a14c304aa751c117c5ef557459c13eddfd4453b5
[pandora-kernel.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  *
43  * The data to be stored is divided into chunks using chunksize.  Each device
44  * is divided into far_copies sections.   In each section, chunks are laid out
45  * in a style similar to raid0, but near_copies copies of each chunk is stored
46  * (each on a different drive).  The starting device for each section is offset
47  * near_copies from the starting device of the previous section.  Thus there
48  * are (near_copies * far_copies) of each chunk, and each is on a different
49  * drive.  near_copies and far_copies must be at least one, and their product
50  * is at most raid_disks.
51  *
52  * If far_offset is true, then the far_copies are handled a bit differently.
53  * The copies are still in different stripes, but instead of being very far
54  * apart on disk, there are adjacent stripes.
55  *
56  * The far and offset algorithms are handled slightly differently if
57  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
58  * sets that are (near_copies * far_copies) in size.  The far copied stripes
59  * are still shifted by 'near_copies' devices, but this shifting stays confined
60  * to the set rather than the entire array.  This is done to improve the number
61  * of device combinations that can fail without causing the array to fail.
62  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
63  * on a device):
64  *    A B C D    A B C D E
65  *      ...         ...
66  *    D A B C    E A B C D
67  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68  *    [A B] [C D]    [A B] [C D E]
69  *    |...| |...|    |...| | ... |
70  *    [B A] [D C]    [B A] [E C D]
71  */
72
73 /*
74  * Number of guaranteed r10bios in case of extreme VM load:
75  */
76 #define NR_RAID10_BIOS 256
77
78 /* when we get a read error on a read-only array, we redirect to another
79  * device without failing the first device, or trying to over-write to
80  * correct the read error.  To keep track of bad blocks on a per-bio
81  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
82  */
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85  * bad-block marking which must be done from process context.  So we record
86  * the success by setting devs[n].bio to IO_MADE_GOOD
87  */
88 #define IO_MADE_GOOD ((struct bio *)2)
89
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
91
92 /* When there are this many requests queued to be written by
93  * the raid10 thread, we become 'congested' to provide back-pressure
94  * for writeback.
95  */
96 static int max_queued_requests = 1024;
97
98 static void allow_barrier(struct r10conf *conf);
99 static void lower_barrier(struct r10conf *conf);
100 static int _enough(struct r10conf *conf, int previous, int ignore);
101 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
102                                 int *skipped);
103 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
104 static void end_reshape_write(struct bio *bio, int error);
105 static void end_reshape(struct r10conf *conf);
106
107 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
108 {
109         struct r10conf *conf = data;
110         int size = offsetof(struct r10bio, devs[conf->copies]);
111
112         /* allocate a r10bio with room for raid_disks entries in the
113          * bios array */
114         return kzalloc(size, gfp_flags);
115 }
116
117 static void r10bio_pool_free(void *r10_bio, void *data)
118 {
119         kfree(r10_bio);
120 }
121
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
129
130 /*
131  * When performing a resync, we need to read and compare, so
132  * we need as many pages are there are copies.
133  * When performing a recovery, we need 2 bios, one for read,
134  * one for write (we recover only one drive per r10buf)
135  *
136  */
137 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
138 {
139         struct r10conf *conf = data;
140         struct page *page;
141         struct r10bio *r10_bio;
142         struct bio *bio;
143         int i, j;
144         int nalloc;
145
146         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
147         if (!r10_bio)
148                 return NULL;
149
150         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
151             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
152                 nalloc = conf->copies; /* resync */
153         else
154                 nalloc = 2; /* recovery */
155
156         /*
157          * Allocate bios.
158          */
159         for (j = nalloc ; j-- ; ) {
160                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
161                 if (!bio)
162                         goto out_free_bio;
163                 r10_bio->devs[j].bio = bio;
164                 if (!conf->have_replacement)
165                         continue;
166                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
167                 if (!bio)
168                         goto out_free_bio;
169                 r10_bio->devs[j].repl_bio = bio;
170         }
171         /*
172          * Allocate RESYNC_PAGES data pages and attach them
173          * where needed.
174          */
175         for (j = 0 ; j < nalloc; j++) {
176                 struct bio *rbio = r10_bio->devs[j].repl_bio;
177                 bio = r10_bio->devs[j].bio;
178                 for (i = 0; i < RESYNC_PAGES; i++) {
179                         if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
180                                                &conf->mddev->recovery)) {
181                                 /* we can share bv_page's during recovery
182                                  * and reshape */
183                                 struct bio *rbio = r10_bio->devs[0].bio;
184                                 page = rbio->bi_io_vec[i].bv_page;
185                                 get_page(page);
186                         } else
187                                 page = alloc_page(gfp_flags);
188                         if (unlikely(!page))
189                                 goto out_free_pages;
190
191                         bio->bi_io_vec[i].bv_page = page;
192                         if (rbio)
193                                 rbio->bi_io_vec[i].bv_page = page;
194                 }
195         }
196
197         return r10_bio;
198
199 out_free_pages:
200         for ( ; i > 0 ; i--)
201                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
202         while (j--)
203                 for (i = 0; i < RESYNC_PAGES ; i++)
204                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
205         j = 0;
206 out_free_bio:
207         for ( ; j < nalloc; j++) {
208                 if (r10_bio->devs[j].bio)
209                         bio_put(r10_bio->devs[j].bio);
210                 if (r10_bio->devs[j].repl_bio)
211                         bio_put(r10_bio->devs[j].repl_bio);
212         }
213         r10bio_pool_free(r10_bio, conf);
214         return NULL;
215 }
216
217 static void r10buf_pool_free(void *__r10_bio, void *data)
218 {
219         int i;
220         struct r10conf *conf = data;
221         struct r10bio *r10bio = __r10_bio;
222         int j;
223
224         for (j=0; j < conf->copies; j++) {
225                 struct bio *bio = r10bio->devs[j].bio;
226                 if (bio) {
227                         for (i = 0; i < RESYNC_PAGES; i++) {
228                                 safe_put_page(bio->bi_io_vec[i].bv_page);
229                                 bio->bi_io_vec[i].bv_page = NULL;
230                         }
231                         bio_put(bio);
232                 }
233                 bio = r10bio->devs[j].repl_bio;
234                 if (bio)
235                         bio_put(bio);
236         }
237         r10bio_pool_free(r10bio, conf);
238 }
239
240 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
241 {
242         int i;
243
244         for (i = 0; i < conf->copies; i++) {
245                 struct bio **bio = & r10_bio->devs[i].bio;
246                 if (!BIO_SPECIAL(*bio))
247                         bio_put(*bio);
248                 *bio = NULL;
249                 bio = &r10_bio->devs[i].repl_bio;
250                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
251                         bio_put(*bio);
252                 *bio = NULL;
253         }
254 }
255
256 static void free_r10bio(struct r10bio *r10_bio)
257 {
258         struct r10conf *conf = r10_bio->mddev->private;
259
260         put_all_bios(conf, r10_bio);
261         mempool_free(r10_bio, conf->r10bio_pool);
262 }
263
264 static void put_buf(struct r10bio *r10_bio)
265 {
266         struct r10conf *conf = r10_bio->mddev->private;
267
268         mempool_free(r10_bio, conf->r10buf_pool);
269
270         lower_barrier(conf);
271 }
272
273 static void reschedule_retry(struct r10bio *r10_bio)
274 {
275         unsigned long flags;
276         struct mddev *mddev = r10_bio->mddev;
277         struct r10conf *conf = mddev->private;
278
279         spin_lock_irqsave(&conf->device_lock, flags);
280         list_add(&r10_bio->retry_list, &conf->retry_list);
281         conf->nr_queued ++;
282         spin_unlock_irqrestore(&conf->device_lock, flags);
283
284         /* wake up frozen array... */
285         wake_up(&conf->wait_barrier);
286
287         md_wakeup_thread(mddev->thread);
288 }
289
290 /*
291  * raid_end_bio_io() is called when we have finished servicing a mirrored
292  * operation and are ready to return a success/failure code to the buffer
293  * cache layer.
294  */
295 static void raid_end_bio_io(struct r10bio *r10_bio)
296 {
297         struct bio *bio = r10_bio->master_bio;
298         int done;
299         struct r10conf *conf = r10_bio->mddev->private;
300
301         if (bio->bi_phys_segments) {
302                 unsigned long flags;
303                 spin_lock_irqsave(&conf->device_lock, flags);
304                 bio->bi_phys_segments--;
305                 done = (bio->bi_phys_segments == 0);
306                 spin_unlock_irqrestore(&conf->device_lock, flags);
307         } else
308                 done = 1;
309         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
310                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
311         if (done) {
312                 bio_endio(bio, 0);
313                 /*
314                  * Wake up any possible resync thread that waits for the device
315                  * to go idle.
316                  */
317                 allow_barrier(conf);
318         }
319         free_r10bio(r10_bio);
320 }
321
322 /*
323  * Update disk head position estimator based on IRQ completion info.
324  */
325 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
326 {
327         struct r10conf *conf = r10_bio->mddev->private;
328
329         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
330                 r10_bio->devs[slot].addr + (r10_bio->sectors);
331 }
332
333 /*
334  * Find the disk number which triggered given bio
335  */
336 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
337                          struct bio *bio, int *slotp, int *replp)
338 {
339         int slot;
340         int repl = 0;
341
342         for (slot = 0; slot < conf->copies; slot++) {
343                 if (r10_bio->devs[slot].bio == bio)
344                         break;
345                 if (r10_bio->devs[slot].repl_bio == bio) {
346                         repl = 1;
347                         break;
348                 }
349         }
350
351         BUG_ON(slot == conf->copies);
352         update_head_pos(slot, r10_bio);
353
354         if (slotp)
355                 *slotp = slot;
356         if (replp)
357                 *replp = repl;
358         return r10_bio->devs[slot].devnum;
359 }
360
361 static void raid10_end_read_request(struct bio *bio, int error)
362 {
363         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364         struct r10bio *r10_bio = bio->bi_private;
365         int slot, dev;
366         struct md_rdev *rdev;
367         struct r10conf *conf = r10_bio->mddev->private;
368
369         slot = r10_bio->read_slot;
370         dev = r10_bio->devs[slot].devnum;
371         rdev = r10_bio->devs[slot].rdev;
372         /*
373          * this branch is our 'one mirror IO has finished' event handler:
374          */
375         update_head_pos(slot, r10_bio);
376
377         if (uptodate) {
378                 /*
379                  * Set R10BIO_Uptodate in our master bio, so that
380                  * we will return a good error code to the higher
381                  * levels even if IO on some other mirrored buffer fails.
382                  *
383                  * The 'master' represents the composite IO operation to
384                  * user-side. So if something waits for IO, then it will
385                  * wait for the 'master' bio.
386                  */
387                 set_bit(R10BIO_Uptodate, &r10_bio->state);
388         } else {
389                 /* If all other devices that store this block have
390                  * failed, we want to return the error upwards rather
391                  * than fail the last device.  Here we redefine
392                  * "uptodate" to mean "Don't want to retry"
393                  */
394                 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
395                              rdev->raid_disk))
396                         uptodate = 1;
397         }
398         if (uptodate) {
399                 raid_end_bio_io(r10_bio);
400                 rdev_dec_pending(rdev, conf->mddev);
401         } else {
402                 /*
403                  * oops, read error - keep the refcount on the rdev
404                  */
405                 char b[BDEVNAME_SIZE];
406                 printk_ratelimited(KERN_ERR
407                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
408                                    mdname(conf->mddev),
409                                    bdevname(rdev->bdev, b),
410                                    (unsigned long long)r10_bio->sector);
411                 set_bit(R10BIO_ReadError, &r10_bio->state);
412                 reschedule_retry(r10_bio);
413         }
414 }
415
416 static void close_write(struct r10bio *r10_bio)
417 {
418         /* clear the bitmap if all writes complete successfully */
419         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
420                         r10_bio->sectors,
421                         !test_bit(R10BIO_Degraded, &r10_bio->state),
422                         0);
423         md_write_end(r10_bio->mddev);
424 }
425
426 static void one_write_done(struct r10bio *r10_bio)
427 {
428         if (atomic_dec_and_test(&r10_bio->remaining)) {
429                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
430                         reschedule_retry(r10_bio);
431                 else {
432                         close_write(r10_bio);
433                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
434                                 reschedule_retry(r10_bio);
435                         else
436                                 raid_end_bio_io(r10_bio);
437                 }
438         }
439 }
440
441 static void raid10_end_write_request(struct bio *bio, int error)
442 {
443         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
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 (!uptodate) {
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  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
677  *      @mddev: the md device
678  *      @bvm: properties of new bio
679  *      @biovec: the request that could be merged to it.
680  *
681  *      Return amount of bytes we can accept at this offset
682  *      This requires checking for end-of-chunk if near_copies != raid_disks,
683  *      and for subordinate merge_bvec_fns if merge_check_needed.
684  */
685 static int raid10_mergeable_bvec(struct mddev *mddev,
686                                  struct bvec_merge_data *bvm,
687                                  struct bio_vec *biovec)
688 {
689         struct r10conf *conf = mddev->private;
690         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
691         int max;
692         unsigned int chunk_sectors;
693         unsigned int bio_sectors = bvm->bi_size >> 9;
694         struct geom *geo = &conf->geo;
695
696         chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
697         if (conf->reshape_progress != MaxSector &&
698             ((sector >= conf->reshape_progress) !=
699              conf->mddev->reshape_backwards))
700                 geo = &conf->prev;
701
702         if (geo->near_copies < geo->raid_disks) {
703                 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
704                                         + bio_sectors)) << 9;
705                 if (max < 0)
706                         /* bio_add cannot handle a negative return */
707                         max = 0;
708                 if (max <= biovec->bv_len && bio_sectors == 0)
709                         return biovec->bv_len;
710         } else
711                 max = biovec->bv_len;
712
713         if (mddev->merge_check_needed) {
714                 struct {
715                         struct r10bio r10_bio;
716                         struct r10dev devs[conf->copies];
717                 } on_stack;
718                 struct r10bio *r10_bio = &on_stack.r10_bio;
719                 int s;
720                 if (conf->reshape_progress != MaxSector) {
721                         /* Cannot give any guidance during reshape */
722                         if (max <= biovec->bv_len && bio_sectors == 0)
723                                 return biovec->bv_len;
724                         return 0;
725                 }
726                 r10_bio->sector = sector;
727                 raid10_find_phys(conf, r10_bio);
728                 rcu_read_lock();
729                 for (s = 0; s < conf->copies; s++) {
730                         int disk = r10_bio->devs[s].devnum;
731                         struct md_rdev *rdev = rcu_dereference(
732                                 conf->mirrors[disk].rdev);
733                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
734                                 struct request_queue *q =
735                                         bdev_get_queue(rdev->bdev);
736                                 if (q->merge_bvec_fn) {
737                                         bvm->bi_sector = r10_bio->devs[s].addr
738                                                 + rdev->data_offset;
739                                         bvm->bi_bdev = rdev->bdev;
740                                         max = min(max, q->merge_bvec_fn(
741                                                           q, bvm, biovec));
742                                 }
743                         }
744                         rdev = rcu_dereference(conf->mirrors[disk].replacement);
745                         if (rdev && !test_bit(Faulty, &rdev->flags)) {
746                                 struct request_queue *q =
747                                         bdev_get_queue(rdev->bdev);
748                                 if (q->merge_bvec_fn) {
749                                         bvm->bi_sector = r10_bio->devs[s].addr
750                                                 + rdev->data_offset;
751                                         bvm->bi_bdev = rdev->bdev;
752                                         max = min(max, q->merge_bvec_fn(
753                                                           q, bvm, biovec));
754                                 }
755                         }
756                 }
757                 rcu_read_unlock();
758         }
759         return max;
760 }
761
762 /*
763  * This routine returns the disk from which the requested read should
764  * be done. There is a per-array 'next expected sequential IO' sector
765  * number - if this matches on the next IO then we use the last disk.
766  * There is also a per-disk 'last know head position' sector that is
767  * maintained from IRQ contexts, both the normal and the resync IO
768  * completion handlers update this position correctly. If there is no
769  * perfect sequential match then we pick the disk whose head is closest.
770  *
771  * If there are 2 mirrors in the same 2 devices, performance degrades
772  * because position is mirror, not device based.
773  *
774  * The rdev for the device selected will have nr_pending incremented.
775  */
776
777 /*
778  * FIXME: possibly should rethink readbalancing and do it differently
779  * depending on near_copies / far_copies geometry.
780  */
781 static struct md_rdev *read_balance(struct r10conf *conf,
782                                     struct r10bio *r10_bio,
783                                     int *max_sectors)
784 {
785         const sector_t this_sector = r10_bio->sector;
786         int disk, slot;
787         int sectors = r10_bio->sectors;
788         int best_good_sectors;
789         sector_t new_distance, best_dist;
790         struct md_rdev *best_rdev, *rdev = NULL;
791         int do_balance;
792         int best_slot;
793         struct geom *geo = &conf->geo;
794
795         raid10_find_phys(conf, r10_bio);
796         rcu_read_lock();
797 retry:
798         sectors = r10_bio->sectors;
799         best_slot = -1;
800         best_rdev = NULL;
801         best_dist = MaxSector;
802         best_good_sectors = 0;
803         do_balance = 1;
804         /*
805          * Check if we can balance. We can balance on the whole
806          * device if no resync is going on (recovery is ok), or below
807          * the resync window. We take the first readable disk when
808          * above the resync window.
809          */
810         if (conf->mddev->recovery_cp < MaxSector
811             && (this_sector + sectors >= conf->next_resync))
812                 do_balance = 0;
813
814         for (slot = 0; slot < conf->copies ; slot++) {
815                 sector_t first_bad;
816                 int bad_sectors;
817                 sector_t dev_sector;
818
819                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
820                         continue;
821                 disk = r10_bio->devs[slot].devnum;
822                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
823                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
824                     test_bit(Unmerged, &rdev->flags) ||
825                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
826                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
827                 if (rdev == NULL ||
828                     test_bit(Faulty, &rdev->flags) ||
829                     test_bit(Unmerged, &rdev->flags))
830                         continue;
831                 if (!test_bit(In_sync, &rdev->flags) &&
832                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
833                         continue;
834
835                 dev_sector = r10_bio->devs[slot].addr;
836                 if (is_badblock(rdev, dev_sector, sectors,
837                                 &first_bad, &bad_sectors)) {
838                         if (best_dist < MaxSector)
839                                 /* Already have a better slot */
840                                 continue;
841                         if (first_bad <= dev_sector) {
842                                 /* Cannot read here.  If this is the
843                                  * 'primary' device, then we must not read
844                                  * beyond 'bad_sectors' from another device.
845                                  */
846                                 bad_sectors -= (dev_sector - first_bad);
847                                 if (!do_balance && sectors > bad_sectors)
848                                         sectors = bad_sectors;
849                                 if (best_good_sectors > sectors)
850                                         best_good_sectors = sectors;
851                         } else {
852                                 sector_t good_sectors =
853                                         first_bad - dev_sector;
854                                 if (good_sectors > best_good_sectors) {
855                                         best_good_sectors = good_sectors;
856                                         best_slot = slot;
857                                         best_rdev = rdev;
858                                 }
859                                 if (!do_balance)
860                                         /* Must read from here */
861                                         break;
862                         }
863                         continue;
864                 } else
865                         best_good_sectors = sectors;
866
867                 if (!do_balance)
868                         break;
869
870                 /* This optimisation is debatable, and completely destroys
871                  * sequential read speed for 'far copies' arrays.  So only
872                  * keep it for 'near' arrays, and review those later.
873                  */
874                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
875                         break;
876
877                 /* for far > 1 always use the lowest address */
878                 if (geo->far_copies > 1)
879                         new_distance = r10_bio->devs[slot].addr;
880                 else
881                         new_distance = abs(r10_bio->devs[slot].addr -
882                                            conf->mirrors[disk].head_position);
883                 if (new_distance < best_dist) {
884                         best_dist = new_distance;
885                         best_slot = slot;
886                         best_rdev = rdev;
887                 }
888         }
889         if (slot >= conf->copies) {
890                 slot = best_slot;
891                 rdev = best_rdev;
892         }
893
894         if (slot >= 0) {
895                 atomic_inc(&rdev->nr_pending);
896                 if (test_bit(Faulty, &rdev->flags)) {
897                         /* Cannot risk returning a device that failed
898                          * before we inc'ed nr_pending
899                          */
900                         rdev_dec_pending(rdev, conf->mddev);
901                         goto retry;
902                 }
903                 r10_bio->read_slot = slot;
904         } else
905                 rdev = NULL;
906         rcu_read_unlock();
907         *max_sectors = best_good_sectors;
908
909         return rdev;
910 }
911
912 static int raid10_congested(struct mddev *mddev, int bits)
913 {
914         struct r10conf *conf = mddev->private;
915         int i, ret = 0;
916
917         if ((bits & (1 << WB_async_congested)) &&
918             conf->pending_count >= max_queued_requests)
919                 return 1;
920
921         rcu_read_lock();
922         for (i = 0;
923              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
924                      && ret == 0;
925              i++) {
926                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
927                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
928                         struct request_queue *q = bdev_get_queue(rdev->bdev);
929
930                         ret |= bdi_congested(&q->backing_dev_info, bits);
931                 }
932         }
933         rcu_read_unlock();
934         return ret;
935 }
936
937 static void flush_pending_writes(struct r10conf *conf)
938 {
939         /* Any writes that have been queued but are awaiting
940          * bitmap updates get flushed here.
941          */
942         spin_lock_irq(&conf->device_lock);
943
944         if (conf->pending_bio_list.head) {
945                 struct bio *bio;
946                 bio = bio_list_get(&conf->pending_bio_list);
947                 conf->pending_count = 0;
948                 spin_unlock_irq(&conf->device_lock);
949                 /* flush any pending bitmap writes to disk
950                  * before proceeding w/ I/O */
951                 bitmap_unplug(conf->mddev->bitmap);
952                 wake_up(&conf->wait_barrier);
953
954                 while (bio) { /* submit pending writes */
955                         struct bio *next = bio->bi_next;
956                         bio->bi_next = NULL;
957                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
958                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
959                                 /* Just ignore it */
960                                 bio_endio(bio, 0);
961                         else
962                                 generic_make_request(bio);
963                         bio = next;
964                 }
965         } else
966                 spin_unlock_irq(&conf->device_lock);
967 }
968
969 /* Barriers....
970  * Sometimes we need to suspend IO while we do something else,
971  * either some resync/recovery, or reconfigure the array.
972  * To do this we raise a 'barrier'.
973  * The 'barrier' is a counter that can be raised multiple times
974  * to count how many activities are happening which preclude
975  * normal IO.
976  * We can only raise the barrier if there is no pending IO.
977  * i.e. if nr_pending == 0.
978  * We choose only to raise the barrier if no-one is waiting for the
979  * barrier to go down.  This means that as soon as an IO request
980  * is ready, no other operations which require a barrier will start
981  * until the IO request has had a chance.
982  *
983  * So: regular IO calls 'wait_barrier'.  When that returns there
984  *    is no backgroup IO happening,  It must arrange to call
985  *    allow_barrier when it has finished its IO.
986  * backgroup IO calls must call raise_barrier.  Once that returns
987  *    there is no normal IO happeing.  It must arrange to call
988  *    lower_barrier when the particular background IO completes.
989  */
990
991 static void raise_barrier(struct r10conf *conf, int force)
992 {
993         BUG_ON(force && !conf->barrier);
994         spin_lock_irq(&conf->resync_lock);
995
996         /* Wait until no block IO is waiting (unless 'force') */
997         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
998                             conf->resync_lock);
999
1000         /* block any new IO from starting */
1001         conf->barrier++;
1002
1003         /* Now wait for all pending IO to complete */
1004         wait_event_lock_irq(conf->wait_barrier,
1005                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1006                             conf->resync_lock);
1007
1008         spin_unlock_irq(&conf->resync_lock);
1009 }
1010
1011 static void lower_barrier(struct r10conf *conf)
1012 {
1013         unsigned long flags;
1014         spin_lock_irqsave(&conf->resync_lock, flags);
1015         conf->barrier--;
1016         spin_unlock_irqrestore(&conf->resync_lock, flags);
1017         wake_up(&conf->wait_barrier);
1018 }
1019
1020 static void wait_barrier(struct r10conf *conf)
1021 {
1022         spin_lock_irq(&conf->resync_lock);
1023         if (conf->barrier) {
1024                 conf->nr_waiting++;
1025                 /* Wait for the barrier to drop.
1026                  * However if there are already pending
1027                  * requests (preventing the barrier from
1028                  * rising completely), and the
1029                  * pre-process bio queue isn't empty,
1030                  * then don't wait, as we need to empty
1031                  * that queue to get the nr_pending
1032                  * count down.
1033                  */
1034                 wait_event_lock_irq(conf->wait_barrier,
1035                                     !conf->barrier ||
1036                                     (conf->nr_pending &&
1037                                      current->bio_list &&
1038                                      !bio_list_empty(current->bio_list)),
1039                                     conf->resync_lock);
1040                 conf->nr_waiting--;
1041         }
1042         conf->nr_pending++;
1043         spin_unlock_irq(&conf->resync_lock);
1044 }
1045
1046 static void allow_barrier(struct r10conf *conf)
1047 {
1048         unsigned long flags;
1049         spin_lock_irqsave(&conf->resync_lock, flags);
1050         conf->nr_pending--;
1051         spin_unlock_irqrestore(&conf->resync_lock, flags);
1052         wake_up(&conf->wait_barrier);
1053 }
1054
1055 static void freeze_array(struct r10conf *conf, int extra)
1056 {
1057         /* stop syncio and normal IO and wait for everything to
1058          * go quiet.
1059          * We increment barrier and nr_waiting, and then
1060          * wait until nr_pending match nr_queued+extra
1061          * This is called in the context of one normal IO request
1062          * that has failed. Thus any sync request that might be pending
1063          * will be blocked by nr_pending, and we need to wait for
1064          * pending IO requests to complete or be queued for re-try.
1065          * Thus the number queued (nr_queued) plus this request (extra)
1066          * must match the number of pending IOs (nr_pending) before
1067          * we continue.
1068          */
1069         spin_lock_irq(&conf->resync_lock);
1070         conf->barrier++;
1071         conf->nr_waiting++;
1072         wait_event_lock_irq_cmd(conf->wait_barrier,
1073                                 conf->nr_pending == conf->nr_queued+extra,
1074                                 conf->resync_lock,
1075                                 flush_pending_writes(conf));
1076
1077         spin_unlock_irq(&conf->resync_lock);
1078 }
1079
1080 static void unfreeze_array(struct r10conf *conf)
1081 {
1082         /* reverse the effect of the freeze */
1083         spin_lock_irq(&conf->resync_lock);
1084         conf->barrier--;
1085         conf->nr_waiting--;
1086         wake_up(&conf->wait_barrier);
1087         spin_unlock_irq(&conf->resync_lock);
1088 }
1089
1090 static sector_t choose_data_offset(struct r10bio *r10_bio,
1091                                    struct md_rdev *rdev)
1092 {
1093         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1094             test_bit(R10BIO_Previous, &r10_bio->state))
1095                 return rdev->data_offset;
1096         else
1097                 return rdev->new_data_offset;
1098 }
1099
1100 struct raid10_plug_cb {
1101         struct blk_plug_cb      cb;
1102         struct bio_list         pending;
1103         int                     pending_cnt;
1104 };
1105
1106 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1107 {
1108         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1109                                                    cb);
1110         struct mddev *mddev = plug->cb.data;
1111         struct r10conf *conf = mddev->private;
1112         struct bio *bio;
1113
1114         if (from_schedule || current->bio_list) {
1115                 spin_lock_irq(&conf->device_lock);
1116                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1117                 conf->pending_count += plug->pending_cnt;
1118                 spin_unlock_irq(&conf->device_lock);
1119                 wake_up(&conf->wait_barrier);
1120                 md_wakeup_thread(mddev->thread);
1121                 kfree(plug);
1122                 return;
1123         }
1124
1125         /* we aren't scheduling, so we can do the write-out directly. */
1126         bio = bio_list_get(&plug->pending);
1127         bitmap_unplug(mddev->bitmap);
1128         wake_up(&conf->wait_barrier);
1129
1130         while (bio) { /* submit pending writes */
1131                 struct bio *next = bio->bi_next;
1132                 bio->bi_next = NULL;
1133                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1134                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1135                         /* Just ignore it */
1136                         bio_endio(bio, 0);
1137                 else
1138                         generic_make_request(bio);
1139                 bio = next;
1140         }
1141         kfree(plug);
1142 }
1143
1144 static void __make_request(struct mddev *mddev, struct bio *bio)
1145 {
1146         struct r10conf *conf = mddev->private;
1147         struct r10bio *r10_bio;
1148         struct bio *read_bio;
1149         int i;
1150         const int rw = bio_data_dir(bio);
1151         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1152         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1153         const unsigned long do_discard = (bio->bi_rw
1154                                           & (REQ_DISCARD | REQ_SECURE));
1155         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1156         unsigned long flags;
1157         struct md_rdev *blocked_rdev;
1158         struct blk_plug_cb *cb;
1159         struct raid10_plug_cb *plug = NULL;
1160         int sectors_handled;
1161         int max_sectors;
1162         int sectors;
1163
1164         /*
1165          * Register the new request and wait if the reconstruction
1166          * thread has put up a bar for new requests.
1167          * Continue immediately if no resync is active currently.
1168          */
1169         wait_barrier(conf);
1170
1171         sectors = bio_sectors(bio);
1172         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1173             bio->bi_iter.bi_sector < conf->reshape_progress &&
1174             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1175                 /* IO spans the reshape position.  Need to wait for
1176                  * reshape to pass
1177                  */
1178                 allow_barrier(conf);
1179                 wait_event(conf->wait_barrier,
1180                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1181                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1182                            sectors);
1183                 wait_barrier(conf);
1184         }
1185         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1186             bio_data_dir(bio) == WRITE &&
1187             (mddev->reshape_backwards
1188              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1189                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1190              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1191                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1192                 /* Need to update reshape_position in metadata */
1193                 mddev->reshape_position = conf->reshape_progress;
1194                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1195                 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1196                 md_wakeup_thread(mddev->thread);
1197                 wait_event(mddev->sb_wait,
1198                            !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1199
1200                 conf->reshape_safe = mddev->reshape_position;
1201         }
1202
1203         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1204
1205         r10_bio->master_bio = bio;
1206         r10_bio->sectors = sectors;
1207
1208         r10_bio->mddev = mddev;
1209         r10_bio->sector = bio->bi_iter.bi_sector;
1210         r10_bio->state = 0;
1211
1212         /* We might need to issue multiple reads to different
1213          * devices if there are bad blocks around, so we keep
1214          * track of the number of reads in bio->bi_phys_segments.
1215          * If this is 0, there is only one r10_bio and no locking
1216          * will be needed when the request completes.  If it is
1217          * non-zero, then it is the number of not-completed requests.
1218          */
1219         bio->bi_phys_segments = 0;
1220         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1221
1222         if (rw == READ) {
1223                 /*
1224                  * read balancing logic:
1225                  */
1226                 struct md_rdev *rdev;
1227                 int slot;
1228
1229 read_again:
1230                 rdev = read_balance(conf, r10_bio, &max_sectors);
1231                 if (!rdev) {
1232                         raid_end_bio_io(r10_bio);
1233                         return;
1234                 }
1235                 slot = r10_bio->read_slot;
1236
1237                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1238                 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1239                          max_sectors);
1240
1241                 r10_bio->devs[slot].bio = read_bio;
1242                 r10_bio->devs[slot].rdev = rdev;
1243
1244                 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1245                         choose_data_offset(r10_bio, rdev);
1246                 read_bio->bi_bdev = rdev->bdev;
1247                 read_bio->bi_end_io = raid10_end_read_request;
1248                 read_bio->bi_rw = READ | do_sync;
1249                 read_bio->bi_private = r10_bio;
1250
1251                 if (max_sectors < r10_bio->sectors) {
1252                         /* Could not read all from this device, so we will
1253                          * need another r10_bio.
1254                          */
1255                         sectors_handled = (r10_bio->sector + max_sectors
1256                                            - bio->bi_iter.bi_sector);
1257                         r10_bio->sectors = max_sectors;
1258                         spin_lock_irq(&conf->device_lock);
1259                         if (bio->bi_phys_segments == 0)
1260                                 bio->bi_phys_segments = 2;
1261                         else
1262                                 bio->bi_phys_segments++;
1263                         spin_unlock_irq(&conf->device_lock);
1264                         /* Cannot call generic_make_request directly
1265                          * as that will be queued in __generic_make_request
1266                          * and subsequent mempool_alloc might block
1267                          * waiting for it.  so hand bio over to raid10d.
1268                          */
1269                         reschedule_retry(r10_bio);
1270
1271                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1272
1273                         r10_bio->master_bio = bio;
1274                         r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1275                         r10_bio->state = 0;
1276                         r10_bio->mddev = mddev;
1277                         r10_bio->sector = bio->bi_iter.bi_sector +
1278                                 sectors_handled;
1279                         goto read_again;
1280                 } else
1281                         generic_make_request(read_bio);
1282                 return;
1283         }
1284
1285         /*
1286          * WRITE:
1287          */
1288         if (conf->pending_count >= max_queued_requests) {
1289                 md_wakeup_thread(mddev->thread);
1290                 wait_event(conf->wait_barrier,
1291                            conf->pending_count < max_queued_requests);
1292         }
1293         /* first select target devices under rcu_lock and
1294          * inc refcount on their rdev.  Record them by setting
1295          * bios[x] to bio
1296          * If there are known/acknowledged bad blocks on any device
1297          * on which we have seen a write error, we want to avoid
1298          * writing to those blocks.  This potentially requires several
1299          * writes to write around the bad blocks.  Each set of writes
1300          * gets its own r10_bio with a set of bios attached.  The number
1301          * of r10_bios is recored in bio->bi_phys_segments just as with
1302          * the read case.
1303          */
1304
1305         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1306         raid10_find_phys(conf, r10_bio);
1307 retry_write:
1308         blocked_rdev = NULL;
1309         rcu_read_lock();
1310         max_sectors = r10_bio->sectors;
1311
1312         for (i = 0;  i < conf->copies; i++) {
1313                 int d = r10_bio->devs[i].devnum;
1314                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1315                 struct md_rdev *rrdev = rcu_dereference(
1316                         conf->mirrors[d].replacement);
1317                 if (rdev == rrdev)
1318                         rrdev = NULL;
1319                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1320                         atomic_inc(&rdev->nr_pending);
1321                         blocked_rdev = rdev;
1322                         break;
1323                 }
1324                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1325                         atomic_inc(&rrdev->nr_pending);
1326                         blocked_rdev = rrdev;
1327                         break;
1328                 }
1329                 if (rdev && (test_bit(Faulty, &rdev->flags)
1330                              || test_bit(Unmerged, &rdev->flags)))
1331                         rdev = NULL;
1332                 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1333                               || test_bit(Unmerged, &rrdev->flags)))
1334                         rrdev = NULL;
1335
1336                 r10_bio->devs[i].bio = NULL;
1337                 r10_bio->devs[i].repl_bio = NULL;
1338
1339                 if (!rdev && !rrdev) {
1340                         set_bit(R10BIO_Degraded, &r10_bio->state);
1341                         continue;
1342                 }
1343                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1344                         sector_t first_bad;
1345                         sector_t dev_sector = r10_bio->devs[i].addr;
1346                         int bad_sectors;
1347                         int is_bad;
1348
1349                         is_bad = is_badblock(rdev, dev_sector,
1350                                              max_sectors,
1351                                              &first_bad, &bad_sectors);
1352                         if (is_bad < 0) {
1353                                 /* Mustn't write here until the bad block
1354                                  * is acknowledged
1355                                  */
1356                                 atomic_inc(&rdev->nr_pending);
1357                                 set_bit(BlockedBadBlocks, &rdev->flags);
1358                                 blocked_rdev = rdev;
1359                                 break;
1360                         }
1361                         if (is_bad && first_bad <= dev_sector) {
1362                                 /* Cannot write here at all */
1363                                 bad_sectors -= (dev_sector - first_bad);
1364                                 if (bad_sectors < max_sectors)
1365                                         /* Mustn't write more than bad_sectors
1366                                          * to other devices yet
1367                                          */
1368                                         max_sectors = bad_sectors;
1369                                 /* We don't set R10BIO_Degraded as that
1370                                  * only applies if the disk is missing,
1371                                  * so it might be re-added, and we want to
1372                                  * know to recover this chunk.
1373                                  * In this case the device is here, and the
1374                                  * fact that this chunk is not in-sync is
1375                                  * recorded in the bad block log.
1376                                  */
1377                                 continue;
1378                         }
1379                         if (is_bad) {
1380                                 int good_sectors = first_bad - dev_sector;
1381                                 if (good_sectors < max_sectors)
1382                                         max_sectors = good_sectors;
1383                         }
1384                 }
1385                 if (rdev) {
1386                         r10_bio->devs[i].bio = bio;
1387                         atomic_inc(&rdev->nr_pending);
1388                 }
1389                 if (rrdev) {
1390                         r10_bio->devs[i].repl_bio = bio;
1391                         atomic_inc(&rrdev->nr_pending);
1392                 }
1393         }
1394         rcu_read_unlock();
1395
1396         if (unlikely(blocked_rdev)) {
1397                 /* Have to wait for this device to get unblocked, then retry */
1398                 int j;
1399                 int d;
1400
1401                 for (j = 0; j < i; j++) {
1402                         if (r10_bio->devs[j].bio) {
1403                                 d = r10_bio->devs[j].devnum;
1404                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1405                         }
1406                         if (r10_bio->devs[j].repl_bio) {
1407                                 struct md_rdev *rdev;
1408                                 d = r10_bio->devs[j].devnum;
1409                                 rdev = conf->mirrors[d].replacement;
1410                                 if (!rdev) {
1411                                         /* Race with remove_disk */
1412                                         smp_mb();
1413                                         rdev = conf->mirrors[d].rdev;
1414                                 }
1415                                 rdev_dec_pending(rdev, mddev);
1416                         }
1417                 }
1418                 allow_barrier(conf);
1419                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1420                 wait_barrier(conf);
1421                 goto retry_write;
1422         }
1423
1424         if (max_sectors < r10_bio->sectors) {
1425                 /* We are splitting this into multiple parts, so
1426                  * we need to prepare for allocating another r10_bio.
1427                  */
1428                 r10_bio->sectors = max_sectors;
1429                 spin_lock_irq(&conf->device_lock);
1430                 if (bio->bi_phys_segments == 0)
1431                         bio->bi_phys_segments = 2;
1432                 else
1433                         bio->bi_phys_segments++;
1434                 spin_unlock_irq(&conf->device_lock);
1435         }
1436         sectors_handled = r10_bio->sector + max_sectors -
1437                 bio->bi_iter.bi_sector;
1438
1439         atomic_set(&r10_bio->remaining, 1);
1440         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1441
1442         for (i = 0; i < conf->copies; i++) {
1443                 struct bio *mbio;
1444                 int d = r10_bio->devs[i].devnum;
1445                 if (r10_bio->devs[i].bio) {
1446                         struct md_rdev *rdev = conf->mirrors[d].rdev;
1447                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1448                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1449                                  max_sectors);
1450                         r10_bio->devs[i].bio = mbio;
1451
1452                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1453                                            choose_data_offset(r10_bio,
1454                                                               rdev));
1455                         mbio->bi_bdev = rdev->bdev;
1456                         mbio->bi_end_io = raid10_end_write_request;
1457                         mbio->bi_rw =
1458                                 WRITE | do_sync | do_fua | do_discard | do_same;
1459                         mbio->bi_private = r10_bio;
1460
1461                         atomic_inc(&r10_bio->remaining);
1462
1463                         cb = blk_check_plugged(raid10_unplug, mddev,
1464                                                sizeof(*plug));
1465                         if (cb)
1466                                 plug = container_of(cb, struct raid10_plug_cb,
1467                                                     cb);
1468                         else
1469                                 plug = NULL;
1470                         spin_lock_irqsave(&conf->device_lock, flags);
1471                         if (plug) {
1472                                 bio_list_add(&plug->pending, mbio);
1473                                 plug->pending_cnt++;
1474                         } else {
1475                                 bio_list_add(&conf->pending_bio_list, mbio);
1476                                 conf->pending_count++;
1477                         }
1478                         spin_unlock_irqrestore(&conf->device_lock, flags);
1479                         if (!plug)
1480                                 md_wakeup_thread(mddev->thread);
1481                 }
1482
1483                 if (r10_bio->devs[i].repl_bio) {
1484                         struct md_rdev *rdev = conf->mirrors[d].replacement;
1485                         if (rdev == NULL) {
1486                                 /* Replacement just got moved to main 'rdev' */
1487                                 smp_mb();
1488                                 rdev = conf->mirrors[d].rdev;
1489                         }
1490                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1491                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1492                                  max_sectors);
1493                         r10_bio->devs[i].repl_bio = mbio;
1494
1495                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1496                                            choose_data_offset(
1497                                                    r10_bio, rdev));
1498                         mbio->bi_bdev = rdev->bdev;
1499                         mbio->bi_end_io = raid10_end_write_request;
1500                         mbio->bi_rw =
1501                                 WRITE | do_sync | do_fua | do_discard | do_same;
1502                         mbio->bi_private = r10_bio;
1503
1504                         atomic_inc(&r10_bio->remaining);
1505                         spin_lock_irqsave(&conf->device_lock, flags);
1506                         bio_list_add(&conf->pending_bio_list, mbio);
1507                         conf->pending_count++;
1508                         spin_unlock_irqrestore(&conf->device_lock, flags);
1509                         if (!mddev_check_plugged(mddev))
1510                                 md_wakeup_thread(mddev->thread);
1511                 }
1512         }
1513
1514         /* Don't remove the bias on 'remaining' (one_write_done) until
1515          * after checking if we need to go around again.
1516          */
1517
1518         if (sectors_handled < bio_sectors(bio)) {
1519                 one_write_done(r10_bio);
1520                 /* We need another r10_bio.  It has already been counted
1521                  * in bio->bi_phys_segments.
1522                  */
1523                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1524
1525                 r10_bio->master_bio = bio;
1526                 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1527
1528                 r10_bio->mddev = mddev;
1529                 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1530                 r10_bio->state = 0;
1531                 goto retry_write;
1532         }
1533         one_write_done(r10_bio);
1534 }
1535
1536 static void make_request(struct mddev *mddev, struct bio *bio)
1537 {
1538         struct r10conf *conf = mddev->private;
1539         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1540         int chunk_sects = chunk_mask + 1;
1541
1542         struct bio *split;
1543
1544         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1545                 md_flush_request(mddev, bio);
1546                 return;
1547         }
1548
1549         md_write_start(mddev, bio);
1550
1551         do {
1552
1553                 /*
1554                  * If this request crosses a chunk boundary, we need to split
1555                  * it.
1556                  */
1557                 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1558                              bio_sectors(bio) > chunk_sects
1559                              && (conf->geo.near_copies < conf->geo.raid_disks
1560                                  || conf->prev.near_copies <
1561                                  conf->prev.raid_disks))) {
1562                         split = bio_split(bio, chunk_sects -
1563                                           (bio->bi_iter.bi_sector &
1564                                            (chunk_sects - 1)),
1565                                           GFP_NOIO, fs_bio_set);
1566                         bio_chain(split, bio);
1567                 } else {
1568                         split = bio;
1569                 }
1570
1571                 __make_request(mddev, split);
1572         } while (split != bio);
1573
1574         /* In case raid10d snuck in to freeze_array */
1575         wake_up(&conf->wait_barrier);
1576 }
1577
1578 static void status(struct seq_file *seq, struct mddev *mddev)
1579 {
1580         struct r10conf *conf = mddev->private;
1581         int i;
1582
1583         if (conf->geo.near_copies < conf->geo.raid_disks)
1584                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1585         if (conf->geo.near_copies > 1)
1586                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1587         if (conf->geo.far_copies > 1) {
1588                 if (conf->geo.far_offset)
1589                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1590                 else
1591                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1592         }
1593         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1594                                         conf->geo.raid_disks - mddev->degraded);
1595         for (i = 0; i < conf->geo.raid_disks; i++)
1596                 seq_printf(seq, "%s",
1597                               conf->mirrors[i].rdev &&
1598                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1599         seq_printf(seq, "]");
1600 }
1601
1602 /* check if there are enough drives for
1603  * every block to appear on atleast one.
1604  * Don't consider the device numbered 'ignore'
1605  * as we might be about to remove it.
1606  */
1607 static int _enough(struct r10conf *conf, int previous, int ignore)
1608 {
1609         int first = 0;
1610         int has_enough = 0;
1611         int disks, ncopies;
1612         if (previous) {
1613                 disks = conf->prev.raid_disks;
1614                 ncopies = conf->prev.near_copies;
1615         } else {
1616                 disks = conf->geo.raid_disks;
1617                 ncopies = conf->geo.near_copies;
1618         }
1619
1620         rcu_read_lock();
1621         do {
1622                 int n = conf->copies;
1623                 int cnt = 0;
1624                 int this = first;
1625                 while (n--) {
1626                         struct md_rdev *rdev;
1627                         if (this != ignore &&
1628                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1629                             test_bit(In_sync, &rdev->flags))
1630                                 cnt++;
1631                         this = (this+1) % disks;
1632                 }
1633                 if (cnt == 0)
1634                         goto out;
1635                 first = (first + ncopies) % disks;
1636         } while (first != 0);
1637         has_enough = 1;
1638 out:
1639         rcu_read_unlock();
1640         return has_enough;
1641 }
1642
1643 static int enough(struct r10conf *conf, int ignore)
1644 {
1645         /* when calling 'enough', both 'prev' and 'geo' must
1646          * be stable.
1647          * This is ensured if ->reconfig_mutex or ->device_lock
1648          * is held.
1649          */
1650         return _enough(conf, 0, ignore) &&
1651                 _enough(conf, 1, ignore);
1652 }
1653
1654 static void error(struct mddev *mddev, struct md_rdev *rdev)
1655 {
1656         char b[BDEVNAME_SIZE];
1657         struct r10conf *conf = mddev->private;
1658         unsigned long flags;
1659
1660         /*
1661          * If it is not operational, then we have already marked it as dead
1662          * else if it is the last working disks, ignore the error, let the
1663          * next level up know.
1664          * else mark the drive as failed
1665          */
1666         spin_lock_irqsave(&conf->device_lock, flags);
1667         if (test_bit(In_sync, &rdev->flags)
1668             && !enough(conf, rdev->raid_disk)) {
1669                 /*
1670                  * Don't fail the drive, just return an IO error.
1671                  */
1672                 spin_unlock_irqrestore(&conf->device_lock, flags);
1673                 return;
1674         }
1675         if (test_and_clear_bit(In_sync, &rdev->flags))
1676                 mddev->degraded++;
1677         /*
1678          * If recovery is running, make sure it aborts.
1679          */
1680         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1681         set_bit(Blocked, &rdev->flags);
1682         set_bit(Faulty, &rdev->flags);
1683         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1684         set_bit(MD_CHANGE_PENDING, &mddev->flags);
1685         spin_unlock_irqrestore(&conf->device_lock, flags);
1686         printk(KERN_ALERT
1687                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1688                "md/raid10:%s: Operation continuing on %d devices.\n",
1689                mdname(mddev), bdevname(rdev->bdev, b),
1690                mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1691 }
1692
1693 static void print_conf(struct r10conf *conf)
1694 {
1695         int i;
1696         struct raid10_info *tmp;
1697
1698         printk(KERN_DEBUG "RAID10 conf printout:\n");
1699         if (!conf) {
1700                 printk(KERN_DEBUG "(!conf)\n");
1701                 return;
1702         }
1703         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1704                 conf->geo.raid_disks);
1705
1706         for (i = 0; i < conf->geo.raid_disks; i++) {
1707                 char b[BDEVNAME_SIZE];
1708                 tmp = conf->mirrors + i;
1709                 if (tmp->rdev)
1710                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1711                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1712                                 !test_bit(Faulty, &tmp->rdev->flags),
1713                                 bdevname(tmp->rdev->bdev,b));
1714         }
1715 }
1716
1717 static void close_sync(struct r10conf *conf)
1718 {
1719         wait_barrier(conf);
1720         allow_barrier(conf);
1721
1722         mempool_destroy(conf->r10buf_pool);
1723         conf->r10buf_pool = NULL;
1724 }
1725
1726 static int raid10_spare_active(struct mddev *mddev)
1727 {
1728         int i;
1729         struct r10conf *conf = mddev->private;
1730         struct raid10_info *tmp;
1731         int count = 0;
1732         unsigned long flags;
1733
1734         /*
1735          * Find all non-in_sync disks within the RAID10 configuration
1736          * and mark them in_sync
1737          */
1738         for (i = 0; i < conf->geo.raid_disks; i++) {
1739                 tmp = conf->mirrors + i;
1740                 if (tmp->replacement
1741                     && tmp->replacement->recovery_offset == MaxSector
1742                     && !test_bit(Faulty, &tmp->replacement->flags)
1743                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1744                         /* Replacement has just become active */
1745                         if (!tmp->rdev
1746                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1747                                 count++;
1748                         if (tmp->rdev) {
1749                                 /* Replaced device not technically faulty,
1750                                  * but we need to be sure it gets removed
1751                                  * and never re-added.
1752                                  */
1753                                 set_bit(Faulty, &tmp->rdev->flags);
1754                                 sysfs_notify_dirent_safe(
1755                                         tmp->rdev->sysfs_state);
1756                         }
1757                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1758                 } else if (tmp->rdev
1759                            && tmp->rdev->recovery_offset == MaxSector
1760                            && !test_bit(Faulty, &tmp->rdev->flags)
1761                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1762                         count++;
1763                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1764                 }
1765         }
1766         spin_lock_irqsave(&conf->device_lock, flags);
1767         mddev->degraded -= count;
1768         spin_unlock_irqrestore(&conf->device_lock, flags);
1769
1770         print_conf(conf);
1771         return count;
1772 }
1773
1774 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1775 {
1776         struct r10conf *conf = mddev->private;
1777         int err = -EEXIST;
1778         int mirror;
1779         int first = 0;
1780         int last = conf->geo.raid_disks - 1;
1781         struct request_queue *q = bdev_get_queue(rdev->bdev);
1782
1783         if (mddev->recovery_cp < MaxSector)
1784                 /* only hot-add to in-sync arrays, as recovery is
1785                  * very different from resync
1786                  */
1787                 return -EBUSY;
1788         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1789                 return -EINVAL;
1790
1791         if (rdev->raid_disk >= 0)
1792                 first = last = rdev->raid_disk;
1793
1794         if (q->merge_bvec_fn) {
1795                 set_bit(Unmerged, &rdev->flags);
1796                 mddev->merge_check_needed = 1;
1797         }
1798
1799         if (rdev->saved_raid_disk >= first &&
1800             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1801                 mirror = rdev->saved_raid_disk;
1802         else
1803                 mirror = first;
1804         for ( ; mirror <= last ; mirror++) {
1805                 struct raid10_info *p = &conf->mirrors[mirror];
1806                 if (p->recovery_disabled == mddev->recovery_disabled)
1807                         continue;
1808                 if (p->rdev) {
1809                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1810                             p->replacement != NULL)
1811                                 continue;
1812                         clear_bit(In_sync, &rdev->flags);
1813                         set_bit(Replacement, &rdev->flags);
1814                         rdev->raid_disk = mirror;
1815                         err = 0;
1816                         if (mddev->gendisk)
1817                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1818                                                   rdev->data_offset << 9);
1819                         conf->fullsync = 1;
1820                         rcu_assign_pointer(p->replacement, rdev);
1821                         break;
1822                 }
1823
1824                 if (mddev->gendisk)
1825                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1826                                           rdev->data_offset << 9);
1827
1828                 p->head_position = 0;
1829                 p->recovery_disabled = mddev->recovery_disabled - 1;
1830                 rdev->raid_disk = mirror;
1831                 err = 0;
1832                 if (rdev->saved_raid_disk != mirror)
1833                         conf->fullsync = 1;
1834                 rcu_assign_pointer(p->rdev, rdev);
1835                 break;
1836         }
1837         if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1838                 /* Some requests might not have seen this new
1839                  * merge_bvec_fn.  We must wait for them to complete
1840                  * before merging the device fully.
1841                  * First we make sure any code which has tested
1842                  * our function has submitted the request, then
1843                  * we wait for all outstanding requests to complete.
1844                  */
1845                 synchronize_sched();
1846                 freeze_array(conf, 0);
1847                 unfreeze_array(conf);
1848                 clear_bit(Unmerged, &rdev->flags);
1849         }
1850         md_integrity_add_rdev(rdev, mddev);
1851         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1852                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1853
1854         print_conf(conf);
1855         return err;
1856 }
1857
1858 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1859 {
1860         struct r10conf *conf = mddev->private;
1861         int err = 0;
1862         int number = rdev->raid_disk;
1863         struct md_rdev **rdevp;
1864         struct raid10_info *p = conf->mirrors + number;
1865
1866         print_conf(conf);
1867         if (rdev == p->rdev)
1868                 rdevp = &p->rdev;
1869         else if (rdev == p->replacement)
1870                 rdevp = &p->replacement;
1871         else
1872                 return 0;
1873
1874         if (test_bit(In_sync, &rdev->flags) ||
1875             atomic_read(&rdev->nr_pending)) {
1876                 err = -EBUSY;
1877                 goto abort;
1878         }
1879         /* Only remove faulty devices if recovery
1880          * is not possible.
1881          */
1882         if (!test_bit(Faulty, &rdev->flags) &&
1883             mddev->recovery_disabled != p->recovery_disabled &&
1884             (!p->replacement || p->replacement == rdev) &&
1885             number < conf->geo.raid_disks &&
1886             enough(conf, -1)) {
1887                 err = -EBUSY;
1888                 goto abort;
1889         }
1890         *rdevp = NULL;
1891         synchronize_rcu();
1892         if (atomic_read(&rdev->nr_pending)) {
1893                 /* lost the race, try later */
1894                 err = -EBUSY;
1895                 *rdevp = rdev;
1896                 goto abort;
1897         } else if (p->replacement) {
1898                 /* We must have just cleared 'rdev' */
1899                 p->rdev = p->replacement;
1900                 clear_bit(Replacement, &p->replacement->flags);
1901                 smp_mb(); /* Make sure other CPUs may see both as identical
1902                            * but will never see neither -- if they are careful.
1903                            */
1904                 p->replacement = NULL;
1905                 clear_bit(WantReplacement, &rdev->flags);
1906         } else
1907                 /* We might have just remove the Replacement as faulty
1908                  * Clear the flag just in case
1909                  */
1910                 clear_bit(WantReplacement, &rdev->flags);
1911
1912         err = md_integrity_register(mddev);
1913
1914 abort:
1915
1916         print_conf(conf);
1917         return err;
1918 }
1919
1920 static void end_sync_read(struct bio *bio, int error)
1921 {
1922         struct r10bio *r10_bio = bio->bi_private;
1923         struct r10conf *conf = r10_bio->mddev->private;
1924         int d;
1925
1926         if (bio == r10_bio->master_bio) {
1927                 /* this is a reshape read */
1928                 d = r10_bio->read_slot; /* really the read dev */
1929         } else
1930                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1931
1932         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1933                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1934         else
1935                 /* The write handler will notice the lack of
1936                  * R10BIO_Uptodate and record any errors etc
1937                  */
1938                 atomic_add(r10_bio->sectors,
1939                            &conf->mirrors[d].rdev->corrected_errors);
1940
1941         /* for reconstruct, we always reschedule after a read.
1942          * for resync, only after all reads
1943          */
1944         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1945         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1946             atomic_dec_and_test(&r10_bio->remaining)) {
1947                 /* we have read all the blocks,
1948                  * do the comparison in process context in raid10d
1949                  */
1950                 reschedule_retry(r10_bio);
1951         }
1952 }
1953
1954 static void end_sync_request(struct r10bio *r10_bio)
1955 {
1956         struct mddev *mddev = r10_bio->mddev;
1957
1958         while (atomic_dec_and_test(&r10_bio->remaining)) {
1959                 if (r10_bio->master_bio == NULL) {
1960                         /* the primary of several recovery bios */
1961                         sector_t s = r10_bio->sectors;
1962                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1963                             test_bit(R10BIO_WriteError, &r10_bio->state))
1964                                 reschedule_retry(r10_bio);
1965                         else
1966                                 put_buf(r10_bio);
1967                         md_done_sync(mddev, s, 1);
1968                         break;
1969                 } else {
1970                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1971                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1972                             test_bit(R10BIO_WriteError, &r10_bio->state))
1973                                 reschedule_retry(r10_bio);
1974                         else
1975                                 put_buf(r10_bio);
1976                         r10_bio = r10_bio2;
1977                 }
1978         }
1979 }
1980
1981 static void end_sync_write(struct bio *bio, int error)
1982 {
1983         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1984         struct r10bio *r10_bio = bio->bi_private;
1985         struct mddev *mddev = r10_bio->mddev;
1986         struct r10conf *conf = mddev->private;
1987         int d;
1988         sector_t first_bad;
1989         int bad_sectors;
1990         int slot;
1991         int repl;
1992         struct md_rdev *rdev = NULL;
1993
1994         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1995         if (repl)
1996                 rdev = conf->mirrors[d].replacement;
1997         else
1998                 rdev = conf->mirrors[d].rdev;
1999
2000         if (!uptodate) {
2001                 if (repl)
2002                         md_error(mddev, rdev);
2003                 else {
2004                         set_bit(WriteErrorSeen, &rdev->flags);
2005                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2006                                 set_bit(MD_RECOVERY_NEEDED,
2007                                         &rdev->mddev->recovery);
2008                         set_bit(R10BIO_WriteError, &r10_bio->state);
2009                 }
2010         } else if (is_badblock(rdev,
2011                              r10_bio->devs[slot].addr,
2012                              r10_bio->sectors,
2013                              &first_bad, &bad_sectors))
2014                 set_bit(R10BIO_MadeGood, &r10_bio->state);
2015
2016         rdev_dec_pending(rdev, mddev);
2017
2018         end_sync_request(r10_bio);
2019 }
2020
2021 /*
2022  * Note: sync and recover and handled very differently for raid10
2023  * This code is for resync.
2024  * For resync, we read through virtual addresses and read all blocks.
2025  * If there is any error, we schedule a write.  The lowest numbered
2026  * drive is authoritative.
2027  * However requests come for physical address, so we need to map.
2028  * For every physical address there are raid_disks/copies virtual addresses,
2029  * which is always are least one, but is not necessarly an integer.
2030  * This means that a physical address can span multiple chunks, so we may
2031  * have to submit multiple io requests for a single sync request.
2032  */
2033 /*
2034  * We check if all blocks are in-sync and only write to blocks that
2035  * aren't in sync
2036  */
2037 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2038 {
2039         struct r10conf *conf = mddev->private;
2040         int i, first;
2041         struct bio *tbio, *fbio;
2042         int vcnt;
2043
2044         atomic_set(&r10_bio->remaining, 1);
2045
2046         /* find the first device with a block */
2047         for (i=0; i<conf->copies; i++)
2048                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2049                         break;
2050
2051         if (i == conf->copies)
2052                 goto done;
2053
2054         first = i;
2055         fbio = r10_bio->devs[i].bio;
2056
2057         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2058         /* now find blocks with errors */
2059         for (i=0 ; i < conf->copies ; i++) {
2060                 int  j, d;
2061
2062                 tbio = r10_bio->devs[i].bio;
2063
2064                 if (tbio->bi_end_io != end_sync_read)
2065                         continue;
2066                 if (i == first)
2067                         continue;
2068                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2069                         /* We know that the bi_io_vec layout is the same for
2070                          * both 'first' and 'i', so we just compare them.
2071                          * All vec entries are PAGE_SIZE;
2072                          */
2073                         int sectors = r10_bio->sectors;
2074                         for (j = 0; j < vcnt; j++) {
2075                                 int len = PAGE_SIZE;
2076                                 if (sectors < (len / 512))
2077                                         len = sectors * 512;
2078                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2079                                            page_address(tbio->bi_io_vec[j].bv_page),
2080                                            len))
2081                                         break;
2082                                 sectors -= len/512;
2083                         }
2084                         if (j == vcnt)
2085                                 continue;
2086                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2087                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2088                                 /* Don't fix anything. */
2089                                 continue;
2090                 }
2091                 /* Ok, we need to write this bio, either to correct an
2092                  * inconsistency or to correct an unreadable block.
2093                  * First we need to fixup bv_offset, bv_len and
2094                  * bi_vecs, as the read request might have corrupted these
2095                  */
2096                 bio_reset(tbio);
2097
2098                 tbio->bi_vcnt = vcnt;
2099                 tbio->bi_iter.bi_size = r10_bio->sectors << 9;
2100                 tbio->bi_rw = WRITE;
2101                 tbio->bi_private = r10_bio;
2102                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2103                 tbio->bi_end_io = end_sync_write;
2104
2105                 bio_copy_data(tbio, fbio);
2106
2107                 d = r10_bio->devs[i].devnum;
2108                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2109                 atomic_inc(&r10_bio->remaining);
2110                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2111
2112                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2113                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2114                 generic_make_request(tbio);
2115         }
2116
2117         /* Now write out to any replacement devices
2118          * that are active
2119          */
2120         for (i = 0; i < conf->copies; i++) {
2121                 int d;
2122
2123                 tbio = r10_bio->devs[i].repl_bio;
2124                 if (!tbio || !tbio->bi_end_io)
2125                         continue;
2126                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2127                     && r10_bio->devs[i].bio != fbio)
2128                         bio_copy_data(tbio, fbio);
2129                 d = r10_bio->devs[i].devnum;
2130                 atomic_inc(&r10_bio->remaining);
2131                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2132                              bio_sectors(tbio));
2133                 generic_make_request(tbio);
2134         }
2135
2136 done:
2137         if (atomic_dec_and_test(&r10_bio->remaining)) {
2138                 md_done_sync(mddev, r10_bio->sectors, 1);
2139                 put_buf(r10_bio);
2140         }
2141 }
2142
2143 /*
2144  * Now for the recovery code.
2145  * Recovery happens across physical sectors.
2146  * We recover all non-is_sync drives by finding the virtual address of
2147  * each, and then choose a working drive that also has that virt address.
2148  * There is a separate r10_bio for each non-in_sync drive.
2149  * Only the first two slots are in use. The first for reading,
2150  * The second for writing.
2151  *
2152  */
2153 static void fix_recovery_read_error(struct r10bio *r10_bio)
2154 {
2155         /* We got a read error during recovery.
2156          * We repeat the read in smaller page-sized sections.
2157          * If a read succeeds, write it to the new device or record
2158          * a bad block if we cannot.
2159          * If a read fails, record a bad block on both old and
2160          * new devices.
2161          */
2162         struct mddev *mddev = r10_bio->mddev;
2163         struct r10conf *conf = mddev->private;
2164         struct bio *bio = r10_bio->devs[0].bio;
2165         sector_t sect = 0;
2166         int sectors = r10_bio->sectors;
2167         int idx = 0;
2168         int dr = r10_bio->devs[0].devnum;
2169         int dw = r10_bio->devs[1].devnum;
2170
2171         while (sectors) {
2172                 int s = sectors;
2173                 struct md_rdev *rdev;
2174                 sector_t addr;
2175                 int ok;
2176
2177                 if (s > (PAGE_SIZE>>9))
2178                         s = PAGE_SIZE >> 9;
2179
2180                 rdev = conf->mirrors[dr].rdev;
2181                 addr = r10_bio->devs[0].addr + sect,
2182                 ok = sync_page_io(rdev,
2183                                   addr,
2184                                   s << 9,
2185                                   bio->bi_io_vec[idx].bv_page,
2186                                   READ, false);
2187                 if (ok) {
2188                         rdev = conf->mirrors[dw].rdev;
2189                         addr = r10_bio->devs[1].addr + sect;
2190                         ok = sync_page_io(rdev,
2191                                           addr,
2192                                           s << 9,
2193                                           bio->bi_io_vec[idx].bv_page,
2194                                           WRITE, false);
2195                         if (!ok) {
2196                                 set_bit(WriteErrorSeen, &rdev->flags);
2197                                 if (!test_and_set_bit(WantReplacement,
2198                                                       &rdev->flags))
2199                                         set_bit(MD_RECOVERY_NEEDED,
2200                                                 &rdev->mddev->recovery);
2201                         }
2202                 }
2203                 if (!ok) {
2204                         /* We don't worry if we cannot set a bad block -
2205                          * it really is bad so there is no loss in not
2206                          * recording it yet
2207                          */
2208                         rdev_set_badblocks(rdev, addr, s, 0);
2209
2210                         if (rdev != conf->mirrors[dw].rdev) {
2211                                 /* need bad block on destination too */
2212                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2213                                 addr = r10_bio->devs[1].addr + sect;
2214                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2215                                 if (!ok) {
2216                                         /* just abort the recovery */
2217                                         printk(KERN_NOTICE
2218                                                "md/raid10:%s: recovery aborted"
2219                                                " due to read error\n",
2220                                                mdname(mddev));
2221
2222                                         conf->mirrors[dw].recovery_disabled
2223                                                 = mddev->recovery_disabled;
2224                                         set_bit(MD_RECOVERY_INTR,
2225                                                 &mddev->recovery);
2226                                         break;
2227                                 }
2228                         }
2229                 }
2230
2231                 sectors -= s;
2232                 sect += s;
2233                 idx++;
2234         }
2235 }
2236
2237 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2238 {
2239         struct r10conf *conf = mddev->private;
2240         int d;
2241         struct bio *wbio, *wbio2;
2242
2243         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2244                 fix_recovery_read_error(r10_bio);
2245                 end_sync_request(r10_bio);
2246                 return;
2247         }
2248
2249         /*
2250          * share the pages with the first bio
2251          * and submit the write request
2252          */
2253         d = r10_bio->devs[1].devnum;
2254         wbio = r10_bio->devs[1].bio;
2255         wbio2 = r10_bio->devs[1].repl_bio;
2256         /* Need to test wbio2->bi_end_io before we call
2257          * generic_make_request as if the former is NULL,
2258          * the latter is free to free wbio2.
2259          */
2260         if (wbio2 && !wbio2->bi_end_io)
2261                 wbio2 = NULL;
2262         if (wbio->bi_end_io) {
2263                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2264                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2265                 generic_make_request(wbio);
2266         }
2267         if (wbio2) {
2268                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2269                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2270                              bio_sectors(wbio2));
2271                 generic_make_request(wbio2);
2272         }
2273 }
2274
2275 /*
2276  * Used by fix_read_error() to decay the per rdev read_errors.
2277  * We halve the read error count for every hour that has elapsed
2278  * since the last recorded read error.
2279  *
2280  */
2281 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2282 {
2283         struct timespec cur_time_mon;
2284         unsigned long hours_since_last;
2285         unsigned int read_errors = atomic_read(&rdev->read_errors);
2286
2287         ktime_get_ts(&cur_time_mon);
2288
2289         if (rdev->last_read_error.tv_sec == 0 &&
2290             rdev->last_read_error.tv_nsec == 0) {
2291                 /* first time we've seen a read error */
2292                 rdev->last_read_error = cur_time_mon;
2293                 return;
2294         }
2295
2296         hours_since_last = (cur_time_mon.tv_sec -
2297                             rdev->last_read_error.tv_sec) / 3600;
2298
2299         rdev->last_read_error = cur_time_mon;
2300
2301         /*
2302          * if hours_since_last is > the number of bits in read_errors
2303          * just set read errors to 0. We do this to avoid
2304          * overflowing the shift of read_errors by hours_since_last.
2305          */
2306         if (hours_since_last >= 8 * sizeof(read_errors))
2307                 atomic_set(&rdev->read_errors, 0);
2308         else
2309                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2310 }
2311
2312 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2313                             int sectors, struct page *page, int rw)
2314 {
2315         sector_t first_bad;
2316         int bad_sectors;
2317
2318         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2319             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2320                 return -1;
2321         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2322                 /* success */
2323                 return 1;
2324         if (rw == WRITE) {
2325                 set_bit(WriteErrorSeen, &rdev->flags);
2326                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2327                         set_bit(MD_RECOVERY_NEEDED,
2328                                 &rdev->mddev->recovery);
2329         }
2330         /* need to record an error - either for the block or the device */
2331         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2332                 md_error(rdev->mddev, rdev);
2333         return 0;
2334 }
2335
2336 /*
2337  * This is a kernel thread which:
2338  *
2339  *      1.      Retries failed read operations on working mirrors.
2340  *      2.      Updates the raid superblock when problems encounter.
2341  *      3.      Performs writes following reads for array synchronising.
2342  */
2343
2344 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2345 {
2346         int sect = 0; /* Offset from r10_bio->sector */
2347         int sectors = r10_bio->sectors;
2348         struct md_rdev*rdev;
2349         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2350         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2351
2352         /* still own a reference to this rdev, so it cannot
2353          * have been cleared recently.
2354          */
2355         rdev = conf->mirrors[d].rdev;
2356
2357         if (test_bit(Faulty, &rdev->flags))
2358                 /* drive has already been failed, just ignore any
2359                    more fix_read_error() attempts */
2360                 return;
2361
2362         check_decay_read_errors(mddev, rdev);
2363         atomic_inc(&rdev->read_errors);
2364         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2365                 char b[BDEVNAME_SIZE];
2366                 bdevname(rdev->bdev, b);
2367
2368                 printk(KERN_NOTICE
2369                        "md/raid10:%s: %s: Raid device exceeded "
2370                        "read_error threshold [cur %d:max %d]\n",
2371                        mdname(mddev), b,
2372                        atomic_read(&rdev->read_errors), max_read_errors);
2373                 printk(KERN_NOTICE
2374                        "md/raid10:%s: %s: Failing raid device\n",
2375                        mdname(mddev), b);
2376                 md_error(mddev, conf->mirrors[d].rdev);
2377                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2378                 return;
2379         }
2380
2381         while(sectors) {
2382                 int s = sectors;
2383                 int sl = r10_bio->read_slot;
2384                 int success = 0;
2385                 int start;
2386
2387                 if (s > (PAGE_SIZE>>9))
2388                         s = PAGE_SIZE >> 9;
2389
2390                 rcu_read_lock();
2391                 do {
2392                         sector_t first_bad;
2393                         int bad_sectors;
2394
2395                         d = r10_bio->devs[sl].devnum;
2396                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2397                         if (rdev &&
2398                             !test_bit(Unmerged, &rdev->flags) &&
2399                             test_bit(In_sync, &rdev->flags) &&
2400                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2401                                         &first_bad, &bad_sectors) == 0) {
2402                                 atomic_inc(&rdev->nr_pending);
2403                                 rcu_read_unlock();
2404                                 success = sync_page_io(rdev,
2405                                                        r10_bio->devs[sl].addr +
2406                                                        sect,
2407                                                        s<<9,
2408                                                        conf->tmppage, READ, false);
2409                                 rdev_dec_pending(rdev, mddev);
2410                                 rcu_read_lock();
2411                                 if (success)
2412                                         break;
2413                         }
2414                         sl++;
2415                         if (sl == conf->copies)
2416                                 sl = 0;
2417                 } while (!success && sl != r10_bio->read_slot);
2418                 rcu_read_unlock();
2419
2420                 if (!success) {
2421                         /* Cannot read from anywhere, just mark the block
2422                          * as bad on the first device to discourage future
2423                          * reads.
2424                          */
2425                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2426                         rdev = conf->mirrors[dn].rdev;
2427
2428                         if (!rdev_set_badblocks(
2429                                     rdev,
2430                                     r10_bio->devs[r10_bio->read_slot].addr
2431                                     + sect,
2432                                     s, 0)) {
2433                                 md_error(mddev, rdev);
2434                                 r10_bio->devs[r10_bio->read_slot].bio
2435                                         = IO_BLOCKED;
2436                         }
2437                         break;
2438                 }
2439
2440                 start = sl;
2441                 /* write it back and re-read */
2442                 rcu_read_lock();
2443                 while (sl != r10_bio->read_slot) {
2444                         char b[BDEVNAME_SIZE];
2445
2446                         if (sl==0)
2447                                 sl = conf->copies;
2448                         sl--;
2449                         d = r10_bio->devs[sl].devnum;
2450                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2451                         if (!rdev ||
2452                             test_bit(Unmerged, &rdev->flags) ||
2453                             !test_bit(In_sync, &rdev->flags))
2454                                 continue;
2455
2456                         atomic_inc(&rdev->nr_pending);
2457                         rcu_read_unlock();
2458                         if (r10_sync_page_io(rdev,
2459                                              r10_bio->devs[sl].addr +
2460                                              sect,
2461                                              s, conf->tmppage, WRITE)
2462                             == 0) {
2463                                 /* Well, this device is dead */
2464                                 printk(KERN_NOTICE
2465                                        "md/raid10:%s: read correction "
2466                                        "write failed"
2467                                        " (%d sectors at %llu on %s)\n",
2468                                        mdname(mddev), s,
2469                                        (unsigned long long)(
2470                                                sect +
2471                                                choose_data_offset(r10_bio,
2472                                                                   rdev)),
2473                                        bdevname(rdev->bdev, b));
2474                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2475                                        "drive\n",
2476                                        mdname(mddev),
2477                                        bdevname(rdev->bdev, b));
2478                         }
2479                         rdev_dec_pending(rdev, mddev);
2480                         rcu_read_lock();
2481                 }
2482                 sl = start;
2483                 while (sl != r10_bio->read_slot) {
2484                         char b[BDEVNAME_SIZE];
2485
2486                         if (sl==0)
2487                                 sl = conf->copies;
2488                         sl--;
2489                         d = r10_bio->devs[sl].devnum;
2490                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2491                         if (!rdev ||
2492                             !test_bit(In_sync, &rdev->flags))
2493                                 continue;
2494
2495                         atomic_inc(&rdev->nr_pending);
2496                         rcu_read_unlock();
2497                         switch (r10_sync_page_io(rdev,
2498                                              r10_bio->devs[sl].addr +
2499                                              sect,
2500                                              s, conf->tmppage,
2501                                                  READ)) {
2502                         case 0:
2503                                 /* Well, this device is dead */
2504                                 printk(KERN_NOTICE
2505                                        "md/raid10:%s: unable to read back "
2506                                        "corrected sectors"
2507                                        " (%d sectors at %llu on %s)\n",
2508                                        mdname(mddev), s,
2509                                        (unsigned long long)(
2510                                                sect +
2511                                                choose_data_offset(r10_bio, rdev)),
2512                                        bdevname(rdev->bdev, b));
2513                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2514                                        "drive\n",
2515                                        mdname(mddev),
2516                                        bdevname(rdev->bdev, b));
2517                                 break;
2518                         case 1:
2519                                 printk(KERN_INFO
2520                                        "md/raid10:%s: read error corrected"
2521                                        " (%d sectors at %llu on %s)\n",
2522                                        mdname(mddev), s,
2523                                        (unsigned long long)(
2524                                                sect +
2525                                                choose_data_offset(r10_bio, rdev)),
2526                                        bdevname(rdev->bdev, b));
2527                                 atomic_add(s, &rdev->corrected_errors);
2528                         }
2529
2530                         rdev_dec_pending(rdev, mddev);
2531                         rcu_read_lock();
2532                 }
2533                 rcu_read_unlock();
2534
2535                 sectors -= s;
2536                 sect += s;
2537         }
2538 }
2539
2540 static int narrow_write_error(struct r10bio *r10_bio, int i)
2541 {
2542         struct bio *bio = r10_bio->master_bio;
2543         struct mddev *mddev = r10_bio->mddev;
2544         struct r10conf *conf = mddev->private;
2545         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2546         /* bio has the data to be written to slot 'i' where
2547          * we just recently had a write error.
2548          * We repeatedly clone the bio and trim down to one block,
2549          * then try the write.  Where the write fails we record
2550          * a bad block.
2551          * It is conceivable that the bio doesn't exactly align with
2552          * blocks.  We must handle this.
2553          *
2554          * We currently own a reference to the rdev.
2555          */
2556
2557         int block_sectors;
2558         sector_t sector;
2559         int sectors;
2560         int sect_to_write = r10_bio->sectors;
2561         int ok = 1;
2562
2563         if (rdev->badblocks.shift < 0)
2564                 return 0;
2565
2566         block_sectors = roundup(1 << rdev->badblocks.shift,
2567                                 bdev_logical_block_size(rdev->bdev) >> 9);
2568         sector = r10_bio->sector;
2569         sectors = ((r10_bio->sector + block_sectors)
2570                    & ~(sector_t)(block_sectors - 1))
2571                 - sector;
2572
2573         while (sect_to_write) {
2574                 struct bio *wbio;
2575                 if (sectors > sect_to_write)
2576                         sectors = sect_to_write;
2577                 /* Write at 'sector' for 'sectors' */
2578                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2579                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2580                 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2581                                    choose_data_offset(r10_bio, rdev) +
2582                                    (sector - r10_bio->sector));
2583                 wbio->bi_bdev = rdev->bdev;
2584                 if (submit_bio_wait(WRITE, wbio) == 0)
2585                         /* Failure! */
2586                         ok = rdev_set_badblocks(rdev, sector,
2587                                                 sectors, 0)
2588                                 && ok;
2589
2590                 bio_put(wbio);
2591                 sect_to_write -= sectors;
2592                 sector += sectors;
2593                 sectors = block_sectors;
2594         }
2595         return ok;
2596 }
2597
2598 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2599 {
2600         int slot = r10_bio->read_slot;
2601         struct bio *bio;
2602         struct r10conf *conf = mddev->private;
2603         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2604         char b[BDEVNAME_SIZE];
2605         unsigned long do_sync;
2606         int max_sectors;
2607
2608         /* we got a read error. Maybe the drive is bad.  Maybe just
2609          * the block and we can fix it.
2610          * We freeze all other IO, and try reading the block from
2611          * other devices.  When we find one, we re-write
2612          * and check it that fixes the read error.
2613          * This is all done synchronously while the array is
2614          * frozen.
2615          */
2616         bio = r10_bio->devs[slot].bio;
2617         bdevname(bio->bi_bdev, b);
2618         bio_put(bio);
2619         r10_bio->devs[slot].bio = NULL;
2620
2621         if (mddev->ro == 0) {
2622                 freeze_array(conf, 1);
2623                 fix_read_error(conf, mddev, r10_bio);
2624                 unfreeze_array(conf);
2625         } else
2626                 r10_bio->devs[slot].bio = IO_BLOCKED;
2627
2628         rdev_dec_pending(rdev, mddev);
2629
2630 read_more:
2631         rdev = read_balance(conf, r10_bio, &max_sectors);
2632         if (rdev == NULL) {
2633                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2634                        " read error for block %llu\n",
2635                        mdname(mddev), b,
2636                        (unsigned long long)r10_bio->sector);
2637                 raid_end_bio_io(r10_bio);
2638                 return;
2639         }
2640
2641         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2642         slot = r10_bio->read_slot;
2643         printk_ratelimited(
2644                 KERN_ERR
2645                 "md/raid10:%s: %s: redirecting "
2646                 "sector %llu to another mirror\n",
2647                 mdname(mddev),
2648                 bdevname(rdev->bdev, b),
2649                 (unsigned long long)r10_bio->sector);
2650         bio = bio_clone_mddev(r10_bio->master_bio,
2651                               GFP_NOIO, mddev);
2652         bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2653         r10_bio->devs[slot].bio = bio;
2654         r10_bio->devs[slot].rdev = rdev;
2655         bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2656                 + choose_data_offset(r10_bio, rdev);
2657         bio->bi_bdev = rdev->bdev;
2658         bio->bi_rw = READ | do_sync;
2659         bio->bi_private = r10_bio;
2660         bio->bi_end_io = raid10_end_read_request;
2661         if (max_sectors < r10_bio->sectors) {
2662                 /* Drat - have to split this up more */
2663                 struct bio *mbio = r10_bio->master_bio;
2664                 int sectors_handled =
2665                         r10_bio->sector + max_sectors
2666                         - mbio->bi_iter.bi_sector;
2667                 r10_bio->sectors = max_sectors;
2668                 spin_lock_irq(&conf->device_lock);
2669                 if (mbio->bi_phys_segments == 0)
2670                         mbio->bi_phys_segments = 2;
2671                 else
2672                         mbio->bi_phys_segments++;
2673                 spin_unlock_irq(&conf->device_lock);
2674                 generic_make_request(bio);
2675
2676                 r10_bio = mempool_alloc(conf->r10bio_pool,
2677                                         GFP_NOIO);
2678                 r10_bio->master_bio = mbio;
2679                 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2680                 r10_bio->state = 0;
2681                 set_bit(R10BIO_ReadError,
2682                         &r10_bio->state);
2683                 r10_bio->mddev = mddev;
2684                 r10_bio->sector = mbio->bi_iter.bi_sector
2685                         + sectors_handled;
2686
2687                 goto read_more;
2688         } else
2689                 generic_make_request(bio);
2690 }
2691
2692 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2693 {
2694         /* Some sort of write request has finished and it
2695          * succeeded in writing where we thought there was a
2696          * bad block.  So forget the bad block.
2697          * Or possibly if failed and we need to record
2698          * a bad block.
2699          */
2700         int m;
2701         struct md_rdev *rdev;
2702
2703         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2704             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2705                 for (m = 0; m < conf->copies; m++) {
2706                         int dev = r10_bio->devs[m].devnum;
2707                         rdev = conf->mirrors[dev].rdev;
2708                         if (r10_bio->devs[m].bio == NULL)
2709                                 continue;
2710                         if (test_bit(BIO_UPTODATE,
2711                                      &r10_bio->devs[m].bio->bi_flags)) {
2712                                 rdev_clear_badblocks(
2713                                         rdev,
2714                                         r10_bio->devs[m].addr,
2715                                         r10_bio->sectors, 0);
2716                         } else {
2717                                 if (!rdev_set_badblocks(
2718                                             rdev,
2719                                             r10_bio->devs[m].addr,
2720                                             r10_bio->sectors, 0))
2721                                         md_error(conf->mddev, rdev);
2722                         }
2723                         rdev = conf->mirrors[dev].replacement;
2724                         if (r10_bio->devs[m].repl_bio == NULL)
2725                                 continue;
2726                         if (test_bit(BIO_UPTODATE,
2727                                      &r10_bio->devs[m].repl_bio->bi_flags)) {
2728                                 rdev_clear_badblocks(
2729                                         rdev,
2730                                         r10_bio->devs[m].addr,
2731                                         r10_bio->sectors, 0);
2732                         } else {
2733                                 if (!rdev_set_badblocks(
2734                                             rdev,
2735                                             r10_bio->devs[m].addr,
2736                                             r10_bio->sectors, 0))
2737                                         md_error(conf->mddev, rdev);
2738                         }
2739                 }
2740                 put_buf(r10_bio);
2741         } else {
2742                 bool fail = false;
2743                 for (m = 0; m < conf->copies; m++) {
2744                         int dev = r10_bio->devs[m].devnum;
2745                         struct bio *bio = r10_bio->devs[m].bio;
2746                         rdev = conf->mirrors[dev].rdev;
2747                         if (bio == IO_MADE_GOOD) {
2748                                 rdev_clear_badblocks(
2749                                         rdev,
2750                                         r10_bio->devs[m].addr,
2751                                         r10_bio->sectors, 0);
2752                                 rdev_dec_pending(rdev, conf->mddev);
2753                         } else if (bio != NULL &&
2754                                    !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2755                                 fail = true;
2756                                 if (!narrow_write_error(r10_bio, m)) {
2757                                         md_error(conf->mddev, rdev);
2758                                         set_bit(R10BIO_Degraded,
2759                                                 &r10_bio->state);
2760                                 }
2761                                 rdev_dec_pending(rdev, conf->mddev);
2762                         }
2763                         bio = r10_bio->devs[m].repl_bio;
2764                         rdev = conf->mirrors[dev].replacement;
2765                         if (rdev && bio == IO_MADE_GOOD) {
2766                                 rdev_clear_badblocks(
2767                                         rdev,
2768                                         r10_bio->devs[m].addr,
2769                                         r10_bio->sectors, 0);
2770                                 rdev_dec_pending(rdev, conf->mddev);
2771                         }
2772                 }
2773                 if (test_bit(R10BIO_WriteError,
2774                              &r10_bio->state))
2775                         close_write(r10_bio);
2776                 if (fail) {
2777                         spin_lock_irq(&conf->device_lock);
2778                         list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2779                         spin_unlock_irq(&conf->device_lock);
2780                         md_wakeup_thread(conf->mddev->thread);
2781                 } else
2782                         raid_end_bio_io(r10_bio);
2783         }
2784 }
2785
2786 static void raid10d(struct md_thread *thread)
2787 {
2788         struct mddev *mddev = thread->mddev;
2789         struct r10bio *r10_bio;
2790         unsigned long flags;
2791         struct r10conf *conf = mddev->private;
2792         struct list_head *head = &conf->retry_list;
2793         struct blk_plug plug;
2794
2795         md_check_recovery(mddev);
2796
2797         if (!list_empty_careful(&conf->bio_end_io_list) &&
2798             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2799                 LIST_HEAD(tmp);
2800                 spin_lock_irqsave(&conf->device_lock, flags);
2801                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2802                         list_add(&tmp, &conf->bio_end_io_list);
2803                         list_del_init(&conf->bio_end_io_list);
2804                 }
2805                 spin_unlock_irqrestore(&conf->device_lock, flags);
2806                 while (!list_empty(&tmp)) {
2807                         r10_bio = list_first_entry(&conf->bio_end_io_list,
2808                                                   struct r10bio, retry_list);
2809                         list_del(&r10_bio->retry_list);
2810                         raid_end_bio_io(r10_bio);
2811                 }
2812         }
2813
2814         blk_start_plug(&plug);
2815         for (;;) {
2816
2817                 flush_pending_writes(conf);
2818
2819                 spin_lock_irqsave(&conf->device_lock, flags);
2820                 if (list_empty(head)) {
2821                         spin_unlock_irqrestore(&conf->device_lock, flags);
2822                         break;
2823                 }
2824                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2825                 list_del(head->prev);
2826                 conf->nr_queued--;
2827                 spin_unlock_irqrestore(&conf->device_lock, flags);
2828
2829                 mddev = r10_bio->mddev;
2830                 conf = mddev->private;
2831                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2832                     test_bit(R10BIO_WriteError, &r10_bio->state))
2833                         handle_write_completed(conf, r10_bio);
2834                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2835                         reshape_request_write(mddev, r10_bio);
2836                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2837                         sync_request_write(mddev, r10_bio);
2838                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2839                         recovery_request_write(mddev, r10_bio);
2840                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2841                         handle_read_error(mddev, r10_bio);
2842                 else {
2843                         /* just a partial read to be scheduled from a
2844                          * separate context
2845                          */
2846                         int slot = r10_bio->read_slot;
2847                         generic_make_request(r10_bio->devs[slot].bio);
2848                 }
2849
2850                 cond_resched();
2851                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2852                         md_check_recovery(mddev);
2853         }
2854         blk_finish_plug(&plug);
2855 }
2856
2857 static int init_resync(struct r10conf *conf)
2858 {
2859         int buffs;
2860         int i;
2861
2862         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2863         BUG_ON(conf->r10buf_pool);
2864         conf->have_replacement = 0;
2865         for (i = 0; i < conf->geo.raid_disks; i++)
2866                 if (conf->mirrors[i].replacement)
2867                         conf->have_replacement = 1;
2868         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2869         if (!conf->r10buf_pool)
2870                 return -ENOMEM;
2871         conf->next_resync = 0;
2872         return 0;
2873 }
2874
2875 /*
2876  * perform a "sync" on one "block"
2877  *
2878  * We need to make sure that no normal I/O request - particularly write
2879  * requests - conflict with active sync requests.
2880  *
2881  * This is achieved by tracking pending requests and a 'barrier' concept
2882  * that can be installed to exclude normal IO requests.
2883  *
2884  * Resync and recovery are handled very differently.
2885  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2886  *
2887  * For resync, we iterate over virtual addresses, read all copies,
2888  * and update if there are differences.  If only one copy is live,
2889  * skip it.
2890  * For recovery, we iterate over physical addresses, read a good
2891  * value for each non-in_sync drive, and over-write.
2892  *
2893  * So, for recovery we may have several outstanding complex requests for a
2894  * given address, one for each out-of-sync device.  We model this by allocating
2895  * a number of r10_bio structures, one for each out-of-sync device.
2896  * As we setup these structures, we collect all bio's together into a list
2897  * which we then process collectively to add pages, and then process again
2898  * to pass to generic_make_request.
2899  *
2900  * The r10_bio structures are linked using a borrowed master_bio pointer.
2901  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2902  * has its remaining count decremented to 0, the whole complex operation
2903  * is complete.
2904  *
2905  */
2906
2907 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2908                              int *skipped)
2909 {
2910         struct r10conf *conf = mddev->private;
2911         struct r10bio *r10_bio;
2912         struct bio *biolist = NULL, *bio;
2913         sector_t max_sector, nr_sectors;
2914         int i;
2915         int max_sync;
2916         sector_t sync_blocks;
2917         sector_t sectors_skipped = 0;
2918         int chunks_skipped = 0;
2919         sector_t chunk_mask = conf->geo.chunk_mask;
2920
2921         if (!conf->r10buf_pool)
2922                 if (init_resync(conf))
2923                         return 0;
2924
2925         /*
2926          * Allow skipping a full rebuild for incremental assembly
2927          * of a clean array, like RAID1 does.
2928          */
2929         if (mddev->bitmap == NULL &&
2930             mddev->recovery_cp == MaxSector &&
2931             mddev->reshape_position == MaxSector &&
2932             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2933             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2934             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2935             conf->fullsync == 0) {
2936                 *skipped = 1;
2937                 return mddev->dev_sectors - sector_nr;
2938         }
2939
2940  skipped:
2941         max_sector = mddev->dev_sectors;
2942         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2943             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2944                 max_sector = mddev->resync_max_sectors;
2945         if (sector_nr >= max_sector) {
2946                 /* If we aborted, we need to abort the
2947                  * sync on the 'current' bitmap chucks (there can
2948                  * be several when recovering multiple devices).
2949                  * as we may have started syncing it but not finished.
2950                  * We can find the current address in
2951                  * mddev->curr_resync, but f