1652         }
1653         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1654                  conf->raid_disks);
1655
1656         rcu_read_lock();
1657         for (i = 0; i < conf->raid_disks; i++) {
1658                 char b[BDEVNAME_SIZE];
1659                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1660                 if (rdev)
1661                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1662                                  i, !test_bit(In_sync, &rdev->flags),
1663                                  !test_bit(Faulty, &rdev->flags),
1664                                  bdevname(rdev->bdev,b));
1665         }
1666         rcu_read_unlock();
1667 }
1668
1669 static void close_sync(struct r1conf *conf)
1670 {
1671         wait_all_barriers(conf);
1672         allow_all_barriers(conf);
1673
1674         mempool_destroy(conf->r1buf_pool);
1675         conf->r1buf_pool = NULL;
1676 }
1677
1678 static int raid1_spare_active(struct mddev *mddev)
1679 {
1680         int i;
1681         struct r1conf *conf = mddev->private;
1682         int count = 0;
1683         unsigned long flags;
1684
1685         /*
1686          * Find all failed disks within the RAID1 configuration
1687          * and mark them readable.
1688          * Called under mddev lock, so rcu protection not needed.
1689          * device_lock used to avoid races with raid1_end_read_request
1690          * which expects 'In_sync' flags and ->degraded to be consistent.
1691          */
1692         spin_lock_irqsave(&conf->device_lock, flags);
1693         for (i = 0; i < conf->raid_disks; i++) {
1694                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1695                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1696                 if (repl
1697                     && !test_bit(Candidate, &repl->flags)
1698                     && repl->recovery_offset == MaxSector
1699                     && !test_bit(Faulty, &repl->flags)
1700                     && !test_and_set_bit(In_sync, &repl->flags)) {
1701                         /* replacement has just become active */
1702                         if (!rdev ||
1703                             !test_and_clear_bit(In_sync, &rdev->flags))
1704                                 count++;
1705                         if (rdev) {
1706                                 /* Replaced device not technically
1707                                  * faulty, but we need to be sure
1708                                  * it gets removed and never re-added
1709                                  */
1710                                 set_bit(Faulty, &rdev->flags);
1711                                 sysfs_notify_dirent_safe(
1712                                         rdev->sysfs_state);
1713                         }
1714                 }
1715                 if (rdev
1716                     && rdev->recovery_offset == MaxSector
1717                     && !test_bit(Faulty, &rdev->flags)
1718                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1719                         count++;
1720                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1721                 }
1722         }
1723         mddev->degraded -= count;
1724         spin_unlock_irqrestore(&conf->device_lock, flags);
1725
1726         print_conf(conf);
1727         return count;
1728 }
1729
1730 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1731 {
1732         struct r1conf *conf = mddev->private;
1733         int err = -EEXIST;
1734         int mirror = 0;
1735         struct raid1_info *p;
1736         int first = 0;
1737         int last = conf->raid_disks - 1;
1738
1739         if (mddev->recovery_disabled == conf->recovery_disabled)
1740                 return -EBUSY;
1741
1742         if (md_integrity_add_rdev(rdev, mddev))
1743                 return -ENXIO;
1744
1745         if (rdev->raid_disk >= 0)
1746                 first = last = rdev->raid_disk;
1747
1748         /*
1749          * find the disk ... but prefer rdev->saved_raid_disk
1750          * if possible.
1751          */
1752         if (rdev->saved_raid_disk >= 0 &&
1753             rdev->saved_raid_disk >= first &&
1754             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1755                 first = last = rdev->saved_raid_disk;
1756
1757         for (mirror = first; mirror <= last; mirror++) {
1758                 p = conf->mirrors+mirror;
1759                 if (!p->rdev) {
1760
1761                         if (mddev->gendisk)
1762                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1763                                                   rdev->data_offset << 9);
1764
1765                         p->head_position = 0;
1766                         rdev->raid_disk = mirror;
1767                         err = 0;
1768                         /* As all devices are equivalent, we don't need a full recovery
1769                          * if this was recently any drive of the array
1770                          */
1771                         if (rdev->saved_raid_disk < 0)
1772                                 conf->fullsync = 1;
1773                         rcu_assign_pointer(p->rdev, rdev);
1774                         break;
1775                 }
1776                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1777                     p[conf->raid_disks].rdev == NULL) {
1778                         /* Add this device as a replacement */
1779                         clear_bit(In_sync, &rdev->flags);
1780                         set_bit(Replacement, &rdev->flags);
1781                         rdev->raid_disk = mirror;
1782                         err = 0;
1783                         conf->fullsync = 1;
1784                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1785                         break;
1786                 }
1787         }
1788         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1789                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1790         print_conf(conf);
1791         return err;
1792 }
1793
1794 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1795 {
1796         struct r1conf *conf = mddev->private;
1797         int err = 0;
1798         int number = rdev->raid_disk;
1799         struct raid1_info *p = conf->mirrors + number;
1800
1801         if (rdev != p->rdev)
1802                 p = conf->mirrors + conf->raid_disks + number;
1803
1804         print_conf(conf);
1805         if (rdev == p->rdev) {
1806                 if (test_bit(In_sync, &rdev->flags) ||
1807                     atomic_read(&rdev->nr_pending)) {
1808                         err = -EBUSY;
1809                         goto abort;
1810                 }
1811                 /* Only remove non-faulty devices if recovery
1812                  * is not possible.
1813                  */
1814                 if (!test_bit(Faulty, &rdev->flags) &&
1815                     mddev->recovery_disabled != conf->recovery_disabled &&
1816                     mddev->degraded < conf->raid_disks) {
1817                         err = -EBUSY;
1818                         goto abort;
1819                 }
1820                 p->rdev = NULL;
1821                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1822                         synchronize_rcu();
1823                         if (atomic_read(&rdev->nr_pending)) {
1824                                 /* lost the race, try later */
1825                                 err = -EBUSY;
1826                                 p->rdev = rdev;
1827                                 goto abort;
1828                         }
1829                 }
1830                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1831                         /* We just removed a device that is being replaced.
1832                          * Move down the replacement.  We drain all IO before
1833                          * doing this to avoid confusion.
1834                          */
1835                         struct md_rdev *repl =
1836                                 conf->mirrors[conf->raid_disks + number].rdev;
1837                         freeze_array(conf, 0);
1838                         clear_bit(Replacement, &repl->flags);
1839                         p->rdev = repl;
1840                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1841                         unfreeze_array(conf);
1842                 }
1843
1844                 clear_bit(WantReplacement, &rdev->flags);
1845                 err = md_integrity_register(mddev);
1846         }
1847 abort:
1848
1849         print_conf(conf);
1850         return err;
1851 }
1852
1853 static void end_sync_read(struct bio *bio)
1854 {
1855         struct r1bio *r1_bio = get_resync_r1bio(bio);
1856
1857         update_head_pos(r1_bio->read_disk, r1_bio);
1858
1859         /*
1860          * we have read a block, now it needs to be re-written,
1861          * or re-read if the read failed.
1862          * We don't do much here, just schedule handling by raid1d
1863          */
1864         if (!bio->bi_status)
1865                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1866
1867         if (atomic_dec_and_test(&r1_bio->remaining))
1868                 reschedule_retry(r1_bio);
1869 }
1870
1871 static void end_sync_write(struct bio *bio)
1872 {
1873         int uptodate = !bio->bi_status;
1874         struct r1bio *r1_bio = get_resync_r1bio(bio);
1875         struct mddev *mddev = r1_bio->mddev;
1876         struct r1conf *conf = mddev->private;
1877         sector_t first_bad;
1878         int bad_sectors;
1879         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1880
1881         if (!uptodate) {
1882                 sector_t sync_blocks = 0;
1883                 sector_t s = r1_bio->sector;
1884                 long sectors_to_go = r1_bio->sectors;
1885                 /* make sure these bits doesn't get cleared. */
1886                 do {
1887                         bitmap_end_sync(mddev->bitmap, s,
1888                                         &sync_blocks, 1);
1889                         s += sync_blocks;
1890                         sectors_to_go -= sync_blocks;
1891                 } while (sectors_to_go > 0);
1892                 set_bit(WriteErrorSeen, &rdev->flags);
1893                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1894                         set_bit(MD_RECOVERY_NEEDED, &
1895                                 mddev->recovery);
1896                 set_bit(R1BIO_WriteError, &r1_bio->state);
1897         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1898                                &first_bad, &bad_sectors) &&
1899                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1900                                 r1_bio->sector,
1901                                 r1_bio->sectors,
1902                                 &first_bad, &bad_sectors)
1903                 )
1904                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1905
1906         if (atomic_dec_and_test(&r1_bio->remaining)) {
1907                 int s = r1_bio->sectors;
1908                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1909                     test_bit(R1BIO_WriteError, &r1_bio->state))
1910                         reschedule_retry(r1_bio);
1911                 else {
1912                         put_buf(r1_bio);
1913                         md_done_sync(mddev, s, uptodate);
1914                 }
1915         }
1916 }
1917
1918 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1919                             int sectors, struct page *page, int rw)
1920 {
1921         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1922                 /* success */
1923                 return 1;
1924         if (rw == WRITE) {
1925                 set_bit(WriteErrorSeen, &rdev->flags);
1926                 if (!test_and_set_bit(WantReplacement,
1927                                       &rdev->flags))
1928                         set_bit(MD_RECOVERY_NEEDED, &
1929                                 rdev->mddev->recovery);
1930         }
1931         /* need to record an error - either for the block or the device */
1932         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1933                 md_error(rdev->mddev, rdev);
1934         return 0;
1935 }
1936
1937 static int fix_sync_read_error(struct r1bio *r1_bio)
1938 {
1939         /* Try some synchronous reads of other devices to get
1940          * good data, much like with normal read errors.  Only
1941          * read into the pages we already have so we don't
1942          * need to re-issue the read request.
1943          * We don't need to freeze the array, because being in an
1944          * active sync request, there is no normal IO, and
1945          * no overlapping syncs.
1946          * We don't need to check is_badblock() again as we
1947          * made sure that anything with a bad block in range
1948          * will have bi_end_io clear.
1949          */
1950         struct mddev *mddev = r1_bio->mddev;
1951         struct r1conf *conf = mddev->private;
1952         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1953         struct page **pages = get_resync_pages(bio)->pages;
1954         sector_t sect = r1_bio->sector;
1955         int sectors = r1_bio->sectors;
1956         int idx = 0;
1957         struct md_rdev *rdev;
1958
1959         rdev = conf->mirrors[r1_bio->read_disk].rdev;
1960         if (test_bit(FailFast, &rdev->flags)) {
1961                 /* Don't try recovering from here - just fail it
1962                  * ... unless it is the last working device of course */
1963                 md_error(mddev, rdev);
1964                 if (test_bit(Faulty, &rdev->flags))
1965                         /* Don't try to read from here, but make sure
1966                          * put_buf does it's thing
1967                          */
1968                         bio->bi_end_io = end_sync_write;
1969         }
1970
1971         while(sectors) {
1972                 int s = sectors;
1973                 int d = r1_bio->read_disk;
1974                 int success = 0;
1975                 int start;
1976
1977                 if (s > (PAGE_SIZE>>9))
1978                         s = PAGE_SIZE >> 9;
1979                 do {
1980                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1981                                 /* No rcu protection needed here devices
1982                                  * can only be removed when no resync is
1983                                  * active, and resync is currently active
1984                                  */
1985                                 rdev = conf->mirrors[d].rdev;
1986                                 if (sync_page_io(rdev, sect, s<<9,
1987                                                  pages[idx],
1988                                                  REQ_OP_READ, 0, false)) {
1989                                         success = 1;
1990                                         break;
1991                                 }
1992                         }
1993                         d++;
1994                         if (d == conf->raid_disks * 2)
1995                                 d = 0;
1996                 } while (!success && d != r1_bio->read_disk);
1997
1998                 if (!success) {
1999                         char b[BDEVNAME_SIZE];
2000                         int abort = 0;
2001                         /* Cannot read from anywhere, this block is lost.
2002                          * Record a bad block on each device.  If that doesn't
2003                          * work just disable and interrupt the recovery.
2004                          * Don't fail devices as that won't really help.
2005                          */
2006                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2007                                             mdname(mddev),
2008                                             bdevname(bio->bi_bdev, b),
2009                                             (unsigned long long)r1_bio->sector);
2010                         for (d = 0; d < conf->raid_disks * 2; d++) {
2011                                 rdev = conf->mirrors[d].rdev;
2012                                 if (!rdev || test_bit(Faulty, &rdev->flags))
2013                                         continue;
2014                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
2015                                         abort = 1;
2016                         }
2017                         if (abort) {
2018                                 conf->recovery_disabled =
2019                                         mddev->recovery_disabled;
2020                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2021                                 md_done_sync(mddev, r1_bio->sectors, 0);
2022                                 put_buf(r1_bio);
2023                                 return 0;
2024                         }
2025                         /* Try next page */
2026                         sectors -= s;
2027                         sect += s;
2028                         idx++;
2029                         continue;
2030                 }
2031
2032                 start = d;
2033                 /* write it back and re-read */
2034                 while (d != r1_bio->read_disk) {
2035                         if (d == 0)
2036                                 d = conf->raid_disks * 2;
2037                         d--;
2038                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2039                                 continue;
2040                         rdev = conf->mirrors[d].rdev;
2041                         if (r1_sync_page_io(rdev, sect, s,
2042                                             pages[idx],
2043                                             WRITE) == 0) {
2044                                 r1_bio->bios[d]->bi_end_io = NULL;
2045                                 rdev_dec_pending(rdev, mddev);
2046                         }
2047                 }
2048                 d = start;
2049                 while (d != r1_bio->read_disk) {
2050                         if (d == 0)
2051                                 d = conf->raid_disks * 2;
2052                         d--;
2053                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2054                                 continue;
2055                         rdev = conf->mirrors[d].rdev;
2056                         if (r1_sync_page_io(rdev, sect, s,
2057                                             pages[idx],
2058                                             READ) != 0)
2059                                 atomic_add(s, &rdev->corrected_errors);
2060                 }
2061                 sectors -= s;
2062                 sect += s;
2063                 idx ++;
2064         }
2065         set_bit(R1BIO_Uptodate, &r1_bio->state);
2066         bio->bi_status = 0;
2067         return 1;
2068 }
2069
2070 static void process_checks(struct r1bio *r1_bio)
2071 {
2072         /* We have read all readable devices.  If we haven't
2073          * got the block, then there is no hope left.
2074          * If we have, then we want to do a comparison
2075          * and skip the write if everything is the same.
2076          * If any blocks failed to read, then we need to
2077          * attempt an over-write
2078          */
2079         struct mddev *mddev = r1_bio->mddev;
2080         struct r1conf *conf = mddev->private;
2081         int primary;
2082         int i;
2083         int vcnt;
2084
2085         /* Fix variable parts of all bios */
2086         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2087         for (i = 0; i < conf->raid_disks * 2; i++) {
2088                 int j;
2089                 int size;
2090                 blk_status_t status;
2091                 struct bio_vec *bi;
2092                 struct bio *b = r1_bio->bios[i];
2093                 struct resync_pages *rp = get_resync_pages(b);
2094                 if (b->bi_end_io != end_sync_read)
2095                         continue;
2096                 /* fixup the bio for reuse, but preserve errno */
2097                 status = b->bi_status;
2098                 bio_reset(b);
2099                 b->bi_status = status;
2100                 b->bi_vcnt = vcnt;
2101                 b->bi_iter.bi_size = r1_bio->sectors << 9;
2102                 b->bi_iter.bi_sector = r1_bio->sector +
2103                         conf->mirrors[i].rdev->data_offset;
2104                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
2105                 b->bi_end_io = end_sync_read;
2106                 rp->raid_bio = r1_bio;
2107                 b->bi_private = rp;
2108
2109                 size = b->bi_iter.bi_size;
2110                 bio_for_each_segment_all(bi, b, j) {
2111                         bi->bv_offset = 0;
2112                         if (size > PAGE_SIZE)
2113                                 bi->bv_len = PAGE_SIZE;
2114                         else
2115                                 bi->bv_len = size;
2116                         size -= PAGE_SIZE;
2117                 }
2118         }
2119         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2120                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2121                     !r1_bio->bios[primary]->bi_status) {
2122                         r1_bio->bios[primary]->bi_end_io = NULL;
2123                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2124                         break;
2125                 }
2126         r1_bio->read_disk = primary;
2127         for (i = 0; i < conf->raid_disks * 2; i++) {
2128                 int j;
2129                 struct bio *pbio = r1_bio->bios[primary];
2130                 struct bio *sbio = r1_bio->bios[i];
2131                 blk_status_t status = sbio->bi_status;
2132                 struct page **ppages = get_resync_pages(pbio)->pages;
2133                 struct page **spages = get_resync_pages(sbio)->pages;
2134                 struct bio_vec *bi;
2135                 int page_len[RESYNC_PAGES] = { 0 };
2136
2137                 if (sbio->bi_end_io != end_sync_read)
2138                         continue;
2139                 /* Now we can 'fixup' the error value */
2140                 sbio->bi_status = 0;
2141
2142                 bio_for_each_segment_all(bi, sbio, j)
2143                         page_len[j] = bi->bv_len;
2144
2145                 if (!status) {
2146                         for (j = vcnt; j-- ; ) {
2147                                 if (memcmp(page_address(ppages[j]),
2148                                            page_address(spages[j]),
2149                                            page_len[j]))
2150                                         break;
2151                         }
2152                 } else
2153                         j = 0;
2154                 if (j >= 0)
2155                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2156                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2157                               && !status)) {
2158                         /* No need to write to this device. */
2159                         sbio->bi_end_io = NULL;
2160                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2161                         continue;
2162                 }
2163
2164                 bio_copy_data(sbio, pbio);
2165         }
2166 }
2167
2168 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2169 {
2170         struct r1conf *conf = mddev->private;
2171         int i;
2172         int disks = conf->raid_disks * 2;
2173         struct bio *wbio;
2174
2175         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2176                 /* ouch - failed to read all of that. */
2177                 if (!fix_sync_read_error(r1_bio))
2178                         return;
2179
2180         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2181                 process_checks(r1_bio);
2182
2183         /*
2184          * schedule writes
2185          */
2186         atomic_set(&r1_bio->remaining, 1);
2187         for (i = 0; i < disks ; i++) {
2188                 wbio = r1_bio->bios[i];
2189                 if (wbio->bi_end_io == NULL ||
2190                     (wbio->bi_end_io == end_sync_read &&
2191                      (i == r1_bio->read_disk ||
2192                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2193                         continue;
2194                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2195                         continue;
2196
2197                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2198                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2199                         wbio->bi_opf |= MD_FAILFAST;
2200
2201                 wbio->bi_end_io = end_sync_write;
2202                 atomic_inc(&r1_bio->remaining);
2203                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2204
2205                 generic_make_request(wbio);
2206         }
2207
2208         if (atomic_dec_and_test(&r1_bio->remaining)) {
2209                 /* if we're here, all write(s) have completed, so clean up */
2210                 int s = r1_bio->sectors;
2211                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2212                     test_bit(R1BIO_WriteError, &r1_bio->state))
2213                         reschedule_retry(r1_bio);
2214                 else {
2215                         put_buf(r1_bio);
2216                         md_done_sync(mddev, s, 1);
2217                 }
2218         }
2219 }
2220
2221 /*
2222  * This is a kernel thread which:
2223  *
2224  *      1.      Retries failed read operations on working mirrors.
2225  *      2.      Updates the raid superblock when problems encounter.
2226  *      3.      Performs writes following reads for array synchronising.
2227  */
2228
2229 static void fix_read_error(struct r1conf *conf, int read_disk,
2230                            sector_t sect, int sectors)
2231 {
2232         struct mddev *mddev = conf->mddev;
2233         while(sectors) {
2234                 int s = sectors;
2235                 int d = read_disk;
2236                 int success = 0;
2237                 int start;
2238                 struct md_rdev *rdev;
2239
2240                 if (s > (PAGE_SIZE>>9))
2241                         s = PAGE_SIZE >> 9;
2242
2243                 do {
2244                         sector_t first_bad;
2245                         int bad_sectors;
2246
2247                         rcu_read_lock();
2248                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2249                         if (rdev &&
2250                             (test_bit(In_sync, &rdev->flags) ||
2251                              (!test_bit(Faulty, &rdev->flags) &&
2252                               rdev->recovery_offset >= sect + s)) &&
2253                             is_badblock(rdev, sect, s,
2254                                         &first_bad, &bad_sectors) == 0) {
2255                                 atomic_inc(&rdev->nr_pending);
2256                                 rcu_read_unlock();
2257                                 if (sync_page_io(rdev, sect, s<<9,
2258                                          conf->tmppage, REQ_OP_READ, 0, false))
2259                                         success = 1;
2260                                 rdev_dec_pending(rdev, mddev);
2261                                 if (success)
2262                                         break;
2263                         } else
2264                                 rcu_read_unlock();
2265                         d++;
2266                         if (d == conf->raid_disks * 2)
2267                                 d = 0;
2268                 } while (!success && d != read_disk);
2269
2270                 if (!success) {
2271                         /* Cannot read from anywhere - mark it bad */
2272                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2273                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2274                                 md_error(mddev, rdev);
2275                         break;
2276                 }
2277                 /* write it back and re-read */
2278                 start = d;
2279                 while (d != read_disk) {
2280                         if (d==0)
2281                                 d = conf->raid_disks * 2;
2282                         d--;
2283                         rcu_read_lock();
2284                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2285                         if (rdev &&
2286                             !test_bit(Faulty, &rdev->flags)) {
2287                                 atomic_inc(&rdev->nr_pending);
2288                                 rcu_read_unlock();
2289                                 r1_sync_page_io(rdev, sect, s,
2290                                                 conf->tmppage, WRITE);
2291                                 rdev_dec_pending(rdev, mddev);
2292                         } else
2293                                 rcu_read_unlock();
2294                 }
2295                 d = start;
2296                 while (d != read_disk) {
2297                         char b[BDEVNAME_SIZE];
2298                         if (d==0)
2299                                 d = conf->raid_disks * 2;
2300                         d--;
2301                         rcu_read_lock();
2302                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2303                         if (rdev &&
2304                             !test_bit(Faulty, &rdev->flags)) {
2305                                 atomic_inc(&rdev->nr_pending);
2306                                 rcu_read_unlock();
2307                                 if (r1_sync_page_io(rdev, sect, s,
2308                                                     conf->tmppage, READ)) {
2309                                         atomic_add(s, &rdev->corrected_errors);
2310                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2311                                                 mdname(mddev), s,
2312                                                 (unsigned long long)(sect +
2313                                                                      rdev->data_offset),
2314                                                 bdevname(rdev->bdev, b));
2315                                 }
2316                                 rdev_dec_pending(rdev, mddev);
2317                         } else
2318                                 rcu_read_unlock();
2319                 }
2320                 sectors -= s;
2321                 sect += s;
2322         }
2323 }
2324
2325 static int narrow_write_error(struct r1bio *r1_bio, int i)
2326 {
2327         struct mddev *mddev = r1_bio->mddev;
2328         struct r1conf *conf = mddev->private;
2329         struct md_rdev *rdev = conf->mirrors[i].rdev;
2330
2331         /* bio has the data to be written to device 'i' where
2332          * we just recently had a write error.
2333          * We repeatedly clone the bio and trim down to one block,
2334          * then try the write.  Where the write fails we record
2335          * a bad block.
2336          * It is conceivable that the bio doesn't exactly align with
2337          * blocks.  We must handle this somehow.
2338          *
2339          * We currently own a reference on the rdev.
2340          */
2341
2342         int block_sectors;
2343         sector_t sector;
2344         int sectors;
2345         int sect_to_write = r1_bio->sectors;
2346         int ok = 1;
2347
2348         if (rdev->badblocks.shift < 0)
2349                 return 0;
2350
2351         block_sectors = roundup(1 << rdev->badblocks.shift,
2352                                 bdev_logical_block_size(rdev->bdev) >> 9);
2353         sector = r1_bio->sector;
2354         sectors = ((sector + block_sectors)
2355                    & ~(sector_t)(block_sectors - 1))
2356                 - sector;
2357
2358         while (sect_to_write) {
2359                 struct bio *wbio;
2360                 if (sectors > sect_to_write)
2361                         sectors = sect_to_write;
2362                 /* Write at 'sector' for 'sectors'*/
2363
2364                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2365                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2366                                               GFP_NOIO,
2367                                               mddev->bio_set);
2368                         /* We really need a _all clone */
2369                         wbio->bi_iter = (struct bvec_iter){ 0 };
2370                 } else {
2371                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2372                                               mddev->bio_set);
2373                 }
2374
2375                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2376                 wbio->bi_iter.bi_sector = r1_bio->sector;
2377                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2378
2379                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2380                 wbio->bi_iter.bi_sector += rdev->data_offset;
2381                 wbio->bi_bdev = rdev->bdev;
2382
2383                 if (submit_bio_wait(wbio) < 0)
2384                         /* failure! */
2385                         ok = rdev_set_badblocks(rdev, sector,
2386                                                 sectors, 0)
2387                                 && ok;
2388
2389                 bio_put(wbio);
2390                 sect_to_write -= sectors;
2391                 sector += sectors;
2392                 sectors = block_sectors;
2393         }
2394         return ok;
2395 }
2396
2397 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2398 {
2399         int m;
2400         int s = r1_bio->sectors;
2401         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2402                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2403                 struct bio *bio = r1_bio->bios[m];
2404                 if (bio->bi_end_io == NULL)
2405                         continue;
2406                 if (!bio->bi_status &&
2407                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2408                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2409                 }
2410                 if (bio->bi_status &&
2411                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2412                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2413                                 md_error(conf->mddev, rdev);
2414                 }
2415         }
2416         put_buf(r1_bio);
2417         md_done_sync(conf->mddev, s, 1);
2418 }
2419
2420 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2421 {
2422         int m, idx;
2423         bool fail = false;
2424
2425         for (m = 0; m < conf->raid_disks * 2 ; m++)
2426                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2427                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2428                         rdev_clear_badblocks(rdev,
2429                                              r1_bio->sector,
2430                                              r1_bio->sectors, 0);
2431                         rdev_dec_pending(rdev, conf->mddev);
2432                 } else if (r1_bio->bios[m] != NULL) {
2433                         /* This drive got a write error.  We need to
2434                          * narrow down and record precise write
2435                          * errors.
2436                          */
2437                         fail = true;
2438                         if (!narrow_write_error(r1_bio, m)) {
2439                                 md_error(conf->mddev,
2440                                          conf->mirrors[m].rdev);
2441                                 /* an I/O failed, we can't clear the bitmap */
2442                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2443                         }
2444                         rdev_dec_pending(conf->mirrors[m].rdev,
2445                                          conf->mddev);
2446                 }
2447         if (fail) {
2448                 spin_lock_irq(&conf->device_lock);
2449                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2450                 idx = sector_to_idx(r1_bio->sector);
2451                 atomic_inc(&conf->nr_queued[idx]);
2452                 spin_unlock_irq(&conf->device_lock);
2453                 /*
2454                  * In case freeze_array() is waiting for condition
2455                  * get_unqueued_pending() == extra to be true.
2456                  */
2457                 wake_up(&conf->wait_barrier);
2458                 md_wakeup_thread(conf->mddev->thread);
2459         } else {
2460                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2461                         close_write(r1_bio);
2462                 raid_end_bio_io(r1_bio);
2463         }
2464 }
2465
2466 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2467 {
2468         struct mddev *mddev = conf->mddev;
2469         struct bio *bio;
2470         struct md_rdev *rdev;
2471         dev_t bio_dev;
2472         sector_t bio_sector;
2473
2474         clear_bit(R1BIO_ReadError, &r1_bio->state);
2475         /* we got a read error. Maybe the drive is bad.  Maybe just
2476          * the block and we can fix it.
2477          * We freeze all other IO, and try reading the block from
2478          * other devices.  When we find one, we re-write
2479          * and check it that fixes the read error.
2480          * This is all done synchronously while the array is
2481          * frozen
2482          */
2483
2484         bio = r1_bio->bios[r1_bio->read_disk];
2485         bio_dev = bio->bi_bdev->bd_dev;
2486         bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2487         bio_put(bio);
2488         r1_bio->bios[r1_bio->read_disk] = NULL;
2489
2490         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2491         if (mddev->ro == 0
2492             && !test_bit(FailFast, &rdev->flags)) {
2493                 freeze_array(conf, 1);
2494                 fix_read_error(conf, r1_bio->read_disk,
2495                                r1_bio->sector, r1_bio->sectors);
2496                 unfreeze_array(conf);
2497         } else {
2498                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2499         }
2500
2501         rdev_dec_pending(rdev, conf->mddev);
2502         allow_barrier(conf, r1_bio->sector);
2503         bio = r1_bio->master_bio;
2504
2505         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2506         r1_bio->state = 0;
2507         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2508 }
2509
2510 static void raid1d(struct md_thread *thread)
2511 {
2512         struct mddev *mddev = thread->mddev;
2513         struct r1bio *r1_bio;
2514         unsigned long flags;
2515         struct r1conf *conf = mddev->private;
2516         struct list_head *head = &conf->retry_list;
2517         struct blk_plug plug;
2518         int idx;
2519
2520         md_check_recovery(mddev);
2521
2522         if (!list_empty_careful(&conf->bio_end_io_list) &&
2523             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2524                 LIST_HEAD(tmp);
2525                 spin_lock_irqsave(&conf->device_lock, flags);
2526                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2527                         list_splice_init(&conf->bio_end_io_list, &tmp);
2528                 spin_unlock_irqrestore(&conf->device_lock, flags);
2529                 while (!list_empty(&tmp)) {
2530                         r1_bio = list_first_entry(&tmp, struct r1bio,
2531                                                   retry_list);
2532                         list_del(&r1_bio->retry_list);
2533                         idx = sector_to_idx(r1_bio->sector);
2534                         atomic_dec(&conf->nr_queued[idx]);
2535                         if (mddev->degraded)
2536                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2537                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2538                                 close_write(r1_bio);
2539                         raid_end_bio_io(r1_bio);
2540                 }
2541         }
2542
2543         blk_start_plug(&plug);
2544         for (;;) {
2545
2546                 flush_pending_writes(conf);
2547
2548                 spin_lock_irqsave(&conf->device_lock, flags);
2549                 if (list_empty(head)) {
2550                         spin_unlock_irqrestore(&conf->device_lock, flags);
2551                         break;
2552                 }
2553                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2554                 list_del(head->prev);
2555                 idx = sector_to_idx(r1_bio->sector);
2556                 atomic_dec(&conf->nr_queued[idx]);
2557                 spin_unlock_irqrestore(&conf->device_lock, flags);
2558
2559                 mddev = r1_bio->mddev;
2560                 conf = mddev->private;
2561                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2562                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2563                             test_bit(R1BIO_WriteError, &r1_bio->state))
2564                                 handle_sync_write_finished(conf, r1_bio);
2565                         else
2566                                 sync_request_write(mddev, r1_bio);
2567                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2568                            test_bit(R1BIO_WriteError, &r1_bio->state))
2569                         handle_write_finished(conf, r1_bio);
2570                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2571                         handle_read_error(conf, r1_bio);
2572                 else
2573                         WARN_ON_ONCE(1);
2574
2575                 cond_resched();
2576                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2577                         md_check_recovery(mddev);
2578         }
2579         blk_finish_plug(&plug);
2580 }
2581
2582 static int init_resync(struct r1conf *conf)
2583 {
2584         int buffs;
2585
2586         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2587         BUG_ON(conf->r1buf_pool);
2588         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2589                                           conf->poolinfo);
2590         if (!conf->r1buf_pool)
2591                 return -ENOMEM;
2592         return 0;
2593 }
2594
2595 /*
2596  * perform a "sync" on one "block"
2597  *
2598  * We need to make sure that no normal I/O request - particularly write
2599  * requests - conflict with active sync requests.
2600  *
2601  * This is achieved by tracking pending requests and a 'barrier' concept
2602  * that can be installed to exclude normal IO requests.
2603  */
2604
2605 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2606                                    int *skipped)
2607 {
2608         struct r1conf *conf = mddev->private;
2609         struct r1bio *r1_bio;
2610         struct bio *bio;
2611         sector_t max_sector, nr_sectors;
2612         int disk = -1;
2613         int i;
2614         int wonly = -1;
2615         int write_targets = 0, read_targets = 0;
2616         sector_t sync_blocks;
2617         int still_degraded = 0;
2618         int good_sectors = RESYNC_SECTORS;
2619         int min_bad = 0; /* number of sectors that are bad in all devices */
2620         int idx = sector_to_idx(sector_nr);
2621         int page_idx = 0;
2622
2623         if (!conf->r1buf_pool)
2624                 if (init_resync(conf))
2625                         return 0;
2626
2627         max_sector = mddev->dev_sectors;
2628         if (sector_nr >= max_sector) {
2629                 /* If we aborted, we need to abort the
2630                  * sync on the 'current' bitmap chunk (there will
2631                  * only be one in raid1 resync.
2632                  * We can find the current addess in mddev->curr_resync
2633                  */
2634                 if (mddev->curr_resync < max_sector) /* aborted */
2635                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2636                                                 &sync_blocks, 1);
2637                 else /* completed sync */
2638                         conf->fullsync = 0;
2639
2640                 bitmap_close_sync(mddev->bitmap);
2641                 close_sync(conf);
2642
2643                 if (mddev_is_clustered(mddev)) {
2644                         conf->cluster_sync_low = 0;
2645                         conf->cluster_sync_high = 0;
2646                 }
2647                 return 0;
2648         }
2649
2650         if (mddev->bitmap == NULL &&
2651             mddev->recovery_cp == MaxSector &&
2652             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2653             conf->fullsync == 0) {
2654                 *skipped = 1;
2655                 return max_sector - sector_nr;
2656         }
2657         /* before building a request, check if we can skip these blocks..
2658          * This call the bitmap_start_sync doesn't actually record anything
2659          */
2660         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2661             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2662                 /* We can skip this block, and probably several more */
2663                 *skipped = 1;
2664                 return sync_blocks;
2665         }
2666
2667         /*
2668          * If there is non-resync activity waiting for a turn, then let it
2669          * though before starting on this new sync request.
2670          */
2671         if (atomic_read(&conf->nr_waiting[idx]))
2672                 schedule_timeout_uninterruptible(1);
2673
2674         /* we are incrementing sector_nr below. To be safe, we check against
2675          * sector_nr + two times RESYNC_SECTORS
2676          */
2677
2678         bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2679                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2680         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2681
2682         raise_barrier(conf, sector_nr);
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->bi_bdev = 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->bi_bdev, 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->bi_bdev, 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
2902 static 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
2910 static 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
3051 static void raid1_free(struct mddev *mddev, void *priv);
3052 static 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                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3129                                                 mddev->queue);
3130                 else
3131                         queue_flag_clear_unlocked(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
3143 static 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
3160 static 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
3189 static 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
3294 static void raid1_quiesce(struct mddev *mddev, int state)
3295 {
3296         struct r1conf *conf = mddev->private;
3297
3298         switch(state) {
3299         case 2: /* wake for suspend */
3300                 wake_up(&conf->wait_barrier);
3301                 break;
3302         case 1:
3303                 freeze_array(conf, 0);
3304                 break;
3305         case 0:
3306                 unfreeze_array(conf);
3307                 break;
3308         }
3309 }
3310
3311 static void *raid1_takeover(struct mddev *mddev)
3312 {
3313         /* raid1 can take over:
3314          *  raid5 with 2 devices, any layout or chunk size
3315          */
3316         if (mddev->level == 5 && mddev->raid_disks == 2) {
3317                 struct r1conf *conf;
3318                 mddev->new_level = 1;
3319                 mddev->new_layout = 0;
3320                 mddev->new_chunk_sectors = 0;
3321                 conf = setup_conf(mddev);
3322                 if (!IS_ERR(conf)) {
3323                         /* Array must appear to be quiesced */
3324                         conf->array_frozen = 1;
3325                         mddev_clear_unsupported_flags(mddev,
3326                                 UNSUPPORTED_MDDEV_FLAGS);
3327                 }
3328                 return conf;
3329         }
3330         return ERR_PTR(-EINVAL);
3331 }
3332
3333 static struct md_personality raid1_personality =
3334 {
3335         .name           = "raid1",
3336         .level          = 1,
3337         .owner          = THIS_MODULE,
3338         .make_request   = raid1_make_request,
3339         .run            = raid1_run,
3340         .free           = raid1_free,
3341         .status         = raid1_status,
3342         .error_handler  = raid1_error,
3343         .hot_add_disk   = raid1_add_disk,
3344         .hot_remove_disk= raid1_remove_disk,
3345         .spare_active   = raid1_spare_active,
3346         .sync_request   = raid1_sync_request,
3347         .resize         = raid1_resize,
3348         .size           = raid1_size,
3349         .check_reshape  = raid1_reshape,
3350         .quiesce        = raid1_quiesce,
3351         .takeover       = raid1_takeover,
3352         .congested      = raid1_congested,
3353 };
3354
3355 static int __init raid_init(void)
3356 {
3357         return register_md_personality(&raid1_personality);
3358 }
3359
3360 static void raid_exit(void)
3361 {
3362         unregister_md_personality(&raid1_personality);
3363 }
3364
3365 module_init(raid_init);
3366 module_exit(raid_exit);
3367 MODULE_LICENSE("GPL");
3368 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3369 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3370 MODULE_ALIAS("md-raid1");
3371 MODULE_ALIAS("md-level-1");
3372
3373 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);