raid1/raid10: slow down resync if there is non-resync activity pending
[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 #include "md.h"
41 #include "raid1.h"
42 #include "bitmap.h"
43
44 /*
45  * Number of guaranteed r1bios in case of extreme VM load:
46  */
47 #define NR_RAID1_BIOS 256
48
49 /* when we get a read error on a read-only array, we redirect to another
50  * device without failing the first device, or trying to over-write to
51  * correct the read error.  To keep track of bad blocks on a per-bio
52  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
53  */
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56  * bad-block marking which must be done from process context.  So we record
57  * the success by setting devs[n].bio to IO_MADE_GOOD
58  */
59 #define IO_MADE_GOOD ((struct bio *)2)
60
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
62
63 /* When there are this many requests queue to be written by
64  * the raid1 thread, we become 'congested' to provide back-pressure
65  * for writeback.
66  */
67 static int max_queued_requests = 1024;
68
69 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
70                           sector_t bi_sector);
71 static void lower_barrier(struct r1conf *conf);
72
73 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
74 {
75         struct pool_info *pi = data;
76         int size = offsetof(struct r1bio, bios[pi->raid_disks]);
77
78         /* allocate a r1bio with room for raid_disks entries in the bios array */
79         return kzalloc(size, gfp_flags);
80 }
81
82 static void r1bio_pool_free(void *r1_bio, void *data)
83 {
84         kfree(r1_bio);
85 }
86
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_DEPTH 32
89 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
90 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
91 #define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
92 #define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
93 #define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
94 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
95 #define NEXT_NORMALIO_DISTANCE (3 * RESYNC_WINDOW_SECTORS)
96
97 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
98 {
99         struct pool_info *pi = data;
100         struct r1bio *r1_bio;
101         struct bio *bio;
102         int need_pages;
103         int i, j;
104
105         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
106         if (!r1_bio)
107                 return NULL;
108
109         /*
110          * Allocate bios : 1 for reading, n-1 for writing
111          */
112         for (j = pi->raid_disks ; j-- ; ) {
113                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
114                 if (!bio)
115                         goto out_free_bio;
116                 r1_bio->bios[j] = bio;
117         }
118         /*
119          * Allocate RESYNC_PAGES data pages and attach them to
120          * the first bio.
121          * If this is a user-requested check/repair, allocate
122          * RESYNC_PAGES for each bio.
123          */
124         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
125                 need_pages = pi->raid_disks;
126         else
127                 need_pages = 1;
128         for (j = 0; j < need_pages; j++) {
129                 bio = r1_bio->bios[j];
130                 bio->bi_vcnt = RESYNC_PAGES;
131
132                 if (bio_alloc_pages(bio, gfp_flags))
133                         goto out_free_pages;
134         }
135         /* If not user-requests, copy the page pointers to all bios */
136         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
137                 for (i=0; i<RESYNC_PAGES ; i++)
138                         for (j=1; j<pi->raid_disks; j++)
139                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
140                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
141         }
142
143         r1_bio->master_bio = NULL;
144
145         return r1_bio;
146
147 out_free_pages:
148         while (--j >= 0) {
149                 struct bio_vec *bv;
150
151                 bio_for_each_segment_all(bv, r1_bio->bios[j], i)
152                         __free_page(bv->bv_page);
153         }
154
155 out_free_bio:
156         while (++j < pi->raid_disks)
157                 bio_put(r1_bio->bios[j]);
158         r1bio_pool_free(r1_bio, data);
159         return NULL;
160 }
161
162 static void r1buf_pool_free(void *__r1_bio, void *data)
163 {
164         struct pool_info *pi = data;
165         int i,j;
166         struct r1bio *r1bio = __r1_bio;
167
168         for (i = 0; i < RESYNC_PAGES; i++)
169                 for (j = pi->raid_disks; j-- ;) {
170                         if (j == 0 ||
171                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
172                             r1bio->bios[0]->bi_io_vec[i].bv_page)
173                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
174                 }
175         for (i=0 ; i < pi->raid_disks; i++)
176                 bio_put(r1bio->bios[i]);
177
178         r1bio_pool_free(r1bio, data);
179 }
180
181 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
182 {
183         int i;
184
185         for (i = 0; i < conf->raid_disks * 2; i++) {
186                 struct bio **bio = r1_bio->bios + i;
187                 if (!BIO_SPECIAL(*bio))
188                         bio_put(*bio);
189                 *bio = NULL;
190         }
191 }
192
193 static void free_r1bio(struct r1bio *r1_bio)
194 {
195         struct r1conf *conf = r1_bio->mddev->private;
196
197         put_all_bios(conf, r1_bio);
198         mempool_free(r1_bio, conf->r1bio_pool);
199 }
200
201 static void put_buf(struct r1bio *r1_bio)
202 {
203         struct r1conf *conf = r1_bio->mddev->private;
204         int i;
205
206         for (i = 0; i < conf->raid_disks * 2; i++) {
207                 struct bio *bio = r1_bio->bios[i];
208                 if (bio->bi_end_io)
209                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
210         }
211
212         mempool_free(r1_bio, conf->r1buf_pool);
213
214         lower_barrier(conf);
215 }
216
217 static void reschedule_retry(struct r1bio *r1_bio)
218 {
219         unsigned long flags;
220         struct mddev *mddev = r1_bio->mddev;
221         struct r1conf *conf = mddev->private;
222
223         spin_lock_irqsave(&conf->device_lock, flags);
224         list_add(&r1_bio->retry_list, &conf->retry_list);
225         conf->nr_queued ++;
226         spin_unlock_irqrestore(&conf->device_lock, flags);
227
228         wake_up(&conf->wait_barrier);
229         md_wakeup_thread(mddev->thread);
230 }
231
232 /*
233  * raid_end_bio_io() is called when we have finished servicing a mirrored
234  * operation and are ready to return a success/failure code to the buffer
235  * cache layer.
236  */
237 static void call_bio_endio(struct r1bio *r1_bio)
238 {
239         struct bio *bio = r1_bio->master_bio;
240         int done;
241         struct r1conf *conf = r1_bio->mddev->private;
242         sector_t start_next_window = r1_bio->start_next_window;
243         sector_t bi_sector = bio->bi_iter.bi_sector;
244
245         if (bio->bi_phys_segments) {
246                 unsigned long flags;
247                 spin_lock_irqsave(&conf->device_lock, flags);
248                 bio->bi_phys_segments--;
249                 done = (bio->bi_phys_segments == 0);
250                 spin_unlock_irqrestore(&conf->device_lock, flags);
251                 /*
252                  * make_request() might be waiting for
253                  * bi_phys_segments to decrease
254                  */
255                 wake_up(&conf->wait_barrier);
256         } else
257                 done = 1;
258
259         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
260                 bio->bi_error = -EIO;
261
262         if (done) {
263                 bio_endio(bio);
264                 /*
265                  * Wake up any possible resync thread that waits for the device
266                  * to go idle.
267                  */
268                 allow_barrier(conf, start_next_window, bi_sector);
269         }
270 }
271
272 static void raid_end_bio_io(struct r1bio *r1_bio)
273 {
274         struct bio *bio = r1_bio->master_bio;
275
276         /* if nobody has done the final endio yet, do it now */
277         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
278                 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
279                          (bio_data_dir(bio) == WRITE) ? "write" : "read",
280                          (unsigned long long) bio->bi_iter.bi_sector,
281                          (unsigned long long) bio_end_sector(bio) - 1);
282
283                 call_bio_endio(r1_bio);
284         }
285         free_r1bio(r1_bio);
286 }
287
288 /*
289  * Update disk head position estimator based on IRQ completion info.
290  */
291 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
292 {
293         struct r1conf *conf = r1_bio->mddev->private;
294
295         conf->mirrors[disk].head_position =
296                 r1_bio->sector + (r1_bio->sectors);
297 }
298
299 /*
300  * Find the disk number which triggered given bio
301  */
302 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
303 {
304         int mirror;
305         struct r1conf *conf = r1_bio->mddev->private;
306         int raid_disks = conf->raid_disks;
307
308         for (mirror = 0; mirror < raid_disks * 2; mirror++)
309                 if (r1_bio->bios[mirror] == bio)
310                         break;
311
312         BUG_ON(mirror == raid_disks * 2);
313         update_head_pos(mirror, r1_bio);
314
315         return mirror;
316 }
317
318 static void raid1_end_read_request(struct bio *bio)
319 {
320         int uptodate = !bio->bi_error;
321         struct r1bio *r1_bio = bio->bi_private;
322         int mirror;
323         struct r1conf *conf = r1_bio->mddev->private;
324
325         mirror = r1_bio->read_disk;
326         /*
327          * this branch is our 'one mirror IO has finished' event handler:
328          */
329         update_head_pos(mirror, r1_bio);
330
331         if (uptodate)
332                 set_bit(R1BIO_Uptodate, &r1_bio->state);
333         else {
334                 /* If all other devices have failed, we want to return
335                  * the error upwards rather than fail the last device.
336                  * Here we redefine "uptodate" to mean "Don't want to retry"
337                  */
338                 unsigned long flags;
339                 spin_lock_irqsave(&conf->device_lock, flags);
340                 if (r1_bio->mddev->degraded == conf->raid_disks ||
341                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
342                      test_bit(In_sync, &conf->mirrors[mirror].rdev->flags)))
343                         uptodate = 1;
344                 spin_unlock_irqrestore(&conf->device_lock, flags);
345         }
346
347         if (uptodate) {
348                 raid_end_bio_io(r1_bio);
349                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
350         } else {
351                 /*
352                  * oops, read error:
353                  */
354                 char b[BDEVNAME_SIZE];
355                 printk_ratelimited(
356                         KERN_ERR "md/raid1:%s: %s: "
357                         "rescheduling sector %llu\n",
358                         mdname(conf->mddev),
359                         bdevname(conf->mirrors[mirror].rdev->bdev,
360                                  b),
361                         (unsigned long long)r1_bio->sector);
362                 set_bit(R1BIO_ReadError, &r1_bio->state);
363                 reschedule_retry(r1_bio);
364                 /* don't drop the reference on read_disk yet */
365         }
366 }
367
368 static void close_write(struct r1bio *r1_bio)
369 {
370         /* it really is the end of this request */
371         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
372                 /* free extra copy of the data pages */
373                 int i = r1_bio->behind_page_count;
374                 while (i--)
375                         safe_put_page(r1_bio->behind_bvecs[i].bv_page);
376                 kfree(r1_bio->behind_bvecs);
377                 r1_bio->behind_bvecs = NULL;
378         }
379         /* clear the bitmap if all writes complete successfully */
380         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
381                         r1_bio->sectors,
382                         !test_bit(R1BIO_Degraded, &r1_bio->state),
383                         test_bit(R1BIO_BehindIO, &r1_bio->state));
384         md_write_end(r1_bio->mddev);
385 }
386
387 static void r1_bio_write_done(struct r1bio *r1_bio)
388 {
389         if (!atomic_dec_and_test(&r1_bio->remaining))
390                 return;
391
392         if (test_bit(R1BIO_WriteError, &r1_bio->state))
393                 reschedule_retry(r1_bio);
394         else {
395                 close_write(r1_bio);
396                 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
397                         reschedule_retry(r1_bio);
398                 else
399                         raid_end_bio_io(r1_bio);
400         }
401 }
402
403 static void raid1_end_write_request(struct bio *bio)
404 {
405         struct r1bio *r1_bio = bio->bi_private;
406         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
407         struct r1conf *conf = r1_bio->mddev->private;
408         struct bio *to_put = NULL;
409
410         mirror = find_bio_disk(r1_bio, bio);
411
412         /*
413          * 'one mirror IO has finished' event handler:
414          */
415         if (bio->bi_error) {
416                 set_bit(WriteErrorSeen,
417                         &conf->mirrors[mirror].rdev->flags);
418                 if (!test_and_set_bit(WantReplacement,
419                                       &conf->mirrors[mirror].rdev->flags))
420                         set_bit(MD_RECOVERY_NEEDED, &
421                                 conf->mddev->recovery);
422
423                 set_bit(R1BIO_WriteError, &r1_bio->state);
424         } else {
425                 /*
426                  * Set R1BIO_Uptodate in our master bio, so that we
427                  * will return a good error code for to the higher
428                  * levels even if IO on some other mirrored buffer
429                  * fails.
430                  *
431                  * The 'master' represents the composite IO operation
432                  * to user-side. So if something waits for IO, then it
433                  * will wait for the 'master' bio.
434                  */
435                 sector_t first_bad;
436                 int bad_sectors;
437
438                 r1_bio->bios[mirror] = NULL;
439                 to_put = bio;
440                 /*
441                  * Do not set R1BIO_Uptodate if the current device is
442                  * rebuilding or Faulty. This is because we cannot use
443                  * such device for properly reading the data back (we could
444                  * potentially use it, if the current write would have felt
445                  * before rdev->recovery_offset, but for simplicity we don't
446                  * check this here.
447                  */
448                 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
449                     !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
450                         set_bit(R1BIO_Uptodate, &r1_bio->state);
451
452                 /* Maybe we can clear some bad blocks. */
453                 if (is_badblock(conf->mirrors[mirror].rdev,
454                                 r1_bio->sector, r1_bio->sectors,
455                                 &first_bad, &bad_sectors)) {
456                         r1_bio->bios[mirror] = IO_MADE_GOOD;
457                         set_bit(R1BIO_MadeGood, &r1_bio->state);
458                 }
459         }
460
461         if (behind) {
462                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
463                         atomic_dec(&r1_bio->behind_remaining);
464
465                 /*
466                  * In behind mode, we ACK the master bio once the I/O
467                  * has safely reached all non-writemostly
468                  * disks. Setting the Returned bit ensures that this
469                  * gets done only once -- we don't ever want to return
470                  * -EIO here, instead we'll wait
471                  */
472                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
473                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
474                         /* Maybe we can return now */
475                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
476                                 struct bio *mbio = r1_bio->master_bio;
477                                 pr_debug("raid1: behind end write sectors"
478                                          " %llu-%llu\n",
479                                          (unsigned long long) mbio->bi_iter.bi_sector,
480                                          (unsigned long long) bio_end_sector(mbio) - 1);
481                                 call_bio_endio(r1_bio);
482                         }
483                 }
484         }
485         if (r1_bio->bios[mirror] == NULL)
486                 rdev_dec_pending(conf->mirrors[mirror].rdev,
487                                  conf->mddev);
488
489         /*
490          * Let's see if all mirrored write operations have finished
491          * already.
492          */
493         r1_bio_write_done(r1_bio);
494
495         if (to_put)
496                 bio_put(to_put);
497 }
498
499 /*
500  * This routine returns the disk from which the requested read should
501  * be done. There is a per-array 'next expected sequential IO' sector
502  * number - if this matches on the next IO then we use the last disk.
503  * There is also a per-disk 'last know head position' sector that is
504  * maintained from IRQ contexts, both the normal and the resync IO
505  * completion handlers update this position correctly. If there is no
506  * perfect sequential match then we pick the disk whose head is closest.
507  *
508  * If there are 2 mirrors in the same 2 devices, performance degrades
509  * because position is mirror, not device based.
510  *
511  * The rdev for the device selected will have nr_pending incremented.
512  */
513 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
514 {
515         const sector_t this_sector = r1_bio->sector;
516         int sectors;
517         int best_good_sectors;
518         int best_disk, best_dist_disk, best_pending_disk;
519         int has_nonrot_disk;
520         int disk;
521         sector_t best_dist;
522         unsigned int min_pending;
523         struct md_rdev *rdev;
524         int choose_first;
525         int choose_next_idle;
526
527         rcu_read_lock();
528         /*
529          * Check if we can balance. We can balance on the whole
530          * device if no resync is going on, or below the resync window.
531          * We take the first readable disk when above the resync window.
532          */
533  retry:
534         sectors = r1_bio->sectors;
535         best_disk = -1;
536         best_dist_disk = -1;
537         best_dist = MaxSector;
538         best_pending_disk = -1;
539         min_pending = UINT_MAX;
540         best_good_sectors = 0;
541         has_nonrot_disk = 0;
542         choose_next_idle = 0;
543
544         if ((conf->mddev->recovery_cp < this_sector + sectors) ||
545             (mddev_is_clustered(conf->mddev) &&
546             md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
547                     this_sector + sectors)))
548                 choose_first = 1;
549         else
550                 choose_first = 0;
551
552         for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
553                 sector_t dist;
554                 sector_t first_bad;
555                 int bad_sectors;
556                 unsigned int pending;
557                 bool nonrot;
558
559                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
560                 if (r1_bio->bios[disk] == IO_BLOCKED
561                     || rdev == NULL
562                     || test_bit(Faulty, &rdev->flags))
563                         continue;
564                 if (!test_bit(In_sync, &rdev->flags) &&
565                     rdev->recovery_offset < this_sector + sectors)
566                         continue;
567                 if (test_bit(WriteMostly, &rdev->flags)) {
568                         /* Don't balance among write-mostly, just
569                          * use the first as a last resort */
570                         if (best_dist_disk < 0) {
571                                 if (is_badblock(rdev, this_sector, sectors,
572                                                 &first_bad, &bad_sectors)) {
573                                         if (first_bad <= this_sector)
574                                                 /* Cannot use this */
575                                                 continue;
576                                         best_good_sectors = first_bad - this_sector;
577                                 } else
578                                         best_good_sectors = sectors;
579                                 best_dist_disk = disk;
580                                 best_pending_disk = disk;
581                         }
582                         continue;
583                 }
584                 /* This is a reasonable device to use.  It might
585                  * even be best.
586                  */
587                 if (is_badblock(rdev, this_sector, sectors,
588                                 &first_bad, &bad_sectors)) {
589                         if (best_dist < MaxSector)
590                                 /* already have a better device */
591                                 continue;
592                         if (first_bad <= this_sector) {
593                                 /* cannot read here. If this is the 'primary'
594                                  * device, then we must not read beyond
595                                  * bad_sectors from another device..
596                                  */
597                                 bad_sectors -= (this_sector - first_bad);
598                                 if (choose_first && sectors > bad_sectors)
599                                         sectors = bad_sectors;
600                                 if (best_good_sectors > sectors)
601                                         best_good_sectors = sectors;
602
603                         } else {
604                                 sector_t good_sectors = first_bad - this_sector;
605                                 if (good_sectors > best_good_sectors) {
606                                         best_good_sectors = good_sectors;
607                                         best_disk = disk;
608                                 }
609                                 if (choose_first)
610                                         break;
611                         }
612                         continue;
613                 } else
614                         best_good_sectors = sectors;
615
616                 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
617                 has_nonrot_disk |= nonrot;
618                 pending = atomic_read(&rdev->nr_pending);
619                 dist = abs(this_sector - conf->mirrors[disk].head_position);
620                 if (choose_first) {
621                         best_disk = disk;
622                         break;
623                 }
624                 /* Don't change to another disk for sequential reads */
625                 if (conf->mirrors[disk].next_seq_sect == this_sector
626                     || dist == 0) {
627                         int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
628                         struct raid1_info *mirror = &conf->mirrors[disk];
629
630                         best_disk = disk;
631                         /*
632                          * If buffered sequential IO size exceeds optimal
633                          * iosize, check if there is idle disk. If yes, choose
634                          * the idle disk. read_balance could already choose an
635                          * idle disk before noticing it's a sequential IO in
636                          * this disk. This doesn't matter because this disk
637                          * will idle, next time it will be utilized after the
638                          * first disk has IO size exceeds optimal iosize. In
639                          * this way, iosize of the first disk will be optimal
640                          * iosize at least. iosize of the second disk might be
641                          * small, but not a big deal since when the second disk
642                          * starts IO, the first disk is likely still busy.
643                          */
644                         if (nonrot && opt_iosize > 0 &&
645                             mirror->seq_start != MaxSector &&
646                             mirror->next_seq_sect > opt_iosize &&
647                             mirror->next_seq_sect - opt_iosize >=
648                             mirror->seq_start) {
649                                 choose_next_idle = 1;
650                                 continue;
651                         }
652                         break;
653                 }
654                 /* If device is idle, use it */
655                 if (pending == 0) {
656                         best_disk = disk;
657                         break;
658                 }
659
660                 if (choose_next_idle)
661                         continue;
662
663                 if (min_pending > pending) {
664                         min_pending = pending;
665                         best_pending_disk = disk;
666                 }
667
668                 if (dist < best_dist) {
669                         best_dist = dist;
670                         best_dist_disk = disk;
671                 }
672         }
673
674         /*
675          * If all disks are rotational, choose the closest disk. If any disk is
676          * non-rotational, choose the disk with less pending request even the
677          * disk is rotational, which might/might not be optimal for raids with
678          * mixed ratation/non-rotational disks depending on workload.
679          */
680         if (best_disk == -1) {
681                 if (has_nonrot_disk)
682                         best_disk = best_pending_disk;
683                 else
684                         best_disk = best_dist_disk;
685         }
686
687         if (best_disk >= 0) {
688                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
689                 if (!rdev)
690                         goto retry;
691                 atomic_inc(&rdev->nr_pending);
692                 if (test_bit(Faulty, &rdev->flags)) {
693                         /* cannot risk returning a device that failed
694                          * before we inc'ed nr_pending
695                          */
696                         rdev_dec_pending(rdev, conf->mddev);
697                         goto retry;
698                 }
699                 sectors = best_good_sectors;
700
701                 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
702                         conf->mirrors[best_disk].seq_start = this_sector;
703
704                 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
705         }
706         rcu_read_unlock();
707         *max_sectors = sectors;
708
709         return best_disk;
710 }
711
712 static int raid1_congested(struct mddev *mddev, int bits)
713 {
714         struct r1conf *conf = mddev->private;
715         int i, ret = 0;
716
717         if ((bits & (1 << WB_async_congested)) &&
718             conf->pending_count >= max_queued_requests)
719                 return 1;
720
721         rcu_read_lock();
722         for (i = 0; i < conf->raid_disks * 2; i++) {
723                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
724                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
725                         struct request_queue *q = bdev_get_queue(rdev->bdev);
726
727                         BUG_ON(!q);
728
729                         /* Note the '|| 1' - when read_balance prefers
730                          * non-congested targets, it can be removed
731                          */
732                         if ((bits & (1 << WB_async_congested)) || 1)
733                                 ret |= bdi_congested(&q->backing_dev_info, bits);
734                         else
735                                 ret &= bdi_congested(&q->backing_dev_info, bits);
736                 }
737         }
738         rcu_read_unlock();
739         return ret;
740 }
741
742 static void flush_pending_writes(struct r1conf *conf)
743 {
744         /* Any writes that have been queued but are awaiting
745          * bitmap updates get flushed here.
746          */
747         spin_lock_irq(&conf->device_lock);
748
749         if (conf->pending_bio_list.head) {
750                 struct bio *bio;
751                 bio = bio_list_get(&conf->pending_bio_list);
752                 conf->pending_count = 0;
753                 spin_unlock_irq(&conf->device_lock);
754                 /* flush any pending bitmap writes to
755                  * disk before proceeding w/ I/O */
756                 bitmap_unplug(conf->mddev->bitmap);
757                 wake_up(&conf->wait_barrier);
758
759                 while (bio) { /* submit pending writes */
760                         struct bio *next = bio->bi_next;
761                         bio->bi_next = NULL;
762                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
763                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
764                                 /* Just ignore it */
765                                 bio_endio(bio);
766                         else
767                                 generic_make_request(bio);
768                         bio = next;
769                 }
770         } else
771                 spin_unlock_irq(&conf->device_lock);
772 }
773
774 /* Barriers....
775  * Sometimes we need to suspend IO while we do something else,
776  * either some resync/recovery, or reconfigure the array.
777  * To do this we raise a 'barrier'.
778  * The 'barrier' is a counter that can be raised multiple times
779  * to count how many activities are happening which preclude
780  * normal IO.
781  * We can only raise the barrier if there is no pending IO.
782  * i.e. if nr_pending == 0.
783  * We choose only to raise the barrier if no-one is waiting for the
784  * barrier to go down.  This means that as soon as an IO request
785  * is ready, no other operations which require a barrier will start
786  * until the IO request has had a chance.
787  *
788  * So: regular IO calls 'wait_barrier'.  When that returns there
789  *    is no backgroup IO happening,  It must arrange to call
790  *    allow_barrier when it has finished its IO.
791  * backgroup IO calls must call raise_barrier.  Once that returns
792  *    there is no normal IO happeing.  It must arrange to call
793  *    lower_barrier when the particular background IO completes.
794  */
795 static void raise_barrier(struct r1conf *conf, sector_t sector_nr)
796 {
797         spin_lock_irq(&conf->resync_lock);
798
799         /* Wait until no block IO is waiting */
800         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
801                             conf->resync_lock);
802
803         /* block any new IO from starting */
804         conf->barrier++;
805         conf->next_resync = sector_nr;
806
807         /* For these conditions we must wait:
808          * A: while the array is in frozen state
809          * B: while barrier >= RESYNC_DEPTH, meaning resync reach
810          *    the max count which allowed.
811          * C: next_resync + RESYNC_SECTORS > start_next_window, meaning
812          *    next resync will reach to the window which normal bios are
813          *    handling.
814          * D: while there are any active requests in the current window.
815          */
816         wait_event_lock_irq(conf->wait_barrier,
817                             !conf->array_frozen &&
818                             conf->barrier < RESYNC_DEPTH &&
819                             conf->current_window_requests == 0 &&
820                             (conf->start_next_window >=
821                              conf->next_resync + RESYNC_SECTORS),
822                             conf->resync_lock);
823
824         conf->nr_pending++;
825         spin_unlock_irq(&conf->resync_lock);
826 }
827
828 static void lower_barrier(struct r1conf *conf)
829 {
830         unsigned long flags;
831         BUG_ON(conf->barrier <= 0);
832         spin_lock_irqsave(&conf->resync_lock, flags);
833         conf->barrier--;
834         conf->nr_pending--;
835         spin_unlock_irqrestore(&conf->resync_lock, flags);
836         wake_up(&conf->wait_barrier);
837 }
838
839 static bool need_to_wait_for_sync(struct r1conf *conf, struct bio *bio)
840 {
841         bool wait = false;
842
843         if (conf->array_frozen || !bio)
844                 wait = true;
845         else if (conf->barrier && bio_data_dir(bio) == WRITE) {
846                 if ((conf->mddev->curr_resync_completed
847                      >= bio_end_sector(bio)) ||
848                     (conf->next_resync + NEXT_NORMALIO_DISTANCE
849                      <= bio->bi_iter.bi_sector))
850                         wait = false;
851                 else
852                         wait = true;
853         }
854
855         return wait;
856 }
857
858 static sector_t wait_barrier(struct r1conf *conf, struct bio *bio)
859 {
860         sector_t sector = 0;
861
862         spin_lock_irq(&conf->resync_lock);
863         if (need_to_wait_for_sync(conf, bio)) {
864                 conf->nr_waiting++;
865                 /* Wait for the barrier to drop.
866                  * However if there are already pending
867                  * requests (preventing the barrier from
868                  * rising completely), and the
869                  * per-process bio queue isn't empty,
870                  * then don't wait, as we need to empty
871                  * that queue to allow conf->start_next_window
872                  * to increase.
873                  */
874                 wait_event_lock_irq(conf->wait_barrier,
875                                     !conf->array_frozen &&
876                                     (!conf->barrier ||
877                                      ((conf->start_next_window <
878                                        conf->next_resync + RESYNC_SECTORS) &&
879                                       current->bio_list &&
880                                       !bio_list_empty(current->bio_list))),
881                                     conf->resync_lock);
882                 conf->nr_waiting--;
883         }
884
885         if (bio && bio_data_dir(bio) == WRITE) {
886                 if (bio->bi_iter.bi_sector >= conf->next_resync) {
887                         if (conf->start_next_window == MaxSector)
888                                 conf->start_next_window =
889                                         conf->next_resync +
890                                         NEXT_NORMALIO_DISTANCE;
891
892                         if ((conf->start_next_window + NEXT_NORMALIO_DISTANCE)
893                             <= bio->bi_iter.bi_sector)
894                                 conf->next_window_requests++;
895                         else
896                                 conf->current_window_requests++;
897                         sector = conf->start_next_window;
898                 }
899         }
900
901         conf->nr_pending++;
902         spin_unlock_irq(&conf->resync_lock);
903         return sector;
904 }
905
906 static void allow_barrier(struct r1conf *conf, sector_t start_next_window,
907                           sector_t bi_sector)
908 {
909         unsigned long flags;
910
911         spin_lock_irqsave(&conf->resync_lock, flags);
912         conf->nr_pending--;
913         if (start_next_window) {
914                 if (start_next_window == conf->start_next_window) {
915                         if (conf->start_next_window + NEXT_NORMALIO_DISTANCE
916                             <= bi_sector)
917                                 conf->next_window_requests--;
918                         else
919                                 conf->current_window_requests--;
920                 } else
921                         conf->current_window_requests--;
922
923                 if (!conf->current_window_requests) {
924                         if (conf->next_window_requests) {
925                                 conf->current_window_requests =
926                                         conf->next_window_requests;
927                                 conf->next_window_requests = 0;
928                                 conf->start_next_window +=
929                                         NEXT_NORMALIO_DISTANCE;
930                         } else
931                                 conf->start_next_window = MaxSector;
932                 }
933         }
934         spin_unlock_irqrestore(&conf->resync_lock, flags);
935         wake_up(&conf->wait_barrier);
936 }
937
938 static void freeze_array(struct r1conf *conf, int extra)
939 {
940         /* stop syncio and normal IO and wait for everything to
941          * go quite.
942          * We wait until nr_pending match nr_queued+extra
943          * This is called in the context of one normal IO request
944          * that has failed. Thus any sync request that might be pending
945          * will be blocked by nr_pending, and we need to wait for
946          * pending IO requests to complete or be queued for re-try.
947          * Thus the number queued (nr_queued) plus this request (extra)
948          * must match the number of pending IOs (nr_pending) before
949          * we continue.
950          */
951         spin_lock_irq(&conf->resync_lock);
952         conf->array_frozen = 1;
953         wait_event_lock_irq_cmd(conf->wait_barrier,
954                                 conf->nr_pending == conf->nr_queued+extra,
955                                 conf->resync_lock,
956                                 flush_pending_writes(conf));
957         spin_unlock_irq(&conf->resync_lock);
958 }
959 static void unfreeze_array(struct r1conf *conf)
960 {
961         /* reverse the effect of the freeze */
962         spin_lock_irq(&conf->resync_lock);
963         conf->array_frozen = 0;
964         wake_up(&conf->wait_barrier);
965         spin_unlock_irq(&conf->resync_lock);
966 }
967
968 /* duplicate the data pages for behind I/O
969  */
970 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
971 {
972         int i;
973         struct bio_vec *bvec;
974         struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
975                                         GFP_NOIO);
976         if (unlikely(!bvecs))
977                 return;
978
979         bio_for_each_segment_all(bvec, bio, i) {
980                 bvecs[i] = *bvec;
981                 bvecs[i].bv_page = alloc_page(GFP_NOIO);
982                 if (unlikely(!bvecs[i].bv_page))
983                         goto do_sync_io;
984                 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
985                        kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
986                 kunmap(bvecs[i].bv_page);
987                 kunmap(bvec->bv_page);
988         }
989         r1_bio->behind_bvecs = bvecs;
990         r1_bio->behind_page_count = bio->bi_vcnt;
991         set_bit(R1BIO_BehindIO, &r1_bio->state);
992         return;
993
994 do_sync_io:
995         for (i = 0; i < bio->bi_vcnt; i++)
996                 if (bvecs[i].bv_page)
997                         put_page(bvecs[i].bv_page);
998         kfree(bvecs);
999         pr_debug("%dB behind alloc failed, doing sync I/O\n",
1000                  bio->bi_iter.bi_size);
1001 }
1002
1003 struct raid1_plug_cb {
1004         struct blk_plug_cb      cb;
1005         struct bio_list         pending;
1006         int                     pending_cnt;
1007 };
1008
1009 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1010 {
1011         struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1012                                                   cb);
1013         struct mddev *mddev = plug->cb.data;
1014         struct r1conf *conf = mddev->private;
1015         struct bio *bio;
1016
1017         if (from_schedule || current->bio_list) {
1018                 spin_lock_irq(&conf->device_lock);
1019                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1020                 conf->pending_count += plug->pending_cnt;
1021                 spin_unlock_irq(&conf->device_lock);
1022                 wake_up(&conf->wait_barrier);
1023                 md_wakeup_thread(mddev->thread);
1024                 kfree(plug);
1025                 return;
1026         }
1027
1028         /* we aren't scheduling, so we can do the write-out directly. */
1029         bio = bio_list_get(&plug->pending);
1030         bitmap_unplug(mddev->bitmap);
1031         wake_up(&conf->wait_barrier);
1032
1033         while (bio) { /* submit pending writes */
1034                 struct bio *next = bio->bi_next;
1035                 bio->bi_next = NULL;
1036                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1037                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1038                         /* Just ignore it */
1039                         bio_endio(bio);
1040                 else
1041                         generic_make_request(bio);
1042                 bio = next;
1043         }
1044         kfree(plug);
1045 }
1046
1047 static void raid1_make_request(struct mddev *mddev, struct bio * bio)
1048 {
1049         struct r1conf *conf = mddev->private;
1050         struct raid1_info *mirror;
1051         struct r1bio *r1_bio;
1052         struct bio *read_bio;
1053         int i, disks;
1054         struct bitmap *bitmap;
1055         unsigned long flags;
1056         const int rw = bio_data_dir(bio);
1057         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1058         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1059         const unsigned long do_discard = (bio->bi_rw
1060                                           & (REQ_DISCARD | REQ_SECURE));
1061         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1062         struct md_rdev *blocked_rdev;
1063         struct blk_plug_cb *cb;
1064         struct raid1_plug_cb *plug = NULL;
1065         int first_clone;
1066         int sectors_handled;
1067         int max_sectors;
1068         sector_t start_next_window;
1069
1070         /*
1071          * Register the new request and wait if the reconstruction
1072          * thread has put up a bar for new requests.
1073          * Continue immediately if no resync is active currently.
1074          */
1075
1076         md_write_start(mddev, bio); /* wait on superblock update early */
1077
1078         if (bio_data_dir(bio) == WRITE &&
1079             ((bio_end_sector(bio) > mddev->suspend_lo &&
1080             bio->bi_iter.bi_sector < mddev->suspend_hi) ||
1081             (mddev_is_clustered(mddev) &&
1082              md_cluster_ops->area_resyncing(mddev, WRITE,
1083                      bio->bi_iter.bi_sector, bio_end_sector(bio))))) {
1084                 /* As the suspend_* range is controlled by
1085                  * userspace, we want an interruptible
1086                  * wait.
1087                  */
1088                 DEFINE_WAIT(w);
1089                 for (;;) {
1090                         flush_signals(current);
1091                         prepare_to_wait(&conf->wait_barrier,
1092                                         &w, TASK_INTERRUPTIBLE);
1093                         if (bio_end_sector(bio) <= mddev->suspend_lo ||
1094                             bio->bi_iter.bi_sector >= mddev->suspend_hi ||
1095                             (mddev_is_clustered(mddev) &&
1096                              !md_cluster_ops->area_resyncing(mddev, WRITE,
1097                                      bio->bi_iter.bi_sector, bio_end_sector(bio))))
1098                                 break;
1099                         schedule();
1100                 }
1101                 finish_wait(&conf->wait_barrier, &w);
1102         }
1103
1104         start_next_window = wait_barrier(conf, bio);
1105
1106         bitmap = mddev->bitmap;
1107
1108         /*
1109          * make_request() can abort the operation when READA is being
1110          * used and no empty request is available.
1111          *
1112          */
1113         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1114
1115         r1_bio->master_bio = bio;
1116         r1_bio->sectors = bio_sectors(bio);
1117         r1_bio->state = 0;
1118         r1_bio->mddev = mddev;
1119         r1_bio->sector = bio->bi_iter.bi_sector;
1120
1121         /* We might need to issue multiple reads to different
1122          * devices if there are bad blocks around, so we keep
1123          * track of the number of reads in bio->bi_phys_segments.
1124          * If this is 0, there is only one r1_bio and no locking
1125          * will be needed when requests complete.  If it is
1126          * non-zero, then it is the number of not-completed requests.
1127          */
1128         bio->bi_phys_segments = 0;
1129         bio_clear_flag(bio, BIO_SEG_VALID);
1130
1131         if (rw == READ) {
1132                 /*
1133                  * read balancing logic:
1134                  */
1135                 int rdisk;
1136
1137 read_again:
1138                 rdisk = read_balance(conf, r1_bio, &max_sectors);
1139
1140                 if (rdisk < 0) {
1141                         /* couldn't find anywhere to read from */
1142                         raid_end_bio_io(r1_bio);
1143                         return;
1144                 }
1145                 mirror = conf->mirrors + rdisk;
1146
1147                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1148                     bitmap) {
1149                         /* Reading from a write-mostly device must
1150                          * take care not to over-take any writes
1151                          * that are 'behind'
1152                          */
1153                         wait_event(bitmap->behind_wait,
1154                                    atomic_read(&bitmap->behind_writes) == 0);
1155                 }
1156                 r1_bio->read_disk = rdisk;
1157                 r1_bio->start_next_window = 0;
1158
1159                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1160                 bio_trim(read_bio, r1_bio->sector - bio->bi_iter.bi_sector,
1161                          max_sectors);
1162
1163                 r1_bio->bios[rdisk] = read_bio;
1164
1165                 read_bio->bi_iter.bi_sector = r1_bio->sector +
1166                         mirror->rdev->data_offset;
1167                 read_bio->bi_bdev = mirror->rdev->bdev;
1168                 read_bio->bi_end_io = raid1_end_read_request;
1169                 read_bio->bi_rw = READ | do_sync;
1170                 read_bio->bi_private = r1_bio;
1171
1172                 if (max_sectors < r1_bio->sectors) {
1173                         /* could not read all from this device, so we will
1174                          * need another r1_bio.
1175                          */
1176
1177                         sectors_handled = (r1_bio->sector + max_sectors
1178                                            - bio->bi_iter.bi_sector);
1179                         r1_bio->sectors = max_sectors;
1180                         spin_lock_irq(&conf->device_lock);
1181                         if (bio->bi_phys_segments == 0)
1182                                 bio->bi_phys_segments = 2;
1183                         else
1184                                 bio->bi_phys_segments++;
1185                         spin_unlock_irq(&conf->device_lock);
1186                         /* Cannot call generic_make_request directly
1187                          * as that will be queued in __make_request
1188                          * and subsequent mempool_alloc might block waiting
1189                          * for it.  So hand bio over to raid1d.
1190                          */
1191                         reschedule_retry(r1_bio);
1192
1193                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1194
1195                         r1_bio->master_bio = bio;
1196                         r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1197                         r1_bio->state = 0;
1198                         r1_bio->mddev = mddev;
1199                         r1_bio->sector = bio->bi_iter.bi_sector +
1200                                 sectors_handled;
1201                         goto read_again;
1202                 } else
1203                         generic_make_request(read_bio);
1204                 return;
1205         }
1206
1207         /*
1208          * WRITE:
1209          */
1210         if (conf->pending_count >= max_queued_requests) {
1211                 md_wakeup_thread(mddev->thread);
1212                 wait_event(conf->wait_barrier,
1213                            conf->pending_count < max_queued_requests);
1214         }
1215         /* first select target devices under rcu_lock and
1216          * inc refcount on their rdev.  Record them by setting
1217          * bios[x] to bio
1218          * If there are known/acknowledged bad blocks on any device on
1219          * which we have seen a write error, we want to avoid writing those
1220          * blocks.
1221          * This potentially requires several writes to write around
1222          * the bad blocks.  Each set of writes gets it's own r1bio
1223          * with a set of bios attached.
1224          */
1225
1226         disks = conf->raid_disks * 2;
1227  retry_write:
1228         r1_bio->start_next_window = start_next_window;
1229         blocked_rdev = NULL;
1230         rcu_read_lock();
1231         max_sectors = r1_bio->sectors;
1232         for (i = 0;  i < disks; i++) {
1233                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1234                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1235                         atomic_inc(&rdev->nr_pending);
1236                         blocked_rdev = rdev;
1237                         break;
1238                 }
1239                 r1_bio->bios[i] = NULL;
1240                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1241                         if (i < conf->raid_disks)
1242                                 set_bit(R1BIO_Degraded, &r1_bio->state);
1243                         continue;
1244                 }
1245
1246                 atomic_inc(&rdev->nr_pending);
1247                 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1248                         sector_t first_bad;
1249                         int bad_sectors;
1250                         int is_bad;
1251
1252                         is_bad = is_badblock(rdev, r1_bio->sector,
1253                                              max_sectors,
1254                                              &first_bad, &bad_sectors);
1255                         if (is_bad < 0) {
1256                                 /* mustn't write here until the bad block is
1257                                  * acknowledged*/
1258                                 set_bit(BlockedBadBlocks, &rdev->flags);
1259                                 blocked_rdev = rdev;
1260                                 break;
1261                         }
1262                         if (is_bad && first_bad <= r1_bio->sector) {
1263                                 /* Cannot write here at all */
1264                                 bad_sectors -= (r1_bio->sector - first_bad);
1265                                 if (bad_sectors < max_sectors)
1266                                         /* mustn't write more than bad_sectors
1267                                          * to other devices yet
1268                                          */
1269                                         max_sectors = bad_sectors;
1270                                 rdev_dec_pending(rdev, mddev);
1271                                 /* We don't set R1BIO_Degraded as that
1272                                  * only applies if the disk is
1273                                  * missing, so it might be re-added,
1274                                  * and we want to know to recover this
1275                                  * chunk.
1276                                  * In this case the device is here,
1277                                  * and the fact that this chunk is not
1278                                  * in-sync is recorded in the bad
1279                                  * block log
1280                                  */
1281                                 continue;
1282                         }
1283                         if (is_bad) {
1284                                 int good_sectors = first_bad - r1_bio->sector;
1285                                 if (good_sectors < max_sectors)
1286                                         max_sectors = good_sectors;
1287                         }
1288                 }
1289                 r1_bio->bios[i] = bio;
1290         }
1291         rcu_read_unlock();
1292
1293         if (unlikely(blocked_rdev)) {
1294                 /* Wait for this device to become unblocked */
1295                 int j;
1296                 sector_t old = start_next_window;
1297
1298                 for (j = 0; j < i; j++)
1299                         if (r1_bio->bios[j])
1300                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1301                 r1_bio->state = 0;
1302                 allow_barrier(conf, start_next_window, bio->bi_iter.bi_sector);
1303                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1304                 start_next_window = wait_barrier(conf, bio);
1305                 /*
1306                  * We must make sure the multi r1bios of bio have
1307                  * the same value of bi_phys_segments
1308                  */
1309                 if (bio->bi_phys_segments && old &&
1310                     old != start_next_window)
1311                         /* Wait for the former r1bio(s) to complete */
1312                         wait_event(conf->wait_barrier,
1313                                    bio->bi_phys_segments == 1);
1314                 goto retry_write;
1315         }
1316
1317         if (max_sectors < r1_bio->sectors) {
1318                 /* We are splitting this write into multiple parts, so
1319                  * we need to prepare for allocating another r1_bio.
1320                  */
1321                 r1_bio->sectors = max_sectors;
1322                 spin_lock_irq(&conf->device_lock);
1323                 if (bio->bi_phys_segments == 0)
1324                         bio->bi_phys_segments = 2;
1325                 else
1326                         bio->bi_phys_segments++;
1327                 spin_unlock_irq(&conf->device_lock);
1328         }
1329         sectors_handled = r1_bio->sector + max_sectors - bio->bi_iter.bi_sector;
1330
1331         atomic_set(&r1_bio->remaining, 1);
1332         atomic_set(&r1_bio->behind_remaining, 0);
1333
1334         first_clone = 1;
1335         for (i = 0; i < disks; i++) {
1336                 struct bio *mbio;
1337                 if (!r1_bio->bios[i])
1338                         continue;
1339
1340                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1341                 bio_trim(mbio, r1_bio->sector - bio->bi_iter.bi_sector, max_sectors);
1342
1343                 if (first_clone) {
1344                         /* do behind I/O ?
1345                          * Not if there are too many, or cannot
1346                          * allocate memory, or a reader on WriteMostly
1347                          * is waiting for behind writes to flush */
1348                         if (bitmap &&
1349                             (atomic_read(&bitmap->behind_writes)
1350                              < mddev->bitmap_info.max_write_behind) &&
1351                             !waitqueue_active(&bitmap->behind_wait))
1352                                 alloc_behind_pages(mbio, r1_bio);
1353
1354                         bitmap_startwrite(bitmap, r1_bio->sector,
1355                                           r1_bio->sectors,
1356                                           test_bit(R1BIO_BehindIO,
1357                                                    &r1_bio->state));
1358                         first_clone = 0;
1359                 }
1360                 if (r1_bio->behind_bvecs) {
1361                         struct bio_vec *bvec;
1362                         int j;
1363
1364                         /*
1365                          * We trimmed the bio, so _all is legit
1366                          */
1367                         bio_for_each_segment_all(bvec, mbio, j)
1368                                 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1369                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1370                                 atomic_inc(&r1_bio->behind_remaining);
1371                 }
1372
1373                 r1_bio->bios[i] = mbio;
1374
1375                 mbio->bi_iter.bi_sector = (r1_bio->sector +
1376                                    conf->mirrors[i].rdev->data_offset);
1377                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1378                 mbio->bi_end_io = raid1_end_write_request;
1379                 mbio->bi_rw =
1380                         WRITE | do_flush_fua | do_sync | do_discard | do_same;
1381                 mbio->bi_private = r1_bio;
1382
1383                 atomic_inc(&r1_bio->remaining);
1384
1385                 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1386                 if (cb)
1387                         plug = container_of(cb, struct raid1_plug_cb, cb);
1388                 else
1389                         plug = NULL;
1390                 spin_lock_irqsave(&conf->device_lock, flags);
1391                 if (plug) {
1392                         bio_list_add(&plug->pending, mbio);
1393                         plug->pending_cnt++;
1394                 } else {
1395                         bio_list_add(&conf->pending_bio_list, mbio);
1396                         conf->pending_count++;
1397                 }
1398                 spin_unlock_irqrestore(&conf->device_lock, flags);
1399                 if (!plug)
1400                         md_wakeup_thread(mddev->thread);
1401         }
1402         /* Mustn't call r1_bio_write_done before this next test,
1403          * as it could result in the bio being freed.
1404          */
1405         if (sectors_handled < bio_sectors(bio)) {
1406                 r1_bio_write_done(r1_bio);
1407                 /* We need another r1_bio.  It has already been counted
1408                  * in bio->bi_phys_segments
1409                  */
1410                 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1411                 r1_bio->master_bio = bio;
1412                 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1413                 r1_bio->state = 0;
1414                 r1_bio->mddev = mddev;
1415                 r1_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1416                 goto retry_write;
1417         }
1418
1419         r1_bio_write_done(r1_bio);
1420
1421         /* In case raid1d snuck in to freeze_array */
1422         wake_up(&conf->wait_barrier);
1423 }
1424
1425 static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1426 {
1427         struct r1conf *conf = mddev->private;
1428         int i;
1429
1430         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1431                    conf->raid_disks - mddev->degraded);
1432         rcu_read_lock();
1433         for (i = 0; i < conf->raid_disks; i++) {
1434                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1435                 seq_printf(seq, "%s",
1436                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1437         }
1438         rcu_read_unlock();
1439         seq_printf(seq, "]");
1440 }
1441
1442 static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1443 {
1444         char b[BDEVNAME_SIZE];
1445         struct r1conf *conf = mddev->private;
1446         unsigned long flags;
1447
1448         /*
1449          * If it is not operational, then we have already marked it as dead
1450          * else if it is the last working disks, ignore the error, let the
1451          * next level up know.
1452          * else mark the drive as failed
1453          */
1454         if (test_bit(In_sync, &rdev->flags)
1455             && (conf->raid_disks - mddev->degraded) == 1) {
1456                 /*
1457                  * Don't fail the drive, act as though we were just a
1458                  * normal single drive.
1459                  * However don't try a recovery from this drive as
1460                  * it is very likely to fail.
1461                  */
1462                 conf->recovery_disabled = mddev->recovery_disabled;
1463                 return;
1464         }
1465         set_bit(Blocked, &rdev->flags);
1466         spin_lock_irqsave(&conf->device_lock, flags);
1467         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1468                 mddev->degraded++;
1469                 set_bit(Faulty, &rdev->flags);
1470         } else
1471                 set_bit(Faulty, &rdev->flags);
1472         spin_unlock_irqrestore(&conf->device_lock, flags);
1473         /*
1474          * if recovery is running, make sure it aborts.
1475          */
1476         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1477         set_mask_bits(&mddev->flags, 0,
1478                       BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1479         printk(KERN_ALERT
1480                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1481                "md/raid1:%s: Operation continuing on %d devices.\n",
1482                mdname(mddev), bdevname(rdev->bdev, b),
1483                mdname(mddev), conf->raid_disks - mddev->degraded);
1484 }
1485
1486 static void print_conf(struct r1conf *conf)
1487 {
1488         int i;
1489
1490         printk(KERN_DEBUG "RAID1 conf printout:\n");
1491         if (!conf) {
1492                 printk(KERN_DEBUG "(!conf)\n");
1493                 return;
1494         }
1495         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1496                 conf->raid_disks);
1497
1498         rcu_read_lock();
1499         for (i = 0; i < conf->raid_disks; i++) {
1500                 char b[BDEVNAME_SIZE];
1501                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1502                 if (rdev)
1503                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1504                                i, !test_bit(In_sync, &rdev->flags),
1505                                !test_bit(Faulty, &rdev->flags),
1506                                bdevname(rdev->bdev,b));
1507         }
1508         rcu_read_unlock();
1509 }
1510
1511 static void close_sync(struct r1conf *conf)
1512 {
1513         wait_barrier(conf, NULL);
1514         allow_barrier(conf, 0, 0);
1515
1516         mempool_destroy(conf->r1buf_pool);
1517         conf->r1buf_pool = NULL;
1518
1519         spin_lock_irq(&conf->resync_lock);
1520         conf->next_resync = MaxSector - 2 * NEXT_NORMALIO_DISTANCE;
1521         conf->start_next_window = MaxSector;
1522         conf->current_window_requests +=
1523                 conf->next_window_requests;
1524         conf->next_window_requests = 0;
1525         spin_unlock_irq(&conf->resync_lock);
1526 }
1527
1528 static int raid1_spare_active(struct mddev *mddev)
1529 {
1530         int i;
1531         struct r1conf *conf = mddev->private;
1532         int count = 0;
1533         unsigned long flags;
1534
1535         /*
1536          * Find all failed disks within the RAID1 configuration
1537          * and mark them readable.
1538          * Called under mddev lock, so rcu protection not needed.
1539          * device_lock used to avoid races with raid1_end_read_request
1540          * which expects 'In_sync' flags and ->degraded to be consistent.
1541          */
1542         spin_lock_irqsave(&conf->device_lock, flags);
1543         for (i = 0; i < conf->raid_disks; i++) {
1544                 struct md_rdev *rdev = conf->mirrors[i].rdev;
1545                 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1546                 if (repl
1547                     && !test_bit(Candidate, &repl->flags)
1548                     && repl->recovery_offset == MaxSector
1549                     && !test_bit(Faulty, &repl->flags)
1550                     && !test_and_set_bit(In_sync, &repl->flags)) {
1551                         /* replacement has just become active */
1552                         if (!rdev ||
1553                             !test_and_clear_bit(In_sync, &rdev->flags))
1554                                 count++;
1555                         if (rdev) {
1556                                 /* Replaced device not technically
1557                                  * faulty, but we need to be sure
1558                                  * it gets removed and never re-added
1559                                  */
1560                                 set_bit(Faulty, &rdev->flags);
1561                                 sysfs_notify_dirent_safe(
1562                                         rdev->sysfs_state);
1563                         }
1564                 }
1565                 if (rdev
1566                     && rdev->recovery_offset == MaxSector
1567                     && !test_bit(Faulty, &rdev->flags)
1568                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1569                         count++;
1570                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1571                 }
1572         }
1573         mddev->degraded -= count;
1574         spin_unlock_irqrestore(&conf->device_lock, flags);
1575
1576         print_conf(conf);
1577         return count;
1578 }
1579
1580 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1581 {
1582         struct r1conf *conf = mddev->private;
1583         int err = -EEXIST;
1584         int mirror = 0;
1585         struct raid1_info *p;
1586         int first = 0;
1587         int last = conf->raid_disks - 1;
1588
1589         if (mddev->recovery_disabled == conf->recovery_disabled)
1590                 return -EBUSY;
1591
1592         if (md_integrity_add_rdev(rdev, mddev))
1593                 return -ENXIO;
1594
1595         if (rdev->raid_disk >= 0)
1596                 first = last = rdev->raid_disk;
1597
1598         /*
1599          * find the disk ... but prefer rdev->saved_raid_disk
1600          * if possible.
1601          */
1602         if (rdev->saved_raid_disk >= 0 &&
1603             rdev->saved_raid_disk >= first &&
1604             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1605                 first = last = rdev->saved_raid_disk;
1606
1607         for (mirror = first; mirror <= last; mirror++) {
1608                 p = conf->mirrors+mirror;
1609                 if (!p->rdev) {
1610
1611                         if (mddev->gendisk)
1612                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1613                                                   rdev->data_offset << 9);
1614
1615                         p->head_position = 0;
1616                         rdev->raid_disk = mirror;
1617                         err = 0;
1618                         /* As all devices are equivalent, we don't need a full recovery
1619                          * if this was recently any drive of the array
1620                          */
1621                         if (rdev->saved_raid_disk < 0)
1622                                 conf->fullsync = 1;
1623                         rcu_assign_pointer(p->rdev, rdev);
1624                         break;
1625                 }
1626                 if (test_bit(WantReplacement, &p->rdev->flags) &&
1627                     p[conf->raid_disks].rdev == NULL) {
1628                         /* Add this device as a replacement */
1629                         clear_bit(In_sync, &rdev->flags);
1630                         set_bit(Replacement, &rdev->flags);
1631                         rdev->raid_disk = mirror;
1632                         err = 0;
1633                         conf->fullsync = 1;
1634                         rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1635                         break;
1636                 }
1637         }
1638         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1639                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1640         print_conf(conf);
1641         return err;
1642 }
1643
1644 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1645 {
1646         struct r1conf *conf = mddev->private;
1647         int err = 0;
1648         int number = rdev->raid_disk;
1649         struct raid1_info *p = conf->mirrors + number;
1650
1651         if (rdev != p->rdev)
1652                 p = conf->mirrors + conf->raid_disks + number;
1653
1654         print_conf(conf);
1655         if (rdev == p->rdev) {
1656                 if (test_bit(In_sync, &rdev->flags) ||
1657                     atomic_read(&rdev->nr_pending)) {
1658                         err = -EBUSY;
1659                         goto abort;
1660                 }
1661                 /* Only remove non-faulty devices if recovery
1662                  * is not possible.
1663                  */
1664                 if (!test_bit(Faulty, &rdev->flags) &&
1665                     mddev->recovery_disabled != conf->recovery_disabled &&
1666                     mddev->degraded < conf->raid_disks) {
1667                         err = -EBUSY;
1668                         goto abort;
1669                 }
1670                 p->rdev = NULL;
1671                 synchronize_rcu();
1672                 if (atomic_read(&rdev->nr_pending)) {
1673                         /* lost the race, try later */
1674                         err = -EBUSY;
1675                         p->rdev = rdev;
1676                         goto abort;
1677                 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1678                         /* We just removed a device that is being replaced.
1679                          * Move down the replacement.  We drain all IO before
1680                          * doing this to avoid confusion.
1681                          */
1682                         struct md_rdev *repl =
1683                                 conf->mirrors[conf->raid_disks + number].rdev;
1684                         freeze_array(conf, 0);
1685                         clear_bit(Replacement, &repl->flags);
1686                         p->rdev = repl;
1687                         conf->mirrors[conf->raid_disks + number].rdev = NULL;
1688                         unfreeze_array(conf);
1689                         clear_bit(WantReplacement, &rdev->flags);
1690                 } else
1691                         clear_bit(WantReplacement, &rdev->flags);
1692                 err = md_integrity_register(mddev);
1693         }
1694 abort:
1695
1696         print_conf(conf);
1697         return err;
1698 }
1699
1700 static void end_sync_read(struct bio *bio)
1701 {
1702         struct r1bio *r1_bio = bio->bi_private;
1703
1704         update_head_pos(r1_bio->read_disk, r1_bio);
1705
1706         /*
1707          * we have read a block, now it needs to be re-written,
1708          * or re-read if the read failed.
1709          * We don't do much here, just schedule handling by raid1d
1710          */
1711         if (!bio->bi_error)
1712                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1713
1714         if (atomic_dec_and_test(&r1_bio->remaining))
1715                 reschedule_retry(r1_bio);
1716 }
1717
1718 static void end_sync_write(struct bio *bio)
1719 {
1720         int uptodate = !bio->bi_error;
1721         struct r1bio *r1_bio = bio->bi_private;
1722         struct mddev *mddev = r1_bio->mddev;
1723         struct r1conf *conf = mddev->private;
1724         int mirror=0;
1725         sector_t first_bad;
1726         int bad_sectors;
1727
1728         mirror = find_bio_disk(r1_bio, bio);
1729
1730         if (!uptodate) {
1731                 sector_t sync_blocks = 0;
1732                 sector_t s = r1_bio->sector;
1733                 long sectors_to_go = r1_bio->sectors;
1734                 /* make sure these bits doesn't get cleared. */
1735                 do {
1736                         bitmap_end_sync(mddev->bitmap, s,
1737                                         &sync_blocks, 1);
1738                         s += sync_blocks;
1739                         sectors_to_go -= sync_blocks;
1740                 } while (sectors_to_go > 0);
1741                 set_bit(WriteErrorSeen,
1742                         &conf->mirrors[mirror].rdev->flags);
1743                 if (!test_and_set_bit(WantReplacement,
1744                                       &conf->mirrors[mirror].rdev->flags))
1745                         set_bit(MD_RECOVERY_NEEDED, &
1746                                 mddev->recovery);
1747                 set_bit(R1BIO_WriteError, &r1_bio->state);
1748         } else if (is_badblock(conf->mirrors[mirror].rdev,
1749                                r1_bio->sector,
1750                                r1_bio->sectors,
1751                                &first_bad, &bad_sectors) &&
1752                    !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1753                                 r1_bio->sector,
1754                                 r1_bio->sectors,
1755                                 &first_bad, &bad_sectors)
1756                 )
1757                 set_bit(R1BIO_MadeGood, &r1_bio->state);
1758
1759         if (atomic_dec_and_test(&r1_bio->remaining)) {
1760                 int s = r1_bio->sectors;
1761                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1762                     test_bit(R1BIO_WriteError, &r1_bio->state))
1763                         reschedule_retry(r1_bio);
1764                 else {
1765                         put_buf(r1_bio);
1766                         md_done_sync(mddev, s, uptodate);
1767                 }
1768         }
1769 }
1770
1771 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1772                             int sectors, struct page *page, int rw)
1773 {
1774         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1775                 /* success */
1776                 return 1;
1777         if (rw == WRITE) {
1778                 set_bit(WriteErrorSeen, &rdev->flags);
1779                 if (!test_and_set_bit(WantReplacement,
1780                                       &rdev->flags))
1781                         set_bit(MD_RECOVERY_NEEDED, &
1782                                 rdev->mddev->recovery);
1783         }
1784         /* need to record an error - either for the block or the device */
1785         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1786                 md_error(rdev->mddev, rdev);
1787         return 0;
1788 }
1789
1790 static int fix_sync_read_error(struct r1bio *r1_bio)
1791 {
1792         /* Try some synchronous reads of other devices to get
1793          * good data, much like with normal read errors.  Only
1794          * read into the pages we already have so we don't
1795          * need to re-issue the read request.
1796          * We don't need to freeze the array, because being in an
1797          * active sync request, there is no normal IO, and
1798          * no overlapping syncs.
1799          * We don't need to check is_badblock() again as we
1800          * made sure that anything with a bad block in range
1801          * will have bi_end_io clear.
1802          */
1803         struct mddev *mddev = r1_bio->mddev;
1804         struct r1conf *conf = mddev->private;
1805         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1806         sector_t sect = r1_bio->sector;
1807         int sectors = r1_bio->sectors;
1808         int idx = 0;
1809
1810         while(sectors) {
1811                 int s = sectors;
1812                 int d = r1_bio->read_disk;
1813                 int success = 0;
1814                 struct md_rdev *rdev;
1815                 int start;
1816
1817                 if (s > (PAGE_SIZE>>9))
1818                         s = PAGE_SIZE >> 9;
1819                 do {
1820                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1821                                 /* No rcu protection needed here devices
1822                                  * can only be removed when no resync is
1823                                  * active, and resync is currently active
1824                                  */
1825                                 rdev = conf->mirrors[d].rdev;
1826                                 if (sync_page_io(rdev, sect, s<<9,
1827                                                  bio->bi_io_vec[idx].bv_page,
1828                                                  READ, false)) {
1829                                         success = 1;
1830                                         break;
1831                                 }
1832                         }
1833                         d++;
1834                         if (d == conf->raid_disks * 2)
1835                                 d = 0;
1836                 } while (!success && d != r1_bio->read_disk);
1837
1838                 if (!success) {
1839                         char b[BDEVNAME_SIZE];
1840                         int abort = 0;
1841                         /* Cannot read from anywhere, this block is lost.
1842                          * Record a bad block on each device.  If that doesn't
1843                          * work just disable and interrupt the recovery.
1844                          * Don't fail devices as that won't really help.
1845                          */
1846                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1847                                " for block %llu\n",
1848                                mdname(mddev),
1849                                bdevname(bio->bi_bdev, b),
1850                                (unsigned long long)r1_bio->sector);
1851                         for (d = 0; d < conf->raid_disks * 2; d++) {
1852                                 rdev = conf->mirrors[d].rdev;
1853                                 if (!rdev || test_bit(Faulty, &rdev->flags))
1854                                         continue;
1855                                 if (!rdev_set_badblocks(rdev, sect, s, 0))
1856                                         abort = 1;
1857                         }
1858                         if (abort) {
1859                                 conf->recovery_disabled =
1860                                         mddev->recovery_disabled;
1861                                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1862                                 md_done_sync(mddev, r1_bio->sectors, 0);
1863                                 put_buf(r1_bio);
1864                                 return 0;
1865                         }
1866                         /* Try next page */
1867                         sectors -= s;
1868                         sect += s;
1869                         idx++;
1870                         continue;
1871                 }
1872
1873                 start = d;
1874                 /* write it back and re-read */
1875                 while (d != r1_bio->read_disk) {
1876                         if (d == 0)
1877                                 d = conf->raid_disks * 2;
1878                         d--;
1879                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1880                                 continue;
1881                         rdev = conf->mirrors[d].rdev;
1882                         if (r1_sync_page_io(rdev, sect, s,
1883                                             bio->bi_io_vec[idx].bv_page,
1884                                             WRITE) == 0) {
1885                                 r1_bio->bios[d]->bi_end_io = NULL;
1886                                 rdev_dec_pending(rdev, mddev);
1887                         }
1888                 }
1889                 d = start;
1890                 while (d != r1_bio->read_disk) {
1891                         if (d == 0)
1892                                 d = conf->raid_disks * 2;
1893                         d--;
1894                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1895                                 continue;
1896                         rdev = conf->mirrors[d].rdev;
1897                         if (r1_sync_page_io(rdev, sect, s,
1898                                             bio->bi_io_vec[idx].bv_page,
1899                                             READ) != 0)
1900                                 atomic_add(s, &rdev->corrected_errors);
1901                 }
1902                 sectors -= s;
1903                 sect += s;
1904                 idx ++;
1905         }
1906         set_bit(R1BIO_Uptodate, &r1_bio->state);
1907         bio->bi_error = 0;
1908         return 1;
1909 }
1910
1911 static void process_checks(struct r1bio *r1_bio)
1912 {
1913         /* We have read all readable devices.  If we haven't
1914          * got the block, then there is no hope left.
1915          * If we have, then we want to do a comparison
1916          * and skip the write if everything is the same.
1917          * If any blocks failed to read, then we need to
1918          * attempt an over-write
1919          */
1920         struct mddev *mddev = r1_bio->mddev;
1921         struct r1conf *conf = mddev->private;
1922         int primary;
1923         int i;
1924         int vcnt;
1925
1926         /* Fix variable parts of all bios */
1927         vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1928         for (i = 0; i < conf->raid_disks * 2; i++) {
1929                 int j;
1930                 int size;
1931                 int error;
1932                 struct bio *b = r1_bio->bios[i];
1933                 if (b->bi_end_io != end_sync_read)
1934                         continue;
1935                 /* fixup the bio for reuse, but preserve errno */
1936                 error = b->bi_error;
1937                 bio_reset(b);
1938                 b->bi_error = error;
1939                 b->bi_vcnt = vcnt;
1940                 b->bi_iter.bi_size = r1_bio->sectors << 9;
1941                 b->bi_iter.bi_sector = r1_bio->sector +
1942                         conf->mirrors[i].rdev->data_offset;
1943                 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1944                 b->bi_end_io = end_sync_read;
1945                 b->bi_private = r1_bio;
1946
1947                 size = b->bi_iter.bi_size;
1948                 for (j = 0; j < vcnt ; j++) {
1949                         struct bio_vec *bi;
1950                         bi = &b->bi_io_vec[j];
1951                         bi->bv_offset = 0;
1952                         if (size > PAGE_SIZE)
1953                                 bi->bv_len = PAGE_SIZE;
1954                         else
1955                                 bi->bv_len = size;
1956                         size -= PAGE_SIZE;
1957                 }
1958         }
1959         for (primary = 0; primary < conf->raid_disks * 2; primary++)
1960                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1961                     !r1_bio->bios[primary]->bi_error) {
1962                         r1_bio->bios[primary]->bi_end_io = NULL;
1963                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1964                         break;
1965                 }
1966         r1_bio->read_disk = primary;
1967         for (i = 0; i < conf->raid_disks * 2; i++) {
1968                 int j;
1969                 struct bio *pbio = r1_bio->bios[primary];
1970                 struct bio *sbio = r1_bio->bios[i];
1971                 int error = sbio->bi_error;
1972
1973                 if (sbio->bi_end_io != end_sync_read)
1974                         continue;
1975                 /* Now we can 'fixup' the error value */
1976                 sbio->bi_error = 0;
1977
1978                 if (!error) {
1979                         for (j = vcnt; j-- ; ) {
1980                                 struct page *p, *s;
1981                                 p = pbio->bi_io_vec[j].bv_page;
1982                                 s = sbio->bi_io_vec[j].bv_page;
1983                                 if (memcmp(page_address(p),
1984                                            page_address(s),
1985                                            sbio->bi_io_vec[j].bv_len))
1986                                         break;
1987                         }
1988                 } else
1989                         j = 0;
1990                 if (j >= 0)
1991                         atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1992                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1993                               && !error)) {
1994                         /* No need to write to this device. */
1995                         sbio->bi_end_io = NULL;
1996                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1997                         continue;
1998                 }
1999
2000                 bio_copy_data(sbio, pbio);
2001         }
2002 }
2003
2004 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2005 {
2006         struct r1conf *conf = mddev->private;
2007         int i;
2008         int disks = conf->raid_disks * 2;
2009         struct bio *bio, *wbio;
2010
2011         bio = r1_bio->bios[r1_bio->read_disk];
2012
2013         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2014                 /* ouch - failed to read all of that. */
2015                 if (!fix_sync_read_error(r1_bio))
2016                         return;
2017
2018         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2019                 process_checks(r1_bio);
2020
2021         /*
2022          * schedule writes
2023          */
2024         atomic_set(&r1_bio->remaining, 1);
2025         for (i = 0; i < disks ; i++) {
2026                 wbio = r1_bio->bios[i];
2027                 if (wbio->bi_end_io == NULL ||
2028                     (wbio->bi_end_io == end_sync_read &&
2029                      (i == r1_bio->read_disk ||
2030                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2031                         continue;
2032
2033                 wbio->bi_rw = WRITE;
2034                 wbio->bi_end_io = end_sync_write;
2035                 atomic_inc(&r1_bio->remaining);
2036                 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2037
2038                 generic_make_request(wbio);
2039         }
2040
2041         if (atomic_dec_and_test(&r1_bio->remaining)) {
2042                 /* if we're here, all write(s) have completed, so clean up */
2043                 int s = r1_bio->sectors;
2044                 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2045                     test_bit(R1BIO_WriteError, &r1_bio->state))
2046                         reschedule_retry(r1_bio);
2047                 else {
2048                         put_buf(r1_bio);
2049                         md_done_sync(mddev, s, 1);
2050                 }
2051         }
2052 }
2053
2054 /*
2055  * This is a kernel thread which:
2056  *
2057  *      1.      Retries failed read operations on working mirrors.
2058  *      2.      Updates the raid superblock when problems encounter.
2059  *      3.      Performs writes following reads for array synchronising.
2060  */
2061
2062 static void fix_read_error(struct r1conf *conf, int read_disk,
2063                            sector_t sect, int sectors)
2064 {
2065         struct mddev *mddev = conf->mddev;
2066         while(sectors) {
2067                 int s = sectors;
2068                 int d = read_disk;
2069                 int success = 0;
2070                 int start;
2071                 struct md_rdev *rdev;
2072
2073                 if (s > (PAGE_SIZE>>9))
2074                         s = PAGE_SIZE >> 9;
2075
2076                 do {
2077                         /* Note: no rcu protection needed here
2078                          * as this is synchronous in the raid1d thread
2079                          * which is the thread that might remove
2080                          * a device.  If raid1d ever becomes multi-threaded....
2081                          */
2082                         sector_t first_bad;
2083                         int bad_sectors;
2084
2085                         rdev = conf->mirrors[d].rdev;
2086                         if (rdev &&
2087                             (test_bit(In_sync, &rdev->flags) ||
2088                              (!test_bit(Faulty, &rdev->flags) &&
2089                               rdev->recovery_offset >= sect + s)) &&
2090                             is_badblock(rdev, sect, s,
2091                                         &first_bad, &bad_sectors) == 0 &&
2092                             sync_page_io(rdev, sect, s<<9,
2093                                          conf->tmppage, READ, false))
2094                                 success = 1;
2095                         else {
2096                                 d++;
2097                                 if (d == conf->raid_disks * 2)
2098                                         d = 0;
2099                         }
2100                 } while (!success && d != read_disk);
2101
2102                 if (!success) {
2103                         /* Cannot read from anywhere - mark it bad */
2104                         struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2105                         if (!rdev_set_badblocks(rdev, sect, s, 0))
2106                                 md_error(mddev, rdev);
2107                         break;
2108                 }
2109                 /* write it back and re-read */
2110                 start = d;
2111                 while (d != read_disk) {
2112                         if (d==0)
2113                                 d = conf->raid_disks * 2;
2114                         d--;
2115                         rdev = conf->mirrors[d].rdev;
2116                         if (rdev &&
2117                             !test_bit(Faulty, &rdev->flags))
2118                                 r1_sync_page_io(rdev, sect, s,
2119                                                 conf->tmppage, WRITE);
2120                 }
2121                 d = start;
2122                 while (d != read_disk) {
2123                         char b[BDEVNAME_SIZE];
2124                         if (d==0)
2125                                 d = conf->raid_disks * 2;
2126                         d--;
2127                         rdev = conf->mirrors[d].rdev;
2128                         if (rdev &&
2129                             !test_bit(Faulty, &rdev->flags)) {
2130                                 if (r1_sync_page_io(rdev, sect, s,
2131                                                     conf->tmppage, READ)) {
2132                                         atomic_add(s, &rdev->corrected_errors);
2133                                         printk(KERN_INFO
2134                                                "md/raid1:%s: read error corrected "
2135                                                "(%d sectors at %llu on %s)\n",
2136                                                mdname(mddev), s,
2137                                                (unsigned long long)(sect +
2138                                                    rdev->data_offset),
2139                                                bdevname(rdev->bdev, b));
2140                                 }
2141                         }
2142                 }
2143                 sectors -= s;
2144                 sect += s;
2145         }
2146 }
2147
2148 static int narrow_write_error(struct r1bio *r1_bio, int i)
2149 {
2150         struct mddev *mddev = r1_bio->mddev;
2151         struct r1conf *conf = mddev->private;
2152         struct md_rdev *rdev = conf->mirrors[i].rdev;
2153
2154         /* bio has the data to be written to device 'i' where
2155          * we just recently had a write error.
2156          * We repeatedly clone the bio and trim down to one block,
2157          * then try the write.  Where the write fails we record
2158          * a bad block.
2159          * It is conceivable that the bio doesn't exactly align with
2160          * blocks.  We must handle this somehow.
2161          *
2162          * We currently own a reference on the rdev.
2163          */
2164
2165         int block_sectors;
2166         sector_t sector;
2167         int sectors;
2168         int sect_to_write = r1_bio->sectors;
2169         int ok = 1;
2170
2171         if (rdev->badblocks.shift < 0)
2172                 return 0;
2173
2174         block_sectors = roundup(1 << rdev->badblocks.shift,
2175                                 bdev_logical_block_size(rdev->bdev) >> 9);
2176         sector = r1_bio->sector;
2177         sectors = ((sector + block_sectors)
2178                    & ~(sector_t)(block_sectors - 1))
2179                 - sector;
2180
2181         while (sect_to_write) {
2182                 struct bio *wbio;
2183                 if (sectors > sect_to_write)
2184                         sectors = sect_to_write;
2185                 /* Write at 'sector' for 'sectors'*/
2186
2187                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2188                         unsigned vcnt = r1_bio->behind_page_count;
2189                         struct bio_vec *vec = r1_bio->behind_bvecs;
2190
2191                         while (!vec->bv_page) {
2192                                 vec++;
2193                                 vcnt--;
2194                         }
2195
2196                         wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2197                         memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2198
2199                         wbio->bi_vcnt = vcnt;
2200                 } else {
2201                         wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2202                 }
2203
2204                 wbio->bi_rw = WRITE;
2205                 wbio->bi_iter.bi_sector = r1_bio->sector;
2206                 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2207
2208                 bio_trim(wbio, sector - r1_bio->sector, sectors);
2209                 wbio->bi_iter.bi_sector += rdev->data_offset;
2210                 wbio->bi_bdev = rdev->bdev;
2211                 if (submit_bio_wait(WRITE, wbio) < 0)
2212                         /* failure! */
2213                         ok = rdev_set_badblocks(rdev, sector,
2214                                                 sectors, 0)
2215                                 && ok;
2216
2217                 bio_put(wbio);
2218                 sect_to_write -= sectors;
2219                 sector += sectors;
2220                 sectors = block_sectors;
2221         }
2222         return ok;
2223 }
2224
2225 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2226 {
2227         int m;
2228         int s = r1_bio->sectors;
2229         for (m = 0; m < conf->raid_disks * 2 ; m++) {
2230                 struct md_rdev *rdev = conf->mirrors[m].rdev;
2231                 struct bio *bio = r1_bio->bios[m];
2232                 if (bio->bi_end_io == NULL)
2233                         continue;
2234                 if (!bio->bi_error &&
2235                     test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2236                         rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2237                 }
2238                 if (bio->bi_error &&
2239                     test_bit(R1BIO_WriteError, &r1_bio->state)) {
2240                         if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2241                                 md_error(conf->mddev, rdev);
2242                 }
2243         }
2244         put_buf(r1_bio);
2245         md_done_sync(conf->mddev, s, 1);
2246 }
2247
2248 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2249 {
2250         int m;
2251         bool fail = false;
2252         for (m = 0; m < conf->raid_disks * 2 ; m++)
2253                 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2254                         struct md_rdev *rdev = conf->mirrors[m].rdev;
2255                         rdev_clear_badblocks(rdev,
2256                                              r1_bio->sector,
2257                                              r1_bio->sectors, 0);
2258                         rdev_dec_pending(rdev, conf->mddev);
2259                 } else if (r1_bio->bios[m] != NULL) {
2260                         /* This drive got a write error.  We need to
2261                          * narrow down and record precise write
2262                          * errors.
2263                          */
2264                         fail = true;
2265                         if (!narrow_write_error(r1_bio, m)) {
2266                                 md_error(conf->mddev,
2267                                          conf->mirrors[m].rdev);
2268                                 /* an I/O failed, we can't clear the bitmap */
2269                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2270                         }
2271                         rdev_dec_pending(conf->mirrors[m].rdev,
2272                                          conf->mddev);
2273                 }
2274         if (fail) {
2275                 spin_lock_irq(&conf->device_lock);
2276                 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2277                 conf->nr_queued++;
2278                 spin_unlock_irq(&conf->device_lock);
2279                 md_wakeup_thread(conf->mddev->thread);
2280         } else {
2281                 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2282                         close_write(r1_bio);
2283                 raid_end_bio_io(r1_bio);
2284         }
2285 }
2286
2287 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2288 {
2289         int disk;
2290         int max_sectors;
2291         struct mddev *mddev = conf->mddev;
2292         struct bio *bio;
2293         char b[BDEVNAME_SIZE];
2294         struct md_rdev *rdev;
2295
2296         clear_bit(R1BIO_ReadError, &r1_bio->state);
2297         /* we got a read error. Maybe the drive is bad.  Maybe just
2298          * the block and we can fix it.
2299          * We freeze all other IO, and try reading the block from
2300          * other devices.  When we find one, we re-write
2301          * and check it that fixes the read error.
2302          * This is all done synchronously while the array is
2303          * frozen
2304          */
2305         if (mddev->ro == 0) {
2306                 freeze_array(conf, 1);
2307                 fix_read_error(conf, r1_bio->read_disk,
2308                                r1_bio->sector, r1_bio->sectors);
2309                 unfreeze_array(conf);
2310         } else
2311                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2312         rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2313
2314         bio = r1_bio->bios[r1_bio->read_disk];
2315         bdevname(bio->bi_bdev, b);
2316 read_more:
2317         disk = read_balance(conf, r1_bio, &max_sectors);
2318         if (disk == -1) {
2319                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2320                        " read error for block %llu\n",
2321                        mdname(mddev), b, (unsigned long long)r1_bio->sector);
2322                 raid_end_bio_io(r1_bio);
2323         } else {
2324                 const unsigned long do_sync
2325                         = r1_bio->master_bio->bi_rw & REQ_SYNC;
2326                 if (bio) {
2327                         r1_bio->bios[r1_bio->read_disk] =
2328                                 mddev->ro ? IO_BLOCKED : NULL;
2329                         bio_put(bio);
2330                 }
2331                 r1_bio->read_disk = disk;
2332                 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2333                 bio_trim(bio, r1_bio->sector - bio->bi_iter.bi_sector,
2334                          max_sectors);
2335                 r1_bio->bios[r1_bio->read_disk] = bio;
2336                 rdev = conf->mirrors[disk].rdev;
2337                 printk_ratelimited(KERN_ERR
2338                                    "md/raid1:%s: redirecting sector %llu"
2339                                    " to other mirror: %s\n",
2340                                    mdname(mddev),
2341                                    (unsigned long long)r1_bio->sector,
2342                                    bdevname(rdev->bdev, b));
2343                 bio->bi_iter.bi_sector = r1_bio->sector + rdev->data_offset;
2344                 bio->bi_bdev = rdev->bdev;
2345                 bio->bi_end_io = raid1_end_read_request;
2346                 bio->bi_rw = READ | do_sync;
2347                 bio->bi_private = r1_bio;
2348                 if (max_sectors < r1_bio->sectors) {
2349                         /* Drat - have to split this up more */
2350                         struct bio *mbio = r1_bio->master_bio;
2351                         int sectors_handled = (r1_bio->sector + max_sectors
2352                                                - mbio->bi_iter.bi_sector);
2353                         r1_bio->sectors = max_sectors;
2354                         spin_lock_irq(&conf->device_lock);
2355                         if (mbio->bi_phys_segments == 0)
2356                                 mbio->bi_phys_segments = 2;
2357                         else
2358                                 mbio->bi_phys_segments++;
2359                         spin_unlock_irq(&conf->device_lock);
2360                         generic_make_request(bio);
2361                         bio = NULL;
2362
2363                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2364
2365                         r1_bio->master_bio = mbio;
2366                         r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2367                         r1_bio->state = 0;
2368                         set_bit(R1BIO_ReadError, &r1_bio->state);
2369                         r1_bio->mddev = mddev;
2370                         r1_bio->sector = mbio->bi_iter.bi_sector +
2371                                 sectors_handled;
2372
2373                         goto read_more;
2374                 } else
2375                         generic_make_request(bio);
2376         }
2377 }
2378
2379 static void raid1d(struct md_thread *thread)
2380 {
2381         struct mddev *mddev = thread->mddev;
2382         struct r1bio *r1_bio;
2383         unsigned long flags;
2384         struct r1conf *conf = mddev->private;
2385         struct list_head *head = &conf->retry_list;
2386         struct blk_plug plug;
2387
2388         md_check_recovery(mddev);
2389
2390         if (!list_empty_careful(&conf->bio_end_io_list) &&
2391             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2392                 LIST_HEAD(tmp);
2393                 spin_lock_irqsave(&conf->device_lock, flags);
2394                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2395                         while (!list_empty(&conf->bio_end_io_list)) {
2396                                 list_move(conf->bio_end_io_list.prev, &tmp);
2397                                 conf->nr_queued--;
2398                         }
2399                 }
2400                 spin_unlock_irqrestore(&conf->device_lock, flags);
2401                 while (!list_empty(&tmp)) {
2402                         r1_bio = list_first_entry(&tmp, struct r1bio,
2403                                                   retry_list);
2404                         list_del(&r1_bio->retry_list);
2405                         if (mddev->degraded)
2406                                 set_bit(R1BIO_Degraded, &r1_bio->state);
2407                         if (test_bit(R1BIO_WriteError, &r1_bio->state))
2408                                 close_write(r1_bio);
2409                         raid_end_bio_io(r1_bio);
2410                 }
2411         }
2412
2413         blk_start_plug(&plug);
2414         for (;;) {
2415
2416                 flush_pending_writes(conf);
2417
2418                 spin_lock_irqsave(&conf->device_lock, flags);
2419                 if (list_empty(head)) {
2420                         spin_unlock_irqrestore(&conf->device_lock, flags);
2421                         break;
2422                 }
2423                 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2424                 list_del(head->prev);
2425                 conf->nr_queued--;
2426                 spin_unlock_irqrestore(&conf->device_lock, flags);
2427
2428                 mddev = r1_bio->mddev;
2429                 conf = mddev->private;
2430                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2431                         if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2432                             test_bit(R1BIO_WriteError, &r1_bio->state))
2433                                 handle_sync_write_finished(conf, r1_bio);
2434                         else
2435                                 sync_request_write(mddev, r1_bio);
2436                 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2437                            test_bit(R1BIO_WriteError, &r1_bio->state))
2438                         handle_write_finished(conf, r1_bio);
2439                 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2440                         handle_read_error(conf, r1_bio);
2441                 else
2442                         /* just a partial read to be scheduled from separate
2443                          * context
2444                          */
2445                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2446
2447                 cond_resched();
2448                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2449                         md_check_recovery(mddev);
2450         }
2451         blk_finish_plug(&plug);
2452 }
2453
2454 static int init_resync(struct r1conf *conf)
2455 {
2456         int buffs;
2457
2458         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2459         BUG_ON(conf->r1buf_pool);
2460         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2461                                           conf->poolinfo);
2462         if (!conf->r1buf_pool)
2463                 return -ENOMEM;
2464         conf->next_resync = 0;
2465         return 0;
2466 }
2467
2468 /*
2469  * perform a "sync" on one "block"
2470  *
2471  * We need to make sure that no normal I/O request - particularly write
2472  * requests - conflict with active sync requests.
2473  *
2474  * This is achieved by tracking pending requests and a 'barrier' concept
2475  * that can be installed to exclude normal IO requests.
2476  */
2477
2478 static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2479                                    int *skipped)
2480 {
2481         struct r1conf *conf = mddev->private;
2482         struct r1bio *r1_bio;
2483         struct bio *bio;
2484         sector_t max_sector, nr_sectors;
2485         int disk = -1;
2486         int i;
2487         int wonly = -1;
2488         int write_targets = 0, read_targets = 0;
2489         sector_t sync_blocks;
2490         int still_degraded = 0;
2491         int good_sectors = RESYNC_SECTORS;
2492         int min_bad = 0; /* number of sectors that are bad in all devices */
2493
2494         if (!conf->r1buf_pool)
2495                 if (init_resync(conf))
2496                         return 0;
2497
2498         max_sector = mddev->dev_sectors;
2499         if (sector_nr >= max_sector) {
2500                 /* If we aborted, we need to abort the
2501                  * sync on the 'current' bitmap chunk (there will
2502                  * only be one in raid1 resync.
2503                  * We can find the current addess in mddev->curr_resync
2504                  */
2505                 if (mddev->curr_resync < max_sector) /* aborted */
2506                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2507                                                 &sync_blocks, 1);
2508                 else /* completed sync */
2509                         conf->fullsync = 0;
2510
2511                 bitmap_close_sync(mddev->bitmap);
2512                 close_sync(conf);
2513
2514                 if (mddev_is_clustered(mddev)) {
2515                         conf->cluster_sync_low = 0;
2516                         conf->cluster_sync_high = 0;
2517                 }
2518                 return 0;
2519         }
2520
2521         if (mddev->bitmap == NULL &&
2522             mddev->recovery_cp == MaxSector &&
2523             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2524             conf->fullsync == 0) {
2525                 *skipped = 1;
2526                 return max_sector - sector_nr;
2527         }
2528         /* before building a request, check if we can skip these blocks..
2529          * This call the bitmap_start_sync doesn't actually record anything
2530          */
2531         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2532             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2533                 /* We can skip this block, and probably several more */
2534                 *skipped = 1;
2535                 return sync_blocks;
2536         }
2537
2538         /*
2539          * If there is non-resync activity waiting for a turn, then let it
2540          * though before starting on this new sync request.
2541          */
2542         if (conf->nr_waiting)
2543                 schedule_timeout_uninterruptible(1);
2544
2545         /* we are incrementing sector_nr below. To be safe, we check against
2546          * sector_nr + two times RESYNC_SECTORS
2547          */
2548
2549         bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2550                 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2551         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2552
2553         raise_barrier(conf, sector_nr);
2554
2555         rcu_read_lock();
2556         /*
2557          * If we get a correctably read error during resync or recovery,
2558          * we might want to read from a different device.  So we
2559          * flag all drives that could conceivably be read from for READ,
2560          * and any others (which will be non-In_sync devices) for WRITE.
2561          * If a read fails, we try reading from something else for which READ
2562          * is OK.
2563          */
2564
2565         r1_bio->mddev = mddev;
2566         r1_bio->sector = sector_nr;
2567         r1_bio->state = 0;
2568         set_bit(R1BIO_IsSync, &r1_bio->state);
2569
2570         for (i = 0; i < conf->raid_disks * 2; i++) {
2571                 struct md_rdev *rdev;
2572                 bio = r1_bio->bios[i];
2573                 bio_reset(bio);
2574
2575                 rdev = rcu_dereference(conf->mirrors[i].rdev);
2576                 if (rdev == NULL ||
2577                     test_bit(Faulty, &rdev->flags)) {
2578                         if (i < conf->raid_disks)
2579                                 still_degraded = 1;
2580                 } else if (!test_bit(In_sync, &rdev->flags)) {
2581                         bio->bi_rw = WRITE;
2582                         bio->bi_end_io = end_sync_write;
2583                         write_targets ++;
2584                 } else {
2585                         /* may need to read from here */
2586                         sector_t first_bad = MaxSector;
2587                         int bad_sectors;
2588
2589                         if (is_badblock(rdev, sector_nr, good_sectors,
2590                                         &first_bad, &bad_sectors)) {
2591                                 if (first_bad > sector_nr)
2592                                         good_sectors = first_bad - sector_nr;
2593                                 else {
2594                                         bad_sectors -= (sector_nr - first_bad);
2595                                         if (min_bad == 0 ||
2596                                             min_bad > bad_sectors)
2597                                                 min_bad = bad_sectors;
2598                                 }
2599                         }
2600                         if (sector_nr < first_bad) {
2601                                 if (test_bit(WriteMostly, &rdev->flags)) {
2602                                         if (wonly < 0)
2603                                                 wonly = i;
2604                                 } else {
2605                                         if (disk < 0)
2606                                                 disk = i;
2607                                 }
2608                                 bio->bi_rw = READ;
2609                                 bio->bi_end_io = end_sync_read;
2610                                 read_targets++;
2611                         } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2612                                 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2613                                 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2614                                 /*
2615                                  * The device is suitable for reading (InSync),
2616                                  * but has bad block(s) here. Let's try to correct them,
2617                                  * if we are doing resync or repair. Otherwise, leave
2618                                  * this device alone for this sync request.
2619                                  */
2620                                 bio->bi_rw = WRITE;
2621                                 bio->bi_end_io = end_sync_write;
2622                                 write_targets++;
2623                         }
2624                 }
2625                 if (bio->bi_end_io) {
2626                         atomic_inc(&rdev->nr_pending);
2627                         bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2628                         bio->bi_bdev = rdev->bdev;
2629                         bio->bi_private = r1_bio;
2630                 }
2631         }
2632         rcu_read_unlock();
2633         if (disk < 0)
2634                 disk = wonly;
2635         r1_bio->read_disk = disk;
2636
2637         if (read_targets == 0 && min_bad > 0) {
2638                 /* These sectors are bad on all InSync devices, so we
2639                  * need to mark them bad on all write targets
2640                  */
2641                 int ok = 1;
2642                 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2643                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2644                                 struct md_rdev *rdev = conf->mirrors[i].rdev;
2645                                 ok = rdev_set_badblocks(rdev, sector_nr,
2646                                                         min_bad, 0
2647                                         ) && ok;
2648                         }
2649                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2650                 *skipped = 1;
2651                 put_buf(r1_bio);
2652
2653                 if (!ok) {
2654                         /* Cannot record the badblocks, so need to
2655                          * abort the resync.
2656                          * If there are multiple read targets, could just
2657                          * fail the really bad ones ???
2658                          */
2659                         conf->recovery_disabled = mddev->recovery_disabled;
2660                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2661                         return 0;
2662                 } else
2663                         return min_bad;
2664
2665         }
2666         if (min_bad > 0 && min_bad < good_sectors) {
2667                 /* only resync enough to reach the next bad->good
2668                  * transition */
2669                 good_sectors = min_bad;
2670         }
2671
2672         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2673                 /* extra read targets are also write targets */
2674                 write_targets += read_targets-1;
2675
2676         if (write_targets == 0 || read_targets == 0) {
2677                 /* There is nowhere to write, so all non-sync
2678                  * drives must be failed - so we are finished
2679                  */
2680                 sector_t rv;
2681                 if (min_bad > 0)
2682                         max_sector = sector_nr + min_bad;
2683                 rv = max_sector - sector_nr;
2684                 *skipped = 1;
2685                 put_buf(r1_bio);
2686                 return rv;
2687         }
2688
2689         if (max_sector > mddev->resync_max)
2690                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2691         if (max_sector > sector_nr + good_sectors)
2692                 max_sector = sector_nr + good_sectors;
2693         nr_sectors = 0;
2694         sync_blocks = 0;
2695         do {
2696                 struct page *page;
2697                 int len = PAGE_SIZE;
2698                 if (sector_nr + (len>>9) > max_sector)
2699                         len = (max_sector - sector_nr) << 9;
2700                 if (len == 0)
2701                         break;
2702                 if (sync_blocks == 0) {
2703                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2704                                                &sync_blocks, still_degraded) &&
2705                             !conf->fullsync &&
2706                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2707                                 break;
2708                         if ((len >> 9) > sync_blocks)
2709                                 len = sync_blocks<<9;
2710                 }
2711
2712                 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2713                         bio = r1_bio->bios[i];
2714                         if (bio->bi_end_io) {
2715                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2716                                 if (bio_add_page(bio, page, len, 0) == 0) {
2717                                         /* stop here */
2718                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2719                                         while (i > 0) {
2720                                                 i--;
2721                                                 bio = r1_bio->bios[i];
2722                                                 if (bio->bi_end_io==NULL)
2723                                                         continue;
2724                                                 /* remove last page from this bio */
2725                                                 bio->bi_vcnt--;
2726                                                 bio->bi_iter.bi_size -= len;
2727                                                 bio_clear_flag(bio, BIO_SEG_VALID);
2728                                         }
2729                                         goto bio_full;
2730                                 }
2731                         }
2732                 }
2733                 nr_sectors += len>>9;
2734                 sector_nr += len>>9;
2735                 sync_blocks -= (len>>9);
2736         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2737  bio_full:
2738         r1_bio->sectors = nr_sectors;
2739
2740         if (mddev_is_clustered(mddev) &&
2741                         conf->cluster_sync_high < sector_nr + nr_sectors) {
2742                 conf->cluster_sync_low = mddev->curr_resync_completed;
2743                 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2744                 /* Send resync message */
2745                 md_cluster_ops->resync_info_update(mddev,
2746                                 conf->cluster_sync_low,
2747                                 conf->cluster_sync_high);
2748         }
2749
2750         /* For a user-requested sync, we read all readable devices and do a
2751          * compare
2752          */
2753         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2754                 atomic_set(&r1_bio->remaining, read_targets);
2755                 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2756                         bio = r1_bio->bios[i];
2757                         if (bio->bi_end_io == end_sync_read) {
2758                                 read_targets--;
2759                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2760                                 generic_make_request(bio);
2761                         }
2762                 }
2763         } else {
2764                 atomic_set(&r1_bio->remaining, 1);
2765                 bio = r1_bio->bios[r1_bio->read_disk];
2766                 md_sync_acct(bio->bi_bdev, nr_sectors);
2767                 generic_make_request(bio);
2768
2769         }
2770         return nr_sectors;
2771 }
2772
2773 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2774 {
2775         if (sectors)
2776                 return sectors;
2777
2778         return mddev->dev_sectors;
2779 }
2780
2781 static struct r1conf *setup_conf(struct mddev *mddev)
2782 {
2783         struct r1conf *conf;
2784         int i;
2785         struct raid1_info *disk;
2786         struct md_rdev *rdev;
2787         int err = -ENOMEM;
2788
2789         conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2790         if (!conf)
2791                 goto abort;
2792
2793         conf->mirrors = kzalloc(sizeof(struct raid1_info)
2794                                 * mddev->raid_disks * 2,
2795                                  GFP_KERNEL);
2796         if (!conf->mirrors)
2797                 goto abort;
2798
2799         conf->tmppage = alloc_page(GFP_KERNEL);
2800         if (!conf->tmppage)
2801                 goto abort;
2802
2803         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2804         if (!conf->poolinfo)
2805                 goto abort;
2806         conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2807         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2808                                           r1bio_pool_free,
2809                                           conf->poolinfo);
2810         if (!conf->r1bio_pool)
2811                 goto abort;
2812
2813         conf->poolinfo->mddev = mddev;
2814
2815         err = -EINVAL;
2816         spin_lock_init(&conf->device_lock);
2817         rdev_for_each(rdev, mddev) {
2818                 struct request_queue *q;
2819                 int disk_idx = rdev->raid_disk;
2820                 if (disk_idx >= mddev->raid_disks
2821                     || disk_idx < 0)
2822                         continue;
2823                 if (test_bit(Replacement, &rdev->flags))
2824                         disk = conf->mirrors + mddev->raid_disks + disk_idx;
2825                 else
2826                         disk = conf->mirrors + disk_idx;
2827
2828                 if (disk->rdev)
2829                         goto abort;
2830                 disk->rdev = rdev;
2831                 q = bdev_get_queue(rdev->bdev);
2832
2833                 disk->head_position = 0;
2834                 disk->seq_start = MaxSector;
2835         }
2836         conf->raid_disks = mddev->raid_disks;
2837         conf->mddev = mddev;
2838         INIT_LIST_HEAD(&conf->retry_list);
2839         INIT_LIST_HEAD(&conf->bio_end_io_list);
2840
2841         spin_lock_init(&conf->resync_lock);
2842         init_waitqueue_head(&conf->wait_barrier);
2843
2844         bio_list_init(&conf->pending_bio_list);
2845         conf->pending_count = 0;
2846         conf->recovery_disabled = mddev->recovery_disabled - 1;
2847
2848         conf->start_next_window = MaxSector;
2849         conf->current_window_requests = conf->next_window_requests = 0;
2850
2851         err = -EIO;
2852         for (i = 0; i < conf->raid_disks * 2; i++) {
2853
2854                 disk = conf->mirrors + i;
2855
2856                 if (i < conf->raid_disks &&
2857                     disk[conf->raid_disks].rdev) {
2858                         /* This slot has a replacement. */
2859                         if (!disk->rdev) {
2860                                 /* No original, just make the replacement
2861                                  * a recovering spare
2862                                  */
2863                                 disk->rdev =
2864                                         disk[conf->raid_disks].rdev;
2865                                 disk[conf->raid_disks].rdev = NULL;
2866                         } else if (!test_bit(In_sync, &disk->rdev->flags))
2867                                 /* Original is not in_sync - bad */
2868                                 goto abort;
2869                 }
2870
2871                 if (!disk->rdev ||
2872                     !test_bit(In_sync, &disk->rdev->flags)) {
2873                         disk->head_position = 0;
2874                         if (disk->rdev &&
2875                             (disk->rdev->saved_raid_disk < 0))
2876                                 conf->fullsync = 1;
2877                 }
2878         }
2879
2880         err = -ENOMEM;
2881         conf->thread = md_register_thread(raid1d, mddev, "raid1");
2882         if (!conf->thread) {
2883                 printk(KERN_ERR
2884                        "md/raid1:%s: couldn't allocate thread\n",
2885                        mdname(mddev));
2886                 goto abort;
2887         }
2888
2889         return conf;
2890
2891  abort:
2892         if (conf) {
2893                 mempool_destroy(conf->r1bio_pool);
2894                 kfree(conf->mirrors);
2895                 safe_put_page(conf->tmppage);
2896                 kfree(conf->poolinfo);
2897                 kfree(conf);
2898         }
2899         return ERR_PTR(err);
2900 }
2901
2902 static void raid1_free(struct mddev *mddev, void *priv);
2903 static int raid1_run(struct mddev *mddev)
2904 {
2905         struct r1conf *conf;
2906         int i;
2907         struct md_rdev *rdev;
2908         int ret;
2909         bool discard_supported = false;
2910
2911         if (mddev->level != 1) {
2912                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2913                        mdname(mddev), mddev->level);
2914                 return -EIO;
2915         }
2916         if (mddev->reshape_position != MaxSector) {
2917                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2918                        mdname(mddev));
2919                 return -EIO;
2920         }
2921         /*
2922          * copy the already verified devices into our private RAID1
2923          * bookkeeping area. [whatever we allocate in run(),
2924          * should be freed in raid1_free()]
2925          */
2926         if (mddev->private == NULL)
2927                 conf = setup_conf(mddev);
2928         else
2929                 conf = mddev->private;
2930
2931         if (IS_ERR(conf))
2932                 return PTR_ERR(conf);
2933
2934         if (mddev->queue)
2935                 blk_queue_max_write_same_sectors(mddev->queue, 0);
2936
2937         rdev_for_each(rdev, mddev) {
2938                 if (!mddev->gendisk)
2939                         continue;
2940                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2941                                   rdev->data_offset << 9);
2942                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2943                         discard_supported = true;
2944         }
2945
2946         mddev->degraded = 0;
2947         for (i=0; i < conf->raid_disks; i++)
2948                 if (conf->mirrors[i].rdev == NULL ||
2949                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2950                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2951                         mddev->degraded++;
2952
2953         if (conf->raid_disks - mddev->degraded == 1)
2954                 mddev->recovery_cp = MaxSector;
2955
2956         if (mddev->recovery_cp != MaxSector)
2957                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2958                        " -- starting background reconstruction\n",
2959                        mdname(mddev));
2960         printk(KERN_INFO
2961                 "md/raid1:%s: active with %d out of %d mirrors\n",
2962                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2963                 mddev->raid_disks);
2964
2965         /*
2966          * Ok, everything is just fine now
2967          */
2968         mddev->thread = conf->thread;
2969         conf->thread = NULL;
2970         mddev->private = conf;
2971
2972         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2973
2974         if (mddev->queue) {
2975                 if (discard_supported)
2976                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2977                                                 mddev->queue);
2978                 else
2979                         queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2980                                                   mddev->queue);
2981         }
2982
2983         ret =  md_integrity_register(mddev);
2984         if (ret) {
2985                 md_unregister_thread(&mddev->thread);
2986                 raid1_free(mddev, conf);
2987         }
2988         return ret;