f7f3c8a634195a7804e08fb887c0de51b763760c
[sfrench/cifs-2.6.git] / drivers / md / raid10.c
1 /*
2  * raid10.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 2000-2004 Neil Brown
5  *
6  * RAID-10 support for md.
7  *
8  * Base on code in raid1.c.  See raid1.c for further copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include "md.h"
29 #include "raid10.h"
30 #include "raid0.h"
31 #include "bitmap.h"
32
33 /*
34  * RAID10 provides a combination of RAID0 and RAID1 functionality.
35  * The layout of data is defined by
36  *    chunk_size
37  *    raid_disks
38  *    near_copies (stored in low byte of layout)
39  *    far_copies (stored in second byte of layout)
40  *    far_offset (stored in bit 16 of layout )
41  *    use_far_sets (stored in bit 17 of layout )
42  *    use_far_sets_bugfixed (stored in bit 18 of layout )
43  *
44  * The data to be stored is divided into chunks using chunksize.  Each device
45  * is divided into far_copies sections.   In each section, chunks are laid out
46  * in a style similar to raid0, but near_copies copies of each chunk is stored
47  * (each on a different drive).  The starting device for each section is offset
48  * near_copies from the starting device of the previous section.  Thus there
49  * are (near_copies * far_copies) of each chunk, and each is on a different
50  * drive.  near_copies and far_copies must be at least one, and their product
51  * is at most raid_disks.
52  *
53  * If far_offset is true, then the far_copies are handled a bit differently.
54  * The copies are still in different stripes, but instead of being very far
55  * apart on disk, there are adjacent stripes.
56  *
57  * The far and offset algorithms are handled slightly differently if
58  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
59  * sets that are (near_copies * far_copies) in size.  The far copied stripes
60  * are still shifted by 'near_copies' devices, but this shifting stays confined
61  * to the set rather than the entire array.  This is done to improve the number
62  * of device combinations that can fail without causing the array to fail.
63  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
64  * on a device):
65  *    A B C D    A B C D E
66  *      ...         ...
67  *    D A B C    E A B C D
68  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69  *    [A B] [C D]    [A B] [C D E]
70  *    |...| |...|    |...| | ... |
71  *    [B A] [D C]    [B A] [E C D]
72  */
73
74 /*
75  * Number of guaranteed r10bios in case of extreme VM load:
76  */
77 #define NR_RAID10_BIOS 256
78
79 /* when we get a read error on a read-only array, we redirect to another
80  * device without failing the first device, or trying to over-write to
81  * correct the read error.  To keep track of bad blocks on a per-bio
82  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83  */
84 #define IO_BLOCKED ((struct bio *)1)
85 /* When we successfully write to a known bad-block, we need to remove the
86  * bad-block marking which must be done from process context.  So we record
87  * the success by setting devs[n].bio to IO_MADE_GOOD
88  */
89 #define IO_MADE_GOOD ((struct bio *)2)
90
91 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92
93 /* When there are this many requests queued to be written by
94  * the raid10 thread, we become 'congested' to provide back-pressure
95  * for writeback.
96  */
97 static int max_queued_requests = 1024;
98
99 static void allow_barrier(struct r10conf *conf);
100 static void lower_barrier(struct r10conf *conf);
101 static int _enough(struct r10conf *conf, int previous, int ignore);
102 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103                                 int *skipped);
104 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
105 static void end_reshape_write(struct bio *bio);
106 static void end_reshape(struct r10conf *conf);
107
108 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
109 {
110         struct r10conf *conf = data;
111         int size = offsetof(struct r10bio, devs[conf->copies]);
112
113         /* allocate a r10bio with room for raid_disks entries in the
114          * bios array */
115         return kzalloc(size, gfp_flags);
116 }
117
118 static void r10bio_pool_free(void *r10_bio, void *data)
119 {
120         kfree(r10_bio);
121 }
122
123 /* Maximum size of each resync request */
124 #define RESYNC_BLOCK_SIZE (64*1024)
125 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
126 /* amount of memory to reserve for resync requests */
127 #define RESYNC_WINDOW (1024*1024)
128 /* maximum number of concurrent requests, memory permitting */
129 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
130
131 /*
132  * When performing a resync, we need to read and compare, so
133  * we need as many pages are there are copies.
134  * When performing a recovery, we need 2 bios, one for read,
135  * one for write (we recover only one drive per r10buf)
136  *
137  */
138 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
139 {
140         struct r10conf *conf = data;
141         struct page *page;
142         struct r10bio *r10_bio;
143         struct bio *bio;
144         int i, j;
145         int nalloc;
146
147         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
148         if (!r10_bio)
149                 return NULL;
150
151         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
152             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
153                 nalloc = conf->copies; /* resync */
154         else
155                 nalloc = 2; /* recovery */
156
157         /*
158          * Allocate bios.
159          */
160         for (j = nalloc ; j-- ; ) {
161                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
162                 if (!bio)
163                         goto out_free_bio;
164                 r10_bio->devs[j].bio = bio;
165                 if (!conf->have_replacement)
166                         continue;
167                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
168                 if (!bio)
169                         goto out_free_bio;
170                 r10_bio->devs[j].repl_bio = bio;
171         }
172         /*
173          * Allocate RESYNC_PAGES data pages and attach them
174          * where needed.
175          */
176         for (j = 0 ; j < nalloc; j++) {
177                 struct bio *rbio = r10_bio->devs[j].repl_bio;
178                 bio = r10_bio->devs[j].bio;
179                 for (i = 0; i < RESYNC_PAGES; i++) {
180                         if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
181                                                &conf->mddev->recovery)) {
182                                 /* we can share bv_page's during recovery
183                                  * and reshape */
184                                 struct bio *rbio = r10_bio->devs[0].bio;
185                                 page = rbio->bi_io_vec[i].bv_page;
186                                 get_page(page);
187                         } else
188                                 page = alloc_page(gfp_flags);
189                         if (unlikely(!page))
190                                 goto out_free_pages;
191
192                         bio->bi_io_vec[i].bv_page = page;
193                         if (rbio)
194                                 rbio->bi_io_vec[i].bv_page = page;
195                 }
196         }
197
198         return r10_bio;
199
200 out_free_pages:
201         for ( ; i > 0 ; i--)
202                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
203         while (j--)
204                 for (i = 0; i < RESYNC_PAGES ; i++)
205                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
206         j = 0;
207 out_free_bio:
208         for ( ; j < nalloc; j++) {
209                 if (r10_bio->devs[j].bio)
210                         bio_put(r10_bio->devs[j].bio);
211                 if (r10_bio->devs[j].repl_bio)
212                         bio_put(r10_bio->devs[j].repl_bio);
213         }
214         r10bio_pool_free(r10_bio, conf);
215         return NULL;
216 }
217
218 static void r10buf_pool_free(void *__r10_bio, void *data)
219 {
220         int i;
221         struct r10conf *conf = data;
222         struct r10bio *r10bio = __r10_bio;
223         int j;
224
225         for (j=0; j < conf->copies; j++) {
226                 struct bio *bio = r10bio->devs[j].bio;
227                 if (bio) {
228                         for (i = 0; i < RESYNC_PAGES; i++) {
229                                 safe_put_page(bio->bi_io_vec[i].bv_page);
230                                 bio->bi_io_vec[i].bv_page = NULL;
231                         }
232                         bio_put(bio);
233                 }
234                 bio = r10bio->devs[j].repl_bio;
235                 if (bio)
236                         bio_put(bio);
237         }
238         r10bio_pool_free(r10bio, conf);
239 }
240
241 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
242 {
243         int i;
244
245         for (i = 0; i < conf->copies; i++) {
246                 struct bio **bio = & r10_bio->devs[i].bio;
247                 if (!BIO_SPECIAL(*bio))
248                         bio_put(*bio);
249                 *bio = NULL;
250                 bio = &r10_bio->devs[i].repl_bio;
251                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
252                         bio_put(*bio);
253                 *bio = NULL;
254         }
255 }
256
257 static void free_r10bio(struct r10bio *r10_bio)
258 {
259         struct r10conf *conf = r10_bio->mddev->private;
260
261         put_all_bios(conf, r10_bio);
262         mempool_free(r10_bio, conf->r10bio_pool);
263 }
264
265 static void put_buf(struct r10bio *r10_bio)
266 {
267         struct r10conf *conf = r10_bio->mddev->private;
268
269         mempool_free(r10_bio, conf->r10buf_pool);
270
271         lower_barrier(conf);
272 }
273
274 static void reschedule_retry(struct r10bio *r10_bio)
275 {
276         unsigned long flags;
277         struct mddev *mddev = r10_bio->mddev;
278         struct r10conf *conf = mddev->private;
279
280         spin_lock_irqsave(&conf->device_lock, flags);
281         list_add(&r10_bio->retry_list, &conf->retry_list);
282         conf->nr_queued ++;
283         spin_unlock_irqrestore(&conf->device_lock, flags);
284
285         /* wake up frozen array... */
286         wake_up(&conf->wait_barrier);
287
288         md_wakeup_thread(mddev->thread);
289 }
290
291 /*
292  * raid_end_bio_io() is called when we have finished servicing a mirrored
293  * operation and are ready to return a success/failure code to the buffer
294  * cache layer.
295  */
296 static void raid_end_bio_io(struct r10bio *r10_bio)
297 {
298         struct bio *bio = r10_bio->master_bio;
299         int done;
300         struct r10conf *conf = r10_bio->mddev->private;
301
302         if (bio->bi_phys_segments) {
303                 unsigned long flags;
304                 spin_lock_irqsave(&conf->device_lock, flags);
305                 bio->bi_phys_segments--;
306                 done = (bio->bi_phys_segments == 0);
307                 spin_unlock_irqrestore(&conf->device_lock, flags);
308         } else
309                 done = 1;
310         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
311                 bio->bi_error = -EIO;
312         if (done) {
313                 bio_endio(bio);
314                 /*
315                  * Wake up any possible resync thread that waits for the device
316                  * to go idle.
317                  */
318                 allow_barrier(conf);
319         }
320         free_r10bio(r10_bio);
321 }
322
323 /*
324  * Update disk head position estimator based on IRQ completion info.
325  */
326 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
327 {
328         struct r10conf *conf = r10_bio->mddev->private;
329
330         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
331                 r10_bio->devs[slot].addr + (r10_bio->sectors);
332 }
333
334 /*
335  * Find the disk number which triggered given bio
336  */
337 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
338                          struct bio *bio, int *slotp, int *replp)
339 {
340         int slot;
341         int repl = 0;
342
343         for (slot = 0; slot < conf->copies; slot++) {
344                 if (r10_bio->devs[slot].bio == bio)
345                         break;
346                 if (r10_bio->devs[slot].repl_bio == bio) {
347                         repl = 1;
348                         break;
349                 }
350         }
351
352         BUG_ON(slot == conf->copies);
353         update_head_pos(slot, r10_bio);
354
355         if (slotp)
356                 *slotp = slot;
357         if (replp)
358                 *replp = repl;
359         return r10_bio->devs[slot].devnum;
360 }
361
362 static void raid10_end_read_request(struct bio *bio)
363 {
364         int uptodate = !bio->bi_error;
365         struct r10bio *r10_bio = bio->bi_private;
366         int slot, dev;
367         struct md_rdev *rdev;
368         struct r10conf *conf = r10_bio->mddev->private;
369
370         slot = r10_bio->read_slot;
371         dev = r10_bio->devs[slot].devnum;
372         rdev = r10_bio->devs[slot].rdev;
373         /*
374          * this branch is our 'one mirror IO has finished' event handler:
375          */
376         update_head_pos(slot, r10_bio);
377
378         if (uptodate) {
379                 /*
380                  * Set R10BIO_Uptodate in our master bio, so that
381                  * we will return a good error code to the higher
382                  * levels even if IO on some other mirrored buffer fails.
383                  *
384                  * The 'master' represents the composite IO operation to
385                  * user-side. So if something waits for IO, then it will
386                  * wait for the 'master' bio.
387                  */
388                 set_bit(R10BIO_Uptodate, &r10_bio->state);
389         } else {
390                 /* If all other devices that store this block have
391                  * failed, we want to return the error upwards rather
392                  * than fail the last device.  Here we redefine
393                  * "uptodate" to mean "Don't want to retry"
394                  */
395                 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
396                              rdev->raid_disk))
397                         uptodate = 1;
398         }
399         if (uptodate) {
400                 raid_end_bio_io(r10_bio);
401                 rdev_dec_pending(rdev, conf->mddev);
402         } else {
403                 /*
404                  * oops, read error - keep the refcount on the rdev
405                  */
406                 char b[BDEVNAME_SIZE];
407                 printk_ratelimited(KERN_ERR
408                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
409                                    mdname(conf->mddev),
410                                    bdevname(rdev->bdev, b),
411                                    (unsigned long long)r10_bio->sector);
412                 set_bit(R10BIO_ReadError, &r10_bio->state);
413                 reschedule_retry(r10_bio);
414         }
415 }
416
417 static void close_write(struct r10bio *r10_bio)
418 {
419         /* clear the bitmap if all writes complete successfully */
420         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421                         r10_bio->sectors,
422                         !test_bit(R10BIO_Degraded, &r10_bio->state),
423                         0);
424         md_write_end(r10_bio->mddev);
425 }
426
427 static void one_write_done(struct r10bio *r10_bio)
428 {
429         if (atomic_dec_and_test(&r10_bio->remaining)) {
430                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431                         reschedule_retry(r10_bio);
432                 else {
433                         close_write(r10_bio);
434                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435                                 reschedule_retry(r10_bio);
436                         else
437                                 raid_end_bio_io(r10_bio);
438                 }
439         }
440 }
441
442 static void raid10_end_write_request(struct bio *bio)
443 {
444         struct r10bio *r10_bio = bio->bi_private;
445         int dev;
446         int dec_rdev = 1;
447         struct r10conf *conf = r10_bio->mddev->private;
448         int slot, repl;
449         struct md_rdev *rdev = NULL;
450
451         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
452
453         if (repl)
454                 rdev = conf->mirrors[dev].replacement;
455         if (!rdev) {
456                 smp_rmb();
457                 repl = 0;
458                 rdev = conf->mirrors[dev].rdev;
459         }
460         /*
461          * this branch is our 'one mirror IO has finished' event handler:
462          */
463         if (bio->bi_error) {
464                 if (repl)
465                         /* Never record new bad blocks to replacement,
466                          * just fail it.
467                          */
468                         md_error(rdev->mddev, rdev);
469                 else {
470                         set_bit(WriteErrorSeen, &rdev->flags);
471                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
472                                 set_bit(MD_RECOVERY_NEEDED,
473                                         &rdev->mddev->recovery);
474                         set_bit(R10BIO_WriteError, &r10_bio->state);
475                         dec_rdev = 0;
476                 }
477         } else {
478                 /*
479                  * Set R10BIO_Uptodate in our master bio, so that
480                  * we will return a good error code for to the higher
481                  * levels even if IO on some other mirrored buffer fails.
482                  *
483                  * The 'master' represents the composite IO operation to
484                  * user-side. So if something waits for IO, then it will
485                  * wait for the 'master' bio.
486                  */
487                 sector_t first_bad;
488                 int bad_sectors;
489
490                 /*
491                  * Do not set R10BIO_Uptodate if the current device is
492                  * rebuilding or Faulty. This is because we cannot use
493                  * such device for properly reading the data back (we could
494                  * potentially use it, if the current write would have felt
495                  * before rdev->recovery_offset, but for simplicity we don't
496                  * check this here.
497                  */
498                 if (test_bit(In_sync, &rdev->flags) &&
499                     !test_bit(Faulty, &rdev->flags))
500                         set_bit(R10BIO_Uptodate, &r10_bio->state);
501
502                 /* Maybe we can clear some bad blocks. */
503                 if (is_badblock(rdev,
504                                 r10_bio->devs[slot].addr,
505                                 r10_bio->sectors,
506                                 &first_bad, &bad_sectors)) {
507                         bio_put(bio);
508                         if (repl)
509                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
510                         else
511                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
512                         dec_rdev = 0;
513                         set_bit(R10BIO_MadeGood, &r10_bio->state);
514                 }
515         }
516
517         /*
518          *
519          * Let's see if all mirrored write operations have finished
520          * already.
521          */
522         one_write_done(r10_bio);
523         if (dec_rdev)
524                 rdev_dec_pending(rdev, conf->mddev);
525 }
526
527 /*
528  * RAID10 layout manager
529  * As well as the chunksize and raid_disks count, there are two
530  * parameters: near_copies and far_copies.
531  * near_copies * far_copies must be <= raid_disks.
532  * Normally one of these will be 1.
533  * If both are 1, we get raid0.
534  * If near_copies == raid_disks, we get raid1.
535  *
536  * Chunks are laid out in raid0 style with near_copies copies of the
537  * first chunk, followed by near_copies copies of the next chunk and
538  * so on.
539  * If far_copies > 1, then after 1/far_copies of the array has been assigned
540  * as described above, we start again with a device offset of near_copies.
541  * So we effectively have another copy of the whole array further down all
542  * the drives, but with blocks on different drives.
543  * With this layout, and block is never stored twice on the one device.
544  *
545  * raid10_find_phys finds the sector offset of a given virtual sector
546  * on each device that it is on.
547  *
548  * raid10_find_virt does the reverse mapping, from a device and a
549  * sector offset to a virtual address
550  */
551
552 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
553 {
554         int n,f;
555         sector_t sector;
556         sector_t chunk;
557         sector_t stripe;
558         int dev;
559         int slot = 0;
560         int last_far_set_start, last_far_set_size;
561
562         last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
563         last_far_set_start *= geo->far_set_size;
564
565         last_far_set_size = geo->far_set_size;
566         last_far_set_size += (geo->raid_disks % geo->far_set_size);
567
568         /* now calculate first sector/dev */
569         chunk = r10bio->sector >> geo->chunk_shift;
570         sector = r10bio->sector & geo->chunk_mask;
571
572         chunk *= geo->near_copies;
573         stripe = chunk;
574         dev = sector_div(stripe, geo->raid_disks);
575         if (geo->far_offset)
576                 stripe *= geo->far_copies;
577
578         sector += stripe << geo->chunk_shift;
579
580         /* and calculate all the others */
581         for (n = 0; n < geo->near_copies; n++) {
582                 int d = dev;
583                 int set;
584                 sector_t s = sector;
585                 r10bio->devs[slot].devnum = d;
586                 r10bio->devs[slot].addr = s;
587                 slot++;
588
589                 for (f = 1; f < geo->far_copies; f++) {
590                         set = d / geo->far_set_size;
591                         d += geo->near_copies;
592
593                         if ((geo->raid_disks % geo->far_set_size) &&
594                             (d > last_far_set_start)) {
595                                 d -= last_far_set_start;
596                                 d %= last_far_set_size;
597                                 d += last_far_set_start;
598                         } else {
599                                 d %= geo->far_set_size;
600                                 d += geo->far_set_size * set;
601                         }
602                         s += geo->stride;
603                         r10bio->devs[slot].devnum = d;
604                         r10bio->devs[slot].addr = s;
605                         slot++;
606                 }
607                 dev++;
608                 if (dev >= geo->raid_disks) {
609                         dev = 0;
610                         sector += (geo->chunk_mask + 1);
611                 }
612         }
613 }
614
615 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
616 {
617         struct geom *geo = &conf->geo;
618
619         if (conf->reshape_progress != MaxSector &&
620             ((r10bio->sector >= conf->reshape_progress) !=
621              conf->mddev->reshape_backwards)) {
622                 set_bit(R10BIO_Previous, &r10bio->state);
623                 geo = &conf->prev;
624         } else
625                 clear_bit(R10BIO_Previous, &r10bio->state);
626
627         __raid10_find_phys(geo, r10bio);
628 }
629
630 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
631 {
632         sector_t offset, chunk, vchunk;
633         /* Never use conf->prev as this is only called during resync
634          * or recovery, so reshape isn't happening
635          */
636         struct geom *geo = &conf->geo;
637         int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
638         int far_set_size = geo->far_set_size;
639         int last_far_set_start;
640
641         if (geo->raid_disks % geo->far_set_size) {
642                 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
643                 last_far_set_start *= geo->far_set_size;
644
645                 if (dev >= last_far_set_start) {
646                         far_set_size = geo->far_set_size;
647                         far_set_size += (geo->raid_disks % geo->far_set_size);
648                         far_set_start = last_far_set_start;
649                 }
650         }
651
652         offset = sector & geo->chunk_mask;
653         if (geo->far_offset) {
654                 int fc;
655                 chunk = sector >> geo->chunk_shift;
656                 fc = sector_div(chunk, geo->far_copies);
657                 dev -= fc * geo->near_copies;
658                 if (dev < far_set_start)
659                         dev += far_set_size;
660         } else {
661                 while (sector >= geo->stride) {
662                         sector -= geo->stride;
663                         if (dev < (geo->near_copies + far_set_start))
664                                 dev += far_set_size - geo->near_copies;
665                         else
666                                 dev -= geo->near_copies;
667                 }
668                 chunk = sector >> geo->chunk_shift;
669         }
670         vchunk = chunk * geo->raid_disks + dev;
671         sector_div(vchunk, geo->near_copies);
672         return (vchunk << geo->chunk_shift) + offset;
673 }
674
675 /*
676  * This routine returns the disk from which the requested read should
677  * be done. There is a per-array 'next expected sequential IO' sector
678  * number - if this matches on the next IO then we use the last disk.
679  * There is also a per-disk 'last know head position' sector that is
680  * maintained from IRQ contexts, both the normal and the resync IO
681  * completion handlers update this position correctly. If there is no
682  * perfect sequential match then we pick the disk whose head is closest.
683  *
684  * If there are 2 mirrors in the same 2 devices, performance degrades
685  * because position is mirror, not device based.
686  *
687  * The rdev for the device selected will have nr_pending incremented.
688  */
689
690 /*
691  * FIXME: possibly should rethink readbalancing and do it differently
692  * depending on near_copies / far_copies geometry.
693  */
694 static struct md_rdev *read_balance(struct r10conf *conf,
695                                     struct r10bio *r10_bio,
696                                     int *max_sectors)
697 {
698         const sector_t this_sector = r10_bio->sector;
699         int disk, slot;
700         int sectors = r10_bio->sectors;
701         int best_good_sectors;
702         sector_t new_distance, best_dist;
703         struct md_rdev *best_rdev, *rdev = NULL;
704         int do_balance;
705         int best_slot;
706         struct geom *geo = &conf->geo;
707
708         raid10_find_phys(conf, r10_bio);
709         rcu_read_lock();
710         sectors = r10_bio->sectors;
711         best_slot = -1;
712         best_rdev = NULL;
713         best_dist = MaxSector;
714         best_good_sectors = 0;
715         do_balance = 1;
716         /*
717          * Check if we can balance. We can balance on the whole
718          * device if no resync is going on (recovery is ok), or below
719          * the resync window. We take the first readable disk when
720          * above the resync window.
721          */
722         if (conf->mddev->recovery_cp < MaxSector
723             && (this_sector + sectors >= conf->next_resync))
724                 do_balance = 0;
725
726         for (slot = 0; slot < conf->copies ; slot++) {
727                 sector_t first_bad;
728                 int bad_sectors;
729                 sector_t dev_sector;
730
731                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
732                         continue;
733                 disk = r10_bio->devs[slot].devnum;
734                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
735                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
736                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
737                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
738                 if (rdev == NULL ||
739                     test_bit(Faulty, &rdev->flags))
740                         continue;
741                 if (!test_bit(In_sync, &rdev->flags) &&
742                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
743                         continue;
744
745                 dev_sector = r10_bio->devs[slot].addr;
746                 if (is_badblock(rdev, dev_sector, sectors,
747                                 &first_bad, &bad_sectors)) {
748                         if (best_dist < MaxSector)
749                                 /* Already have a better slot */
750                                 continue;
751                         if (first_bad <= dev_sector) {
752                                 /* Cannot read here.  If this is the
753                                  * 'primary' device, then we must not read
754                                  * beyond 'bad_sectors' from another device.
755                                  */
756                                 bad_sectors -= (dev_sector - first_bad);
757                                 if (!do_balance && sectors > bad_sectors)
758                                         sectors = bad_sectors;
759                                 if (best_good_sectors > sectors)
760                                         best_good_sectors = sectors;
761                         } else {
762                                 sector_t good_sectors =
763                                         first_bad - dev_sector;
764                                 if (good_sectors > best_good_sectors) {
765                                         best_good_sectors = good_sectors;
766                                         best_slot = slot;
767                                         best_rdev = rdev;
768                                 }
769                                 if (!do_balance)
770                                         /* Must read from here */
771                                         break;
772                         }
773                         continue;
774                 } else
775                         best_good_sectors = sectors;
776
777                 if (!do_balance)
778                         break;
779
780                 /* This optimisation is debatable, and completely destroys
781                  * sequential read speed for 'far copies' arrays.  So only
782                  * keep it for 'near' arrays, and review those later.
783                  */
784                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
785                         break;
786
787                 /* for far > 1 always use the lowest address */
788                 if (geo->far_copies > 1)
789                         new_distance = r10_bio->devs[slot].addr;
790                 else
791                         new_distance = abs(r10_bio->devs[slot].addr -
792                                            conf->mirrors[disk].head_position);
793                 if (new_distance < best_dist) {
794                         best_dist = new_distance;
795                         best_slot = slot;
796                         best_rdev = rdev;
797                 }
798         }
799         if (slot >= conf->copies) {
800                 slot = best_slot;
801                 rdev = best_rdev;
802         }
803
804         if (slot >= 0) {
805                 atomic_inc(&rdev->nr_pending);
806                 r10_bio->read_slot = slot;
807         } else
808                 rdev = NULL;
809         rcu_read_unlock();
810         *max_sectors = best_good_sectors;
811
812         return rdev;
813 }
814
815 static int raid10_congested(struct mddev *mddev, int bits)
816 {
817         struct r10conf *conf = mddev->private;
818         int i, ret = 0;
819
820         if ((bits & (1 << WB_async_congested)) &&
821             conf->pending_count >= max_queued_requests)
822                 return 1;
823
824         rcu_read_lock();
825         for (i = 0;
826              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
827                      && ret == 0;
828              i++) {
829                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
830                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
831                         struct request_queue *q = bdev_get_queue(rdev->bdev);
832
833                         ret |= bdi_congested(&q->backing_dev_info, bits);
834                 }
835         }
836         rcu_read_unlock();
837         return ret;
838 }
839
840 static void flush_pending_writes(struct r10conf *conf)
841 {
842         /* Any writes that have been queued but are awaiting
843          * bitmap updates get flushed here.
844          */
845         spin_lock_irq(&conf->device_lock);
846
847         if (conf->pending_bio_list.head) {
848                 struct bio *bio;
849                 bio = bio_list_get(&conf->pending_bio_list);
850                 conf->pending_count = 0;
851                 spin_unlock_irq(&conf->device_lock);
852                 /* flush any pending bitmap writes to disk
853                  * before proceeding w/ I/O */
854                 bitmap_unplug(conf->mddev->bitmap);
855                 wake_up(&conf->wait_barrier);
856
857                 while (bio) { /* submit pending writes */
858                         struct bio *next = bio->bi_next;
859                         bio->bi_next = NULL;
860                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
861                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
862                                 /* Just ignore it */
863                                 bio_endio(bio);
864                         else
865                                 generic_make_request(bio);
866                         bio = next;
867                 }
868         } else
869                 spin_unlock_irq(&conf->device_lock);
870 }
871
872 /* Barriers....
873  * Sometimes we need to suspend IO while we do something else,
874  * either some resync/recovery, or reconfigure the array.
875  * To do this we raise a 'barrier'.
876  * The 'barrier' is a counter that can be raised multiple times
877  * to count how many activities are happening which preclude
878  * normal IO.
879  * We can only raise the barrier if there is no pending IO.
880  * i.e. if nr_pending == 0.
881  * We choose only to raise the barrier if no-one is waiting for the
882  * barrier to go down.  This means that as soon as an IO request
883  * is ready, no other operations which require a barrier will start
884  * until the IO request has had a chance.
885  *
886  * So: regular IO calls 'wait_barrier'.  When that returns there
887  *    is no backgroup IO happening,  It must arrange to call
888  *    allow_barrier when it has finished its IO.
889  * backgroup IO calls must call raise_barrier.  Once that returns
890  *    there is no normal IO happeing.  It must arrange to call
891  *    lower_barrier when the particular background IO completes.
892  */
893
894 static void raise_barrier(struct r10conf *conf, int force)
895 {
896         BUG_ON(force && !conf->barrier);
897         spin_lock_irq(&conf->resync_lock);
898
899         /* Wait until no block IO is waiting (unless 'force') */
900         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
901                             conf->resync_lock);
902
903         /* block any new IO from starting */
904         conf->barrier++;
905
906         /* Now wait for all pending IO to complete */
907         wait_event_lock_irq(conf->wait_barrier,
908                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
909                             conf->resync_lock);
910
911         spin_unlock_irq(&conf->resync_lock);
912 }
913
914 static void lower_barrier(struct r10conf *conf)
915 {
916         unsigned long flags;
917         spin_lock_irqsave(&conf->resync_lock, flags);
918         conf->barrier--;
919         spin_unlock_irqrestore(&conf->resync_lock, flags);
920         wake_up(&conf->wait_barrier);
921 }
922
923 static void wait_barrier(struct r10conf *conf)
924 {
925         spin_lock_irq(&conf->resync_lock);
926         if (conf->barrier) {
927                 conf->nr_waiting++;
928                 /* Wait for the barrier to drop.
929                  * However if there are already pending
930                  * requests (preventing the barrier from
931                  * rising completely), and the
932                  * pre-process bio queue isn't empty,
933                  * then don't wait, as we need to empty
934                  * that queue to get the nr_pending
935                  * count down.
936                  */
937                 wait_event_lock_irq(conf->wait_barrier,
938                                     !conf->barrier ||
939                                     (conf->nr_pending &&
940                                      current->bio_list &&
941                                      !bio_list_empty(current->bio_list)),
942                                     conf->resync_lock);
943                 conf->nr_waiting--;
944         }
945         conf->nr_pending++;
946         spin_unlock_irq(&conf->resync_lock);
947 }
948
949 static void allow_barrier(struct r10conf *conf)
950 {
951         unsigned long flags;
952         spin_lock_irqsave(&conf->resync_lock, flags);
953         conf->nr_pending--;
954         spin_unlock_irqrestore(&conf->resync_lock, flags);
955         wake_up(&conf->wait_barrier);
956 }
957
958 static void freeze_array(struct r10conf *conf, int extra)
959 {
960         /* stop syncio and normal IO and wait for everything to
961          * go quiet.
962          * We increment barrier and nr_waiting, and then
963          * wait until nr_pending match nr_queued+extra
964          * This is called in the context of one normal IO request
965          * that has failed. Thus any sync request that might be pending
966          * will be blocked by nr_pending, and we need to wait for
967          * pending IO requests to complete or be queued for re-try.
968          * Thus the number queued (nr_queued) plus this request (extra)
969          * must match the number of pending IOs (nr_pending) before
970          * we continue.
971          */
972         spin_lock_irq(&conf->resync_lock);
973         conf->barrier++;
974         conf->nr_waiting++;
975         wait_event_lock_irq_cmd(conf->wait_barrier,
976                                 conf->nr_pending == conf->nr_queued+extra,
977                                 conf->resync_lock,
978                                 flush_pending_writes(conf));
979
980         spin_unlock_irq(&conf->resync_lock);
981 }
982
983 static void unfreeze_array(struct r10conf *conf)
984 {
985         /* reverse the effect of the freeze */
986         spin_lock_irq(&conf->resync_lock);
987         conf->barrier--;
988         conf->nr_waiting--;
989         wake_up(&conf->wait_barrier);
990         spin_unlock_irq(&conf->resync_lock);
991 }
992
993 static sector_t choose_data_offset(struct r10bio *r10_bio,
994                                    struct md_rdev *rdev)
995 {
996         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
997             test_bit(R10BIO_Previous, &r10_bio->state))
998                 return rdev->data_offset;
999         else
1000                 return rdev->new_data_offset;
1001 }
1002
1003 struct raid10_plug_cb {
1004         struct blk_plug_cb      cb;
1005         struct bio_list         pending;
1006         int                     pending_cnt;
1007 };
1008
1009 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1010 {
1011         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1012                                                    cb);
1013         struct mddev *mddev = plug->cb.data;
1014         struct r10conf *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 __make_request(struct mddev *mddev, struct bio *bio)
1048 {
1049         struct r10conf *conf = mddev->private;
1050         struct r10bio *r10_bio;
1051         struct bio *read_bio;
1052         int i;
1053         const int rw = bio_data_dir(bio);
1054         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1055         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1056         const unsigned long do_discard = (bio->bi_rw
1057                                           & (REQ_DISCARD | REQ_SECURE));
1058         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1059         unsigned long flags;
1060         struct md_rdev *blocked_rdev;
1061         struct blk_plug_cb *cb;
1062         struct raid10_plug_cb *plug = NULL;
1063         int sectors_handled;
1064         int max_sectors;
1065         int sectors;
1066
1067         /*
1068          * Register the new request and wait if the reconstruction
1069          * thread has put up a bar for new requests.
1070          * Continue immediately if no resync is active currently.
1071          */
1072         wait_barrier(conf);
1073
1074         sectors = bio_sectors(bio);
1075         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1076             bio->bi_iter.bi_sector < conf->reshape_progress &&
1077             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1078                 /* IO spans the reshape position.  Need to wait for
1079                  * reshape to pass
1080                  */
1081                 allow_barrier(conf);
1082                 wait_event(conf->wait_barrier,
1083                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1084                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1085                            sectors);
1086                 wait_barrier(conf);
1087         }
1088         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1089             bio_data_dir(bio) == WRITE &&
1090             (mddev->reshape_backwards
1091              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1092                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1093              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1094                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1095                 /* Need to update reshape_position in metadata */
1096                 mddev->reshape_position = conf->reshape_progress;
1097                 set_mask_bits(&mddev->flags, 0,
1098                               BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1099                 md_wakeup_thread(mddev->thread);
1100                 wait_event(mddev->sb_wait,
1101                            !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1102
1103                 conf->reshape_safe = mddev->reshape_position;
1104         }
1105
1106         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1107
1108         r10_bio->master_bio = bio;
1109         r10_bio->sectors = sectors;
1110
1111         r10_bio->mddev = mddev;
1112         r10_bio->sector = bio->bi_iter.bi_sector;
1113         r10_bio->state = 0;
1114
1115         /* We might need to issue multiple reads to different
1116          * devices if there are bad blocks around, so we keep
1117          * track of the number of reads in bio->bi_phys_segments.
1118          * If this is 0, there is only one r10_bio and no locking
1119          * will be needed when the request completes.  If it is
1120          * non-zero, then it is the number of not-completed requests.
1121          */
1122         bio->bi_phys_segments = 0;
1123         bio_clear_flag(bio, BIO_SEG_VALID);
1124
1125         if (rw == READ) {
1126                 /*
1127                  * read balancing logic:
1128                  */
1129                 struct md_rdev *rdev;
1130                 int slot;
1131
1132 read_again:
1133                 rdev = read_balance(conf, r10_bio, &max_sectors);
1134                 if (!rdev) {
1135                         raid_end_bio_io(r10_bio);
1136                         return;
1137                 }
1138                 slot = r10_bio->read_slot;
1139
1140                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1141                 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1142                          max_sectors);
1143
1144                 r10_bio->devs[slot].bio = read_bio;
1145                 r10_bio->devs[slot].rdev = rdev;
1146
1147                 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1148                         choose_data_offset(r10_bio, rdev);
1149                 read_bio->bi_bdev = rdev->bdev;
1150                 read_bio->bi_end_io = raid10_end_read_request;
1151                 read_bio->bi_rw = READ | do_sync;
1152                 read_bio->bi_private = r10_bio;
1153
1154                 if (max_sectors < r10_bio->sectors) {
1155                         /* Could not read all from this device, so we will
1156                          * need another r10_bio.
1157                          */
1158                         sectors_handled = (r10_bio->sector + max_sectors
1159                                            - bio->bi_iter.bi_sector);
1160                         r10_bio->sectors = max_sectors;
1161                         spin_lock_irq(&conf->device_lock);
1162                         if (bio->bi_phys_segments == 0)
1163                                 bio->bi_phys_segments = 2;
1164                         else
1165                                 bio->bi_phys_segments++;
1166                         spin_unlock_irq(&conf->device_lock);
1167                         /* Cannot call generic_make_request directly
1168                          * as that will be queued in __generic_make_request
1169                          * and subsequent mempool_alloc might block
1170                          * waiting for it.  so hand bio over to raid10d.
1171                          */
1172                         reschedule_retry(r10_bio);
1173
1174                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1175
1176                         r10_bio->master_bio = bio;
1177                         r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1178                         r10_bio->state = 0;
1179                         r10_bio->mddev = mddev;
1180                         r10_bio->sector = bio->bi_iter.bi_sector +
1181                                 sectors_handled;
1182                         goto read_again;
1183                 } else
1184                         generic_make_request(read_bio);
1185                 return;
1186         }
1187
1188         /*
1189          * WRITE:
1190          */
1191         if (conf->pending_count >= max_queued_requests) {
1192                 md_wakeup_thread(mddev->thread);
1193                 wait_event(conf->wait_barrier,
1194                            conf->pending_count < max_queued_requests);
1195         }
1196         /* first select target devices under rcu_lock and
1197          * inc refcount on their rdev.  Record them by setting
1198          * bios[x] to bio
1199          * If there are known/acknowledged bad blocks on any device
1200          * on which we have seen a write error, we want to avoid
1201          * writing to those blocks.  This potentially requires several
1202          * writes to write around the bad blocks.  Each set of writes
1203          * gets its own r10_bio with a set of bios attached.  The number
1204          * of r10_bios is recored in bio->bi_phys_segments just as with
1205          * the read case.
1206          */
1207
1208         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1209         raid10_find_phys(conf, r10_bio);
1210 retry_write:
1211         blocked_rdev = NULL;
1212         rcu_read_lock();
1213         max_sectors = r10_bio->sectors;
1214
1215         for (i = 0;  i < conf->copies; i++) {
1216                 int d = r10_bio->devs[i].devnum;
1217                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1218                 struct md_rdev *rrdev = rcu_dereference(
1219                         conf->mirrors[d].replacement);
1220                 if (rdev == rrdev)
1221                         rrdev = NULL;
1222                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1223                         atomic_inc(&rdev->nr_pending);
1224                         blocked_rdev = rdev;
1225                         break;
1226                 }
1227                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1228                         atomic_inc(&rrdev->nr_pending);
1229                         blocked_rdev = rrdev;
1230                         break;
1231                 }
1232                 if (rdev && (test_bit(Faulty, &rdev->flags)))
1233                         rdev = NULL;
1234                 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1235                         rrdev = NULL;
1236
1237                 r10_bio->devs[i].bio = NULL;
1238                 r10_bio->devs[i].repl_bio = NULL;
1239
1240                 if (!rdev && !rrdev) {
1241                         set_bit(R10BIO_Degraded, &r10_bio->state);
1242                         continue;
1243                 }
1244                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1245                         sector_t first_bad;
1246                         sector_t dev_sector = r10_bio->devs[i].addr;
1247                         int bad_sectors;
1248                         int is_bad;
1249
1250                         is_bad = is_badblock(rdev, dev_sector,
1251                                              max_sectors,
1252                                              &first_bad, &bad_sectors);
1253                         if (is_bad < 0) {
1254                                 /* Mustn't write here until the bad block
1255                                  * is acknowledged
1256                                  */
1257                                 atomic_inc(&rdev->nr_pending);
1258                                 set_bit(BlockedBadBlocks, &rdev->flags);
1259                                 blocked_rdev = rdev;
1260                                 break;
1261                         }
1262                         if (is_bad && first_bad <= dev_sector) {
1263                                 /* Cannot write here at all */
1264                                 bad_sectors -= (dev_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                                 /* We don't set R10BIO_Degraded as that
1271                                  * only applies if the disk is missing,
1272                                  * so it might be re-added, and we want to
1273                                  * know to recover this chunk.
1274                                  * In this case the device is here, and the
1275                                  * fact that this chunk is not in-sync is
1276                                  * recorded in the bad block log.
1277                                  */
1278                                 continue;
1279                         }
1280                         if (is_bad) {
1281                                 int good_sectors = first_bad - dev_sector;
1282                                 if (good_sectors < max_sectors)
1283                                         max_sectors = good_sectors;
1284                         }
1285                 }
1286                 if (rdev) {
1287                         r10_bio->devs[i].bio = bio;
1288                         atomic_inc(&rdev->nr_pending);
1289                 }
1290                 if (rrdev) {
1291                         r10_bio->devs[i].repl_bio = bio;
1292                         atomic_inc(&rrdev->nr_pending);
1293                 }
1294         }
1295         rcu_read_unlock();
1296
1297         if (unlikely(blocked_rdev)) {
1298                 /* Have to wait for this device to get unblocked, then retry */
1299                 int j;
1300                 int d;
1301
1302                 for (j = 0; j < i; j++) {
1303                         if (r10_bio->devs[j].bio) {
1304                                 d = r10_bio->devs[j].devnum;
1305                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1306                         }
1307                         if (r10_bio->devs[j].repl_bio) {
1308                                 struct md_rdev *rdev;
1309                                 d = r10_bio->devs[j].devnum;
1310                                 rdev = conf->mirrors[d].replacement;
1311                                 if (!rdev) {
1312                                         /* Race with remove_disk */
1313                                         smp_mb();
1314                                         rdev = conf->mirrors[d].rdev;
1315                                 }
1316                                 rdev_dec_pending(rdev, mddev);
1317                         }
1318                 }
1319                 allow_barrier(conf);
1320                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1321                 wait_barrier(conf);
1322                 goto retry_write;
1323         }
1324
1325         if (max_sectors < r10_bio->sectors) {
1326                 /* We are splitting this into multiple parts, so
1327                  * we need to prepare for allocating another r10_bio.
1328                  */
1329                 r10_bio->sectors = max_sectors;
1330                 spin_lock_irq(&conf->device_lock);
1331                 if (bio->bi_phys_segments == 0)
1332                         bio->bi_phys_segments = 2;
1333                 else
1334                         bio->bi_phys_segments++;
1335                 spin_unlock_irq(&conf->device_lock);
1336         }
1337         sectors_handled = r10_bio->sector + max_sectors -
1338                 bio->bi_iter.bi_sector;
1339
1340         atomic_set(&r10_bio->remaining, 1);
1341         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1342
1343         for (i = 0; i < conf->copies; i++) {
1344                 struct bio *mbio;
1345                 int d = r10_bio->devs[i].devnum;
1346                 if (r10_bio->devs[i].bio) {
1347                         struct md_rdev *rdev = conf->mirrors[d].rdev;
1348                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1349                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1350                                  max_sectors);
1351                         r10_bio->devs[i].bio = mbio;
1352
1353                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1354                                            choose_data_offset(r10_bio,
1355                                                               rdev));
1356                         mbio->bi_bdev = rdev->bdev;
1357                         mbio->bi_end_io = raid10_end_write_request;
1358                         mbio->bi_rw =
1359                                 WRITE | do_sync | do_fua | do_discard | do_same;
1360                         mbio->bi_private = r10_bio;
1361
1362                         atomic_inc(&r10_bio->remaining);
1363
1364                         cb = blk_check_plugged(raid10_unplug, mddev,
1365                                                sizeof(*plug));
1366                         if (cb)
1367                                 plug = container_of(cb, struct raid10_plug_cb,
1368                                                     cb);
1369                         else
1370                                 plug = NULL;
1371                         spin_lock_irqsave(&conf->device_lock, flags);
1372                         if (plug) {
1373                                 bio_list_add(&plug->pending, mbio);
1374                                 plug->pending_cnt++;
1375                         } else {
1376                                 bio_list_add(&conf->pending_bio_list, mbio);
1377                                 conf->pending_count++;
1378                         }
1379                         spin_unlock_irqrestore(&conf->device_lock, flags);
1380                         if (!plug)
1381                                 md_wakeup_thread(mddev->thread);
1382                 }
1383
1384                 if (r10_bio->devs[i].repl_bio) {
1385                         struct md_rdev *rdev = conf->mirrors[d].replacement;
1386                         if (rdev == NULL) {
1387                                 /* Replacement just got moved to main 'rdev' */
1388                                 smp_mb();
1389                                 rdev = conf->mirrors[d].rdev;
1390                         }
1391                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1392                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1393                                  max_sectors);
1394                         r10_bio->devs[i].repl_bio = mbio;
1395
1396                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1397                                            choose_data_offset(
1398                                                    r10_bio, rdev));
1399                         mbio->bi_bdev = rdev->bdev;
1400                         mbio->bi_end_io = raid10_end_write_request;
1401                         mbio->bi_rw =
1402                                 WRITE | do_sync | do_fua | do_discard | do_same;
1403                         mbio->bi_private = r10_bio;
1404
1405                         atomic_inc(&r10_bio->remaining);
1406                         spin_lock_irqsave(&conf->device_lock, flags);
1407                         bio_list_add(&conf->pending_bio_list, mbio);
1408                         conf->pending_count++;
1409                         spin_unlock_irqrestore(&conf->device_lock, flags);
1410                         if (!mddev_check_plugged(mddev))
1411                                 md_wakeup_thread(mddev->thread);
1412                 }
1413         }
1414
1415         /* Don't remove the bias on 'remaining' (one_write_done) until
1416          * after checking if we need to go around again.
1417          */
1418
1419         if (sectors_handled < bio_sectors(bio)) {
1420                 one_write_done(r10_bio);
1421                 /* We need another r10_bio.  It has already been counted
1422                  * in bio->bi_phys_segments.
1423                  */
1424                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1425
1426                 r10_bio->master_bio = bio;
1427                 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1428
1429                 r10_bio->mddev = mddev;
1430                 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1431                 r10_bio->state = 0;
1432                 goto retry_write;
1433         }
1434         one_write_done(r10_bio);
1435 }
1436
1437 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1438 {
1439         struct r10conf *conf = mddev->private;
1440         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1441         int chunk_sects = chunk_mask + 1;
1442
1443         struct bio *split;
1444
1445         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1446                 md_flush_request(mddev, bio);
1447                 return;
1448         }
1449
1450         md_write_start(mddev, bio);
1451
1452         do {
1453
1454                 /*
1455                  * If this request crosses a chunk boundary, we need to split
1456                  * it.
1457                  */
1458                 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1459                              bio_sectors(bio) > chunk_sects
1460                              && (conf->geo.near_copies < conf->geo.raid_disks
1461                                  || conf->prev.near_copies <
1462                                  conf->prev.raid_disks))) {
1463                         split = bio_split(bio, chunk_sects -
1464                                           (bio->bi_iter.bi_sector &
1465                                            (chunk_sects - 1)),
1466                                           GFP_NOIO, fs_bio_set);
1467                         bio_chain(split, bio);
1468                 } else {
1469                         split = bio;
1470                 }
1471
1472                 __make_request(mddev, split);
1473         } while (split != bio);
1474
1475         /* In case raid10d snuck in to freeze_array */
1476         wake_up(&conf->wait_barrier);
1477 }
1478
1479 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1480 {
1481         struct r10conf *conf = mddev->private;
1482         int i;
1483
1484         if (conf->geo.near_copies < conf->geo.raid_disks)
1485                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1486         if (conf->geo.near_copies > 1)
1487                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1488         if (conf->geo.far_copies > 1) {
1489                 if (conf->geo.far_offset)
1490                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1491                 else
1492                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1493                 if (conf->geo.far_set_size != conf->geo.raid_disks)
1494                         seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1495         }
1496         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1497                                         conf->geo.raid_disks - mddev->degraded);
1498         rcu_read_lock();
1499         for (i = 0; i < conf->geo.raid_disks; i++) {
1500                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1501                 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1502         }
1503         rcu_read_unlock();
1504         seq_printf(seq, "]");
1505 }
1506
1507 /* check if there are enough drives for
1508  * every block to appear on atleast one.
1509  * Don't consider the device numbered 'ignore'
1510  * as we might be about to remove it.
1511  */
1512 static int _enough(struct r10conf *conf, int previous, int ignore)
1513 {
1514         int first = 0;
1515         int has_enough = 0;
1516         int disks, ncopies;
1517         if (previous) {
1518                 disks = conf->prev.raid_disks;
1519                 ncopies = conf->prev.near_copies;
1520         } else {
1521                 disks = conf->geo.raid_disks;
1522                 ncopies = conf->geo.near_copies;
1523         }
1524
1525         rcu_read_lock();
1526         do {
1527                 int n = conf->copies;
1528                 int cnt = 0;
1529                 int this = first;
1530                 while (n--) {
1531                         struct md_rdev *rdev;
1532                         if (this != ignore &&
1533                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1534                             test_bit(In_sync, &rdev->flags))
1535                                 cnt++;
1536                         this = (this+1) % disks;
1537                 }
1538                 if (cnt == 0)
1539                         goto out;
1540                 first = (first + ncopies) % disks;
1541         } while (first != 0);
1542         has_enough = 1;
1543 out:
1544         rcu_read_unlock();
1545         return has_enough;
1546 }
1547
1548 static int enough(struct r10conf *conf, int ignore)
1549 {
1550         /* when calling 'enough', both 'prev' and 'geo' must
1551          * be stable.
1552          * This is ensured if ->reconfig_mutex or ->device_lock
1553          * is held.
1554          */
1555         return _enough(conf, 0, ignore) &&
1556                 _enough(conf, 1, ignore);
1557 }
1558
1559 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1560 {
1561         char b[BDEVNAME_SIZE];
1562         struct r10conf *conf = mddev->private;
1563         unsigned long flags;
1564
1565         /*
1566          * If it is not operational, then we have already marked it as dead
1567          * else if it is the last working disks, ignore the error, let the
1568          * next level up know.
1569          * else mark the drive as failed
1570          */
1571         spin_lock_irqsave(&conf->device_lock, flags);
1572         if (test_bit(In_sync, &rdev->flags)
1573             && !enough(conf, rdev->raid_disk)) {
1574                 /*
1575                  * Don't fail the drive, just return an IO error.
1576                  */
1577                 spin_unlock_irqrestore(&conf->device_lock, flags);
1578                 return;
1579         }
1580         if (test_and_clear_bit(In_sync, &rdev->flags))
1581                 mddev->degraded++;
1582         /*
1583          * If recovery is running, make sure it aborts.
1584          */
1585         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1586         set_bit(Blocked, &rdev->flags);
1587         set_bit(Faulty, &rdev->flags);
1588         set_mask_bits(&mddev->flags, 0,
1589                       BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1590         spin_unlock_irqrestore(&conf->device_lock, flags);
1591         printk(KERN_ALERT
1592                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1593                "md/raid10:%s: Operation continuing on %d devices.\n",
1594                mdname(mddev), bdevname(rdev->bdev, b),
1595                mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1596 }
1597
1598 static void print_conf(struct r10conf *conf)
1599 {
1600         int i;
1601         struct md_rdev *rdev;
1602
1603         printk(KERN_DEBUG "RAID10 conf printout:\n");
1604         if (!conf) {
1605                 printk(KERN_DEBUG "(!conf)\n");
1606                 return;
1607         }
1608         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1609                 conf->geo.raid_disks);
1610
1611         /* This is only called with ->reconfix_mutex held, so
1612          * rcu protection of rdev is not needed */
1613         for (i = 0; i < conf->geo.raid_disks; i++) {
1614                 char b[BDEVNAME_SIZE];
1615                 rdev = conf->mirrors[i].rdev;
1616                 if (rdev)
1617                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1618                                 i, !test_bit(In_sync, &rdev->flags),
1619                                 !test_bit(Faulty, &rdev->flags),
1620                                 bdevname(rdev->bdev,b));
1621         }
1622 }
1623
1624 static void close_sync(struct r10conf *conf)
1625 {
1626         wait_barrier(conf);
1627         allow_barrier(conf);
1628
1629         mempool_destroy(conf->r10buf_pool);
1630         conf->r10buf_pool = NULL;
1631 }
1632
1633 static int raid10_spare_active(struct mddev *mddev)
1634 {
1635         int i;
1636         struct r10conf *conf = mddev->private;
1637         struct raid10_info *tmp;
1638         int count = 0;
1639         unsigned long flags;
1640
1641         /*
1642          * Find all non-in_sync disks within the RAID10 configuration
1643          * and mark them in_sync
1644          */
1645         for (i = 0; i < conf->geo.raid_disks; i++) {
1646                 tmp = conf->mirrors + i;
1647                 if (tmp->replacement
1648                     && tmp->replacement->recovery_offset == MaxSector
1649                     && !test_bit(Faulty, &tmp->replacement->flags)
1650                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1651                         /* Replacement has just become active */
1652                         if (!tmp->rdev
1653                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1654                                 count++;
1655                         if (tmp->rdev) {
1656                                 /* Replaced device not technically faulty,
1657                                  * but we need to be sure it gets removed
1658                                  * and never re-added.
1659                                  */
1660                                 set_bit(Faulty, &tmp->rdev->flags);
1661                                 sysfs_notify_dirent_safe(
1662                                         tmp->rdev->sysfs_state);
1663                         }
1664                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1665                 } else if (tmp->rdev
1666                            && tmp->rdev->recovery_offset == MaxSector
1667                            && !test_bit(Faulty, &tmp->rdev->flags)
1668                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1669                         count++;
1670                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1671                 }
1672         }
1673         spin_lock_irqsave(&conf->device_lock, flags);
1674         mddev->degraded -= count;
1675         spin_unlock_irqrestore(&conf->device_lock, flags);
1676
1677         print_conf(conf);
1678         return count;
1679 }
1680
1681 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1682 {
1683         struct r10conf *conf = mddev->private;
1684         int err = -EEXIST;
1685         int mirror;
1686         int first = 0;
1687         int last = conf->geo.raid_disks - 1;
1688
1689         if (mddev->recovery_cp < MaxSector)
1690                 /* only hot-add to in-sync arrays, as recovery is
1691                  * very different from resync
1692                  */
1693                 return -EBUSY;
1694         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1695                 return -EINVAL;
1696
1697         if (md_integrity_add_rdev(rdev, mddev))
1698                 return -ENXIO;
1699
1700         if (rdev->raid_disk >= 0)
1701                 first = last = rdev->raid_disk;
1702
1703         if (rdev->saved_raid_disk >= first &&
1704             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1705                 mirror = rdev->saved_raid_disk;
1706         else
1707                 mirror = first;
1708         for ( ; mirror <= last ; mirror++) {
1709                 struct raid10_info *p = &conf->mirrors[mirror];
1710                 if (p->recovery_disabled == mddev->recovery_disabled)
1711                         continue;
1712                 if (p->rdev) {
1713                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1714                             p->replacement != NULL)
1715                                 continue;
1716                         clear_bit(In_sync, &rdev->flags);
1717                         set_bit(Replacement, &rdev->flags);
1718                         rdev->raid_disk = mirror;
1719                         err = 0;
1720                         if (mddev->gendisk)
1721                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1722                                                   rdev->data_offset << 9);
1723                         conf->fullsync = 1;
1724                         rcu_assign_pointer(p->replacement, rdev);
1725                         break;
1726                 }
1727
1728                 if (mddev->gendisk)
1729                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1730                                           rdev->data_offset << 9);
1731
1732                 p->head_position = 0;
1733                 p->recovery_disabled = mddev->recovery_disabled - 1;
1734                 rdev->raid_disk = mirror;
1735                 err = 0;
1736                 if (rdev->saved_raid_disk != mirror)
1737                         conf->fullsync = 1;
1738                 rcu_assign_pointer(p->rdev, rdev);
1739                 break;
1740         }
1741         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1742                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1743
1744         print_conf(conf);
1745         return err;
1746 }
1747
1748 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1749 {
1750         struct r10conf *conf = mddev->private;
1751         int err = 0;
1752         int number = rdev->raid_disk;
1753         struct md_rdev **rdevp;
1754         struct raid10_info *p = conf->mirrors + number;
1755
1756         print_conf(conf);
1757         if (rdev == p->rdev)
1758                 rdevp = &p->rdev;
1759         else if (rdev == p->replacement)
1760                 rdevp = &p->replacement;
1761         else
1762                 return 0;
1763
1764         if (test_bit(In_sync, &rdev->flags) ||
1765             atomic_read(&rdev->nr_pending)) {
1766                 err = -EBUSY;
1767                 goto abort;
1768         }
1769         /* Only remove non-faulty devices if recovery
1770          * is not possible.
1771          */
1772         if (!test_bit(Faulty, &rdev->flags) &&
1773             mddev->recovery_disabled != p->recovery_disabled &&
1774             (!p->replacement || p->replacement == rdev) &&
1775             number < conf->geo.raid_disks &&
1776             enough(conf, -1)) {
1777                 err = -EBUSY;
1778                 goto abort;
1779         }
1780         *rdevp = NULL;
1781         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1782                 synchronize_rcu();
1783                 if (atomic_read(&rdev->nr_pending)) {
1784                         /* lost the race, try later */
1785                         err = -EBUSY;
1786                         *rdevp = rdev;
1787                         goto abort;
1788                 }
1789         }
1790         if (p->replacement) {
1791                 /* We must have just cleared 'rdev' */
1792                 p->rdev = p->replacement;
1793                 clear_bit(Replacement, &p->replacement->flags);
1794                 smp_mb(); /* Make sure other CPUs may see both as identical
1795                            * but will never see neither -- if they are careful.
1796                            */
1797                 p->replacement = NULL;
1798                 clear_bit(WantReplacement, &rdev->flags);
1799         } else
1800                 /* We might have just remove the Replacement as faulty
1801                  * Clear the flag just in case
1802                  */
1803                 clear_bit(WantReplacement, &rdev->flags);
1804
1805         err = md_integrity_register(mddev);
1806
1807 abort:
1808
1809         print_conf(conf);
1810         return err;
1811 }
1812
1813 static void end_sync_read(struct bio *bio)
1814 {
1815         struct r10bio *r10_bio = bio->bi_private;
1816         struct r10conf *conf = r10_bio->mddev->private;
1817         int d;
1818
1819         if (bio == r10_bio->master_bio) {
1820                 /* this is a reshape read */
1821                 d = r10_bio->read_slot; /* really the read dev */
1822         } else
1823                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1824
1825         if (!bio->bi_error)
1826                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1827         else
1828                 /* The write handler will notice the lack of
1829                  * R10BIO_Uptodate and record any errors etc
1830                  */
1831                 atomic_add(r10_bio->sectors,
1832                            &conf->mirrors[d].rdev->corrected_errors);
1833
1834         /* for reconstruct, we always reschedule after a read.
1835          * for resync, only after all reads
1836          */
1837         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1838         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1839             atomic_dec_and_test(&r10_bio->remaining)) {
1840                 /* we have read all the blocks,
1841                  * do the comparison in process context in raid10d
1842                  */
1843                 reschedule_retry(r10_bio);
1844         }
1845 }
1846
1847 static void end_sync_request(struct r10bio *r10_bio)
1848 {
1849         struct mddev *mddev = r10_bio->mddev;
1850
1851         while (atomic_dec_and_test(&r10_bio->remaining)) {
1852                 if (r10_bio->master_bio == NULL) {
1853                         /* the primary of several recovery bios */
1854                         sector_t s = r10_bio->sectors;
1855                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1856                             test_bit(R10BIO_WriteError, &r10_bio->state))
1857                                 reschedule_retry(r10_bio);
1858                         else
1859                                 put_buf(r10_bio);
1860                         md_done_sync(mddev, s, 1);
1861                         break;
1862                 } else {
1863                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1864                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1865                             test_bit(R10BIO_WriteError, &r10_bio->state))
1866                                 reschedule_retry(r10_bio);
1867                         else
1868                                 put_buf(r10_bio);
1869                         r10_bio = r10_bio2;
1870                 }
1871         }
1872 }
1873
1874 static void end_sync_write(struct bio *bio)
1875 {
1876         struct r10bio *r10_bio = bio->bi_private;
1877         struct mddev *mddev = r10_bio->mddev;
1878         struct r10conf *conf = mddev->private;
1879         int d;
1880         sector_t first_bad;
1881         int bad_sectors;
1882         int slot;
1883         int repl;
1884         struct md_rdev *rdev = NULL;
1885
1886         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1887         if (repl)
1888                 rdev = conf->mirrors[d].replacement;
1889         else
1890                 rdev = conf->mirrors[d].rdev;
1891
1892         if (bio->bi_error) {
1893                 if (repl)
1894                         md_error(mddev, rdev);
1895                 else {
1896                         set_bit(WriteErrorSeen, &rdev->flags);
1897                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1898                                 set_bit(MD_RECOVERY_NEEDED,
1899                                         &rdev->mddev->recovery);
1900                         set_bit(R10BIO_WriteError, &r10_bio->state);
1901                 }
1902         } else if (is_badblock(rdev,
1903                              r10_bio->devs[slot].addr,
1904                              r10_bio->sectors,
1905                              &first_bad, &bad_sectors))
1906                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1907
1908         rdev_dec_pending(rdev, mddev);
1909
1910         end_sync_request(r10_bio);
1911 }
1912
1913 /*
1914  * Note: sync and recover and handled very differently for raid10
1915  * This code is for resync.
1916  * For resync, we read through virtual addresses and read all blocks.
1917  * If there is any error, we schedule a write.  The lowest numbered
1918  * drive is authoritative.
1919  * However requests come for physical address, so we need to map.
1920  * For every physical address there are raid_disks/copies virtual addresses,
1921  * which is always are least one, but is not necessarly an integer.
1922  * This means that a physical address can span multiple chunks, so we may
1923  * have to submit multiple io requests for a single sync request.
1924  */
1925 /*
1926  * We check if all blocks are in-sync and only write to blocks that
1927  * aren't in sync
1928  */
1929 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1930 {
1931         struct r10conf *conf = mddev->private;
1932         int i, first;
1933         struct bio *tbio, *fbio;
1934         int vcnt;
1935
1936         atomic_set(&r10_bio->remaining, 1);
1937
1938         /* find the first device with a block */
1939         for (i=0; i<conf->copies; i++)
1940                 if (!r10_bio->devs[i].bio->bi_error)
1941                         break;
1942
1943         if (i == conf->copies)
1944                 goto done;
1945
1946         first = i;
1947         fbio = r10_bio->devs[i].bio;
1948         fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1949         fbio->bi_iter.bi_idx = 0;
1950
1951         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1952         /* now find blocks with errors */
1953         for (i=0 ; i < conf->copies ; i++) {
1954                 int  j, d;
1955
1956                 tbio = r10_bio->devs[i].bio;
1957
1958                 if (tbio->bi_end_io != end_sync_read)
1959                         continue;
1960                 if (i == first)
1961                         continue;
1962                 if (!r10_bio->devs[i].bio->bi_error) {
1963                         /* We know that the bi_io_vec layout is the same for
1964                          * both 'first' and 'i', so we just compare them.
1965                          * All vec entries are PAGE_SIZE;
1966                          */
1967                         int sectors = r10_bio->sectors;
1968                         for (j = 0; j < vcnt; j++) {
1969                                 int len = PAGE_SIZE;
1970                                 if (sectors < (len / 512))
1971                                         len = sectors * 512;
1972                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1973                                            page_address(tbio->bi_io_vec[j].bv_page),
1974                                            len))
1975                                         break;
1976                                 sectors -= len/512;
1977                         }
1978                         if (j == vcnt)
1979                                 continue;
1980                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1981                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1982                                 /* Don't fix anything. */
1983                                 continue;
1984                 }
1985                 /* Ok, we need to write this bio, either to correct an
1986                  * inconsistency or to correct an unreadable block.
1987                  * First we need to fixup bv_offset, bv_len and
1988                  * bi_vecs, as the read request might have corrupted these
1989                  */
1990                 bio_reset(tbio);
1991
1992                 tbio->bi_vcnt = vcnt;
1993                 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
1994                 tbio->bi_rw = WRITE;
1995                 tbio->bi_private = r10_bio;
1996                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1997                 tbio->bi_end_io = end_sync_write;
1998
1999                 bio_copy_data(tbio, fbio);
2000
2001                 d = r10_bio->devs[i].devnum;
2002                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2003                 atomic_inc(&r10_bio->remaining);
2004                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2005
2006                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2007                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2008                 generic_make_request(tbio);
2009         }
2010
2011         /* Now write out to any replacement devices
2012          * that are active
2013          */
2014         for (i = 0; i < conf->copies; i++) {
2015                 int d;
2016
2017                 tbio = r10_bio->devs[i].repl_bio;
2018                 if (!tbio || !tbio->bi_end_io)
2019                         continue;
2020                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2021                     && r10_bio->devs[i].bio != fbio)
2022                         bio_copy_data(tbio, fbio);
2023                 d = r10_bio->devs[i].devnum;
2024                 atomic_inc(&r10_bio->remaining);
2025                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2026                              bio_sectors(tbio));
2027                 generic_make_request(tbio);
2028         }
2029
2030 done:
2031         if (atomic_dec_and_test(&r10_bio->remaining)) {
2032                 md_done_sync(mddev, r10_bio->sectors, 1);
2033                 put_buf(r10_bio);
2034         }
2035 }
2036
2037 /*
2038  * Now for the recovery code.
2039  * Recovery happens across physical sectors.
2040  * We recover all non-is_sync drives by finding the virtual address of
2041  * each, and then choose a working drive that also has that virt address.
2042  * There is a separate r10_bio for each non-in_sync drive.
2043  * Only the first two slots are in use. The first for reading,
2044  * The second for writing.
2045  *
2046  */
2047 static void fix_recovery_read_error(struct r10bio *r10_bio)
2048 {
2049         /* We got a read error during recovery.
2050          * We repeat the read in smaller page-sized sections.
2051          * If a read succeeds, write it to the new device or record
2052          * a bad block if we cannot.
2053          * If a read fails, record a bad block on both old and
2054          * new devices.
2055          */
2056         struct mddev *mddev = r10_bio->mddev;
2057         struct r10conf *conf = mddev->private;
2058         struct bio *bio = r10_bio->devs[0].bio;
2059         sector_t sect = 0;
2060         int sectors = r10_bio->sectors;
2061         int idx = 0;
2062         int dr = r10_bio->devs[0].devnum;
2063         int dw = r10_bio->devs[1].devnum;
2064
2065         while (sectors) {
2066                 int s = sectors;
2067                 struct md_rdev *rdev;
2068                 sector_t addr;
2069                 int ok;
2070
2071                 if (s > (PAGE_SIZE>>9))
2072                         s = PAGE_SIZE >> 9;
2073
2074                 rdev = conf->mirrors[dr].rdev;
2075                 addr = r10_bio->devs[0].addr + sect,
2076                 ok = sync_page_io(rdev,
2077                                   addr,
2078                                   s << 9,
2079                                   bio->bi_io_vec[idx].bv_page,
2080                                   READ, false);
2081                 if (ok) {
2082                         rdev = conf->mirrors[dw].rdev;
2083                         addr = r10_bio->devs[1].addr + sect;
2084                         ok = sync_page_io(rdev,
2085                                           addr,
2086                                           s << 9,
2087                                           bio->bi_io_vec[idx].bv_page,
2088                                           WRITE, false);
2089                         if (!ok) {
2090                                 set_bit(WriteErrorSeen, &rdev->flags);
2091                                 if (!test_and_set_bit(WantReplacement,
2092                                                       &rdev->flags))
2093                                         set_bit(MD_RECOVERY_NEEDED,
2094                                                 &rdev->mddev->recovery);
2095                         }
2096                 }
2097                 if (!ok) {
2098                         /* We don't worry if we cannot set a bad block -
2099                          * it really is bad so there is no loss in not
2100                          * recording it yet
2101                          */
2102                         rdev_set_badblocks(rdev, addr, s, 0);
2103
2104                         if (rdev != conf->mirrors[dw].rdev) {
2105                                 /* need bad block on destination too */
2106                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2107                                 addr = r10_bio->devs[1].addr + sect;
2108                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2109                                 if (!ok) {
2110                                         /* just abort the recovery */
2111                                         printk(KERN_NOTICE
2112                                                "md/raid10:%s: recovery aborted"
2113                                                " due to read error\n",
2114                                                mdname(mddev));
2115
2116                                         conf->mirrors[dw].recovery_disabled
2117                                                 = mddev->recovery_disabled;
2118                                         set_bit(MD_RECOVERY_INTR,
2119                                                 &mddev->recovery);
2120                                         break;
2121                                 }
2122                         }
2123                 }
2124
2125                 sectors -= s;
2126                 sect += s;
2127                 idx++;
2128         }
2129 }
2130
2131 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2132 {
2133         struct r10conf *conf = mddev->private;
2134         int d;
2135         struct bio *wbio, *wbio2;
2136
2137         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2138                 fix_recovery_read_error(r10_bio);
2139                 end_sync_request(r10_bio);
2140                 return;
2141         }
2142
2143         /*
2144          * share the pages with the first bio
2145          * and submit the write request
2146          */
2147         d = r10_bio->devs[1].devnum;
2148         wbio = r10_bio->devs[1].bio;
2149         wbio2 = r10_bio->devs[1].repl_bio;
2150         /* Need to test wbio2->bi_end_io before we call
2151          * generic_make_request as if the former is NULL,
2152          * the latter is free to free wbio2.
2153          */
2154         if (wbio2 && !wbio2->bi_end_io)
2155                 wbio2 = NULL;
2156         if (wbio->bi_end_io) {
2157                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2158                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2159                 generic_make_request(wbio);
2160         }
2161         if (wbio2) {
2162                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2163                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2164                              bio_sectors(wbio2));
2165                 generic_make_request(wbio2);
2166         }
2167 }
2168
2169 /*
2170  * Used by fix_read_error() to decay the per rdev read_errors.
2171  * We halve the read error count for every hour that has elapsed
2172  * since the last recorded read error.
2173  *
2174  */
2175 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2176 {
2177         long cur_time_mon;
2178         unsigned long hours_since_last;
2179         unsigned int read_errors = atomic_read(&rdev->read_errors);
2180
2181         cur_time_mon = ktime_get_seconds();
2182
2183         if (rdev->last_read_error == 0) {
2184                 /* first time we've seen a read error */
2185                 rdev->last_read_error = cur_time_mon;
2186                 return;
2187         }
2188
2189         hours_since_last = (long)(cur_time_mon -
2190                             rdev->last_read_error) / 3600;
2191
2192         rdev->last_read_error = cur_time_mon;
2193
2194         /*
2195          * if hours_since_last is > the number of bits in read_errors
2196          * just set read errors to 0. We do this to avoid
2197          * overflowing the shift of read_errors by hours_since_last.
2198          */
2199         if (hours_since_last >= 8 * sizeof(read_errors))
2200                 atomic_set(&rdev->read_errors, 0);
2201         else
2202                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2203 }
2204
2205 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2206                             int sectors, struct page *page, int rw)
2207 {
2208         sector_t first_bad;
2209         int bad_sectors;
2210
2211         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2212             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2213                 return -1;
2214         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2215                 /* success */
2216                 return 1;
2217         if (rw == WRITE) {
2218                 set_bit(WriteErrorSeen, &rdev->flags);
2219                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2220                         set_bit(MD_RECOVERY_NEEDED,
2221                                 &rdev->mddev->recovery);
2222         }
2223         /* need to record an error - either for the block or the device */
2224         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2225                 md_error(rdev->mddev, rdev);
2226         return 0;
2227 }
2228
2229 /*
2230  * This is a kernel thread which:
2231  *
2232  *      1.      Retries failed read operations on working mirrors.
2233  *      2.      Updates the raid superblock when problems encounter.
2234  *      3.      Performs writes following reads for array synchronising.
2235  */
2236
2237 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2238 {
2239         int sect = 0; /* Offset from r10_bio->sector */
2240         int sectors = r10_bio->sectors;
2241         struct md_rdev*rdev;
2242         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2243         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2244
2245         /* still own a reference to this rdev, so it cannot
2246          * have been cleared recently.
2247          */
2248         rdev = conf->mirrors[d].rdev;
2249
2250         if (test_bit(Faulty, &rdev->flags))
2251                 /* drive has already been failed, just ignore any
2252                    more fix_read_error() attempts */
2253                 return;
2254
2255         check_decay_read_errors(mddev, rdev);
2256         atomic_inc(&rdev->read_errors);
2257         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2258                 char b[BDEVNAME_SIZE];
2259                 bdevname(rdev->bdev, b);
2260
2261                 printk(KERN_NOTICE
2262                        "md/raid10:%s: %s: Raid device exceeded "
2263                        "read_error threshold [cur %d:max %d]\n",
2264                        mdname(mddev), b,
2265                        atomic_read(&rdev->read_errors), max_read_errors);
2266                 printk(KERN_NOTICE
2267                        "md/raid10:%s: %s: Failing raid device\n",
2268                        mdname(mddev), b);
2269                 md_error(mddev, rdev);
2270                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2271                 return;
2272         }
2273
2274         while(sectors) {
2275                 int s = sectors;
2276                 int sl = r10_bio->read_slot;
2277                 int success = 0;
2278                 int start;
2279
2280                 if (s > (PAGE_SIZE>>9))
2281                         s = PAGE_SIZE >> 9;
2282
2283                 rcu_read_lock();
2284                 do {
2285                         sector_t first_bad;
2286                         int bad_sectors;
2287
2288                         d = r10_bio->devs[sl].devnum;
2289                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2290                         if (rdev &&
2291                             test_bit(In_sync, &rdev->flags) &&
2292                             !test_bit(Faulty, &rdev->flags) &&
2293                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2294                                         &first_bad, &bad_sectors) == 0) {
2295                                 atomic_inc(&rdev->nr_pending);
2296                                 rcu_read_unlock();
2297                                 success = sync_page_io(rdev,
2298                                                        r10_bio->devs[sl].addr +
2299                                                        sect,
2300                                                        s<<9,
2301                                                        conf->tmppage, READ, false);
2302                                 rdev_dec_pending(rdev, mddev);
2303                                 rcu_read_lock();
2304                                 if (success)
2305                                         break;
2306                         }
2307                         sl++;
2308                         if (sl == conf->copies)
2309                                 sl = 0;
2310                 } while (!success && sl != r10_bio->read_slot);
2311                 rcu_read_unlock();
2312
2313                 if (!success) {
2314                         /* Cannot read from anywhere, just mark the block
2315                          * as bad on the first device to discourage future
2316                          * reads.
2317                          */
2318                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2319                         rdev = conf->mirrors[dn].rdev;
2320
2321                         if (!rdev_set_badblocks(
2322                                     rdev,
2323                                     r10_bio->devs[r10_bio->read_slot].addr
2324                                     + sect,
2325                                     s, 0)) {
2326                                 md_error(mddev, rdev);
2327                                 r10_bio->devs[r10_bio->read_slot].bio
2328                                         = IO_BLOCKED;
2329                         }
2330                         break;
2331                 }
2332
2333                 start = sl;
2334                 /* write it back and re-read */
2335                 rcu_read_lock();
2336                 while (sl != r10_bio->read_slot) {
2337                         char b[BDEVNAME_SIZE];
2338
2339                         if (sl==0)
2340                                 sl = conf->copies;
2341                         sl--;
2342                         d = r10_bio->devs[sl].devnum;
2343                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2344                         if (!rdev ||
2345                             test_bit(Faulty, &rdev->flags) ||
2346                             !test_bit(In_sync, &rdev->flags))
2347                                 continue;
2348
2349                         atomic_inc(&rdev->nr_pending);
2350                         rcu_read_unlock();
2351                         if (r10_sync_page_io(rdev,
2352                                              r10_bio->devs[sl].addr +
2353                                              sect,
2354                                              s, conf->tmppage, WRITE)
2355                             == 0) {
2356                                 /* Well, this device is dead */
2357                                 printk(KERN_NOTICE
2358                                        "md/raid10:%s: read correction "
2359                                        "write failed"
2360                                        " (%d sectors at %llu on %s)\n",
2361                                        mdname(mddev), s,
2362                                        (unsigned long long)(
2363                                                sect +
2364                                                choose_data_offset(r10_bio,
2365                                                                   rdev)),
2366                                        bdevname(rdev->bdev, b));
2367                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2368                                        "drive\n",
2369                                        mdname(mddev),
2370                                        bdevname(rdev->bdev, b));
2371                         }
2372                         rdev_dec_pending(rdev, mddev);
2373                         rcu_read_lock();
2374                 }
2375                 sl = start;
2376                 while (sl != r10_bio->read_slot) {
2377                         char b[BDEVNAME_SIZE];
2378
2379                         if (sl==0)
2380                                 sl = conf->copies;
2381                         sl--;
2382                         d = r10_bio->devs[sl].devnum;
2383                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2384                         if (!rdev ||
2385                             test_bit(Faulty, &rdev->flags) ||
2386                             !test_bit(In_sync, &rdev->flags))
2387                                 continue;
2388
2389                         atomic_inc(&rdev->nr_pending);
2390                         rcu_read_unlock();
2391                         switch (r10_sync_page_io(rdev,
2392                                              r10_bio->devs[sl].addr +
2393                                              sect,
2394                                              s, conf->tmppage,
2395                                                  READ)) {
2396                         case 0:
2397                                 /* Well, this device is dead */
2398                                 printk(KERN_NOTICE
2399                                        "md/raid10:%s: unable to read back "
2400                                        "corrected sectors"
2401                                        " (%d sectors at %llu on %s)\n",
2402                                        mdname(mddev), s,
2403                                        (unsigned long long)(
2404                                                sect +
2405                                                choose_data_offset(r10_bio, rdev)),
2406                                        bdevname(rdev->bdev, b));
2407                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2408                                        "drive\n",
2409                                        mdname(mddev),
2410                                        bdevname(rdev->bdev, b));
2411                                 break;
2412                         case 1:
2413                                 printk(KERN_INFO
2414                                        "md/raid10:%s: read error corrected"
2415                                        " (%d sectors at %llu on %s)\n",
2416                                        mdname(mddev), s,
2417                                        (unsigned long long)(
2418                                                sect +
2419                                                choose_data_offset(r10_bio, rdev)),
2420                                        bdevname(rdev->bdev, b));
2421                                 atomic_add(s, &rdev->corrected_errors);
2422                         }
2423
2424                         rdev_dec_pending(rdev, mddev);
2425                         rcu_read_lock();
2426                 }
2427                 rcu_read_unlock();
2428
2429                 sectors -= s;
2430                 sect += s;
2431         }
2432 }
2433
2434 static int narrow_write_error(struct r10bio *r10_bio, int i)
2435 {
2436         struct bio *bio = r10_bio->master_bio;
2437         struct mddev *mddev = r10_bio->mddev;
2438         struct r10conf *conf = mddev->private;
2439         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2440         /* bio has the data to be written to slot 'i' where
2441          * we just recently had a write error.
2442          * We repeatedly clone the bio and trim down to one block,
2443          * then try the write.  Where the write fails we record
2444          * a bad block.
2445          * It is conceivable that the bio doesn't exactly align with
2446          * blocks.  We must handle this.
2447          *
2448          * We currently own a reference to the rdev.
2449          */
2450
2451         int block_sectors;
2452         sector_t sector;
2453         int sectors;
2454         int sect_to_write = r10_bio->sectors;
2455         int ok = 1;
2456
2457         if (rdev->badblocks.shift < 0)
2458                 return 0;
2459
2460         block_sectors = roundup(1 << rdev->badblocks.shift,
2461                                 bdev_logical_block_size(rdev->bdev) >> 9);
2462         sector = r10_bio->sector;
2463         sectors = ((r10_bio->sector + block_sectors)
2464                    & ~(sector_t)(block_sectors - 1))
2465                 - sector;
2466
2467         while (sect_to_write) {
2468                 struct bio *wbio;
2469                 if (sectors > sect_to_write)
2470                         sectors = sect_to_write;
2471                 /* Write at 'sector' for 'sectors' */
2472                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2473                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2474                 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2475                                    choose_data_offset(r10_bio, rdev) +
2476                                    (sector - r10_bio->sector));
2477                 wbio->bi_bdev = rdev->bdev;
2478                 if (submit_bio_wait(WRITE, wbio) < 0)
2479                         /* Failure! */
2480                         ok = rdev_set_badblocks(rdev, sector,
2481                                                 sectors, 0)
2482                                 && ok;
2483
2484                 bio_put(wbio);
2485                 sect_to_write -= sectors;
2486                 sector += sectors;
2487                 sectors = block_sectors;
2488         }
2489         return ok;
2490 }
2491
2492 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2493 {
2494         int slot = r10_bio->read_slot;
2495         struct bio *bio;
2496         struct r10conf *conf = mddev->private;
2497         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2498         char b[BDEVNAME_SIZE];
2499         unsigned long do_sync;
2500         int max_sectors;
2501
2502         /* we got a read error. Maybe the drive is bad.  Maybe just
2503          * the block and we can fix it.
2504          * We freeze all other IO, and try reading the block from
2505          * other devices.  When we find one, we re-write
2506          * and check it that fixes the read error.
2507          * This is all done synchronously while the array is
2508          * frozen.
2509          */
2510         bio = r10_bio->devs[slot].bio;
2511         bdevname(bio->bi_bdev, b);
2512         bio_put(bio);
2513         r10_bio->devs[slot].bio = NULL;
2514
2515         if (mddev->ro == 0) {
2516                 freeze_array(conf, 1);
2517                 fix_read_error(conf, mddev, r10_bio);
2518                 unfreeze_array(conf);
2519         } else
2520                 r10_bio->devs[slot].bio = IO_BLOCKED;
2521
2522         rdev_dec_pending(rdev, mddev);
2523
2524 read_more:
2525         rdev = read_balance(conf, r10_bio, &max_sectors);
2526         if (rdev == NULL) {
2527                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2528                        " read error for block %llu\n",
2529                        mdname(mddev), b,
2530                        (unsigned long long)r10_bio->sector);
2531                 raid_end_bio_io(r10_bio);
2532                 return;
2533         }
2534
2535         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2536         slot = r10_bio->read_slot;
2537         printk_ratelimited(
2538                 KERN_ERR
2539                 "md/raid10:%s: %s: redirecting "
2540                 "sector %llu to another mirror\n",
2541                 mdname(mddev),
2542                 bdevname(rdev->bdev, b),
2543                 (unsigned long long)r10_bio->sector);
2544         bio = bio_clone_mddev(r10_bio->master_bio,
2545                               GFP_NOIO, mddev);
2546         bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2547         r10_bio->devs[slot].bio = bio;
2548         r10_bio->devs[slot].rdev = rdev;
2549         bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2550                 + choose_data_offset(r10_bio, rdev);
2551         bio->bi_bdev = rdev->bdev;
2552         bio->bi_rw = READ | do_sync;
2553         bio->bi_private = r10_bio;
2554         bio->bi_end_io = raid10_end_read_request;
2555         if (max_sectors < r10_bio->sectors) {
2556                 /* Drat - have to split this up more */
2557                 struct bio *mbio = r10_bio->master_bio;
2558                 int sectors_handled =
2559                         r10_bio->sector + max_sectors
2560                         - mbio->bi_iter.bi_sector;
2561                 r10_bio->sectors = max_sectors;
2562                 spin_lock_irq(&conf->device_lock);
2563                 if (mbio->bi_phys_segments == 0)
2564                         mbio->bi_phys_segments = 2;
2565                 else
2566                         mbio->bi_phys_segments++;
2567                 spin_unlock_irq(&conf->device_lock);
2568                 generic_make_request(bio);
2569
2570                 r10_bio = mempool_alloc(conf->r10bio_pool,
2571                                         GFP_NOIO);
2572                 r10_bio->master_bio = mbio;
2573                 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2574                 r10_bio->state = 0;
2575                 set_bit(R10BIO_ReadError,
2576                         &r10_bio->state);
2577                 r10_bio->mddev = mddev;
2578                 r10_bio->sector = mbio->bi_iter.bi_sector
2579                         + sectors_handled;
2580
2581                 goto read_more;
2582         } else
2583                 generic_make_request(bio);
2584 }
2585
2586 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2587 {
2588         /* Some sort of write request has finished and it
2589          * succeeded in writing where we thought there was a
2590          * bad block.  So forget the bad block.
2591          * Or possibly if failed and we need to record
2592          * a bad block.
2593          */
2594         int m;
2595         struct md_rdev *rdev;
2596
2597         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2598             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2599                 for (m = 0; m < conf->copies; m++) {
2600                         int dev = r10_bio->devs[m].devnum;
2601                         rdev = conf->mirrors[dev].rdev;
2602                         if (r10_bio->devs[m].bio == NULL)
2603                                 continue;
2604                         if (!r10_bio->devs[m].bio->bi_error) {
2605                                 rdev_clear_badblocks(
2606                                         rdev,
2607                                         r10_bio->devs[m].addr,
2608                                         r10_bio->sectors, 0);
2609                         } else {
2610                                 if (!rdev_set_badblocks(
2611                                             rdev,
2612                                             r10_bio->devs[m].addr,
2613                                             r10_bio->sectors, 0))
2614                                         md_error(conf->mddev, rdev);
2615                         }
2616                         rdev = conf->mirrors[dev].replacement;
2617                         if (r10_bio->devs[m].repl_bio == NULL)
2618                                 continue;
2619
2620                         if (!r10_bio->devs[m].repl_bio->bi_error) {
2621                                 rdev_clear_badblocks(
2622                                         rdev,
2623                                         r10_bio->devs[m].addr,
2624                                         r10_bio->sectors, 0);
2625                         } else {
2626                                 if (!rdev_set_badblocks(
2627                                             rdev,
2628                                             r10_bio->devs[m].addr,
2629                                             r10_bio->sectors, 0))
2630                                         md_error(conf->mddev, rdev);
2631                         }
2632                 }
2633                 put_buf(r10_bio);
2634         } else {
2635                 bool fail = false;
2636                 for (m = 0; m < conf->copies; m++) {
2637                         int dev = r10_bio->devs[m].devnum;
2638                         struct bio *bio = r10_bio->devs[m].bio;
2639                         rdev = conf->mirrors[dev].rdev;
2640                         if (bio == IO_MADE_GOOD) {
2641                                 rdev_clear_badblocks(
2642                                         rdev,
2643                                         r10_bio->devs[m].addr,
2644                                         r10_bio->sectors, 0);
2645                                 rdev_dec_pending(rdev, conf->mddev);
2646                         } else if (bio != NULL && bio->bi_error) {
2647                                 fail = true;
2648                                 if (!narrow_write_error(r10_bio, m)) {
2649                                         md_error(conf->mddev, rdev);
2650                                         set_bit(R10BIO_Degraded,
2651                                                 &r10_bio->state);
2652                                 }
2653                                 rdev_dec_pending(rdev, conf->mddev);
2654                         }
2655                         bio = r10_bio->devs[m].repl_bio;
2656                         rdev = conf->mirrors[dev].replacement;
2657                         if (rdev && bio == IO_MADE_GOOD) {
2658                                 rdev_clear_badblocks(
2659                                         rdev,
2660                                         r10_bio->devs[m].addr,
2661                                         r10_bio->sectors, 0);
2662                                 rdev_dec_pending(rdev, conf->mddev);
2663                         }
2664                 }
2665                 if (fail) {
2666                         spin_lock_irq(&conf->device_lock);
2667                         list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2668                         conf->nr_queued++;
2669                         spin_unlock_irq(&conf->device_lock);
2670                         md_wakeup_thread(conf->mddev->thread);
2671                 } else {
2672                         if (test_bit(R10BIO_WriteError,
2673                                      &r10_bio->state))
2674                                 close_write(r10_bio);
2675                         raid_end_bio_io(r10_bio);
2676                 }
2677         }
2678 }
2679
2680 static void raid10d(struct md_thread *thread)
2681 {
2682         struct mddev *mddev = thread->mddev;
2683         struct r10bio *r10_bio;
2684         unsigned long flags;
2685         struct r10conf *conf = mddev->private;
2686         struct list_head *head = &conf->retry_list;
2687         struct blk_plug plug;
2688
2689         md_check_recovery(mddev);
2690
2691         if (!list_empty_careful(&conf->bio_end_io_list) &&
2692             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2693                 LIST_HEAD(tmp);
2694                 spin_lock_irqsave(&conf->device_lock, flags);
2695                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2696                         while (!list_empty(&conf->bio_end_io_list)) {
2697                                 list_move(conf->bio_end_io_list.prev, &tmp);
2698                                 conf->nr_queued--;
2699                         }
2700                 }
2701                 spin_unlock_irqrestore(&conf->device_lock, flags);
2702                 while (!list_empty(&tmp)) {
2703                         r10_bio = list_first_entry(&tmp, struct r10bio,
2704                                                    retry_list);
2705                         list_del(&r10_bio->retry_list);
2706                         if (mddev->degraded)
2707                                 set_bit(R10BIO_Degraded, &r10_bio->state);
2708
2709                         if (test_bit(R10BIO_WriteError,
2710                                      &r10_bio->state))
2711                                 close_write(r10_bio);
2712                         raid_end_bio_io(r10_bio);
2713                 }
2714         }
2715
2716         blk_start_plug(&plug);
2717         for (;;) {
2718
2719                 flush_pending_writes(conf);
2720
2721                 spin_lock_irqsave(&conf->device_lock, flags);
2722                 if (list_empty(head)) {
2723                         spin_unlock_irqrestore(&conf->device_lock, flags);
2724                         break;
2725                 }
2726                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2727                 list_del(head->prev);
2728                 conf->nr_queued--;
2729                 spin_unlock_irqrestore(&conf->device_lock, flags);
2730
2731                 mddev = r10_bio->mddev;
2732                 conf = mddev->private;
2733                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2734                     test_bit(R10BIO_WriteError, &r10_bio->state))
2735                         handle_write_completed(conf, r10_bio);
2736                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2737                         reshape_request_write(mddev, r10_bio);
2738                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2739                         sync_request_write(mddev, r10_bio);
2740                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2741                         recovery_request_write(mddev, r10_bio);
2742                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2743                         handle_read_error(mddev, r10_bio);
2744                 else {
2745                         /* just a partial read to be scheduled from a
2746                          * separate context
2747                          */
2748                         int slot = r10_bio->read_slot;
2749                         generic_make_request(r10_bio->devs[slot].bio);
2750                 }
2751
2752                 cond_resched();
2753                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2754                         md_check_recovery(mddev);
2755         }
2756         blk_finish_plug(&plug);
2757 }
2758
2759 static int init_resync(struct r10conf *conf)
2760 {
2761         int buffs;
2762         int i;
2763
2764         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2765         BUG_ON(conf->r10buf_pool);
2766         conf->have_replacement = 0;
2767         for (i = 0; i < conf->geo.raid_disks; i++)
2768                 if (conf->mirrors[i].replacement)
2769                         conf->have_replacement = 1;
2770         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2771         if (!conf->r10buf_pool)
2772                 return -ENOMEM;
2773         conf->next_resync = 0;
2774         return 0;
2775 }
2776
2777 /*
2778  * perform a "sync" on one "block"
2779  *
2780  * We need to make sure that no normal I/O request - particularly write
2781  * requests - conflict with active sync requests.
2782  *
2783  * This is achieved by tracking pending requests and a 'barrier' concept
2784  * that can be installed to exclude normal IO requests.
2785  *
2786  * Resync and recovery are handled very differently.
2787  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2788  *
2789  * For resync, we iterate over virtual addresses, read all copies,
2790  * and update if there are differences.  If only one copy is live,
2791  * skip it.
2792  * For recovery, we iterate over physical addresses, read a good
2793  * value for each non-in_sync drive, and over-write.
2794  *
2795  * So, for recovery we may have several outstanding complex requests for a
2796  * given address, one for each out-of-sync device.  We model this by allocating
2797  * a number of r10_bio structures, one for each out-of-sync device.
2798  * As we setup these structures, we collect all bio's together into a list
2799  * which we then process collectively to add pages, and then process again
2800  * to pass to generic_make_request.
2801  *
2802  * The r10_bio structures are linked using a borrowed master_bio pointer.
2803  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2804  * has its remaining count decremented to 0, the whole complex operation
2805  * is complete.
2806  *
2807  */
2808
2809 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2810                              int *skipped)
2811 {
2812         struct r10conf *conf = mddev->private;
2813         struct r10bio *r10_bio;
2814         struct bio *biolist = NULL, *bio;
2815         sector_t max_sector, nr_sectors;
2816         int i;
2817         int max_sync;
2818         sector_t sync_blocks;
2819         sector_t sectors_skipped = 0;
2820         int chunks_skipped = 0;
2821         sector_t chunk_mask = conf->geo.chunk_mask;
2822
2823         if (!conf->r10buf_pool)
2824                 if (init_resync(conf))
2825                         return 0;
2826
2827         /*
2828          * Allow skipping a full rebuild for incremental assembly
2829          * of a clean array, like RAID1 does.
2830          */
2831         if (mddev->bitmap == NULL &&
2832             mddev->recovery_cp == MaxSector &&
2833             mddev->reshape_position == MaxSector &&
2834             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2835             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2836             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2837             conf->fullsync == 0) {
2838                 *skipped = 1;
2839                 return mddev->dev_sectors - sector_nr;
2840         }
2841
2842  skipped:
2843         max_sector = mddev->dev_sectors;
2844         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2845             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2846                 max_sector = mddev->resync_max_sectors;
2847         if (sector_nr >= max_sector) {
2848                 /* If we aborted, we need to abort the
2849                  * sync on the 'current' bitmap chucks (there can
2850                  * be several when recovering multiple devices).
2851                  * as we may have started syncing it but not finished.
2852                  * We can find the current address in
2853                  * mddev->curr_resync, but for recovery,
2854                  * we need to convert that to several
2855                  * virtual addresses.
2856                  */
2857                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2858                         end_reshape(conf);
2859                         close_sync(conf);
2860                         return 0;
2861                 }
2862
2863                 if (mddev->curr_resync < max_sector) { /* aborted */
2864                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2865                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2866                                                 &sync_blocks, 1);
2867                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2868                                 sector_t sect =
2869                                         raid10_find_virt(conf, mddev->curr_resync, i);
2870                                 bitmap_end_sync(mddev->bitmap, sect,
2871                                                 &sync_blocks, 1);
2872                         }
2873                 } else {
2874                         /* completed sync */
2875                         if ((!mddev->bitmap || conf->fullsync)
2876                             && conf->have_replacement
2877                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2878                                 /* Completed a full sync so the replacements
2879                                  * are now fully recovered.
2880                                  */
2881                                 rcu_read_lock();
2882                                 for (i = 0; i < conf->geo.raid_disks; i++) {
2883                                         struct md_rdev *rdev =
2884                                                 rcu_dereference(conf->mirrors[i].replacement);
2885                                         if (rdev)
2886                                                 rdev->recovery_offset = MaxSector;
2887                                 }
2888                                 rcu_read_unlock();
2889                         }
2890                         conf->fullsync = 0;
2891                 }
2892                 bitmap_close_sync(mddev->bitmap);
2893                 close_sync(conf);
2894                 *skipped = 1;
2895                 return sectors_skipped;
2896         }
2897
2898         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2899                 return reshape_request(mddev, sector_nr, skipped);
2900
2901         if (chunks_skipped >= conf->geo.raid_disks) {
2902                 /* if there has been nothing to do on any drive,
2903                  * then there is nothing to do at all..
2904                  */
2905                 *skipped = 1;
2906                 return (max_sector - sector_nr) + sectors_skipped;
2907         }
2908
2909         if (max_sector > mddev->resync_max)
2910                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2911
2912         /* make sure whole request will fit in a chunk - if chunks
2913          * are meaningful
2914          */
2915         if (conf->geo.near_copies < conf->geo.raid_disks &&
2916             max_sector > (sector_nr | chunk_mask))
2917                 max_sector = (sector_nr | chunk_mask) + 1;
2918
2919         /*
2920          * If there is non-resync activity waiting for a turn, then let it
2921          * though before starting on this new sync request.
2922          */
2923         if (conf->nr_waiting)
2924                 schedule_timeout_uninterruptible(1);
2925
2926         /* Again, very different code for resync and recovery.
2927          * Both must result in an r10bio with a list of bios that
2928          * have bi_end_io, bi_sector, bi_bdev set,
2929          * and bi_private set to the r10bio.
2930          * For recovery, we may actually create several r10bios
2931          * with 2 bios in each, that correspond to the bios in the main one.
2932          * In this case, the subordinate r10bios link back through a
2933          * borrowed master_bio pointer, and the counter in the master
2934          * includes a ref from each subordinate.
2935          */
2936         /* First, we decide what to do and set ->bi_end_io
2937          * To end_sync_read if we want to read, and
2938          * end_sync_write if we will want to write.
2939          */
2940
2941         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2942         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2943                 /* recovery... the complicated one */
2944                 int j;
2945                 r10_bio = NULL;
2946
2947                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2948                         int still_degraded;
2949                         struct r10bio *rb2;
2950                         sector_t sect;
2951                         int must_sync;
2952                         int any_working;
2953                         struct raid10_info *mirror = &conf->mirrors[i];
2954                         struct md_rdev *mrdev, *mreplace;
2955
2956                         rcu_read_lock();
2957                         mrdev = rcu_dereference(mirror->rdev);
2958                         mreplace = rcu_dereference(mirror->replacement);
2959
2960                         if ((mrdev == NULL ||
2961                              test_bit(Faulty, &mrdev->flags) ||
2962                              test_bit(In_sync, &mrdev->flags)) &&
2963                             (mreplace == NULL ||
2964                              test_bit(Faulty, &mreplace->flags))) {
2965                                 rcu_read_unlock();
2966                                 continue;
2967                         }
2968
2969                         still_degraded = 0;
2970                         /* want to reconstruct this device */
2971                         rb2 = r10_bio;
2972                         sect = raid10_find_virt(conf, sector_nr, i);
2973                         if (sect >= mddev->resync_max_sectors) {
2974                                 /* last stripe is not complete - don't
2975                                  * try to recover this sector.
2976                                  */
2977                                 rcu_read_unlock();
2978                                 continue;
2979                         }
2980                         if (mreplace && test_bit(Faulty, &mreplace->flags))
2981                                 mreplace = NULL;
2982                         /* Unless we are doing a full sync, or a replacement
2983                          * we only need to recover the block if it is set in
2984                          * the bitmap
2985                          */
2986                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2987                                                       &sync_blocks, 1);
2988                         if (sync_blocks < max_sync)
2989                                 max_sync = sync_blocks;
2990                         if (!must_sync &&
2991                             mreplace == NULL &&
2992                             !conf->fullsync) {
2993                                 /* yep, skip the sync_blocks here, but don't assume
2994                                  * that there will never be anything to do here
2995                                  */
2996                                 chunks_skipped = -1;
2997                                 rcu_read_unlock();
2998                                 continue;
2999                         }
3000                         atomic_inc(&mrdev->nr_pending);
3001                         if (mreplace)
3002                                 atomic_inc(&mreplace->nr_pending);
3003                         rcu_read_unlock();
3004
3005                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3006                         r10_bio->state = 0;
3007                         raise_barrier(conf, rb2 != NULL);
3008                         atomic_set(&r10_bio->remaining, 0);
3009
3010                         r10_bio->master_bio = (struct bio*)rb2;
3011                         if (rb2)
3012                                 atomic_inc(&rb2->remaining);
3013                         r10_bio->mddev = mddev;
3014                         set_bit(R10BIO_IsRecover, &r10_bio->state);
3015                         r10_bio->sector = sect;
3016
3017                         raid10_find_phys(conf, r10_bio);
3018
3019                         /* Need to check if the array will still be
3020                          * degraded
3021                          */
3022                         rcu_read_lock();
3023                         for (j = 0; j < conf->geo.raid_disks; j++) {
3024                                 struct md_rdev *rdev = rcu_dereference(
3025                                         conf->mirrors[j].rdev);
3026                                 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3027                                         still_degraded = 1;
3028                                         break;
3029                                 }
3030                         }
3031
3032                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
3033                                                       &sync_blocks, still_degraded);
3034
3035                         any_working = 0;
3036                         for (j=0; j<conf->copies;j++) {
3037                                 int k;
3038                                 int d = r10_bio->devs[j].devnum;
3039                                 sector_t from_addr, to_addr;
3040                                 struct md_rdev *rdev =
3041                                         rcu_dereference(conf->mirrors[d].rdev);
3042                                 sector_t sector, first_bad;
3043                                 int bad_sectors;
3044                                 if (!rdev ||
3045                                     !test_bit(In_sync, &rdev->flags))
3046                                         continue;
3047                                 /* This is where we read from */
3048                                 any_working = 1;
3049                                 sector = r10_bio->devs[j].addr;
3050
3051                                 if (is_badblock(rdev, sector, max_sync,
3052                                                 &first_bad, &bad_sectors)) {
3053                                         if (first_bad > sector)
3054                                                 max_sync = first_bad - sector;
3055                                         else {
3056                                                 bad_sectors -= (sector
3057                                                                 - first_bad);
3058                                                 if (max_sync > bad_sectors)
3059                                                         max_sync = bad_sectors;
3060                                                 continue;
3061                                         }
3062                                 }
3063                                 bio = r10_bio->devs[0].bio;
3064                                 bio_reset(bio);
3065                                 bio->bi_next = biolist;
3066                                 biolist = bio;
3067                                 bio->bi_private = r10_bio;
3068                                 bio->bi_end_io = end_sync_read;
3069                                 bio->bi_rw = READ;
3070                                 from_addr = r10_bio->devs[j].addr;
3071                                 bio->bi_iter.bi_sector = from_addr +
3072                                         rdev->data_offset;
3073                                 bio->bi_bdev = rdev->bdev;
3074                                 atomic_inc(&rdev->nr_pending);
3075                                 /* and we write to 'i' (if not in_sync) */
3076
3077                                 for (k=0; k<conf->copies; k++)
3078                                         if (r10_bio->devs[k].devnum == i)
3079                                                 break;
3080                                 BUG_ON(k == conf->copies);
3081                                 to_addr = r10_bio->devs[k].addr;
3082                                 r10_bio->devs[0].devnum = d;
3083                                 r10_bio->devs[0].addr = from_addr;
3084                                 r10_bio->devs[1].devnum = i;
3085                                 r10_bio->devs[1].addr = to_addr;
3086
3087                                 if (!test_bit(In_sync, &mrdev->flags)) {
3088                                         bio = r10_bio->devs[1].bio;
3089                                         bio_reset(bio);
3090                                         bio->bi_next = biolist;
3091                                         biolist = bio;
3092                                         bio->bi_private = r10_bio;
3093                                         bio->bi_end_io = end_sync_write;
3094                                         bio->bi_rw = WRITE;
3095                                         bio->bi_iter.bi_sector = to_addr
3096                                                 + mrdev->data_offset;
3097                                         bio->bi_bdev = mrdev->bdev;
3098                                         atomic_inc(&r10_bio->remaining);
3099                                 } else
3100                                         r10_bio->devs[1].bio->bi_end_io = NULL;
3101
3102                                 /* and maybe write to replacement */
3103                                 bio = r10_bio->devs[1].repl_bio;
3104                                 if (bio)
3105                                         bio->bi_end_io = NULL;
3106                                 /* Note: if mreplace != NULL, then bio
3107                                  * cannot be NULL as r10buf_pool_alloc will
3108                                  * have allocated it.
3109                                  * So the second test here is pointless.
3110                                  * But it keeps semantic-checkers happy, and
3111                                  * this comment keeps human reviewers
3112                                  * happy.
3113                                  */
3114                                 if (mreplace == NULL || bio == NULL ||
3115                                     test_bit(Faulty, &mreplace->flags))
3116                                         break;
3117                                 bio_reset(bio);
3118                                 bio->bi_next = biolist;
3119                                 biolist = bio;
3120                                 bio->bi_private = r10_bio;
3121                                 bio->bi_end_io = end_sync_write;
3122                                 bio->bi_rw = WRITE;
3123                                 bio->bi_iter.bi_sector = to_addr +
3124                                         mreplace->data_offset;
3125                                 bio->bi_bdev = mreplace->bdev;
3126                                 atomic_inc(&r10_bio->remaining);
3127                                 break;
3128                         }
3129                         rcu_read_unlock();
3130                         if (j == conf->copies) {
3131                                 /* Cannot recover, so abort the recovery or
3132                                  * record a bad block */
3133                                 if (any_working) {
3134                                         /* problem is that there are bad blocks
3135                                          * on other device(s)
3136                                          */
3137                                         int k;
3138                                         for (k = 0; k < conf->copies; k++)
3139                                                 if (r10_bio->devs[k].devnum == i)
3140                                                         break;
3141                                         if (!test_bit(In_sync,
3142                                                       &mrdev->flags)
3143                                             && !rdev_set_badblocks(
3144                                                     mrdev,
3145                                                     r10_bio->devs[k].addr,
3146                                                     max_sync, 0))
3147                                                 any_working = 0;
3148                                         if (mreplace &&
3149                                             !rdev_set_badblocks(
3150                                                     mreplace,
3151                                                     r10_bio->devs[k].addr,
3152                                                     max_sync, 0))
3153                                                 any_working = 0;
3154                                 }
3155                                 if (!any_working)  {
3156                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
3157                                                               &mddev->recovery))
3158                                                 printk(KERN_INFO "md/raid10:%s: insufficient "
3159                                                        "working devices for recovery.\n",
3160                                                        mdname(mddev));
3161                                         mirror->recovery_disabled
3162                                                 = mddev->recovery_disabled;
3163                                 }
3164                                 put_buf(r10_bio);
3165                                 if (rb2)
3166                                         atomic_dec(&rb2->remaining);
3167                                 r10_bio = rb2;
3168                                 rdev_dec_pending(mrdev, mddev);
3169                                 if (mreplace)
3170                                         rdev_dec_pending(mreplace, mddev);
3171                                 break;
3172                         }
3173                         rdev_dec_pending(mrdev, mddev);
3174                         if (mreplace)
3175                                 rdev_dec_pending(mreplace, mddev);
3176                 }
3177                 if (biolist == NULL) {
3178                         while (r10_bio) {
3179                                 struct r10bio *rb2 = r10_bio;
3180                                 r10_bio = (struct r10bio*) rb2->master_bio;
3181                                 rb2->master_bio = NULL;
3182                                 put_buf(rb2);
3183                         }
3184                         goto giveup;
3185                 }
3186         } else {
3187                 /* resync. Schedule a read for every block at this virt offset */
3188                 int count = 0;
3189
3190                 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3191
3192                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3193                                        &sync_blocks, mddev->degraded) &&
3194                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3195                                                  &mddev->recovery)) {
3196                         /* We can skip this block */
3197                         *skipped = 1;
3198                         return sync_blocks + sectors_skipped;
3199                 }
3200                 if (sync_blocks < max_sync)
3201                         max_sync = sync_blocks;
3202                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3203                 r10_bio->state = 0;
3204
3205                 r10_bio->mddev = mddev;
3206                 atomic_set(&r10_bio->remaining, 0);
3207                 raise_barrier(conf, 0);
3208                 conf->next_resync = sector_nr;
3209
3210                 r10_bio->master_bio = NULL;
3211                 r10_bio->sector = sector_nr;
3212                 set_bit(R10BIO_IsSync, &r10_bio->state);
3213                 raid10_find_phys(conf, r10_bio);
3214                 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3215
3216                 for (i = 0; i < conf->copies; i++) {
3217                         int d = r10_bio->devs[i].devnum;
3218                         sector_t first_bad, sector;
3219                         int bad_sectors;
3220                         struct md_rdev *rdev;
3221
3222                         if (r10_bio->devs[i].repl_bio)
3223                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3224
3225                         bio = r10_bio->devs[i].bio;
3226                         bio_reset(bio);
3227                         bio->bi_error = -EIO;
3228                         rcu_read_lock();
3229                         rdev = rcu_dereference(conf->mirrors[d].rdev);
3230                         if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3231                                 rcu_read_unlock();
3232                                 continue;
3233                         }
3234                         sector = r10_bio->devs[i].addr;
3235                         if (is_badblock(rdev, sector, max_sync,
3236                                         &first_bad, &bad_sectors)) {
3237                                 if (first_bad > sector)
3238                                         max_sync = first_bad - sector;
3239                                 else {
3240                                         bad_sectors -= (sector - first_bad);
3241                                         if (max_sync > bad_sectors)
3242                                                 max_sync = bad_sectors;
3243                                         rcu_read_unlock();
3244                                         continue;
3245                                 }
3246                         }
3247                         atomic_inc(&rdev->nr_pending);
3248                         atomic_inc(&r10_bio->remaining);
3249                         bio->bi_next = biolist;
3250                         biolist = bio;
3251                         bio->bi_private = r10_bio;
3252                         bio->bi_end_io = end_sync_read;
3253                         bio->bi_rw = READ;
3254                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3255                         bio->bi_bdev = rdev->bdev;
3256                         count++;
3257
3258                         rdev = rcu_dereference(conf->mirrors[d].replacement);
3259                         if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3260                                 rcu_read_unlock();
3261                                 continue;
3262                         }
3263                         atomic_inc(&rdev->nr_pending);
3264                         rcu_read_unlock();
3265
3266                         /* Need to set up for writing to the replacement */
3267                         bio = r10_bio->devs[i].repl_bio;
3268                         bio_reset(bio);
3269                         bio->bi_error = -EIO;
3270
3271                         sector = r10_bio->devs[i].addr;
3272                         bio->bi_next = biolist;
3273                         biolist = bio;
3274                         bio->bi_private = r10_bio;
3275                         bio->bi_end_io = end_sync_write;
3276                         bio->bi_rw = WRITE;
3277                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3278                         bio->bi_bdev = rdev->bdev;
3279                         count++;
3280                 }
3281
3282                 if (count < 2) {
3283                         for (i=0; i<conf->copies; i++) {
3284                                 int d = r10_bio->devs[i].devnum;
3285                                 if (r10_bio->devs[i].bio->bi_end_io)
3286                                         rdev_dec_pending(conf->mirrors[d].rdev,
3287                                                          mddev);
3288                                 if (r10_bio->devs[i].repl_bio &&
3289                                     r10_bio->devs[i].repl_bio->bi_end_io)
3290                                         rdev_dec_pending(
3291                                                 conf->mirrors[d].replacement,
3292                                                 mddev);
3293                         }
3294                         put_buf(r10_bio);
3295                         biolist = NULL;
3296                         goto giveup;
3297                 }
3298         }
3299
3300         nr_sectors = 0;
3301         if (sector_nr + max_sync < max_sector)
3302                 max_sector = sector_nr + max_sync;
3303         do {
3304                 struct page *page;
3305                 int len = PAGE_SIZE;
3306                 if (sector_nr + (len>>9) > max_sector)
3307                         len = (max_sector - sector_nr) << 9;
3308                 if (len == 0)
3309                         break;
3310                 for (bio= biolist ; bio ; bio=bio->bi_next) {
3311                         struct bio *bio2;
3312                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3313                         if (bio_add_page(bio, page, len, 0))
3314                                 continue;
3315
3316                         /* stop here */
3317                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3318                         for (bio2 = biolist;
3319                              bio2 && bio2 != bio;
3320                              bio2 = bio2->bi_next) {
3321                                 /* remove last page from this bio */
3322                                 bio2->bi_vcnt--;
3323                                 bio2->bi_iter.bi_size -= len;
3324                                 bio_clear_flag(bio2, BIO_SEG_VALID);
3325                         }
3326                         goto bio_full;
3327                 }
3328                 nr_sectors += len>>9;
3329                 sector_nr += len>>9;
3330         } while (biolist->bi_vcnt < RESYNC_PAGES);
3331  bio_full:
3332         r10_bio->sectors = nr_sectors;
3333
3334         while (biolist) {
3335                 bio = biolist;
3336                 biolist = biolist->bi_next;
3337
3338                 bio->bi_next = NULL;
3339                 r10_bio = bio->bi_private;
3340                 r10_bio->sectors = nr_sectors;
3341