b2eae332e1a29ee585c04ba6d22c2a23b8b99ed3
[sfrench/cifs-2.6.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 Ć˜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
41 #include <trace/events/block.h>
42
43 #include "md.h"
44 #include "raid1.h"
45 #include "md-bitmap.h"
46
47 #define UNSUPPORTED_MDDEV_FLAGS         \
48         ((1L << MD_HAS_JOURNAL) |       \
49          (1L << MD_JOURNAL_CLEAN) |     \
50          (1L << MD_HAS_PPL) |           \
51          (1L << MD_HAS_MULTIPLE_PPLS))
52
53 /*
54  * Number of guaranteed r1bios in case of extreme VM load:
55  */
56 #define NR_RAID1_BIOS 256
57
58 /* when we get a read error on a read-only array, we redirect to another
59  * device without failing the first device, or trying to over-write to
60  * correct the read error.  To keep track of bad blocks on a per-bio
61  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
62  */
63 #define IO_BLOCKED ((struct bio *)1)
64 /* When we successfully write to a known bad-block, we need to remove the
65  * bad-block marking which must be done from process context.  So we record
66  * the success by setting devs[n].bio to IO_MADE_GOOD
67  */
68 #define IO_MADE_GOOD ((struct bio *)2)
69
70 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
71
72 /* When there are this many requests queue to be written by
73  * the raid1 thread, we become 'congested' to provide back-pressure
74  * for writeback.
75  */
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 #include "raid1-10.c"
85
86 /*
87  * for resync bio, r1bio pointer can be retrieved from the per-bio
88  * 'struct resync_pages'.
89  */
90 static inline struct r1bio *get_resync_r1bio(struct bio *bio)
91 {
92         return get_resync_pages(bio)->raid_bio;
93 }
94
95 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
96 {
97         struct pool_info *pi = data;
98         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
99
100         /* allocate a r1bio with room for raid_disks entries in the bios array */
101         return kzalloc(size, gfp_flags);
102 }
103
104 static void r1bio_pool_free(void *r1_bio, void *data)
105 {
106         kfree(r1_bio);
107 }
108
109 #define RESYNC_DEPTH 32
110 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
111 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
112 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
113 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
114 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
115
116 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
117 {
118         struct pool_info *pi = data;
119         struct r1bio *r1_bio;
120         struct bio *bio;
121         int need_pages;
122         int j;
123         struct resync_pages *rps;
124
125         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
126         if (!r1_bio)
127                 return NULL;
128
129         rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
130                       gfp_flags);
131         if (!rps)
132                 goto out_free_r1bio;
133
134         /*
135          * Allocate bios : 1 for reading, n-1 for writing
136          */
137         for (j = pi->raid_disks ; j-- ; ) {
138                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
139                 if (!bio)
140                         goto out_free_bio;
141                 r1_bio->bios[j] = bio;
142         }
143         /*
144          * Allocate RESYNC_PAGES data pages and attach them to
145          * the first bio.
146          * If this is a user-requested check/repair, allocate
147          * RESYNC_PAGES for each bio.
148          */
149         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
150                 need_pages = pi->raid_disks;
151         else
152                 need_pages = 1;
153         for (j = 0; j < pi->raid_disks; j++) {
154                 struct resync_pages *rp = &rps[j];
155
156                 bio = r1_bio->bios[j];
157
158                 if (j < need_pages) {
159                         if (resync_alloc_pages(rp, gfp_flags))
160                                 goto out_free_pages;
161                 } else {
162                         memcpy(rp, &rps[0], sizeof(*rp));
163                         resync_get_all_pages(rp);
164                 }
165
166                 rp->raid_bio = r1_bio;
167                 bio->bi_private = rp;
168         }
169
170         r1_bio->master_bio = NULL;
171
172         return r1_bio;
173
174 out_free_pages:
175         while (--j >= 0)
176                 resync_free_pages(&rps[j]);
177
178 out_free_bio:
179         while (++j < pi->raid_disks)
180                 bio_put(r1_bio->bios[j]);
181         kfree(rps);
182
183 out_free_r1bio:
184         r1bio_pool_free(r1_bio, data);
185         return NULL;
186 }
187
188 static void r1buf_pool_free(void *__r1_bio, void *data)
189 {
190         struct pool_info *pi = data;
191         int i;
192         struct r1bio *r1bio = __r1_bio;
193         struct resync_pages *rp = NULL;
194
195         for (i = pi->raid_disks; i--; ) {
196                 rp = get_resync_pages(r1bio->bios[i]);
197                 resync_free_pages(rp);
198                 bio_put(r1bio->bios[i]);
199         }
200
201         /* resync pages array stored in the 1st bio's .bi_private */
202         kfree(rp);
203
204         r1bio_pool_free(r1bio, data);
205 }
206
207 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
208 {
209         int i;
210
211         for (i = 0; i < conf->raid_disks * 2; i++) {
212                 struct bio **bio = r1_bio->bios + i;
213                 if (!BIO_SPECIAL(*bio))
214                         bio_put(*bio);
215                 *bio = NULL;
216         }
217 }
218
219 static void free_r1bio(struct r1bio *r1_bio)
220 {
221         struct r1conf *conf = r1_bio->mddev->private;
222
223         put_all_bios(conf, r1_bio);
224         mempool_free(r1_bio, conf->r1bio_pool);
225 }
226
227 static void put_buf(struct r1bio *r1_bio)
228 {
229         struct r1conf *conf = r1_bio->mddev->private;
230         sector_t sect = r1_bio->sector;
231         int i;
232
233         for (i = 0; i < conf->raid_disks * 2; i++) {
234                 struct bio *bio = r1_bio->bios[i];
235                 if (bio->bi_end_io)
236                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
237         }
238
239         mempool_free(r1_bio, conf->r1buf_pool);
240
241         lower_barrier(conf, sect);
242 }
243
244 static void reschedule_retry(struct r1bio *r1_bio)
245 {
246         unsigned long flags;
247         struct mddev *mddev = r1_bio->mddev;
248         struct r1conf *conf = mddev->private;
249         int idx;
250
251         idx = sector_to_idx(r1_bio->sector);
252         spin_lock_irqsave(&conf->device_lock, flags);
253         list_add(&r1_bio->retry_list, &conf->retry_list);
254         atomic_inc(&conf->nr_queued[idx]);
255         spin_unlock_irqrestore(&conf->device_lock, flags);
256
257         wake_up(&conf->wait_barrier);
258         md_wakeup_thread(mddev->thread);
259 }
260
261 /*
262  * raid_end_bio_io() is called when we have finished servicing a mirrored
263  * operation and are ready to return a success/failure code to the buffer
264  * cache layer.
265  */
266 static void call_bio_endio(struct r1bio *r1_bio)
267 {
268         struct bio *bio = r1_bio->master_bio;
269         struct r1conf *conf = r1_bio->mddev->private;
270
271         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
272                 bio->bi_status = BLK_STS_IOERR;
273
274         bio_endio(bio);
275         /*
276          * Wake up any possible resync thread that waits for the device
277          * to go idle.
278          */
279         allow_barrier(conf, r1_bio->sector);
280 }
281
282 static void raid_end_bio_io(struct r1bio *r1_bio)
283 {
284         struct bio *bio = r1_bio->master_bio;
285
286         /* if nobody has done the final endio yet, do it now */
287         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
288                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
289                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
290                          (unsigned long long) bio->bi_iter.bi_sector,
291                          (unsigned long long) bio_end_sector(bio) - 1);
292
293                 call_bio_endio(r1_bio);
294         }
295         free_r1bio(r1_bio);
296 }
297
298 /*
299  * Update disk head position estimator based on IRQ completion info.
300  */
301 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
302 {
303         struct r1conf *conf = r1_bio->mddev->private;
304
305         conf->mirrors[disk].head_position =
306                 r1_bio->sector + (r1_bio->sectors);
307 }
308
309 /*
310  * Find the disk number which triggered given bio
311  */
312 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
313 {
314         int mirror;
315         struct r1conf *conf = r1_bio->mddev->private;
316         int raid_disks = conf->raid_disks;
317
318         for (mirror = 0; mirror < raid_disks * 2; mirror++)
319                 if (r1_bio->bios[mirror] == bio)
320                         break;
321
322         BUG_ON(mirror == raid_disks * 2);
323         update_head_pos(mirror, r1_bio);
324
325         return mirror;
326 }
327
328 static void raid1_end_read_request(struct bio *bio)
329 {
330         int uptodate = !bio->bi_status;
331         struct r1bio *r1_bio = bio->bi_private;
332         struct r1conf *conf = r1_bio->mddev->private;
333         struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
334
335         /*
336          * this branch is our 'one mirror IO has finished' event handler:
337          */
338         update_head_pos(r1_bio->read_disk, r1_bio);
339
340         if (uptodate)
341                 set_bit(R1BIO_Uptodate, &r1_bio->state);
342         else if (test_bit(FailFast, &rdev->flags) &&
343                  test_bit(R1BIO_FailFast, &r1_bio->state))
344                 /* This was a fail-fast read so we definitely
345                  * want to retry */
346                 ;
347         else {
348                 /* If all other devices have failed, we want to return
349                  * the error upwards rather than fail the last device.
350                  * Here we redefine "uptodate" to mean "Don't want to retry"
351                  */
352                 unsigned long flags;
353                 spin_lock_irqsave(&conf->device_lock, flags);
354                 if (r1_bio->mddev->degraded == conf->raid_disks ||
355                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
356                      test_bit(In_sync, &rdev->flags)))
357                         uptodate = 1;
358                 spin_unlock_irqrestore(&conf->device_lock, flags);
359         }
360
361         if (uptodate) {
362                 raid_end_bio_io(r1_bio);
363                 rdev_dec_pending(rdev, conf->mddev);
364         } else {
365                 /*
366                  * oops, read error:
367                  */
368                 char b[BDEVNAME_SIZE];
369                 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
370                                    mdname(conf->mddev),
371                                    bdevname(rdev->bdev, b),
372                                    (unsigned long long)r1_bio->sector);
373                 set_bit(R1BIO_ReadError, &r1_bio->state);
374                 reschedule_retry(r1_bio);
375                 /* don't drop the reference on read_disk yet */
376         }
377 }
378
379 static void close_write(struct r1bio *r1_bio)
380 {
381         /* it really is the end of this request */
382         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383                 bio_free_pages(r1_bio->behind_master_bio);
384                 bio_put(r1_bio->behind_master_bio);
385                 r1_bio->behind_master_bio = NULL;
386         }
387         /* clear the bitmap if all writes complete successfully */
388         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
389                         r1_bio->sectors,
390                         !test_bit(R1BIO_Degraded, &r1_bio->state),
391                         test_bit(R1BIO_BehindIO, &r1_bio->state));
392         md_write_end(r1_bio->mddev);
393 }
394
395 static void r1_bio_write_done(struct r1bio *r1_bio)
396 {
397         if (!atomic_dec_and_test(&r1_bio->remaining))
398                 return;
399
400         if (test_bit(R1BIO_WriteError, &r1_bio->state))
401                 reschedule_retry(r1_bio);
402         else {
403                 close_write(r1_bio);
404                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
405                         reschedule_retry(r1_bio);
406                 else
407                         raid_end_bio_io(r1_bio);
408         }
409 }
410
411 static void raid1_end_write_request(struct bio *bio)
412 {
413         struct r1bio *r1_bio = bio->bi_private;
414         int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
415         struct r1conf *conf = r1_bio->mddev->private;
416         struct bio *to_put = NULL;
417         int mirror = find_bio_disk(r1_bio, bio);
418         struct md_rdev *rdev = conf->mirrors[mirror].rdev;
419         bool discard_error;
420
421         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
422
423         /*
424          * 'one mirror IO has finished' event handler:
425          */
426         if (bio->bi_status && !discard_error) {
427                 set_bit(WriteErrorSeen, &rdev->flags);
428                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
429                         set_bit(MD_RECOVERY_NEEDED, &
430                                 conf->mddev->recovery);
431
432                 if (test_bit(FailFast, &rdev->flags) &&
433                     (bio->bi_opf & MD_FAILFAST) &&
434                     /* We never try FailFast to WriteMostly devices */
435                     !test_bit(WriteMostly, &rdev->flags)) {
436                         md_error(r1_bio->mddev, rdev);
437                         if (!test_bit(Faulty, &rdev->flags))
438                                 /* This is the only remaining device,
439                                  * We need to retry the write without
440                                  * FailFast
441                                  */
442                                 set_bit(R1BIO_WriteError, &r1_bio->state);
443                         else {
444                                 /* Finished with this branch */
445                                 r1_bio->bios[mirror] = NULL;
446                                 to_put = bio;
447                         }
448                 } else
449                         set_bit(R1BIO_WriteError, &r1_bio->state);
450         } else {
451                 /*
452                  * Set R1BIO_Uptodate in our master bio, so that we
453                  * will return a good error code for to the higher
454                  * levels even if IO on some other mirrored buffer
455                  * fails.
456                  *
457                  * The 'master' represents the composite IO operation
458                  * to user-side. So if something waits for IO, then it
459                  * will wait for the 'master' bio.
460                  */
461                 sector_t first_bad;
462                 int bad_sectors;
463
464                 r1_bio->bios[mirror] = NULL;
465                 to_put = bio;
466                 /*
467                  * Do not set R1BIO_Uptodate if the current device is
468                  * rebuilding or Faulty. This is because we cannot use
469                  * such device for properly reading the data back (we could
470                  * potentially use it, if the current write would have felt
471                  * before rdev->recovery_offset, but for simplicity we don't
472                  * check this here.
473                  */
474                 if (test_bit(In_sync, &rdev->flags) &&
475                     !test_bit(Faulty, &rdev->flags))
476                         set_bit(R1BIO_Uptodate, &r1_bio->state);
477
478                 /* Maybe we can clear some bad blocks. */
479                 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
480                                 &first_bad, &bad_sectors) && !discard_error) {
481                         r1_bio->bios[mirror] = IO_MADE_GOOD;
482                         set_bit(R1BIO_MadeGood, &r1_bio->state);
483                 }
484         }
485
486         if (behind) {
487                 if (test_bit(WriteMostly, &rdev->flags))
488                         atomic_dec(&r1_bio->behind_remaining);
489
490                 /*
491                  * In behind mode, we ACK the master bio once the I/O
492                  * has safely reached all non-writemostly
493                  * disks. Setting the Returned bit ensures that this
494                  * gets done only once -- we don't ever want to return
495                  * -EIO here, instead we'll wait
496                  */
497                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
498                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
499                         /* Maybe we can return now */
500                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
501                                 struct bio *mbio = r1_bio->master_bio;
502                                 pr_debug("raid1: behind end write sectors"
503                                          " %llu-%llu\n",
504                                          (unsigned long long) mbio->bi_iter.bi_sector,
505                                          (unsigned long long) bio_end_sector(mbio) - 1);
506                                 call_bio_endio(r1_bio);
507                         }
508                 }
509         }
510         if (r1_bio->bios[mirror] == NULL)
511                 rdev_dec_pending(rdev, conf->mddev);
512
513         /*
514          * Let's see if all mirrored write operations have finished
515          * already.
516          */
517         r1_bio_write_done(r1_bio);
518
519         if (to_put)
520                 bio_put(to_put);
521 }
522
523 static sector_t align_to_barrier_unit_end(sector_t start_sector,
524                                           sector_t sectors)
525 {
526         sector_t len;
527
528         WARN_ON(sectors == 0);
529         /*
530          * len is the number of sectors from start_sector to end of the
531          * barrier unit which start_sector belongs to.
532          */
533         len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
534               start_sector;
535
536         if (len > sectors)
537                 len = sectors;
538
539         return len;
540 }
541
542 /*
543  * This routine returns the disk from which the requested read should
544  * be done. There is a per-array 'next expected sequential IO' sector
545  * number - if this matches on the next IO then we use the last disk.
546  * There is also a per-disk 'last know head position' sector that is
547  * maintained from IRQ contexts, both the normal and the resync IO
548  * completion handlers update this position correctly. If there is no
549  * perfect sequential match then we pick the disk whose head is closest.
550  *
551  * If there are 2 mirrors in the same 2 devices, performance degrades
552  * because position is mirror, not device based.
553  *
554  * The rdev for the device selected will have nr_pending incremented.
555  */
556 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
557 {
558         const sector_t this_sector = r1_bio->sector;
559         int sectors;
560         int best_good_sectors;
561         int best_disk, best_dist_disk, best_pending_disk;
562         int has_nonrot_disk;
563         int disk;
564         sector_t best_dist;
565         unsigned int min_pending;
566         struct md_rdev *rdev;
567         int choose_first;
568         int choose_next_idle;
569
570         rcu_read_lock();
571         /*
572          * Check if we can balance. We can balance on the whole
573          * device if no resync is going on, or below the resync window.
574          * We take the first readable disk when above the resync window.
575          */
576  retry:
577         sectors = r1_bio->sectors;
578         best_disk = -1;
579         best_dist_disk = -1;
580         best_dist = MaxSector;
581         best_pending_disk = -1;
582         min_pending = UINT_MAX;
583         best_good_sectors = 0;
584         has_nonrot_disk = 0;
585         choose_next_idle = 0;
586         clear_bit(R1BIO_FailFast, &r1_bio->state);
587
588         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
589             (mddev_is_clustered(conf->mddev) &&
590             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
591                     this_sector + sectors)))
592                 choose_first = 1;
593         else
594                 choose_first = 0;
595
596         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
597                 sector_t dist;
598                 sector_t first_bad;
599                 int bad_sectors;
600                 unsigned int pending;
601                 bool nonrot;
602
603                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
604                 if (r1_bio->bios[disk] == IO_BLOCKED
605                     || rdev == NULL
606                     || test_bit(Faulty, &rdev->flags))
607                         continue;
608                 if (!test_bit(In_sync, &rdev->flags) &&
609                     rdev->recovery_offset < this_sector + sectors)
610                         continue;
611                 if (test_bit(WriteMostly, &rdev->flags)) {
612                         /* Don't balance among write-mostly, just
613                          * use the first as a last resort */
614                         if (best_dist_disk < 0) {
615                                 if (is_badblock(rdev, this_sector, sectors,
616                                                 &first_bad, &bad_sectors)) {
617                                         if (first_bad <= this_sector)
618                                                 /* Cannot use this */
619                                                 continue;
620                                         best_good_sectors = first_bad - this_sector;
621                                 } else
622                                         best_good_sectors = sectors;
623                                 best_dist_disk = disk;
624                                 best_pending_disk = disk;
625                         }
626                         continue;
627                 }
628                 /* This is a reasonable device to use.  It might
629                  * even be best.
630                  */
631                 if (is_badblock(rdev, this_sector, sectors,
632                                 &first_bad, &bad_sectors)) {
633                         if (best_dist < MaxSector)
634                                 /* already have a better device */
635                                 continue;
636                         if (first_bad <= this_sector) {
637                                 /* cannot read here. If this is the 'primary'
638                                  * device, then we must not read beyond
639                                  * bad_sectors from another device..
640                                  */
641                                 bad_sectors -= (this_sector - first_bad);
642                                 if (choose_first && sectors > bad_sectors)
643                                         sectors = bad_sectors;
644                                 if (best_good_sectors > sectors)
645                                         best_good_sectors = sectors;
646
647                         } else {
648                                 sector_t good_sectors = first_bad - this_sector;
649                                 if (good_sectors > best_good_sectors) {
650                                         best_good_sectors = good_sectors;
651                                         best_disk = disk;
652                                 }
653                                 if (choose_first)
654                                         break;
655                         }
656                         continue;
657                 } else {
658                         if ((sectors > best_good_sectors) && (best_disk >= 0))
659                                 best_disk = -1;
660                         best_good_sectors = sectors;
661                 }
662
663                 if (best_disk >= 0)
664                         /* At least two disks to choose from so failfast is OK */
665                         set_bit(R1BIO_FailFast, &r1_bio->state);
666
667                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
668                 has_nonrot_disk |= nonrot;
669                 pending = atomic_read(&rdev->nr_pending);
670                 dist = abs(this_sector - conf->mirrors[disk].head_position);
671                 if (choose_first) {
672                         best_disk = disk;
673                         break;
674                 }
675                 /* Don't change to another disk for sequential reads */
676                 if (conf->mirrors[disk].next_seq_sect == this_sector
677                     || dist == 0) {
678                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
679                         struct raid1_info *mirror = &conf->mirrors[disk];
680
681                         best_disk = disk;
682                         /*
683                          * If buffered sequential IO size exceeds optimal
684                          * iosize, check if there is idle disk. If yes, choose
685                          * the idle disk. read_balance could already choose an
686                          * idle disk before noticing it's a sequential IO in
687                          * this disk. This doesn't matter because this disk
688                          * will idle, next time it will be utilized after the
689                          * first disk has IO size exceeds optimal iosize. In
690                          * this way, iosize of the first disk will be optimal
691                          * iosize at least. iosize of the second disk might be
692                          * small, but not a big deal since when the second disk
693                          * starts IO, the first disk is likely still busy.
694                          */
695                         if (nonrot && opt_iosize > 0 &&
696                             mirror->seq_start != MaxSector &&
697                             mirror->next_seq_sect > opt_iosize &&
698                             mirror->next_seq_sect - opt_iosize >=
699                             mirror->seq_start) {
700                                 choose_next_idle = 1;
701                                 continue;
702                         }
703                         break;
704                 }
705
706                 if (choose_next_idle)
707                         continue;
708
709                 if (min_pending > pending) {
710                         min_pending = pending;
711                         best_pending_disk = disk;
712                 }
713
714                 if (dist < best_dist) {
715                         best_dist = dist;
716                         best_dist_disk = disk;
717                 }
718         }
719
720         /*
721          * If all disks are rotational, choose the closest disk. If any disk is
722          * non-rotational, choose the disk with less pending request even the
723          * disk is rotational, which might/might not be optimal for raids with
724          * mixed ratation/non-rotational disks depending on workload.
725          */
726         if (best_disk == -1) {
727                 if (has_nonrot_disk || min_pending == 0)
728                         best_disk = best_pending_disk;
729                 else
730                         best_disk = best_dist_disk;
731         }
732
733         if (best_disk >= 0) {
734                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
735                 if (!rdev)
736                         goto retry;
737                 atomic_inc(&rdev->nr_pending);
738                 sectors = best_good_sectors;
739
740                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
741                         conf->mirrors[best_disk].seq_start = this_sector;
742
743                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
744         }
745         rcu_read_unlock();
746         *max_sectors = sectors;
747
748         return best_disk;
749 }
750
751 static int raid1_congested(struct mddev *mddev, int bits)
752 {
753         struct r1conf *conf = mddev->private;
754         int i, ret = 0;
755
756         if ((bits & (1 << WB_async_congested)) &&
757             conf->pending_count >= max_queued_requests)
758                 return 1;
759
760         rcu_read_lock();
761         for (i = 0; i < conf->raid_disks * 2; i++) {
762                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
763                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
764                         struct request_queue *q = bdev_get_queue(rdev->bdev);
765
766                         BUG_ON(!q);
767
768                         /* Note the '|| 1' - when read_balance prefers
769                          * non-congested targets, it can be removed
770                          */
771                         if ((bits & (1 << WB_async_congested)) || 1)
772                                 ret |= bdi_congested(q->backing_dev_info, bits);
773                         else
774                                 ret &= bdi_congested(q->backing_dev_info, bits);
775                 }
776         }
777         rcu_read_unlock();
778         return ret;
779 }
780
781 static void flush_bio_list(struct r1conf *conf, struct bio *bio)
782 {
783         /* flush any pending bitmap writes to disk before proceeding w/ I/O */
784         bitmap_unplug(conf->mddev->bitmap);
785         wake_up(&conf->wait_barrier);
786
787         while (bio) { /* submit pending writes */
788                 struct bio *next = bio->bi_next;
789                 struct md_rdev *rdev = (void *)bio->bi_disk;
790                 bio->bi_next = NULL;
791                 bio_set_dev(bio, rdev->bdev);
792                 if (test_bit(Faulty, &rdev->flags)) {
793                         bio_io_error(bio);
794                 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
795                                     !blk_queue_discard(bio->bi_disk->queue)))
796                         /* Just ignore it */
797                         bio_endio(bio);
798                 else
799                         generic_make_request(bio);
800                 bio = next;
801         }
802 }
803
804 static void flush_pending_writes(struct r1conf *conf)
805 {
806         /* Any writes that have been queued but are awaiting
807          * bitmap updates get flushed here.
808          */
809         spin_lock_irq(&conf->device_lock);
810
811         if (conf->pending_bio_list.head) {
812                 struct blk_plug plug;
813                 struct bio *bio;
814
815                 bio = bio_list_get(&conf->pending_bio_list);
816                 conf->pending_count = 0;
817                 spin_unlock_irq(&conf->device_lock);
818
819                 /*
820                  * As this is called in a wait_event() loop (see freeze_array),
821                  * current->state might be TASK_UNINTERRUPTIBLE which will
822                  * cause a warning when we prepare to wait again.  As it is
823                  * rare that this path is taken, it is perfectly safe to force
824                  * us to go around the wait_event() loop again, so the warning
825                  * is a false-positive.  Silence the warning by resetting
826                  * thread state
827                  */
828                 __set_current_state(TASK_RUNNING);
829                 blk_start_plug(&plug);
830                 flush_bio_list(conf, bio);
831                 blk_finish_plug(&plug);
832         } else
833                 spin_unlock_irq(&conf->device_lock);
834 }
835
836 /* Barriers....
837  * Sometimes we need to suspend IO while we do something else,
838  * either some resync/recovery, or reconfigure the array.
839  * To do this we raise a 'barrier'.
840  * The 'barrier' is a counter that can be raised multiple times
841  * to count how many activities are happening which preclude
842  * normal IO.
843  * We can only raise the barrier if there is no pending IO.
844  * i.e. if nr_pending == 0.
845  * We choose only to raise the barrier if no-one is waiting for the
846  * barrier to go down.  This means that as soon as an IO request
847  * is ready, no other operations which require a barrier will start
848  * until the IO request has had a chance.
849  *
850  * So: regular IO calls 'wait_barrier'.  When that returns there
851  *    is no backgroup IO happening,  It must arrange to call
852  *    allow_barrier when it has finished its IO.
853  * backgroup IO calls must call raise_barrier.  Once that returns
854  *    there is no normal IO happeing.  It must arrange to call
855  *    lower_barrier when the particular background IO completes.
856  */
857 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
858 {
859         int idx = sector_to_idx(sector_nr);
860
861         spin_lock_irq(&conf->resync_lock);
862
863         /* Wait until no block IO is waiting */
864         wait_event_lock_irq(conf->wait_barrier,
865                             !atomic_read(&conf->nr_waiting[idx]),
866                             conf->resync_lock);
867
868         /* block any new IO from starting */
869         atomic_inc(&conf->barrier[idx]);
870         /*
871          * In raise_barrier() we firstly increase conf->barrier[idx] then
872          * check conf->nr_pending[idx]. In _wait_barrier() we firstly
873          * increase conf->nr_pending[idx] then check conf->barrier[idx].
874          * A memory barrier here to make sure conf->nr_pending[idx] won't
875          * be fetched before conf->barrier[idx] is increased. Otherwise
876          * there will be a race between raise_barrier() and _wait_barrier().
877          */
878         smp_mb__after_atomic();
879
880         /* For these conditions we must wait:
881          * A: while the array is in frozen state
882          * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
883          *    existing in corresponding I/O barrier bucket.
884          * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
885          *    max resync count which allowed on current I/O barrier bucket.
886          */
887         wait_event_lock_irq(conf->wait_barrier,
888                             !conf->array_frozen &&
889                              !atomic_read(&conf->nr_pending[idx]) &&
890                              atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH,
891                             conf->resync_lock);
892
893         atomic_inc(&conf->nr_sync_pending);
894         spin_unlock_irq(&conf->resync_lock);
895 }
896
897 static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
898 {
899         int idx = sector_to_idx(sector_nr);
900
901         BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
902
903         atomic_dec(&conf->barrier[idx]);
904         atomic_dec(&conf->nr_sync_pending);
905         wake_up(&conf->wait_barrier);
906 }
907
908 static void _wait_barrier(struct r1conf *conf, int idx)
909 {
910         /*
911          * We need to increase conf->nr_pending[idx] very early here,
912          * then raise_barrier() can be blocked when it waits for
913          * conf->nr_pending[idx] to be 0. Then we can avoid holding
914          * conf->resync_lock when there is no barrier raised in same
915          * barrier unit bucket. Also if the array is frozen, I/O
916          * should be blocked until array is unfrozen.
917          */
918         atomic_inc(&conf->nr_pending[idx]);
919         /*
920          * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
921          * check conf->barrier[idx]. In raise_barrier() we firstly increase
922          * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
923          * barrier is necessary here to make sure conf->barrier[idx] won't be
924          * fetched before conf->nr_pending[idx] is increased. Otherwise there
925          * will be a race between _wait_barrier() and raise_barrier().
926          */
927         smp_mb__after_atomic();
928
929         /*
930          * Don't worry about checking two atomic_t variables at same time
931          * here. If during we check conf->barrier[idx], the array is
932          * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
933          * 0, it is safe to return and make the I/O continue. Because the
934          * array is frozen, all I/O returned here will eventually complete
935          * or be queued, no race will happen. See code comment in
936          * frozen_array().
937          */
938         if (!READ_ONCE(conf->array_frozen) &&
939             !atomic_read(&conf->barrier[idx]))
940                 return;
941
942         /*
943          * After holding conf->resync_lock, conf->nr_pending[idx]
944          * should be decreased before waiting for barrier to drop.
945          * Otherwise, we may encounter a race condition because
946          * raise_barrer() might be waiting for conf->nr_pending[idx]
947          * to be 0 at same time.
948          */
949         spin_lock_irq(&conf->resync_lock);
950         atomic_inc(&conf->nr_waiting[idx]);
951         atomic_dec(&conf->nr_pending[idx]);
952         /*
953          * In case freeze_array() is waiting for
954          * get_unqueued_pending() == extra
955          */
956         wake_up(&conf->wait_barrier);
957         /* Wait for the barrier in same barrier unit bucket to drop. */
958         wait_event_lock_irq(conf->wait_barrier,
959                             !conf->array_frozen &&
960                              !atomic_read(&conf->barrier[idx]),
961                             conf->resync_lock);
962         atomic_inc(&conf->nr_pending[idx]);
963         atomic_dec(&conf->nr_waiting[idx]);
964         spin_unlock_irq(&conf->resync_lock);
965 }
966
967 static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
968 {
969         int idx = sector_to_idx(sector_nr);
970
971         /*
972          * Very similar to _wait_barrier(). The difference is, for read
973          * I/O we don't need wait for sync I/O, but if the whole array
974          * is frozen, the read I/O still has to wait until the array is
975          * unfrozen. Since there is no ordering requirement with
976          * conf->barrier[idx] here, memory barrier is unnecessary as well.
977          */
978         atomic_inc(&conf->nr_pending[idx]);
979
980         if (!READ_ONCE(conf->array_frozen))
981                 return;
982
983         spin_lock_irq(&conf->resync_lock);
984         atomic_inc(&conf->nr_waiting[idx]);
985         atomic_dec(&conf->nr_pending[idx]);
986         /*
987          * In case freeze_array() is waiting for
988          * get_unqueued_pending() == extra
989          */
990         wake_up(&conf->wait_barrier);
991         /* Wait for array to be unfrozen */
992         wait_event_lock_irq(conf->wait_barrier,
993                             !conf->array_frozen,
994                             conf->resync_lock);
995         atomic_inc(&conf->nr_pending[idx]);
996         atomic_dec(&conf->nr_waiting[idx]);
997         spin_unlock_irq(&conf->resync_lock);
998 }
999
1000 static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1001 {
1002         int idx = sector_to_idx(sector_nr);
1003
1004         _wait_barrier(conf, idx);
1005 }
1006
1007 static void _allow_barrier(struct r1conf *conf, int idx)
1008 {
1009         atomic_dec(&conf->nr_pending[idx]);
1010         wake_up(&conf->wait_barrier);
1011 }
1012
1013 static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1014 {
1015         int idx = sector_to_idx(sector_nr);
1016
1017         _allow_barrier(conf, idx);
1018 }
1019
1020 /* conf->resync_lock should be held */
1021 static int get_unqueued_pending(struct r1conf *conf)
1022 {
1023         int idx, ret;
1024
1025         ret = atomic_read(&conf->nr_sync_pending);
1026         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1027                 ret += atomic_read(&conf->nr_pending[idx]) -
1028                         atomic_read(&conf->nr_queued[idx]);
1029
1030         return ret;
1031 }
1032
1033 static void freeze_array(struct r1conf *conf, int extra)
1034 {
1035         /* Stop sync I/O and normal I/O and wait for everything to
1036          * go quiet.
1037          * This is called in two situations:
1038          * 1) management command handlers (reshape, remove disk, quiesce).
1039          * 2) one normal I/O request failed.
1040
1041          * After array_frozen is set to 1, new sync IO will be blocked at
1042          * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1043          * or wait_read_barrier(). The flying I/Os will either complete or be
1044          * queued. When everything goes quite, there are only queued I/Os left.
1045
1046          * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1047          * barrier bucket index which this I/O request hits. When all sync and
1048          * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1049          * of all conf->nr_queued[]. But normal I/O failure is an exception,
1050          * in handle_read_error(), we may call freeze_array() before trying to
1051          * fix the read error. In this case, the error read I/O is not queued,
1052          * so get_unqueued_pending() == 1.
1053          *
1054          * Therefore before this function returns, we need to wait until
1055          * get_unqueued_pendings(conf) gets equal to extra. For
1056          * normal I/O context, extra is 1, in rested situations extra is 0.
1057          */
1058         spin_lock_irq(&conf->resync_lock);
1059         conf->array_frozen = 1;
1060         raid1_log(conf->mddev, "wait freeze");
1061         wait_event_lock_irq_cmd(
1062                 conf->wait_barrier,
1063                 get_unqueued_pending(conf) == extra,
1064                 conf->resync_lock,
1065                 flush_pending_writes(conf));
1066         spin_unlock_irq(&conf->resync_lock);
1067 }
1068 static void unfreeze_array(struct r1conf *conf)
1069 {
1070         /* reverse the effect of the freeze */
1071         spin_lock_irq(&conf->resync_lock);
1072         conf->array_frozen = 0;
1073         spin_unlock_irq(&conf->resync_lock);
1074         wake_up(&conf->wait_barrier);
1075 }
1076
1077 static void alloc_behind_master_bio(struct r1bio *r1_bio,
1078                                            struct bio *bio)
1079 {
1080         int size = bio->bi_iter.bi_size;
1081         unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1082         int i = 0;
1083         struct bio *behind_bio = NULL;
1084
1085         behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1086         if (!behind_bio)
1087                 return;
1088
1089         /* discard op, we don't support writezero/writesame yet */
1090         if (!bio_has_data(bio)) {
1091                 behind_bio->bi_iter.bi_size = size;
1092                 goto skip_copy;
1093         }
1094
1095         while (i < vcnt && size) {
1096                 struct page *page;
1097                 int len = min_t(int, PAGE_SIZE, size);
1098
1099                 page = alloc_page(GFP_NOIO);
1100                 if (unlikely(!page))
1101                         goto free_pages;
1102
1103                 bio_add_page(behind_bio, page, len, 0);
1104
1105                 size -= len;
1106                 i++;
1107         }
1108
1109         bio_copy_data(behind_bio, bio);
1110 skip_copy:
1111         r1_bio->behind_master_bio = behind_bio;;
1112         set_bit(R1BIO_BehindIO, &r1_bio->state);
1113
1114         return;
1115
1116 free_pages:
1117         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1118                  bio->bi_iter.bi_size);
1119         bio_free_pages(behind_bio);
1120         bio_put(behind_bio);
1121 }
1122
1123 struct raid1_plug_cb {
1124         struct blk_plug_cb      cb;
1125         struct bio_list         pending;
1126         int                     pending_cnt;
1127 };
1128
1129 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1130 {
1131         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1132                                                   cb);
1133         struct mddev *mddev = plug->cb.data;
1134         struct r1conf *conf = mddev->private;
1135         struct bio *bio;
1136
1137         if (from_schedule || current->bio_list) {
1138                 spin_lock_irq(&conf->device_lock);
1139                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1140                 conf->pending_count += plug->pending_cnt;
1141                 spin_unlock_irq(&conf->device_lock);
1142                 wake_up(&conf->wait_barrier);
1143                 md_wakeup_thread(mddev->thread);
1144                 kfree(plug);
1145                 return;
1146         }
1147
1148         /* we aren't scheduling, so we can do the write-out directly. */
1149         bio = bio_list_get(&plug->pending);
1150         flush_bio_list(conf, bio);
1151         kfree(plug);
1152 }
1153
1154 static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1155 {
1156         r1_bio->master_bio = bio;
1157         r1_bio->sectors = bio_sectors(bio);
1158         r1_bio->state = 0;
1159         r1_bio->mddev = mddev;
1160         r1_bio->sector = bio->bi_iter.bi_sector;
1161 }
1162
1163 static inline struct r1bio *
1164 alloc_r1bio(struct mddev *mddev, struct bio *bio)
1165 {
1166         struct r1conf *conf = mddev->private;
1167         struct r1bio *r1_bio;
1168
1169         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1170         /* Ensure no bio records IO_BLOCKED */
1171         memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1172         init_r1bio(r1_bio, mddev, bio);
1173         return r1_bio;
1174 }
1175
1176 static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1177                                int max_read_sectors, struct r1bio *r1_bio)
1178 {
1179         struct r1conf *conf = mddev->private;
1180         struct raid1_info *mirror;
1181         struct bio *read_bio;
1182         struct bitmap *bitmap = mddev->bitmap;
1183         const int op = bio_op(bio);
1184         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1185         int max_sectors;
1186         int rdisk;
1187         bool print_msg = !!r1_bio;
1188         char b[BDEVNAME_SIZE];
1189
1190         /*
1191          * If r1_bio is set, we are blocking the raid1d thread
1192          * so there is a tiny risk of deadlock.  So ask for
1193          * emergency memory if needed.
1194          */
1195         gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1196
1197         if (print_msg) {
1198                 /* Need to get the block device name carefully */
1199                 struct md_rdev *rdev;
1200                 rcu_read_lock();
1201                 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1202                 if (rdev)
1203                         bdevname(rdev->bdev, b);
1204                 else
1205                         strcpy(b, "???");
1206                 rcu_read_unlock();
1207         }
1208
1209         /*
1210          * Still need barrier for READ in case that whole
1211          * array is frozen.
1212          */
1213         wait_read_barrier(conf, bio->bi_iter.bi_sector);
1214
1215         if (!r1_bio)
1216                 r1_bio = alloc_r1bio(mddev, bio);
1217         else
1218                 init_r1bio(r1_bio, mddev, bio);
1219         r1_bio->sectors = max_read_sectors;
1220
1221         /*
1222          * make_request() can abort the operation when read-ahead is being
1223          * used and no empty request is available.
1224          */
1225         rdisk = read_balance(conf, r1_bio, &max_sectors);
1226
1227         if (rdisk < 0) {
1228                 /* couldn't find anywhere to read from */
1229                 if (print_msg) {
1230                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1231                                             mdname(mddev),
1232                                             b,
1233                                             (unsigned long long)r1_bio->sector);
1234                 }
1235                 raid_end_bio_io(r1_bio);
1236                 return;
1237         }
1238         mirror = conf->mirrors + rdisk;
1239
1240         if (print_msg)
1241                 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1242                                     mdname(mddev),
1243                                     (unsigned long long)r1_bio->sector,
1244                                     bdevname(mirror->rdev->bdev, b));
1245
1246         if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1247             bitmap) {
1248                 /*
1249                  * Reading from a write-mostly device must take care not to
1250                  * over-take any writes that are 'behind'
1251                  */
1252                 raid1_log(mddev, "wait behind writes");
1253                 wait_event(bitmap->behind_wait,
1254                            atomic_read(&bitmap->behind_writes) == 0);
1255         }
1256
1257         if (max_sectors < bio_sectors(bio)) {
1258                 struct bio *split = bio_split(bio, max_sectors,
1259                                               gfp, conf->bio_split);
1260                 bio_chain(split, bio);
1261                 generic_make_request(bio);
1262                 bio = split;
1263                 r1_bio->master_bio = bio;
1264                 r1_bio->sectors = max_sectors;
1265         }
1266
1267         r1_bio->read_disk = rdisk;
1268
1269         read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1270
1271         r1_bio->bios[rdisk] = read_bio;
1272
1273         read_bio->bi_iter.bi_sector = r1_bio->sector +
1274                 mirror->rdev->data_offset;
1275         bio_set_dev(read_bio, mirror->rdev->bdev);
1276         read_bio->bi_end_io = raid1_end_read_request;
1277         bio_set_op_attrs(read_bio, op, do_sync);
1278         if (test_bit(FailFast, &mirror->rdev->flags) &&
1279             test_bit(R1BIO_FailFast, &r1_bio->state))
1280                 read_bio->bi_opf |= MD_FAILFAST;
1281         read_bio->bi_private = r1_bio;
1282
1283         if (mddev->gendisk)
1284                 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1285                                 disk_devt(mddev->gendisk), r1_bio->sector);
1286
1287         generic_make_request(read_bio);
1288 }
1289
1290 static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1291                                 int max_write_sectors)
1292 {
1293         struct r1conf *conf = mddev->private;
1294         struct r1bio *r1_bio;
1295         int i, disks;
1296         struct bitmap *bitmap = mddev->bitmap;
1297         unsigned long flags;
1298         struct md_rdev *blocked_rdev;
1299         struct blk_plug_cb *cb;
1300         struct raid1_plug_cb *plug = NULL;
1301         int first_clone;
1302         int max_sectors;
1303
1304         if (mddev_is_clustered(mddev) &&
1305              md_cluster_ops->area_resyncing(mddev, WRITE,
1306                      bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1307
1308                 DEFINE_WAIT(w);
1309                 for (;;) {
1310                         prepare_to_wait(&conf->wait_barrier,
1311                                         &w, TASK_IDLE);
1312                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1313                                                         bio->bi_iter.bi_sector,
1314                                                         bio_end_sector(bio)))
1315                                 break;
1316                         schedule();
1317                 }
1318                 finish_wait(&conf->wait_barrier, &w);
1319         }
1320
1321         /*
1322          * Register the new request and wait if the reconstruction
1323          * thread has put up a bar for new requests.
1324          * Continue immediately if no resync is active currently.
1325          */
1326         wait_barrier(conf, bio->bi_iter.bi_sector);
1327
1328         r1_bio = alloc_r1bio(mddev, bio);
1329         r1_bio->sectors = max_write_sectors;
1330
1331         if (conf->pending_count >= max_queued_requests) {
1332                 md_wakeup_thread(mddev->thread);
1333                 raid1_log(mddev, "wait queued");
1334                 wait_event(conf->wait_barrier,
1335                            conf->pending_count < max_queued_requests);
1336         }
1337         /* first select target devices under rcu_lock and
1338          * inc refcount on their rdev.  Record them by setting
1339          * bios[x] to bio
1340          * If there are known/acknowledged bad blocks on any device on
1341          * which we have seen a write error, we want to avoid writing those
1342          * blocks.
1343          * This potentially requires several writes to write around
1344          * the bad blocks.  Each set of writes gets it's own r1bio
1345          * with a set of bios attached.
1346          */
1347
1348         disks = conf->raid_disks * 2;
1349  retry_write:
1350         blocked_rdev = NULL;
1351         rcu_read_lock();
1352         max_sectors = r1_bio->sectors;
1353         for (i = 0;  i < disks; i++) {
1354                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1355                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1356                         atomic_inc(&rdev->nr_pending);
1357                         blocked_rdev = rdev;
1358                         break;
1359                 }
1360                 r1_bio->bios[i] = NULL;
1361                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1362                         if (i < conf->raid_disks)
1363                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1364                         continue;
1365                 }
1366
1367                 atomic_inc(&rdev->nr_pending);
1368                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1369                         sector_t first_bad;
1370                         int bad_sectors;
1371                         int is_bad;
1372
1373                         is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1374                                              &first_bad, &bad_sectors);
1375                         if (is_bad < 0) {
1376                                 /* mustn't write here until the bad block is
1377                                  * acknowledged*/
1378                                 set_bit(BlockedBadBlocks, &rdev->flags);
1379                                 blocked_rdev = rdev;
1380                                 break;
1381                         }
1382                         if (is_bad && first_bad <= r1_bio->sector) {
1383                                 /* Cannot write here at all */
1384                                 bad_sectors -= (r1_bio->sector - first_bad);
1385                                 if (bad_sectors < max_sectors)
1386                                         /* mustn't write more than bad_sectors
1387                                          * to other devices yet
1388                                          */
1389                                         max_sectors = bad_sectors;
1390                                 rdev_dec_pending(rdev, mddev);
1391                                 /* We don't set R1BIO_Degraded as that
1392                                  * only applies if the disk is
1393                                  * missing, so it might be re-added,
1394                                  * and we want to know to recover this
1395                                  * chunk.
1396                                  * In this case the device is here,
1397                                  * and the fact that this chunk is not
1398                                  * in-sync is recorded in the bad
1399                                  * block log
1400                                  */
1401                                 continue;
1402                         }
1403                         if (is_bad) {
1404                                 int good_sectors = first_bad - r1_bio->sector;
1405                                 if (good_sectors < max_sectors)
1406                                         max_sectors = good_sectors;
1407                         }
1408                 }
1409                 r1_bio->bios[i] = bio;
1410         }
1411         rcu_read_unlock();
1412
1413         if (unlikely(blocked_rdev)) {
1414                 /* Wait for this device to become unblocked */
1415                 int j;
1416
1417                 for (j = 0; j < i; j++)
1418                         if (r1_bio->bios[j])
1419                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1420                 r1_bio->state = 0;
1421                 allow_barrier(conf, bio->bi_iter.bi_sector);
1422                 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1423                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1424                 wait_barrier(conf, bio->bi_iter.bi_sector);
1425                 goto retry_write;
1426         }
1427
1428         if (max_sectors < bio_sectors(bio)) {
1429                 struct bio *split = bio_split(bio, max_sectors,
1430                                               GFP_NOIO, conf->bio_split);
1431                 bio_chain(split, bio);
1432                 generic_make_request(bio);
1433                 bio = split;
1434                 r1_bio->master_bio = bio;
1435                 r1_bio->sectors = max_sectors;
1436         }
1437
1438         atomic_set(&r1_bio->remaining, 1);
1439         atomic_set(&r1_bio->behind_remaining, 0);
1440
1441         first_clone = 1;
1442
1443         for (i = 0; i < disks; i++) {
1444                 struct bio *mbio = NULL;
1445                 if (!r1_bio->bios[i])
1446                         continue;
1447
1448
1449                 if (first_clone) {
1450                         /* do behind I/O ?
1451                          * Not if there are too many, or cannot
1452                          * allocate memory, or a reader on WriteMostly
1453                          * is waiting for behind writes to flush */
1454                         if (bitmap &&
1455                             (atomic_read(&bitmap->behind_writes)
1456                              < mddev->bitmap_info.max_write_behind) &&
1457                             !waitqueue_active(&bitmap->behind_wait)) {
1458                                 alloc_behind_master_bio(r1_bio, bio);
1459                         }
1460
1461                         bitmap_startwrite(bitmap, r1_bio->sector,
1462                                           r1_bio->sectors,
1463                                           test_bit(R1BIO_BehindIO,
1464                                                    &r1_bio->state));
1465                         first_clone = 0;
1466                 }
1467
1468                 if (r1_bio->behind_master_bio)
1469                         mbio = bio_clone_fast(r1_bio->behind_master_bio,
1470                                               GFP_NOIO, mddev->bio_set);
1471                 else
1472                         mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1473
1474                 if (r1_bio->behind_master_bio) {
1475                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1476                                 atomic_inc(&r1_bio->behind_remaining);
1477                 }
1478
1479                 r1_bio->bios[i] = mbio;
1480
1481                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1482                                    conf->mirrors[i].rdev->data_offset);
1483                 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1484                 mbio->bi_end_io = raid1_end_write_request;
1485                 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1486                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1487                     !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1488                     conf->raid_disks - mddev->degraded > 1)
1489                         mbio->bi_opf |= MD_FAILFAST;
1490                 mbio->bi_private = r1_bio;
1491
1492                 atomic_inc(&r1_bio->remaining);
1493
1494                 if (mddev->gendisk)
1495                         trace_block_bio_remap(mbio->bi_disk->queue,
1496                                               mbio, disk_devt(mddev->gendisk),
1497                                               r1_bio->sector);
1498                 /* flush_pending_writes() needs access to the rdev so...*/
1499                 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1500
1501                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1502                 if (cb)
1503                         plug = container_of(cb, struct raid1_plug_cb, cb);
1504                 else
1505                         plug = NULL;
1506                 if (plug) {
1507                         bio_list_add(&plug->pending, mbio);
1508                         plug->pending_cnt++;
1509                 } else {
1510                         spin_lock_irqsave(&conf->device_lock, flags);
1511                         bio_list_add(&conf->pending_bio_list, mbio);
1512                         conf->pending_count++;
1513                         spin_unlock_irqrestore(&conf->device_lock, flags);
1514                         md_wakeup_thread(mddev->thread);
1515                 }
1516         }
1517
1518         r1_bio_write_done(r1_bio);
1519
1520         /* In case raid1d snuck in to freeze_array */
1521         wake_up(&conf->wait_barrier);
1522 }
1523
1524 static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1525 {
1526         sector_t sectors;
1527
1528         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1529                 md_flush_request(mddev, bio);
1530                 return true;
1531         }
1532
1533         /*
1534          * There is a limit to the maximum size, but
1535          * the read/write handler might find a lower limit
1536          * due to bad blocks.  To avoid multiple splits,
1537          * we pass the maximum number of sectors down
1538          * and let the lower level perform the split.
1539          */
1540         sectors = align_to_barrier_unit_end(
1541                 bio->bi_iter.bi_sector, bio_sectors(bio));
1542
1543         if (bio_data_dir(bio) == READ)
1544                 raid1_read_request(mddev, bio, sectors, NULL);
1545         else {
1546                 if (!md_write_start(mddev,bio))
1547                         return false;
1548                 raid1_write_request(mddev, bio, sectors);
1549         }
1550         return true;
1551 }
1552
1553 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1554 {
1555         struct r1conf *conf = mddev->private;
1556         int i;
1557
1558         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1559                    conf->raid_disks - mddev->degraded);
1560         rcu_read_lock();
1561         for (i = 0; i < conf->raid_disks; i++) {
1562                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1563                 seq_printf(seq, "%s",
1564                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1565         }
1566         rcu_read_unlock();
1567         seq_printf(seq, "]");
1568 }
1569
1570 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1571 {
1572         char b[BDEVNAME_SIZE];
1573         struct r1conf *conf = mddev->private;
1574         unsigned long flags;
1575
1576         /*
1577          * If it is not operational, then we have already marked it as dead
1578          * else if it is the last working disks, ignore the error, let the
1579          * next level up know.
1580          * else mark the drive as failed
1581          */
1582         spin_lock_irqsave(&conf->device_lock, flags);
1583         if (test_bit(In_sync, &rdev->flags)
1584             && (conf->raid_disks - mddev->degraded) == 1) {
1585                 /*
1586                  * Don't fail the drive, act as though we were just a
1587                  * normal single drive.
1588                  * However don't try a recovery from this drive as
1589                  * it is very likely to fail.
1590                  */
1591                 conf->recovery_disabled = mddev->recovery_disabled;
1592                 spin_unlock_irqrestore(&conf->device_lock, flags);
1593                 return;
1594         }
1595         set_bit(Blocked, &rdev->flags);
1596         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1597                 mddev->degraded++;
1598                 set_bit(Faulty, &rdev->flags);
1599         } else
1600                 set_bit(Faulty, &rdev->flags);
1601         spin_unlock_irqrestore(&conf->device_lock, flags);
1602         /*
1603          * if recovery is running, make sure it aborts.
1604          */
1605         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1606         set_mask_bits(&mddev->sb_flags, 0,
1607                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1608         pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1609                 "md/raid1:%s: Operation continuing on %d devices.\n",
1610                 mdname(mddev), bdevname(rdev->bdev, b),
1611                 mdname(mddev), conf->raid_disks - mddev->degraded);
1612 }
1613
1614 static void print_conf(struct r1conf *conf)
1615 {
1616         int i;
1617
1618         pr_debug("RAID1 conf printout:\n");
1619         if (!conf) {
1620                 pr_debug("(!conf)\n");
1621                 return;
1622         }
1623         pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1624                  conf->raid_disks);
1625
1626         rcu_read_lock();
1627         for (i = 0; i < conf->raid_disks; i++) {
1628                 char b[BDEVNAME_SIZE];
1629                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1630                 if (rdev)
1631                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1632                                  i, !test_bit(In_sync, &rdev->flags),
1633                                  !test_bit(Faulty, &rdev->flags),
1634                                  bdevname(rdev->bdev,b));
1635         }
1636         rcu_read_unlock();
1637 }
1638
1639 static void close_sync(struct r1conf *conf)
1640 {
1641         int idx;
1642
1643         for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1644                 _wait_barrier(conf, idx);
1645                 _allow_barrier(conf, idx);
1646         }
1647
1648         mempool_destroy(conf->r1buf_pool);
1649         conf->r1buf_pool = NULL;
1650 }
1651
1652 static int raid1_spare_active(struct mddev *mddev)
1653 {
1654         int i;
1655         struct r1conf *conf = mddev->private;
1656         int count = 0;
1657         unsigned long flags;
1658
1659         /*
1660          * Find all failed disks within the RAID1 configuration
1661          * and mark them readable.
1662          * Called under mddev lock, so rcu protection not needed.
1663          * device_lock used to avoid races with raid1_end_read_request
1664          * which expects 'In_sync' flags and ->degraded to be consistent.
1665          */
1666         spin_lock_irqsave(&conf->device_lock, flags);
1667         for (i = 0; i < conf->raid_disks; i++) {
1668                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1669                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1670                 if (repl
1671                     && !test_bit(Candidate, &repl->flags)
1672                     && repl->recovery_offset == MaxSector
1673                     && !test_bit(Faulty, &repl->flags)
1674                     && !test_and_set_bit(In_sync, &repl->flags)) {
1675                         /* replacement has just become active */
1676                         if (!rdev ||
1677                             !test_and_clear_bit(In_sync, &rdev->flags))
1678                                 count++;
1679                         if (rdev) {
1680                                 /* Replaced device not technically
1681                                  * faulty, but we need to be sure
1682                                  * it gets removed and never re-added
1683                                  */
1684                                 set_bit(Faulty, &rdev->flags);
1685                                 sysfs_notify_dirent_safe(
1686                                         rdev->sysfs_state);
1687                         }
1688                 }
1689                 if (rdev
1690                     && rdev->recovery_offset == MaxSector
1691                     && !test_bit(Faulty, &rdev->flags)
1692                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1693                         count++;
1694                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1695                 }
1696         }
1697         mddev->degraded -= count;
1698         spin_unlock_irqrestore(&conf->device_lock, flags);
1699
1700         print_conf(conf);
1701         return count;
1702 }
1703
1704 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1705 {
1706         struct r1conf *conf = mddev->private;
1707         int err = -EEXIST;
1708         int mirror = 0;
1709         struct raid1_info *p;
1710         int first = 0;
1711         int last = conf->raid_disks - 1;
1712
1713         if (mddev->recovery_disabled == conf->recovery_disabled)
1714                 return -EBUSY;
1715
1716         if (md_integrity_add_rdev(rdev, mddev))
1717                 return -ENXIO;
1718
1719         if (rdev->raid_disk >= 0)
1720                 first = last = rdev->raid_disk;
1721
1722         /*
1723          * find the disk ... but prefer rdev->saved_raid_disk
1724          * if possible.
1725          */
1726         if (rdev->saved_raid_disk >= 0 &&
1727             rdev->saved_raid_disk >= first &&
1728             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1729                 first = last = rdev->saved_raid_disk;
1730
1731         for (mirror = first; mirror <= last; mirror++) {
1732                 p = conf->mirrors+mirror;
1733                 if (!p->rdev) {
1734
1735                         if (mddev->gendisk)
1736                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1737                                                   rdev->data_offset << 9);
1738
1739                         p->head_position = 0;
1740                         rdev->raid_disk = mirror;
1741                         err = 0;
1742                         /* As all devices are equivalent, we don't need a full recovery
1743                          * if this was recently any drive of the array
1744                          */
1745                         if (rdev->saved_raid_disk < 0)
1746                                 conf->fullsync = 1;
1747                         rcu_assign_pointer(p->rdev, rdev);
1748                         break;
1749                 }
1750                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1751                     p[conf->raid_disks].rdev == NULL) {
1752                         /* Add this device as a replacement */
1753                         clear_bit(In_sync, &rdev->flags);
1754                         set_bit(Replacement, &rdev->flags);
1755                         rdev->raid_disk = mirror;
1756                         err = 0;
1757                         conf->fullsync = 1;
1758                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1759                         break;
1760                 }
1761         }
1762         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1763                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1764         print_conf(conf);
1765         return err;
1766 }
1767
1768 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1769 {
1770         struct r1conf *conf = mddev->private;
1771         int err = 0;
1772         int number = rdev->raid_disk;
1773         struct raid1_info *p = conf->mirrors + number;
1774
1775         if (rdev != p->rdev)
1776                 p = conf->mirrors + conf->raid_disks + number;
1777
1778         print_conf(conf);
1779         if (rdev == p->rdev) {
1780                 if (test_bit(In_sync, &rdev->flags) ||
1781                     atomic_read(&rdev->nr_pending)) {
1782                         err = -EBUSY;
1783                         goto abort;
1784                 }
1785                 /* Only remove non-faulty devices if recovery
1786                  * is not possible.
1787                  */
1788                 if (!test_bit(Faulty, &rdev->flags) &&
1789                     mddev->recovery_disabled != conf->recovery_disabled &&
1790                     mddev->degraded < conf->raid_disks) {
1791                         err = -EBUSY;
1792                         goto abort;
1793                 }
1794                 p->rdev = NULL;
1795                 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1796                         synchronize_rcu();
1797                         if (atomic_read(&rdev->nr_pending)) {
1798                                 /* lost the race, try later */
1799                                 err = -EBUSY;
1800                                 p->rdev = rdev;
1801                                 goto abort;
1802                         }
1803                 }
1804                 if (conf->mirrors[conf->raid_disks + number].rdev) {
1805                         /* We just removed a device that is being replaced.
1806                          * Move down the replacement.  We drain all IO before
1807                          * doing this to avoid confusion.
1808                          */
1809                         struct md_rdev *repl =
1810                                 conf->mirrors[conf->raid_disks + number].rdev;
1811                         freeze_array(conf, 0);
1812                         clear_bit(Replacement, &repl->flags);
1813                         p->rdev = repl;
1814                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1815                         unfreeze_array(conf);
1816                 }
1817
1818                 clear_bit(WantReplacement, &rdev->flags);
1819                 err = md_integrity_register(mddev);
1820         }
1821 abort:
1822
1823         print_conf(conf);
1824         return err;
1825 }
1826
1827 static void end_sync_read(struct bio *bio)
1828 {
1829         struct r1bio *r1_bio = get_resync_r1bio(bio);
1830
1831         update_head_pos(r1_bio->read_disk, r1_bio);
1832
1833         /*
1834          * we have read a block, now it needs to be re-written,
1835          * or re-read if the read failed.
1836          * We don't do much here, just schedule handling by raid1d
1837          */
1838         if (!bio->bi_status)
1839                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1840
1841         if (atomic_dec_and_test(&r1_bio->remaining))
1842                 reschedule_retry(r1_bio);
1843 }
1844
1845 static void end_sync_write(struct bio *bio)
1846 {
1847         int uptodate = !bio->bi_status;
1848         struct r1bio *r1_bio = get_resync_r1bio(bio);
1849         struct mddev *mddev = r1_bio->mddev;
1850         struct r1conf *conf = mddev->private;
1851         sector_t first_bad;
1852         int bad_sectors;
1853         struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1854
1855         if (!uptodate) {
1856                 sector_t sync_blocks = 0;
1857                 sector_t s = r1_bio->sector;
1858                 long sectors_to_go = r1_bio->sectors;
1859                 /* make sure these bits doesn't get cleared. */
1860                 do {
1861                         bitmap_end_sync(mddev->bitmap, s,
1862                                         &sync_blocks, 1);
1863                         s += sync_blocks;
1864                         sectors_to_go -= sync_blocks;
1865                 } while (sectors_to_go > 0);
1866                 set_bit(WriteErrorSeen, &rdev->flags);
1867                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1868                         set_bit(MD_RECOVERY_NEEDED, &
1869                                 mddev->recovery);
1870                 set_bit(R1BIO_WriteError, &r1_bio->state);
1871         } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1872                                &first_bad, &bad_sectors) &&
1873                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1874                                 r1_bio->sector,
1875                                 r1_bio->sectors,
1876                                 &first_bad, &bad_sectors)
1877                 )
1878                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1879
1880         if (atomic_dec_and_test(&r1_bio->remaining)) {
1881                 int s = r1_bio->sectors;
1882                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1883                     test_bit(R1BIO_WriteError, &r1_bio->state))
1884                         reschedule_retry(r1_bio);
1885                 else {
1886                         put_buf(r1_bio);
1887                         md_done_sync(mddev, s, uptodate);
1888                 }
1889         }
1890 }
1891
1892 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1893                             int sectors, struct page *page, int rw)
1894 {
1895         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1896                 /* success */
1897                 return 1;
1898         if (rw == WRITE) {
1899                 set_bit(WriteErrorSeen, &rdev->flags);
1900                 if (!test_and_set_bit(WantReplacement,
1901                                       &rdev->flags))
1902                         set_bit(MD_RECOVERY_NEEDED, &
1903                                 rdev->mddev->recovery);
1904         }
1905         /* need to record an error - either for the block or the device */
1906         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1907                 md_error(rdev->mddev, rdev);
1908         return 0;
1909 }
1910
1911 static int fix_sync_read_error(struct r1bio *r1_bio)
1912 {
1913         /* Try some synchronous reads of other devices to get
1914          * good data, much like with normal read errors.  Only
1915          * read into the pages we already have so we don't
1916          * need to re-issue the read request.
1917          * We don't need to freeze the array, because being in an
1918          * active sync request, there is no normal IO, and
1919          * no overlapping syncs.
1920          * We don't need to check is_badblock() again as we
1921          * made sure that anything with a bad block in range
1922          * will have bi_end_io clear.
1923          */
1924         struct mddev *mddev = r1_bio->mddev;
1925         struct r1conf *conf = mddev->private;
1926         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1927         struct page **pages = get_resync_pages(bio)->pages;
1928         sector_t sect = r1_bio->sector;
1929         int sectors = r1_bio->sectors;
1930         int idx = 0;
1931         struct md_rdev *rdev;
1932
1933         rdev = conf->mirrors[r1_bio->read_disk].rdev;
1934         if (test_bit(FailFast, &rdev->flags)) {
1935                 /* Don't try recovering from here - just fail it
1936                  * ... unless it is the last working device of course */
1937                 md_error(mddev, rdev);
1938                 if (test_bit(Faulty, &rdev->flags))
1939                         /* Don't try to read from here, but make sure
1940                          * put_buf does it's thing
1941                          */
1942                         bio->bi_end_io = end_sync_write;
1943         }
1944
1945         while(sectors) {
1946                 int s = sectors;
1947                 int d = r1_bio->read_disk;
1948                 int success = 0;
1949                 int start;
1950
1951                 if (s > (PAGE_SIZE>>9))
1952                         s = PAGE_SIZE >> 9;
1953                 do {
1954                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1955                                 /* No rcu protection needed here devices
1956                                  * can only be removed when no resync is
1957                                  * active, and resync is currently active
1958                                  */
1959                                 rdev = conf->mirrors[d].rdev;
1960                                 if (sync_page_io(rdev, sect, s<<9,
1961                                                  pages[idx],
1962                                                  REQ_OP_READ, 0, false)) {
1963                                         success = 1;
1964                                         break;
1965                                 }
1966                         }
1967                         d++;
1968                         if (d == conf->raid_disks * 2)
1969                                 d = 0;
1970                 } while (!success && d != r1_bio->read_disk);
1971
1972                 if (!success) {
1973                         char b[BDEVNAME_SIZE];
1974                         int abort = 0;
1975                         /* Cannot read from anywhere, this block is lost.
1976                          * Record a bad block on each device.  If that doesn't
1977                          * work just disable and interrupt the recovery.
1978                          * Don't fail devices as that won't really help.
1979                          */
1980                         pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1981                                             mdname(mddev), bio_devname(bio, b),
1982                                             (unsigned long long)r1_bio->sector);
1983                         for (d = 0; d < conf->raid_disks * 2; d++) {
1984                                 rdev = conf->mirrors[d].rdev;
1985                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1986                                         continue;
1987                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1988                                         abort = 1;
1989                         }
1990                         if (abort) {
1991                                 conf->recovery_disabled =
1992                                         mddev->recovery_disabled;
1993                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1994                                 md_done_sync(mddev, r1_bio->sectors, 0);
1995                                 put_buf(r1_bio);
1996                                 return 0;
1997                         }
1998                         /* Try next page */
1999                         sectors -= s;
2000                         sect += s;
2001                         idx++;
2002                         continue;
2003                 }
2004
2005                 start = d;
2006                 /* write it back and re-read */
2007                 while (d != r1_bio->read_disk) {
2008                         if (d == 0)
2009                                 d = conf->raid_disks * 2;
2010                         d--;
2011                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2012                                 continue;
2013                         rdev = conf->mirrors[d].rdev;
2014                         if (r1_sync_page_io(rdev, sect, s,
2015                                             pages[idx],
2016                                             WRITE) == 0) {
2017                                 r1_bio->bios[d]->bi_end_io = NULL;
2018                                 rdev_dec_pending(rdev, mddev);
2019                         }
2020                 }
2021                 d = start;
2022                 while (d != r1_bio->read_disk) {
2023                         if (d == 0)
2024                                 d = conf->raid_disks * 2;
2025                         d--;
2026                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2027                                 continue;
2028                         rdev = conf->mirrors[d].rdev;
2029                         if (r1_sync_page_io(rdev, sect, s,
2030                                             pages[idx],
2031                                             READ) != 0)
2032                                 atomic_add(s, &rdev->corrected_errors);
2033                 }
2034                 sectors -= s;
2035                 sect += s;
2036                 idx ++;
2037         }
2038         set_bit(R1BIO_Uptodate, &r1_bio->state);
2039         bio->bi_status = 0;
2040         return 1;
2041 }
2042
2043 static void process_checks(struct r1bio *r1_bio)
2044 {
2045         /* We have read all readable devices.  If we haven't
2046          * got the block, then there is no hope left.
2047          * If we have, then we want to do a comparison
2048          * and skip the write if everything is the same.
2049          * If any blocks failed to read, then we need to
2050          * attempt an over-write
2051          */
2052         struct mddev *mddev = r1_bio->mddev;
2053         struct r1conf *conf = mddev->private;
2054         int primary;
2055         int i;
2056         int vcnt;
2057
2058         /* Fix variable parts of all bios */
2059         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2060         for (i = 0; i < conf->raid_disks * 2; i++) {
2061                 blk_status_t status;
2062                 struct bio *b = r1_bio->bios[i];
2063                 struct resync_pages *rp = get_resync_pages(b);
2064                 if (b->bi_end_io != end_sync_read)
2065                         continue;
2066                 /* fixup the bio for reuse, but preserve errno */
2067                 status = b->bi_status;
2068                 bio_reset(b);
2069                 b->bi_status = status;
2070                 b->bi_iter.bi_sector = r1_bio->sector +
2071                         conf->mirrors[i].rdev->data_offset;
2072                 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2073                 b->bi_end_io = end_sync_read;
2074                 rp->raid_bio = r1_bio;
2075                 b->bi_private = rp;
2076
2077                 /* initialize bvec table again */
2078                 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2079         }
2080         for (primary = 0; primary < conf->raid_disks * 2; primary++)
2081                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2082                     !r1_bio->bios[primary]->bi_status) {
2083                         r1_bio->bios[primary]->bi_end_io = NULL;
2084                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2085                         break;
2086                 }
2087         r1_bio->read_disk = primary;
2088         for (i = 0; i < conf->raid_disks * 2; i++) {
2089                 int j;
2090                 struct bio *pbio = r1_bio->bios[primary];
2091                 struct bio *sbio = r1_bio->bios[i];
2092                 blk_status_t status = sbio->bi_status;
2093                 struct page **ppages = get_resync_pages(pbio)->pages;
2094                 struct page **spages = get_resync_pages(sbio)->pages;
2095                 struct bio_vec *bi;
2096                 int page_len[RESYNC_PAGES] = { 0 };
2097
2098                 if (sbio->bi_end_io != end_sync_read)
2099                         continue;
2100                 /* Now we can 'fixup' the error value */
2101                 sbio->bi_status = 0;
2102
2103                 bio_for_each_segment_all(bi, sbio, j)
2104                         page_len[j] = bi->bv_len;
2105
2106                 if (!status) {
2107                         for (j = vcnt; j-- ; ) {
2108                                 if (memcmp(page_address(ppages[j]),
2109                                            page_address(spages[j]),
2110                                            page_len[j]))
2111                                         break;
2112                         }
2113                 } else
2114                         j = 0;
2115                 if (j >= 0)
2116                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2117                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2118                               && !status)) {
2119                         /* No need to write to this device. */
2120                         sbio->bi_end_io = NULL;
2121                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2122                         continue;
2123                 }
2124
2125                 bio_copy_data(sbio, pbio);
2126         }
2127 }
2128
2129 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2130 {
2131         struct r1conf *conf = mddev->private;
2132         int i;
2133         int disks = conf->raid_disks * 2;
2134         struct bio *wbio;
2135
2136         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2137                 /* ouch - failed to read all of that. */
2138                 if (!fix_sync_read_error(r1_bio))
2139                         return;
2140
2141         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2142                 process_checks(r1_bio);
2143
2144         /*
2145          * schedule writes
2146          */
2147         atomic_set(&r1_bio->remaining, 1);
2148         for (i = 0; i < disks ; i++) {
2149                 wbio = r1_bio->bios[i];
2150                 if (wbio->bi_end_io == NULL ||
2151                     (wbio->bi_end_io == end_sync_read &&
2152                      (i == r1_bio->read_disk ||
2153                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2154                         continue;
2155                 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2156                         continue;
2157
2158                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2159                 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2160                         wbio->bi_opf |= MD_FAILFAST;
2161
2162                 wbio->bi_end_io = end_sync_write;
2163                 atomic_inc(&r1_bio->remaining);
2164                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2165
2166                 generic_make_request(wbio);
2167         }
2168
2169         if (atomic_dec_and_test(&r1_bio->remaining)) {
2170                 /* if we're here, all write(s) have completed, so clean up */
2171                 int s = r1_bio->sectors;
2172                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2173                     test_bit(R1BIO_WriteError, &r1_bio->state))
2174                         reschedule_retry(r1_bio);
2175                 else {
2176                         put_buf(r1_bio);
2177                         md_done_sync(mddev, s, 1);
2178                 }
2179         }
2180 }
2181
2182 /*
2183  * This is a kernel thread which:
2184  *
2185  *      1.      Retries failed read operations on working mirrors.
2186  *      2.      Updates the raid superblock when problems encounter.
2187  *      3.      Performs writes following reads for array synchronising.
2188  */
2189
2190 static void fix_read_error(struct r1conf *conf, int read_disk,
2191                            sector_t sect, int sectors)
2192 {
2193         struct mddev *mddev = conf->mddev;
2194         while(sectors) {
2195                 int s = sectors;
2196                 int d = read_disk;
2197                 int success = 0;
2198                 int start;
2199                 struct md_rdev *rdev;
2200
2201                 if (s > (PAGE_SIZE>>9))
2202                         s = PAGE_SIZE >> 9;
2203
2204                 do {
2205                         sector_t first_bad;
2206                         int bad_sectors;
2207
2208                         rcu_read_lock();
2209                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2210                         if (rdev &&
2211                             (test_bit(In_sync, &rdev->flags) ||
2212                              (!test_bit(Faulty, &rdev->flags) &&
2213                               rdev->recovery_offset >= sect + s)) &&
2214                             is_badblock(rdev, sect, s,
2215                                         &first_bad, &bad_sectors) == 0) {
2216                                 atomic_inc(&rdev->nr_pending);
2217                                 rcu_read_unlock();
2218                                 if (sync_page_io(rdev, sect, s<<9,
2219                                          conf->tmppage, REQ_OP_READ, 0, false))
2220                                         success = 1;
2221                                 rdev_dec_pending(rdev, mddev);
2222                                 if (success)
2223                                         break;
2224                         } else
2225                                 rcu_read_unlock();
2226                         d++;
2227                         if (d == conf->raid_disks * 2)
2228                                 d = 0;
2229                 } while (!success && d != read_disk);
2230
2231                 if (!success) {
2232                         /* Cannot read from anywhere - mark it bad */
2233                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2234                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2235                                 md_error(mddev, rdev);
2236                         break;
2237                 }
2238                 /* write it back and re-read */
2239                 start = d;
2240                 while (d != read_disk) {
2241                         if (d==0)
2242                                 d = conf->raid_disks * 2;
2243                         d--;
2244                         rcu_read_lock();
2245                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2246                         if (rdev &&
2247                             !test_bit(Faulty, &rdev->flags)) {
2248                                 atomic_inc(&rdev->nr_pending);
2249                                 rcu_read_unlock();
2250                                 r1_sync_page_io(rdev, sect, s,
2251                                                 conf->tmppage, WRITE);
2252                                 rdev_dec_pending(rdev, mddev);
2253                         } else
2254                                 rcu_read_unlock();
2255                 }
2256                 d = start;
2257                 while (d != read_disk) {
2258                         char b[BDEVNAME_SIZE];
2259                         if (d==0)
2260                                 d = conf->raid_disks * 2;
2261                         d--;
2262                         rcu_read_lock();
2263                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2264                         if (rdev &&
2265                             !test_bit(Faulty, &rdev->flags)) {
2266                                 atomic_inc(&rdev->nr_pending);
2267                                 rcu_read_unlock();
2268                                 if (r1_sync_page_io(rdev, sect, s,
2269                                                     conf->tmppage, READ)) {
2270                                         atomic_add(s, &rdev->corrected_errors);
2271                                         pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2272                                                 mdname(mddev), s,
2273                                                 (unsigned long long)(sect +
2274                                                                      rdev->data_offset),
2275                                                 bdevname(rdev->bdev, b));
2276                                 }
2277                                 rdev_dec_pending(rdev, mddev);
2278                         } else
2279                                 rcu_read_unlock();
2280                 }
2281                 sectors -= s;
2282                 sect += s;
2283         }
2284 }
2285
2286 static int narrow_write_error(struct r1bio *r1_bio, int i)
2287 {
2288         struct mddev *mddev = r1_bio->mddev;
2289         struct r1conf *conf = mddev->private;
2290         struct md_rdev *rdev = conf->mirrors[i].rdev;
2291
2292         /* bio has the data to be written to device 'i' where
2293          * we just recently had a write error.
2294          * We repeatedly clone the bio and trim down to one block,
2295          * then try the write.  Where the write fails we record
2296          * a bad block.
2297          * It is conceivable that the bio doesn't exactly align with
2298          * blocks.  We must handle this somehow.
2299          *
2300          * We currently own a reference on the rdev.
2301          */
2302
2303         int block_sectors;
2304         sector_t sector;
2305         int sectors;
2306         int sect_to_write = r1_bio->sectors;
2307         int ok = 1;
2308
2309         if (rdev->badblocks.shift < 0)
2310                 return 0;
2311
2312         block_sectors = roundup(1 << rdev->badblocks.shift,
2313                                 bdev_logical_block_size(rdev->bdev) >> 9);
2314         sector = r1_bio->sector;
2315         sectors = ((sector + block_sectors)
2316                    & ~(sector_t)(block_sectors - 1))
2317                 - sector;
2318
2319         while (sect_to_write) {
2320                 struct bio *wbio;
2321                 if (sectors > sect_to_write)
2322                         sectors = sect_to_write;
2323                 /* Write at 'sector' for 'sectors'*/
2324
2325                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2326                         wbio = bio_clone_fast(r1_bio->behind_master_bio,
2327                                               GFP_NOIO,
2328                                               mddev->bio_set);
2329                 } else {
2330                         wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2331                                               mddev->bio_set);
2332                 }
2333
2334                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2335                 wbio->bi_iter.bi_sector = r1_bio->sector;
2336                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2337
2338                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2339                 wbio->bi_iter.bi_sector += rdev->data_offset;
2340                 bio_set_dev(wbio, rdev->bdev);
2341
2342                 if (submit_bio_wait(wbio) < 0)
2343                         /* failure! */
2344                         ok = rdev_set_badblocks(rdev, sector,
2345                                                 sectors, 0)
2346                                 && ok;
2347
2348                 bio_put(wbio);
2349                 sect_to_write -= sectors;
2350                 sector += sectors;
2351                 sectors = block_sectors;
2352         }
2353         return ok;
2354 }
2355
2356 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2357 {
2358         int m;
2359         int s = r1_bio->sectors;
2360         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2361                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2362                 struct bio *bio = r1_bio->bios[m];
2363                 if (bio->bi_end_io == NULL)
2364                         continue;
2365                 if (!bio->bi_status &&
2366                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2367                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2368                 }
2369                 if (bio->bi_status &&
2370                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2371                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2372                                 md_error(conf->mddev, rdev);
2373                 }
2374         }
2375         put_buf(r1_bio);
2376         md_done_sync(conf->mddev, s, 1);
2377 }
2378
2379 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2380 {
2381         int m, idx;
2382         bool fail = false;
2383
2384         for (m = 0; m < conf->raid_disks * 2 ; m++)
2385                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2386                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2387                         rdev_clear_badblocks(rdev,
2388                                              r1_bio->sector,
2389                                              r1_bio->sectors, 0);
2390                         rdev_dec_pending(rdev, conf->mddev);
2391                 } else if (r1_bio->bios[m] != NULL) {
2392                         /* This drive got a write error.  We need to
2393                          * narrow down and record precise write
2394                          * errors.
2395                          */
2396                         fail = true;
2397                         if (!narrow_write_error(r1_bio, m)) {
2398                                 md_error(conf->mddev,
2399                                          conf->mirrors[m].rdev);
2400                                 /* an I/O failed, we can't clear the bitmap */
2401                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2402                         }
2403                         rdev_dec_pending(conf->mirrors[m].rdev,
2404                                          conf->mddev);
2405                 }
2406         if (fail) {
2407                 spin_lock_irq(&conf->device_lock);
2408                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2409                 idx = sector_to_idx(r1_bio->sector);
2410                 atomic_inc(&conf->nr_queued[idx]);
2411                 spin_unlock_irq(&conf->device_lock);
2412                 /*
2413                  * In case freeze_array() is waiting for condition
2414                  * get_unqueued_pending() == extra to be true.
2415                  */
2416                 wake_up(&conf->wait_barrier);
2417                 md_wakeup_thread(conf->mddev->thread);
2418         } else {
2419                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2420                         close_write(r1_bio);
2421                 raid_end_bio_io(r1_bio);
2422         }
2423 }
2424
2425 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2426 {
2427         struct mddev *mddev = conf->mddev;
2428         struct bio *bio;
2429         struct md_rdev *rdev;
2430         sector_t bio_sector;
2431
2432         clear_bit(R1BIO_ReadError, &r1_bio->state);
2433         /* we got a read error. Maybe the drive is bad.  Maybe just
2434          * the block and we can fix it.
2435          * We freeze all other IO, and try reading the block from
2436          * other devices.  When we find one, we re-write
2437          * and check it that fixes the read error.
2438          * This is all done synchronously while the array is
2439          * frozen
2440          */
2441
2442         bio = r1_bio->bios[r1_bio->read_disk];
2443         bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2444         bio_put(bio);
2445         r1_bio->bios[r1_bio->read_disk] = NULL;
2446
2447         rdev = conf->mirrors[r1_bio->read_disk].rdev;
2448         if (mddev->ro == 0
2449             && !test_bit(FailFast, &rdev->flags)) {
2450                 freeze_array(conf, 1);
2451                 fix_read_error(conf, r1_bio->read_disk,
2452                                r1_bio->sector, r1_bio->sectors);
2453                 unfreeze_array(conf);
2454         } else {
2455                 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2456         }
2457
2458         rdev_dec_pending(rdev, conf->mddev);
2459         allow_barrier(conf, r1_bio->sector);
2460         bio = r1_bio->master_bio;
2461
2462         /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2463         r1_bio->state = 0;
2464         raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2465 }
2466
2467 static void raid1d(struct md_thread *thread)
2468 {
2469         struct mddev *mddev = thread->mddev;
2470         struct r1bio *r1_bio;
2471         unsigned long flags;
2472         struct r1conf *conf = mddev->private;
2473         struct list_head *head = &conf->retry_list;
2474         struct blk_plug plug;
2475         int idx;
2476
2477         md_check_recovery(mddev);
2478
2479         if (!list_empty_careful(&conf->bio_end_io_list) &&
2480             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2481                 LIST_HEAD(tmp);
2482                 spin_lock_irqsave(&conf->device_lock, flags);
2483                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2484                         list_splice_init(&conf->bio_end_io_list, &tmp);
2485                 spin_unlock_irqrestore(&conf->device_lock, flags);
2486                 while (!list_empty(&tmp)) {
2487                         r1_bio = list_first_entry(&tmp, struct r1bio,
2488                                                   retry_list);
2489                         list_del(&r1_bio->retry_list);
2490                         idx = sector_to_idx(r1_bio->sector);
2491                         atomic_dec(&conf->nr_queued[idx]);
2492                         if (mddev->degraded)
2493                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2494                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2495                                 close_write(r1_bio);
2496                         raid_end_bio_io(r1_bio);
2497                 }
2498         }
2499
2500         blk_start_plug(&plug);
2501         for (;;) {
2502
2503                 flush_pending_writes(conf);
2504
2505                 spin_lock_irqsave(&conf->device_lock, flags);
2506                 if (list_empty(head)) {
2507                         spin_unlock_irqrestore(&conf->device_lock, flags);
2508                         break;
2509                 }
2510                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2511                 list_del(head->prev);
2512                 idx = sector_to_idx(r1_bio->sector);
2513                 atomic_dec(&conf->nr_queued[idx]);
2514                 spin_unlock_irqrestore(&conf->device_lock, flags);
2515
2516                 mddev = r1_bio->mddev;
2517                 conf = mddev->private;
2518                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2519                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2520                             test_bit(R1BIO_WriteError, &r1_bio->state))
2521                                 handle_sync_write_finished(conf, r1_bio);
2522                         else
2523                                 sync_request_write(mddev, r1_bio);
2524                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2525                            test_bit(R1BIO_WriteError, &r1_bio->state))
2526                         handle_write_finished(conf, r1_bio);
2527                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2528                         handle_read_error(conf, r1_bio);
2529                 else
2530                         WARN_ON_ONCE(1);
2531
2532                 cond_resched();
2533                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2534                         md_check_recovery(mddev);
2535         }
2536         blk_finish_plug(&plug);
2537 }
2538
2539 static int init_resync(struct r1conf *conf)
2540 {
2541         int buffs;
2542
2543         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2544         BUG_ON(conf->r1buf_pool);
2545         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2546                                           conf->poolinfo);
2547         if (!conf->r1buf_pool)
2548                 return -ENOMEM;
2549         return 0;
2550 }
2551
2552 static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2553 {
2554         struct r1bio *r1bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2555         struct resync_pages *rps;
2556         struct bio *bio;
2557         int i;
2558
2559         for (i = conf->poolinfo->raid_disks; i--; ) {
2560                 bio = r1bio->bios[i];
2561                 rps = bio->bi_private;
2562                 bio_reset(bio);
2563                 bio->bi_private = rps;
2564         }
2565         r1bio->master_bio = NULL;
2566         return r1bio;
2567 }
2568
2569 /*
2570  * perform a "sync" on one "block"
2571  *
2572  * We need to make sure that no normal I/O request - particularly write
2573  * requests - conflict with active sync requests.
2574  *
2575  * This is achieved by tracking pending requests and a 'barrier' concept
2576  * that can be installed to exclude normal IO requests.
2577  */
2578
2579 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2580                                    int *skipped)
2581 {
2582         struct r1conf *conf = mddev->private;
2583         struct r1bio *r1_bio;
2584         struct bio *bio;
2585         sector_t max_sector, nr_sectors;
2586         int disk = -1;
2587         int i;
2588         int wonly = -1;
2589         int write_targets = 0, read_targets = 0;
2590         sector_t sync_blocks;
2591         int still_degraded = 0;
2592         int good_sectors = RESYNC_SECTORS;
2593         int min_bad = 0; /* number of sectors that are bad in all devices */
2594         int idx = sector_to_idx(sector_nr);
2595         int page_idx = 0;
2596
2597         if (!conf->r1buf_pool)
2598                 if (init_resync(conf))
2599                         return 0;
2600
2601         max_sector = mddev->dev_sectors;
2602         if (sector_nr >= max_sector) {
2603                 /* If we aborted, we need to abort the
2604                  * sync on the 'current' bitmap chunk (there will
2605                  * only be one in raid1 resync.
2606                  * We can find the current addess in mddev->curr_resync
2607                  */
2608                 if (mddev->curr_resync < max_sector) /* aborted */
2609                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2610                                                 &sync_blocks, 1);
2611                 else /* completed sync */
2612                         conf->fullsync = 0;
2613
2614                 bitmap_close_sync(mddev->bitmap);
2615                 close_sync(conf);
2616
2617                 if (mddev_is_clustered(mddev)) {
2618                         conf->cluster_sync_low = 0;
2619                         conf->cluster_sync_high = 0;
2620                 }
2621                 return 0;
2622         }
2623
2624         if (mddev->bitmap == NULL &&
2625             mddev->recovery_cp == MaxSector &&
2626             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2627             conf->fullsync == 0) {
2628                 *skipped = 1;
2629                 return max_sector - sector_nr;
2630         }
2631         /* before building a request, check if we can skip these blocks..
2632          * This call the bitmap_start_sync doesn't actually record anything
2633          */
2634         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2635             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2636                 /* We can skip this block, and probably several more */
2637                 *skipped = 1;
2638                 return sync_blocks;
2639         }
2640
2641         /*
2642          * If there is non-resync activity waiting for a turn, then let it
2643          * though before starting on this new sync request.
2644          */
2645         if (atomic_read(&conf->nr_waiting[idx]))
2646                 schedule_timeout_uninterruptible(1);
2647
2648         /* we are incrementing sector_nr below. To be safe, we check against
2649          * sector_nr + two times RESYNC_SECTORS
2650          */
2651
2652         bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2653                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2654         r1_bio = raid1_alloc_init_r1buf(conf);
2655
2656         raise_barrier(conf, sector_nr);
2657
2658         rcu_read_lock();
2659         /*
2660          * If we get a correctably read error during resync or recovery,
2661          * we might want to read from a different device.  So we
2662          * flag all drives that could conceivably be read from for READ,
2663          * and any others (which will be non-In_sync devices) for WRITE.
2664          * If a read fails, we try reading from something else for which READ
2665          * is OK.
2666          */
2667
2668         r1_bio->mddev = mddev;
2669         r1_bio->sector = sector_nr;
2670         r1_bio->state = 0;
2671         set_bit(R1BIO_IsSync, &r1_bio->state);
2672         /* make sure good_sectors won't go across barrier unit boundary */
2673         good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2674
2675         for (i = 0; i < conf->raid_disks * 2; i++) {
2676                 struct md_rdev *rdev;
2677                 bio = r1_bio->bios[i];
2678
2679                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2680                 if (rdev == NULL ||
2681                     test_bit(Faulty, &rdev->flags)) {
2682                         if (i < conf->raid_disks)
2683                                 still_degraded = 1;
2684                 } else if (!test_bit(In_sync, &rdev->flags)) {
2685                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2686                         bio->bi_end_io = end_sync_write;
2687                         write_targets ++;
2688                 } else {
2689                         /* may need to read from here */
2690                         sector_t first_bad = MaxSector;
2691                         int bad_sectors;
2692
2693                         if (is_badblock(rdev, sector_nr, good_sectors,
2694                                         &first_bad, &bad_sectors)) {
2695                                 if (first_bad > sector_nr)
2696                                         good_sectors = first_bad - sector_nr;
2697                                 else {
2698                                         bad_sectors -= (sector_nr - first_bad);
2699                                         if (min_bad == 0 ||
2700                                             min_bad > bad_sectors)
2701                                                 min_bad = bad_sectors;
2702                                 }
2703                         }
2704                         if (sector_nr < first_bad) {
2705                                 if (test_bit(WriteMostly, &rdev->flags)) {
2706                                         if (wonly < 0)
2707                                                 wonly = i;
2708                                 } else {
2709                                         if (disk < 0)
2710                                                 disk = i;
2711                                 }
2712                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2713                                 bio->bi_end_io = end_sync_read;
2714                                 read_targets++;
2715                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2716                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2717                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2718                                 /*
2719                                  * The device is suitable for reading (InSync),
2720                                  * but has bad block(s) here. Let's try to correct them,
2721                                  * if we are doing resync or repair. Otherwise, leave
2722                                  * this device alone for this sync request.
2723                                  */
2724                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2725                                 bio->bi_end_io = end_sync_write;
2726                                 write_targets++;
2727                         }
2728                 }
2729                 if (bio->bi_end_io) {
2730                         atomic_inc(&rdev->nr_pending);
2731                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2732                         bio_set_dev(bio, rdev->bdev);
2733                         if (test_bit(FailFast, &rdev->flags))
2734                                 bio->bi_opf |= MD_FAILFAST;
2735                 }
2736         }
2737         rcu_read_unlock();
2738         if (disk < 0)
2739                 disk = wonly;
2740         r1_bio->read_disk = disk;
2741
2742         if (read_targets == 0 && min_bad > 0) {
2743                 /* These sectors are bad on all InSync devices, so we
2744                  * need to mark them bad on all write targets
2745                  */
2746                 int ok = 1;
2747                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2748                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2749                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2750                                 ok = rdev_set_badblocks(rdev, sector_nr,
2751                                                         min_bad, 0
2752                                         ) && ok;
2753                         }
2754                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2755                 *skipped = 1;
2756                 put_buf(r1_bio);
2757
2758                 if (!ok) {
2759                         /* Cannot record the badblocks, so need to
2760                          * abort the resync.
2761                          * If there are multiple read targets, could just
2762                          * fail the really bad ones ???
2763                          */
2764                         conf->recovery_disabled = mddev->recovery_disabled;
2765                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2766                         return 0;
2767                 } else
2768                         return min_bad;
2769
2770         }
2771         if (min_bad > 0 && min_bad < good_sectors) {
2772                 /* only resync enough to reach the next bad->good
2773                  * transition */
2774                 good_sectors = min_bad;
2775         }
2776
2777         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2778                 /* extra read targets are also write targets */
2779                 write_targets += read_targets-1;
2780
2781         if (write_targets == 0 || read_targets == 0) {
2782                 /* There is nowhere to write, so all non-sync
2783                  * drives must be failed - so we are finished
2784                  */
2785                 sector_t rv;
2786                 if (min_bad > 0)
2787                         max_sector = sector_nr + min_bad;
2788                 rv = max_sector - sector_nr;
2789                 *skipped = 1;
2790                 put_buf(r1_bio);
2791                 return rv;
2792         }
2793
2794         if (max_sector > mddev->resync_max)
2795                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2796         if (max_sector > sector_nr + good_sectors)
2797                 max_sector = sector_nr + good_sectors;
2798         nr_sectors = 0;
2799         sync_blocks = 0;
2800         do {
2801                 struct page *page;
2802                 int len = PAGE_SIZE;
2803                 if (sector_nr + (len>>9) > max_sector)
2804                         len = (max_sector - sector_nr) << 9;
2805                 if (len == 0)
2806                         break;
2807                 if (sync_blocks == 0) {
2808                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2809                                                &sync_blocks, still_degraded) &&
2810                             !conf->fullsync &&
2811                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2812                                 break;
2813                         if ((len >> 9) > sync_blocks)
2814                                 len = sync_blocks<<9;
2815                 }
2816
2817                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2818                         struct resync_pages *rp;
2819
2820                         bio = r1_bio->bios[i];
2821                         rp = get_resync_pages(bio);
2822                         if (bio->bi_end_io) {
2823                                 page = resync_fetch_page(rp, page_idx);
2824
2825                                 /*
2826                                  * won't fail because the vec table is big
2827                                  * enough to hold all these pages
2828                                  */
2829                                 bio_add_page(bio, page, len, 0);
2830                         }
2831                 }
2832                 nr_sectors += len>>9;
2833                 sector_nr += len>>9;
2834                 sync_blocks -= (len>>9);
2835         } while (++page_idx < RESYNC_PAGES);
2836
2837         r1_bio->sectors = nr_sectors;
2838
2839         if (mddev_is_clustered(mddev) &&
2840                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2841                 conf->cluster_sync_low = mddev->curr_resync_completed;
2842                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2843                 /* Send resync message */
2844                 md_cluster_ops->resync_info_update(mddev,
2845                                 conf->cluster_sync_low,
2846                                 conf->cluster_sync_high);
2847         }
2848
2849         /* For a user-requested sync, we read all readable devices and do a
2850          * compare
2851          */
2852         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2853                 atomic_set(&r1_bio->remaining, read_targets);
2854                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2855                         bio = r1_bio->bios[i];
2856                         if (bio->bi_end_io == end_sync_read) {
2857                                 read_targets--;
2858                                 md_sync_acct_bio(bio, nr_sectors);
2859                                 if (read_targets == 1)
2860                                         bio->bi_opf &= ~MD_FAILFAST;
2861                                 generic_make_request(bio);
2862                         }
2863                 }
2864         } else {
2865                 atomic_set(&r1_bio->remaining, 1);
2866                 bio = r1_bio->bios[r1_bio->read_disk];
2867                 md_sync_acct_bio(bio, nr_sectors);
2868                 if (read_targets == 1)
2869                         bio->bi_opf &= ~MD_FAILFAST;
2870                 generic_make_request(bio);
2871
2872         }
2873         return nr_sectors;
2874 }
2875
2876 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2877 {
2878         if (sectors)
2879                 return sectors;
2880
2881         return mddev->dev_sectors;
2882 }
2883
2884 static struct r1conf *setup_conf(struct mddev *mddev)
2885 {
2886         struct r1conf *conf;
2887         int i;
2888         struct raid1_info *disk;
2889         struct md_rdev *rdev;
2890         int err = -ENOMEM;
2891
2892         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2893         if (!conf)
2894                 goto abort;
2895
2896         conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2897                                    sizeof(atomic_t), GFP_KERNEL);
2898         if (!conf->nr_pending)
2899                 goto abort;
2900
2901         conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2902                                    sizeof(atomic_t), GFP_KERNEL);
2903         if (!conf->nr_waiting)
2904                 goto abort;
2905
2906         conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2907                                   sizeof(atomic_t), GFP_KERNEL);
2908         if (!conf->nr_queued)
2909                 goto abort;
2910
2911         conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2912                                 sizeof(atomic_t), GFP_KERNEL);
2913         if (!conf->barrier)
2914                 goto abort;
2915
2916         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2917                                 * mddev->raid_disks * 2,
2918                                  GFP_KERNEL);
2919         if (!conf->mirrors)
2920                 goto abort;
2921
2922         conf->tmppage = alloc_page(GFP_KERNEL);
2923         if (!conf->tmppage)
2924                 goto abort;
2925
2926         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2927         if (!conf->poolinfo)
2928                 goto abort;
2929         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2930         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2931                                           r1bio_pool_free,
2932                                           conf->poolinfo);
2933         if (!conf->r1bio_pool)
2934                 goto abort;
2935
2936         conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2937         if (!conf->bio_split)
2938                 goto abort;
2939
2940         conf->poolinfo->mddev = mddev;
2941
2942         err = -EINVAL;
2943         spin_lock_init(&conf->device_lock);
2944         rdev_for_each(rdev, mddev) {
2945                 int disk_idx = rdev->raid_disk;
2946                 if (disk_idx >= mddev->raid_disks
2947                     || disk_idx < 0)
2948                         continue;
2949                 if (test_bit(Replacement, &rdev->flags))
2950                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2951                 else
2952                         disk = conf->mirrors + disk_idx;
2953
2954                 if (disk->rdev)
2955                         goto abort;
2956                 disk->rdev = rdev;
2957                 disk->head_position = 0;
2958                 disk->seq_start = MaxSector;
2959         }
2960         conf->raid_disks = mddev->raid_disks;
2961         conf->mddev = mddev;
2962         INIT_LIST_HEAD(&conf->retry_list);
2963         INIT_LIST_HEAD(&conf->bio_end_io_list);
2964
2965         spin_lock_init(&conf->resync_lock);
2966         init_waitqueue_head(&conf->wait_barrier);
2967
2968         bio_list_init(&conf->pending_bio_list);
2969         conf->pending_count = 0;
2970         conf->recovery_disabled = mddev->recovery_disabled - 1;
2971
2972         err = -EIO;
2973         for (i = 0; i < conf->raid_disks * 2; i++) {
2974
2975                 disk = conf->mirrors + i;
2976
2977                 if (i < conf->raid_disks &&
2978                     disk[conf->raid_disks].rdev) {
2979                         /* This slot has a replacement. */
2980                         if (!disk->rdev) {
2981                                 /* No original, just make the replacement
2982                                  * a recovering spare
2983                                  */
2984                                 disk->rdev =
2985                                         disk[conf->raid_disks].rdev;
2986                                 disk[conf->raid_disks].rdev = NULL;
2987                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2988                                 /* Original is not in_sync - bad */
2989                                 goto abort;
2990                 }
2991
2992                 if (!disk->rdev ||
2993                     !test_bit(In_sync, &disk->rdev->flags)) {
2994                         disk->head_position = 0;
2995                         if (disk->rdev &&
2996                             (disk->rdev->saved_raid_disk < 0))
2997                                 conf->fullsync = 1;
2998                 }
2999         }
3000
3001         err = -ENOMEM;
3002         conf->thread = md_register_thread(raid1d, mddev, "raid1");
3003         if (!conf->thread)
3004                 goto abort;
3005
3006         return conf;
3007
3008  abort:
3009         if (conf) {
3010                 mempool_destroy(conf->r1bio_pool);
3011                 kfree(conf->mirrors);
3012                 safe_put_page(conf->tmppage);
3013                 kfree(conf->poolinfo);
3014                 kfree(conf->nr_pending);
3015                 kfree(conf->nr_waiting);
3016                 kfree(conf->nr_queued);
3017                 kfree(conf->barrier);
3018                 if (conf->bio_split)
3019                         bioset_free(conf->bio_split);
3020                 kfree(conf);
3021         }
3022         return ERR_PTR(err);
3023 }
3024
3025 static void raid1_free(struct mddev *mddev, void *priv);
3026 static int raid1_run(struct mddev *mddev)
3027 {
3028         struct r1conf *conf;
3029         int i;
3030         struct md_rdev *rdev;
3031         int ret;
3032         bool discard_supported = false;
3033
3034         if (mddev->level != 1) {
3035                 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3036                         mdname(mddev), mddev->level);
3037                 return -EIO;
3038         }
3039         if (mddev->reshape_position != MaxSector) {
3040                 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3041                         mdname(mddev));
3042                 return -EIO;
3043         }
3044         if (mddev_init_writes_pending(mddev) < 0)
3045                 return -ENOMEM;
3046         /*
3047          * copy the already verified devices into our private RAID1
3048          * bookkeeping area. [whatever we allocate in run(),
3049          * should be freed in raid1_free()]
3050          */
3051         if (mddev->private == NULL)
3052                 conf = setup_conf(mddev);
3053         else
3054                 conf = mddev->private;
3055
3056         if (IS_ERR(conf))
3057                 return PTR_ERR(conf);
3058
3059         if (mddev->queue) {
3060                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3061                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3062         }
3063
3064         rdev_for_each(rdev, mddev) {
3065                 if (!mddev->gendisk)
3066                         continue;
3067                 disk_stack_limits(mddev->gendisk, rdev->bdev,
3068                                   rdev->data_offset << 9);
3069                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3070                         discard_supported = true;
3071         }
3072
3073         mddev->degraded = 0;
3074         for (i=0; i < conf->raid_disks; i++)
3075                 if (conf->mirrors[i].rdev == NULL ||
3076                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3077                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3078                         mddev->degraded++;
3079
3080         if (conf->raid_disks - mddev->degraded == 1)
3081                 mddev->recovery_cp = MaxSector;
3082
3083         if (mddev->recovery_cp != MaxSector)
3084                 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3085                         mdname(mddev));
3086         pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3087                 mdname(mddev), mddev->raid_disks - mddev->degraded,
3088                 mddev->raid_disks);
3089
3090         /*
3091          * Ok, everything is just fine now
3092          */
3093         mddev->thread = conf->thread;
3094         conf->thread = NULL;
3095         mddev->private = conf;
3096         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3097
3098         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3099
3100         if (mddev->queue) {
3101                 if (discard_supported)
3102                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3103                                                 mddev->queue);
3104                 else
3105                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3106                                                   mddev->queue);
3107         }
3108
3109         ret =  md_integrity_register(mddev);
3110         if (ret) {
3111                 md_unregister_thread(&mddev->thread);
3112                 raid1_free(mddev, conf);
3113         }
3114         return ret;
3115 }
3116
3117 static void raid1_free(struct mddev *mddev, void *priv)
3118 {
3119         struct r1conf *conf = priv;
3120
3121         mempool_destroy(conf->r1bio_pool);
3122         kfree(conf->mirrors);
3123         safe_put_page(conf->tmppage);
3124         kfree(conf->poolinfo);
3125         kfree(conf->nr_pending);
3126         kfree(conf->nr_waiting);
3127         kfree(conf->nr_queued);
3128         kfree(conf->barrier);
3129         if (conf->bio_split)
3130                 bioset_free(conf->bio_split);
3131         kfree(conf);
3132 }
3133
3134 static int raid1_resize(struct mddev *mddev, sector_t sectors)
3135 {
3136         /* no resync is happening, and there is enough space
3137          * on all devices, so we can resize.
3138          * We need to make sure resync covers any new space.
3139          * If the array is shrinking we should possibly wait until
3140          * any io in the removed space completes, but it hardly seems
3141          * worth it.
3142          */
3143         sector_t newsize = raid1_size(mddev, sectors, 0);
3144         if (mddev->external_size &&
3145             mddev->array_sectors > newsize)
3146                 return -EINVAL;
3147         if (mddev->bitmap) {
3148                 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3149                 if (ret)
3150                         return ret;
3151         }
3152         md_set_array_sectors(mddev, newsize);
3153         if (sectors > mddev->dev_sectors &&
3154             mddev->recovery_cp > mddev->dev_sectors) {
3155                 mddev->recovery_cp = mddev->dev_sectors;
3156                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3157         }
3158         mddev->dev_sectors = sectors;
3159         mddev->resync_max_sectors = sectors;
3160         return 0;
3161 }
3162
3163 static int raid1_reshape(struct mddev *mddev)
3164 {
3165         /* We need to:
3166          * 1/ resize the r1bio_pool
3167          * 2/ resize conf->mirrors
3168          *
3169          * We allocate a new r1bio_pool if we can.
3170          * Then raise a device barrier and wait until all IO stops.
3171          * Then resize conf->mirrors and swap in the new r1bio pool.
3172          *
3173          * At the same time, we "pack" the devices so that all the missing
3174          * devices have the higher raid_disk numbers.
3175          */
3176         mempool_t *newpool, *oldpool;
3177         struct pool_info *newpoolinfo;
3178         struct raid1_info *newmirrors;
3179         struct r1conf *conf = mddev->private;
3180         int cnt, raid_disks;
3181         unsigned long flags;
3182         int d, d2;
3183
3184         /* Cannot change chunk_size, layout, or level */
3185         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3186             mddev->layout != mddev->new_layout ||
3187             mddev->level != mddev->new_level) {
3188                 mddev->new_chunk_sectors = mddev->chunk_sectors;
3189                 mddev->new_layout = mddev->layout;
3190                 mddev->new_level = mddev->level;
3191                 return -EINVAL;
3192         }
3193
3194         if (!mddev_is_clustered(mddev))
3195                 md_allow_write(mddev);
3196
3197         raid_disks = mddev->raid_disks + mddev->delta_disks;
3198
3199         if (raid_disks < conf->raid_disks) {
3200                 cnt=0;
3201                 for (d= 0; d < conf->raid_disks; d++)
3202                         if (conf->mirrors[d].rdev)
3203                                 cnt++;
3204                 if (cnt > raid_disks)
3205                         return -EBUSY;
3206         }
3207
3208         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3209         if (!newpoolinfo)
3210                 return -ENOMEM;
3211         newpoolinfo->mddev = mddev;
3212         newpoolinfo->raid_disks = raid_disks * 2;
3213
3214         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3215                                  r1bio_pool_free, newpoolinfo);
3216         if (!newpool) {
3217                 kfree(newpoolinfo);
3218                 return -ENOMEM;
3219         }
3220         newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3221                              GFP_KERNEL);
3222         if (!newmirrors) {
3223                 kfree(newpoolinfo);
3224                 mempool_destroy(newpool);
3225                 return -ENOMEM;
3226         }
3227
3228         freeze_array(conf, 0);
3229
3230         /* ok, everything is stopped */
3231         oldpool = conf->r1bio_pool;
3232         conf->r1bio_pool = newpool;
3233
3234         for (d = d2 = 0; d < conf->raid_disks; d++) {
3235                 struct md_rdev *rdev = conf->mirrors[d].rdev;
3236                 if (rdev && rdev->raid_disk != d2) {
3237                         sysfs_unlink_rdev(mddev, rdev);
3238                         rdev->raid_disk = d2;
3239                         sysfs_unlink_rdev(mddev, rdev);
3240                         if (sysfs_link_rdev(mddev, rdev))
3241                                 pr_warn("md/raid1:%s: cannot register rd%d\n",
3242                                         mdname(mddev), rdev->raid_disk);
3243                 }
3244                 if (rdev)
3245                         newmirrors[d2++].rdev = rdev;
3246         }
3247         kfree(conf->mirrors);
3248         conf->mirrors = newmirrors;
3249         kfree(conf->poolinfo);
3250         conf->poolinfo = newpoolinfo;
3251
3252         spin_lock_irqsave(&conf->device_lock, flags);
3253         mddev->degraded += (raid_disks - conf->raid_disks);
3254         spin_unlock_irqrestore(&conf->device_lock, flags);
3255         conf->raid_disks = mddev->raid_disks = raid_disks;
3256         mddev->delta_disks = 0;
3257
3258         unfreeze_array(conf);
3259
3260         set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3261         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3262         md_wakeup_thread(mddev->thread);
3263
3264         mempool_destroy(oldpool);
3265         return 0;
3266 }
3267
3268 static void raid1_quiesce(struct mddev *mddev, int quiesce)
3269 {
3270         struct r1conf *conf = mddev->private;
3271
3272         if (quiesce)
3273                 freeze_array(conf, 0);
3274         else
3275                 unfreeze_array(conf);
3276 }
3277
3278 static void *raid1_takeover(struct mddev *mddev)
3279 {
3280         /* raid1 can take over:
3281          *  raid5 with 2 devices, any layout or chunk size
3282          */
3283         if (mddev->level == 5 && mddev->raid_disks == 2) {
3284                 struct r1conf *conf;
3285                 mddev->new_level = 1;
3286                 mddev->new_layout = 0;
3287                 mddev->new_chunk_sectors = 0;
3288                 conf = setup_conf(mddev);
3289                 if (!IS_ERR(conf)) {
3290                         /* Array must appear to be quiesced */
3291                         conf->array_frozen = 1;
3292                         mddev_clear_unsupported_flags(mddev,
3293                                 UNSUPPORTED_MDDEV_FLAGS);
3294                 }
3295                 return conf;
3296         }
3297         return ERR_PTR(-EINVAL);
3298 }
3299
3300 static struct md_personality raid1_personality =
3301 {
3302         .name           = "raid1",
3303         .level          = 1,
3304         .owner          = THIS_MODULE,
3305         .make_request   = raid1_make_request,
3306         .run            = raid1_run,
3307         .free           = raid1_free,
3308         .status         = raid1_status,
3309         .error_handler  = raid1_error,
3310         .hot_add_disk   = raid1_add_disk,
3311         .hot_remove_disk= raid1_remove_disk,
3312         .spare_active   = raid1_spare_active,
3313         .sync_request   = raid1_sync_request,
3314         .resize         = raid1_resize,
3315         .size           = raid1_size,
3316         .check_reshape  = raid1_reshape,
3317         .quiesce        = raid1_quiesce,
3318         .takeover       = raid1_takeover,
3319         .congested      = raid1_congested,
3320 };
3321
3322 static int __init raid_init(void)
3323 {
3324         return register_md_personality(&raid1_personality);
3325 }
3326
3327 static void raid_exit(void)
3328 {
3329         unregister_md_personality(&raid1_personality);
3330 }
3331
3332 module_init(raid_init);
3333 module_exit(raid_exit);
3334 MODULE_LICENSE("GPL");
3335 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3336 MODULE_ALIAS("md-personality-3"); /* RAID1 */
3337 MODULE_ALIAS("md-raid1");
3338 MODULE_ALIAS("md-level-1");
3339
3340 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);