Merge tag 'renesas-fixes-for-v4.18' of https://git.kernel.org/pub/scm/linux/kernel...
[linux-2.6-microblaze.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 <trace/events/block.h>
29 #include "md.h"
30 #include "raid10.h"
31 #include "raid0.h"
32 #include "md-bitmap.h"
33
34 /*
35  * RAID10 provides a combination of RAID0 and RAID1 functionality.
36  * The layout of data is defined by
37  *    chunk_size
38  *    raid_disks
39  *    near_copies (stored in low byte of layout)
40  *    far_copies (stored in second byte of layout)
41  *    far_offset (stored in bit 16 of layout )
42  *    use_far_sets (stored in bit 17 of layout )
43  *    use_far_sets_bugfixed (stored in bit 18 of layout )
44  *
45  * The data to be stored is divided into chunks using chunksize.  Each device
46  * is divided into far_copies sections.   In each section, chunks are laid out
47  * in a style similar to raid0, but near_copies copies of each chunk is stored
48  * (each on a different drive).  The starting device for each section is offset
49  * near_copies from the starting device of the previous section.  Thus there
50  * are (near_copies * far_copies) of each chunk, and each is on a different
51  * drive.  near_copies and far_copies must be at least one, and their product
52  * is at most raid_disks.
53  *
54  * If far_offset is true, then the far_copies are handled a bit differently.
55  * The copies are still in different stripes, but instead of being very far
56  * apart on disk, there are adjacent stripes.
57  *
58  * The far and offset algorithms are handled slightly differently if
59  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
60  * sets that are (near_copies * far_copies) in size.  The far copied stripes
61  * are still shifted by 'near_copies' devices, but this shifting stays confined
62  * to the set rather than the entire array.  This is done to improve the number
63  * of device combinations that can fail without causing the array to fail.
64  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
65  * on a device):
66  *    A B C D    A B C D E
67  *      ...         ...
68  *    D A B C    E A B C D
69  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70  *    [A B] [C D]    [A B] [C D E]
71  *    |...| |...|    |...| | ... |
72  *    [B A] [D C]    [B A] [E C D]
73  */
74
75 /*
76  * Number of guaranteed r10bios in case of extreme VM load:
77  */
78 #define NR_RAID10_BIOS 256
79
80 /* when we get a read error on a read-only array, we redirect to another
81  * device without failing the first device, or trying to over-write to
82  * correct the read error.  To keep track of bad blocks on a per-bio
83  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84  */
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87  * bad-block marking which must be done from process context.  So we record
88  * the success by setting devs[n].bio to IO_MADE_GOOD
89  */
90 #define IO_MADE_GOOD ((struct bio *)2)
91
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
93
94 /* When there are this many requests queued to be written by
95  * the raid10 thread, we become 'congested' to provide back-pressure
96  * for writeback.
97  */
98 static int max_queued_requests = 1024;
99
100 static void allow_barrier(struct r10conf *conf);
101 static void lower_barrier(struct r10conf *conf);
102 static int _enough(struct r10conf *conf, int previous, int ignore);
103 static int enough(struct r10conf *conf, int ignore);
104 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
105                                 int *skipped);
106 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
107 static void end_reshape_write(struct bio *bio);
108 static void end_reshape(struct r10conf *conf);
109
110 #define raid10_log(md, fmt, args...)                            \
111         do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
112
113 #include "raid1-10.c"
114
115 /*
116  * for resync bio, r10bio pointer can be retrieved from the per-bio
117  * 'struct resync_pages'.
118  */
119 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
120 {
121         return get_resync_pages(bio)->raid_bio;
122 }
123
124 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
125 {
126         struct r10conf *conf = data;
127         int size = offsetof(struct r10bio, devs[conf->copies]);
128
129         /* allocate a r10bio with room for raid_disks entries in the
130          * bios array */
131         return kzalloc(size, gfp_flags);
132 }
133
134 static void r10bio_pool_free(void *r10_bio, void *data)
135 {
136         kfree(r10_bio);
137 }
138
139 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
140 /* amount of memory to reserve for resync requests */
141 #define RESYNC_WINDOW (1024*1024)
142 /* maximum number of concurrent requests, memory permitting */
143 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
145 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
146
147 /*
148  * When performing a resync, we need to read and compare, so
149  * we need as many pages are there are copies.
150  * When performing a recovery, we need 2 bios, one for read,
151  * one for write (we recover only one drive per r10buf)
152  *
153  */
154 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
155 {
156         struct r10conf *conf = data;
157         struct r10bio *r10_bio;
158         struct bio *bio;
159         int j;
160         int nalloc, nalloc_rp;
161         struct resync_pages *rps;
162
163         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
164         if (!r10_bio)
165                 return NULL;
166
167         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
168             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
169                 nalloc = conf->copies; /* resync */
170         else
171                 nalloc = 2; /* recovery */
172
173         /* allocate once for all bios */
174         if (!conf->have_replacement)
175                 nalloc_rp = nalloc;
176         else
177                 nalloc_rp = nalloc * 2;
178         rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
179         if (!rps)
180                 goto out_free_r10bio;
181
182         /*
183          * Allocate bios.
184          */
185         for (j = nalloc ; j-- ; ) {
186                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
187                 if (!bio)
188                         goto out_free_bio;
189                 r10_bio->devs[j].bio = bio;
190                 if (!conf->have_replacement)
191                         continue;
192                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
193                 if (!bio)
194                         goto out_free_bio;
195                 r10_bio->devs[j].repl_bio = bio;
196         }
197         /*
198          * Allocate RESYNC_PAGES data pages and attach them
199          * where needed.
200          */
201         for (j = 0; j < nalloc; j++) {
202                 struct bio *rbio = r10_bio->devs[j].repl_bio;
203                 struct resync_pages *rp, *rp_repl;
204
205                 rp = &rps[j];
206                 if (rbio)
207                         rp_repl = &rps[nalloc + j];
208
209                 bio = r10_bio->devs[j].bio;
210
211                 if (!j || test_bit(MD_RECOVERY_SYNC,
212                                    &conf->mddev->recovery)) {
213                         if (resync_alloc_pages(rp, gfp_flags))
214                                 goto out_free_pages;
215                 } else {
216                         memcpy(rp, &rps[0], sizeof(*rp));
217                         resync_get_all_pages(rp);
218                 }
219
220                 rp->raid_bio = r10_bio;
221                 bio->bi_private = rp;
222                 if (rbio) {
223                         memcpy(rp_repl, rp, sizeof(*rp));
224                         rbio->bi_private = rp_repl;
225                 }
226         }
227
228         return r10_bio;
229
230 out_free_pages:
231         while (--j >= 0)
232                 resync_free_pages(&rps[j * 2]);
233
234         j = 0;
235 out_free_bio:
236         for ( ; j < nalloc; j++) {
237                 if (r10_bio->devs[j].bio)
238                         bio_put(r10_bio->devs[j].bio);
239                 if (r10_bio->devs[j].repl_bio)
240                         bio_put(r10_bio->devs[j].repl_bio);
241         }
242         kfree(rps);
243 out_free_r10bio:
244         r10bio_pool_free(r10_bio, conf);
245         return NULL;
246 }
247
248 static void r10buf_pool_free(void *__r10_bio, void *data)
249 {
250         struct r10conf *conf = data;
251         struct r10bio *r10bio = __r10_bio;
252         int j;
253         struct resync_pages *rp = NULL;
254
255         for (j = conf->copies; j--; ) {
256                 struct bio *bio = r10bio->devs[j].bio;
257
258                 if (bio) {
259                         rp = get_resync_pages(bio);
260                         resync_free_pages(rp);
261                         bio_put(bio);
262                 }
263
264                 bio = r10bio->devs[j].repl_bio;
265                 if (bio)
266                         bio_put(bio);
267         }
268
269         /* resync pages array stored in the 1st bio's .bi_private */
270         kfree(rp);
271
272         r10bio_pool_free(r10bio, conf);
273 }
274
275 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
276 {
277         int i;
278
279         for (i = 0; i < conf->copies; i++) {
280                 struct bio **bio = & r10_bio->devs[i].bio;
281                 if (!BIO_SPECIAL(*bio))
282                         bio_put(*bio);
283                 *bio = NULL;
284                 bio = &r10_bio->devs[i].repl_bio;
285                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
286                         bio_put(*bio);
287                 *bio = NULL;
288         }
289 }
290
291 static void free_r10bio(struct r10bio *r10_bio)
292 {
293         struct r10conf *conf = r10_bio->mddev->private;
294
295         put_all_bios(conf, r10_bio);
296         mempool_free(r10_bio, &conf->r10bio_pool);
297 }
298
299 static void put_buf(struct r10bio *r10_bio)
300 {
301         struct r10conf *conf = r10_bio->mddev->private;
302
303         mempool_free(r10_bio, &conf->r10buf_pool);
304
305         lower_barrier(conf);
306 }
307
308 static void reschedule_retry(struct r10bio *r10_bio)
309 {
310         unsigned long flags;
311         struct mddev *mddev = r10_bio->mddev;
312         struct r10conf *conf = mddev->private;
313
314         spin_lock_irqsave(&conf->device_lock, flags);
315         list_add(&r10_bio->retry_list, &conf->retry_list);
316         conf->nr_queued ++;
317         spin_unlock_irqrestore(&conf->device_lock, flags);
318
319         /* wake up frozen array... */
320         wake_up(&conf->wait_barrier);
321
322         md_wakeup_thread(mddev->thread);
323 }
324
325 /*
326  * raid_end_bio_io() is called when we have finished servicing a mirrored
327  * operation and are ready to return a success/failure code to the buffer
328  * cache layer.
329  */
330 static void raid_end_bio_io(struct r10bio *r10_bio)
331 {
332         struct bio *bio = r10_bio->master_bio;
333         struct r10conf *conf = r10_bio->mddev->private;
334
335         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
336                 bio->bi_status = BLK_STS_IOERR;
337
338         bio_endio(bio);
339         /*
340          * Wake up any possible resync thread that waits for the device
341          * to go idle.
342          */
343         allow_barrier(conf);
344
345         free_r10bio(r10_bio);
346 }
347
348 /*
349  * Update disk head position estimator based on IRQ completion info.
350  */
351 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
352 {
353         struct r10conf *conf = r10_bio->mddev->private;
354
355         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
356                 r10_bio->devs[slot].addr + (r10_bio->sectors);
357 }
358
359 /*
360  * Find the disk number which triggered given bio
361  */
362 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
363                          struct bio *bio, int *slotp, int *replp)
364 {
365         int slot;
366         int repl = 0;
367
368         for (slot = 0; slot < conf->copies; slot++) {
369                 if (r10_bio->devs[slot].bio == bio)
370                         break;
371                 if (r10_bio->devs[slot].repl_bio == bio) {
372                         repl = 1;
373                         break;
374                 }
375         }
376
377         BUG_ON(slot == conf->copies);
378         update_head_pos(slot, r10_bio);
379
380         if (slotp)
381                 *slotp = slot;
382         if (replp)
383                 *replp = repl;
384         return r10_bio->devs[slot].devnum;
385 }
386
387 static void raid10_end_read_request(struct bio *bio)
388 {
389         int uptodate = !bio->bi_status;
390         struct r10bio *r10_bio = bio->bi_private;
391         int slot;
392         struct md_rdev *rdev;
393         struct r10conf *conf = r10_bio->mddev->private;
394
395         slot = r10_bio->read_slot;
396         rdev = r10_bio->devs[slot].rdev;
397         /*
398          * this branch is our 'one mirror IO has finished' event handler:
399          */
400         update_head_pos(slot, r10_bio);
401
402         if (uptodate) {
403                 /*
404                  * Set R10BIO_Uptodate in our master bio, so that
405                  * we will return a good error code to the higher
406                  * levels even if IO on some other mirrored buffer fails.
407                  *
408                  * The 'master' represents the composite IO operation to
409                  * user-side. So if something waits for IO, then it will
410                  * wait for the 'master' bio.
411                  */
412                 set_bit(R10BIO_Uptodate, &r10_bio->state);
413         } else {
414                 /* If all other devices that store this block have
415                  * failed, we want to return the error upwards rather
416                  * than fail the last device.  Here we redefine
417                  * "uptodate" to mean "Don't want to retry"
418                  */
419                 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
420                              rdev->raid_disk))
421                         uptodate = 1;
422         }
423         if (uptodate) {
424                 raid_end_bio_io(r10_bio);
425                 rdev_dec_pending(rdev, conf->mddev);
426         } else {
427                 /*
428                  * oops, read error - keep the refcount on the rdev
429                  */
430                 char b[BDEVNAME_SIZE];
431                 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
432                                    mdname(conf->mddev),
433                                    bdevname(rdev->bdev, b),
434                                    (unsigned long long)r10_bio->sector);
435                 set_bit(R10BIO_ReadError, &r10_bio->state);
436                 reschedule_retry(r10_bio);
437         }
438 }
439
440 static void close_write(struct r10bio *r10_bio)
441 {
442         /* clear the bitmap if all writes complete successfully */
443         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
444                         r10_bio->sectors,
445                         !test_bit(R10BIO_Degraded, &r10_bio->state),
446                         0);
447         md_write_end(r10_bio->mddev);
448 }
449
450 static void one_write_done(struct r10bio *r10_bio)
451 {
452         if (atomic_dec_and_test(&r10_bio->remaining)) {
453                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
454                         reschedule_retry(r10_bio);
455                 else {
456                         close_write(r10_bio);
457                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
458                                 reschedule_retry(r10_bio);
459                         else
460                                 raid_end_bio_io(r10_bio);
461                 }
462         }
463 }
464
465 static void raid10_end_write_request(struct bio *bio)
466 {
467         struct r10bio *r10_bio = bio->bi_private;
468         int dev;
469         int dec_rdev = 1;
470         struct r10conf *conf = r10_bio->mddev->private;
471         int slot, repl;
472         struct md_rdev *rdev = NULL;
473         struct bio *to_put = NULL;
474         bool discard_error;
475
476         discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
477
478         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
479
480         if (repl)
481                 rdev = conf->mirrors[dev].replacement;
482         if (!rdev) {
483                 smp_rmb();
484                 repl = 0;
485                 rdev = conf->mirrors[dev].rdev;
486         }
487         /*
488          * this branch is our 'one mirror IO has finished' event handler:
489          */
490         if (bio->bi_status && !discard_error) {
491                 if (repl)
492                         /* Never record new bad blocks to replacement,
493                          * just fail it.
494                          */
495                         md_error(rdev->mddev, rdev);
496                 else {
497                         set_bit(WriteErrorSeen, &rdev->flags);
498                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
499                                 set_bit(MD_RECOVERY_NEEDED,
500                                         &rdev->mddev->recovery);
501
502                         dec_rdev = 0;
503                         if (test_bit(FailFast, &rdev->flags) &&
504                             (bio->bi_opf & MD_FAILFAST)) {
505                                 md_error(rdev->mddev, rdev);
506                                 if (!test_bit(Faulty, &rdev->flags))
507                                         /* This is the only remaining device,
508                                          * We need to retry the write without
509                                          * FailFast
510                                          */
511                                         set_bit(R10BIO_WriteError, &r10_bio->state);
512                                 else {
513                                         r10_bio->devs[slot].bio = NULL;
514                                         to_put = bio;
515                                         dec_rdev = 1;
516                                 }
517                         } else
518                                 set_bit(R10BIO_WriteError, &r10_bio->state);
519                 }
520         } else {
521                 /*
522                  * Set R10BIO_Uptodate in our master bio, so that
523                  * we will return a good error code for to the higher
524                  * levels even if IO on some other mirrored buffer fails.
525                  *
526                  * The 'master' represents the composite IO operation to
527                  * user-side. So if something waits for IO, then it will
528                  * wait for the 'master' bio.
529                  */
530                 sector_t first_bad;
531                 int bad_sectors;
532
533                 /*
534                  * Do not set R10BIO_Uptodate if the current device is
535                  * rebuilding or Faulty. This is because we cannot use
536                  * such device for properly reading the data back (we could
537                  * potentially use it, if the current write would have felt
538                  * before rdev->recovery_offset, but for simplicity we don't
539                  * check this here.
540                  */
541                 if (test_bit(In_sync, &rdev->flags) &&
542                     !test_bit(Faulty, &rdev->flags))
543                         set_bit(R10BIO_Uptodate, &r10_bio->state);
544
545                 /* Maybe we can clear some bad blocks. */
546                 if (is_badblock(rdev,
547                                 r10_bio->devs[slot].addr,
548                                 r10_bio->sectors,
549                                 &first_bad, &bad_sectors) && !discard_error) {
550                         bio_put(bio);
551                         if (repl)
552                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
553                         else
554                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
555                         dec_rdev = 0;
556                         set_bit(R10BIO_MadeGood, &r10_bio->state);
557                 }
558         }
559
560         /*
561          *
562          * Let's see if all mirrored write operations have finished
563          * already.
564          */
565         one_write_done(r10_bio);
566         if (dec_rdev)
567                 rdev_dec_pending(rdev, conf->mddev);
568         if (to_put)
569                 bio_put(to_put);
570 }
571
572 /*
573  * RAID10 layout manager
574  * As well as the chunksize and raid_disks count, there are two
575  * parameters: near_copies and far_copies.
576  * near_copies * far_copies must be <= raid_disks.
577  * Normally one of these will be 1.
578  * If both are 1, we get raid0.
579  * If near_copies == raid_disks, we get raid1.
580  *
581  * Chunks are laid out in raid0 style with near_copies copies of the
582  * first chunk, followed by near_copies copies of the next chunk and
583  * so on.
584  * If far_copies > 1, then after 1/far_copies of the array has been assigned
585  * as described above, we start again with a device offset of near_copies.
586  * So we effectively have another copy of the whole array further down all
587  * the drives, but with blocks on different drives.
588  * With this layout, and block is never stored twice on the one device.
589  *
590  * raid10_find_phys finds the sector offset of a given virtual sector
591  * on each device that it is on.
592  *
593  * raid10_find_virt does the reverse mapping, from a device and a
594  * sector offset to a virtual address
595  */
596
597 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
598 {
599         int n,f;
600         sector_t sector;
601         sector_t chunk;
602         sector_t stripe;
603         int dev;
604         int slot = 0;
605         int last_far_set_start, last_far_set_size;
606
607         last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
608         last_far_set_start *= geo->far_set_size;
609
610         last_far_set_size = geo->far_set_size;
611         last_far_set_size += (geo->raid_disks % geo->far_set_size);
612
613         /* now calculate first sector/dev */
614         chunk = r10bio->sector >> geo->chunk_shift;
615         sector = r10bio->sector & geo->chunk_mask;
616
617         chunk *= geo->near_copies;
618         stripe = chunk;
619         dev = sector_div(stripe, geo->raid_disks);
620         if (geo->far_offset)
621                 stripe *= geo->far_copies;
622
623         sector += stripe << geo->chunk_shift;
624
625         /* and calculate all the others */
626         for (n = 0; n < geo->near_copies; n++) {
627                 int d = dev;
628                 int set;
629                 sector_t s = sector;
630                 r10bio->devs[slot].devnum = d;
631                 r10bio->devs[slot].addr = s;
632                 slot++;
633
634                 for (f = 1; f < geo->far_copies; f++) {
635                         set = d / geo->far_set_size;
636                         d += geo->near_copies;
637
638                         if ((geo->raid_disks % geo->far_set_size) &&
639                             (d > last_far_set_start)) {
640                                 d -= last_far_set_start;
641                                 d %= last_far_set_size;
642                                 d += last_far_set_start;
643                         } else {
644                                 d %= geo->far_set_size;
645                                 d += geo->far_set_size * set;
646                         }
647                         s += geo->stride;
648                         r10bio->devs[slot].devnum = d;
649                         r10bio->devs[slot].addr = s;
650                         slot++;
651                 }
652                 dev++;
653                 if (dev >= geo->raid_disks) {
654                         dev = 0;
655                         sector += (geo->chunk_mask + 1);
656                 }
657         }
658 }
659
660 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
661 {
662         struct geom *geo = &conf->geo;
663
664         if (conf->reshape_progress != MaxSector &&
665             ((r10bio->sector >= conf->reshape_progress) !=
666              conf->mddev->reshape_backwards)) {
667                 set_bit(R10BIO_Previous, &r10bio->state);
668                 geo = &conf->prev;
669         } else
670                 clear_bit(R10BIO_Previous, &r10bio->state);
671
672         __raid10_find_phys(geo, r10bio);
673 }
674
675 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
676 {
677         sector_t offset, chunk, vchunk;
678         /* Never use conf->prev as this is only called during resync
679          * or recovery, so reshape isn't happening
680          */
681         struct geom *geo = &conf->geo;
682         int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
683         int far_set_size = geo->far_set_size;
684         int last_far_set_start;
685
686         if (geo->raid_disks % geo->far_set_size) {
687                 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
688                 last_far_set_start *= geo->far_set_size;
689
690                 if (dev >= last_far_set_start) {
691                         far_set_size = geo->far_set_size;
692                         far_set_size += (geo->raid_disks % geo->far_set_size);
693                         far_set_start = last_far_set_start;
694                 }
695         }
696
697         offset = sector & geo->chunk_mask;
698         if (geo->far_offset) {
699                 int fc;
700                 chunk = sector >> geo->chunk_shift;
701                 fc = sector_div(chunk, geo->far_copies);
702                 dev -= fc * geo->near_copies;
703                 if (dev < far_set_start)
704                         dev += far_set_size;
705         } else {
706                 while (sector >= geo->stride) {
707                         sector -= geo->stride;
708                         if (dev < (geo->near_copies + far_set_start))
709                                 dev += far_set_size - geo->near_copies;
710                         else
711                                 dev -= geo->near_copies;
712                 }
713                 chunk = sector >> geo->chunk_shift;
714         }
715         vchunk = chunk * geo->raid_disks + dev;
716         sector_div(vchunk, geo->near_copies);
717         return (vchunk << geo->chunk_shift) + offset;
718 }
719
720 /*
721  * This routine returns the disk from which the requested read should
722  * be done. There is a per-array 'next expected sequential IO' sector
723  * number - if this matches on the next IO then we use the last disk.
724  * There is also a per-disk 'last know head position' sector that is
725  * maintained from IRQ contexts, both the normal and the resync IO
726  * completion handlers update this position correctly. If there is no
727  * perfect sequential match then we pick the disk whose head is closest.
728  *
729  * If there are 2 mirrors in the same 2 devices, performance degrades
730  * because position is mirror, not device based.
731  *
732  * The rdev for the device selected will have nr_pending incremented.
733  */
734
735 /*
736  * FIXME: possibly should rethink readbalancing and do it differently
737  * depending on near_copies / far_copies geometry.
738  */
739 static struct md_rdev *read_balance(struct r10conf *conf,
740                                     struct r10bio *r10_bio,
741                                     int *max_sectors)
742 {
743         const sector_t this_sector = r10_bio->sector;
744         int disk, slot;
745         int sectors = r10_bio->sectors;
746         int best_good_sectors;
747         sector_t new_distance, best_dist;
748         struct md_rdev *best_rdev, *rdev = NULL;
749         int do_balance;
750         int best_slot;
751         struct geom *geo = &conf->geo;
752
753         raid10_find_phys(conf, r10_bio);
754         rcu_read_lock();
755         best_slot = -1;
756         best_rdev = NULL;
757         best_dist = MaxSector;
758         best_good_sectors = 0;
759         do_balance = 1;
760         clear_bit(R10BIO_FailFast, &r10_bio->state);
761         /*
762          * Check if we can balance. We can balance on the whole
763          * device if no resync is going on (recovery is ok), or below
764          * the resync window. We take the first readable disk when
765          * above the resync window.
766          */
767         if ((conf->mddev->recovery_cp < MaxSector
768              && (this_sector + sectors >= conf->next_resync)) ||
769             (mddev_is_clustered(conf->mddev) &&
770              md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
771                                             this_sector + sectors)))
772                 do_balance = 0;
773
774         for (slot = 0; slot < conf->copies ; slot++) {
775                 sector_t first_bad;
776                 int bad_sectors;
777                 sector_t dev_sector;
778
779                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
780                         continue;
781                 disk = r10_bio->devs[slot].devnum;
782                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
783                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
784                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
785                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
786                 if (rdev == NULL ||
787                     test_bit(Faulty, &rdev->flags))
788                         continue;
789                 if (!test_bit(In_sync, &rdev->flags) &&
790                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
791                         continue;
792
793                 dev_sector = r10_bio->devs[slot].addr;
794                 if (is_badblock(rdev, dev_sector, sectors,
795                                 &first_bad, &bad_sectors)) {
796                         if (best_dist < MaxSector)
797                                 /* Already have a better slot */
798                                 continue;
799                         if (first_bad <= dev_sector) {
800                                 /* Cannot read here.  If this is the
801                                  * 'primary' device, then we must not read
802                                  * beyond 'bad_sectors' from another device.
803                                  */
804                                 bad_sectors -= (dev_sector - first_bad);
805                                 if (!do_balance && sectors > bad_sectors)
806                                         sectors = bad_sectors;
807                                 if (best_good_sectors > sectors)
808                                         best_good_sectors = sectors;
809                         } else {
810                                 sector_t good_sectors =
811                                         first_bad - dev_sector;
812                                 if (good_sectors > best_good_sectors) {
813                                         best_good_sectors = good_sectors;
814                                         best_slot = slot;
815                                         best_rdev = rdev;
816                                 }
817                                 if (!do_balance)
818                                         /* Must read from here */
819                                         break;
820                         }
821                         continue;
822                 } else
823                         best_good_sectors = sectors;
824
825                 if (!do_balance)
826                         break;
827
828                 if (best_slot >= 0)
829                         /* At least 2 disks to choose from so failfast is OK */
830                         set_bit(R10BIO_FailFast, &r10_bio->state);
831                 /* This optimisation is debatable, and completely destroys
832                  * sequential read speed for 'far copies' arrays.  So only
833                  * keep it for 'near' arrays, and review those later.
834                  */
835                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
836                         new_distance = 0;
837
838                 /* for far > 1 always use the lowest address */
839                 else if (geo->far_copies > 1)
840                         new_distance = r10_bio->devs[slot].addr;
841                 else
842                         new_distance = abs(r10_bio->devs[slot].addr -
843                                            conf->mirrors[disk].head_position);
844                 if (new_distance < best_dist) {
845                         best_dist = new_distance;
846                         best_slot = slot;
847                         best_rdev = rdev;
848                 }
849         }
850         if (slot >= conf->copies) {
851                 slot = best_slot;
852                 rdev = best_rdev;
853         }
854
855         if (slot >= 0) {
856                 atomic_inc(&rdev->nr_pending);
857                 r10_bio->read_slot = slot;
858         } else
859                 rdev = NULL;
860         rcu_read_unlock();
861         *max_sectors = best_good_sectors;
862
863         return rdev;
864 }
865
866 static int raid10_congested(struct mddev *mddev, int bits)
867 {
868         struct r10conf *conf = mddev->private;
869         int i, ret = 0;
870
871         if ((bits & (1 << WB_async_congested)) &&
872             conf->pending_count >= max_queued_requests)
873                 return 1;
874
875         rcu_read_lock();
876         for (i = 0;
877              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
878                      && ret == 0;
879              i++) {
880                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
881                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
882                         struct request_queue *q = bdev_get_queue(rdev->bdev);
883
884                         ret |= bdi_congested(q->backing_dev_info, bits);
885                 }
886         }
887         rcu_read_unlock();
888         return ret;
889 }
890
891 static void flush_pending_writes(struct r10conf *conf)
892 {
893         /* Any writes that have been queued but are awaiting
894          * bitmap updates get flushed here.
895          */
896         spin_lock_irq(&conf->device_lock);
897
898         if (conf->pending_bio_list.head) {
899                 struct blk_plug plug;
900                 struct bio *bio;
901
902                 bio = bio_list_get(&conf->pending_bio_list);
903                 conf->pending_count = 0;
904                 spin_unlock_irq(&conf->device_lock);
905
906                 /*
907                  * As this is called in a wait_event() loop (see freeze_array),
908                  * current->state might be TASK_UNINTERRUPTIBLE which will
909                  * cause a warning when we prepare to wait again.  As it is
910                  * rare that this path is taken, it is perfectly safe to force
911                  * us to go around the wait_event() loop again, so the warning
912                  * is a false-positive. Silence the warning by resetting
913                  * thread state
914                  */
915                 __set_current_state(TASK_RUNNING);
916
917                 blk_start_plug(&plug);
918                 /* flush any pending bitmap writes to disk
919                  * before proceeding w/ I/O */
920                 bitmap_unplug(conf->mddev->bitmap);
921                 wake_up(&conf->wait_barrier);
922
923                 while (bio) { /* submit pending writes */
924                         struct bio *next = bio->bi_next;
925                         struct md_rdev *rdev = (void*)bio->bi_disk;
926                         bio->bi_next = NULL;
927                         bio_set_dev(bio, rdev->bdev);
928                         if (test_bit(Faulty, &rdev->flags)) {
929                                 bio_io_error(bio);
930                         } else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
931                                             !blk_queue_discard(bio->bi_disk->queue)))
932                                 /* Just ignore it */
933                                 bio_endio(bio);
934                         else
935                                 generic_make_request(bio);
936                         bio = next;
937                 }
938                 blk_finish_plug(&plug);
939         } else
940                 spin_unlock_irq(&conf->device_lock);
941 }
942
943 /* Barriers....
944  * Sometimes we need to suspend IO while we do something else,
945  * either some resync/recovery, or reconfigure the array.
946  * To do this we raise a 'barrier'.
947  * The 'barrier' is a counter that can be raised multiple times
948  * to count how many activities are happening which preclude
949  * normal IO.
950  * We can only raise the barrier if there is no pending IO.
951  * i.e. if nr_pending == 0.
952  * We choose only to raise the barrier if no-one is waiting for the
953  * barrier to go down.  This means that as soon as an IO request
954  * is ready, no other operations which require a barrier will start
955  * until the IO request has had a chance.
956  *
957  * So: regular IO calls 'wait_barrier'.  When that returns there
958  *    is no backgroup IO happening,  It must arrange to call
959  *    allow_barrier when it has finished its IO.
960  * backgroup IO calls must call raise_barrier.  Once that returns
961  *    there is no normal IO happeing.  It must arrange to call
962  *    lower_barrier when the particular background IO completes.
963  */
964
965 static void raise_barrier(struct r10conf *conf, int force)
966 {
967         BUG_ON(force && !conf->barrier);
968         spin_lock_irq(&conf->resync_lock);
969
970         /* Wait until no block IO is waiting (unless 'force') */
971         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
972                             conf->resync_lock);
973
974         /* block any new IO from starting */
975         conf->barrier++;
976
977         /* Now wait for all pending IO to complete */
978         wait_event_lock_irq(conf->wait_barrier,
979                             !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
980                             conf->resync_lock);
981
982         spin_unlock_irq(&conf->resync_lock);
983 }
984
985 static void lower_barrier(struct r10conf *conf)
986 {
987         unsigned long flags;
988         spin_lock_irqsave(&conf->resync_lock, flags);
989         conf->barrier--;
990         spin_unlock_irqrestore(&conf->resync_lock, flags);
991         wake_up(&conf->wait_barrier);
992 }
993
994 static void wait_barrier(struct r10conf *conf)
995 {
996         spin_lock_irq(&conf->resync_lock);
997         if (conf->barrier) {
998                 conf->nr_waiting++;
999                 /* Wait for the barrier to drop.
1000                  * However if there are already pending
1001                  * requests (preventing the barrier from
1002                  * rising completely), and the
1003                  * pre-process bio queue isn't empty,
1004                  * then don't wait, as we need to empty
1005                  * that queue to get the nr_pending
1006                  * count down.
1007                  */
1008                 raid10_log(conf->mddev, "wait barrier");
1009                 wait_event_lock_irq(conf->wait_barrier,
1010                                     !conf->barrier ||
1011                                     (atomic_read(&conf->nr_pending) &&
1012                                      current->bio_list &&
1013                                      (!bio_list_empty(&current->bio_list[0]) ||
1014                                       !bio_list_empty(&current->bio_list[1]))),
1015                                     conf->resync_lock);
1016                 conf->nr_waiting--;
1017                 if (!conf->nr_waiting)
1018                         wake_up(&conf->wait_barrier);
1019         }
1020         atomic_inc(&conf->nr_pending);
1021         spin_unlock_irq(&conf->resync_lock);
1022 }
1023
1024 static void allow_barrier(struct r10conf *conf)
1025 {
1026         if ((atomic_dec_and_test(&conf->nr_pending)) ||
1027                         (conf->array_freeze_pending))
1028                 wake_up(&conf->wait_barrier);
1029 }
1030
1031 static void freeze_array(struct r10conf *conf, int extra)
1032 {
1033         /* stop syncio and normal IO and wait for everything to
1034          * go quiet.
1035          * We increment barrier and nr_waiting, and then
1036          * wait until nr_pending match nr_queued+extra
1037          * This is called in the context of one normal IO request
1038          * that has failed. Thus any sync request that might be pending
1039          * will be blocked by nr_pending, and we need to wait for
1040          * pending IO requests to complete or be queued for re-try.
1041          * Thus the number queued (nr_queued) plus this request (extra)
1042          * must match the number of pending IOs (nr_pending) before
1043          * we continue.
1044          */
1045         spin_lock_irq(&conf->resync_lock);
1046         conf->array_freeze_pending++;
1047         conf->barrier++;
1048         conf->nr_waiting++;
1049         wait_event_lock_irq_cmd(conf->wait_barrier,
1050                                 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1051                                 conf->resync_lock,
1052                                 flush_pending_writes(conf));
1053
1054         conf->array_freeze_pending--;
1055         spin_unlock_irq(&conf->resync_lock);
1056 }
1057
1058 static void unfreeze_array(struct r10conf *conf)
1059 {
1060         /* reverse the effect of the freeze */
1061         spin_lock_irq(&conf->resync_lock);
1062         conf->barrier--;
1063         conf->nr_waiting--;
1064         wake_up(&conf->wait_barrier);
1065         spin_unlock_irq(&conf->resync_lock);
1066 }
1067
1068 static sector_t choose_data_offset(struct r10bio *r10_bio,
1069                                    struct md_rdev *rdev)
1070 {
1071         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1072             test_bit(R10BIO_Previous, &r10_bio->state))
1073                 return rdev->data_offset;
1074         else
1075                 return rdev->new_data_offset;
1076 }
1077
1078 struct raid10_plug_cb {
1079         struct blk_plug_cb      cb;
1080         struct bio_list         pending;
1081         int                     pending_cnt;
1082 };
1083
1084 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1085 {
1086         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1087                                                    cb);
1088         struct mddev *mddev = plug->cb.data;
1089         struct r10conf *conf = mddev->private;
1090         struct bio *bio;
1091
1092         if (from_schedule || current->bio_list) {
1093                 spin_lock_irq(&conf->device_lock);
1094                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1095                 conf->pending_count += plug->pending_cnt;
1096                 spin_unlock_irq(&conf->device_lock);
1097                 wake_up(&conf->wait_barrier);
1098                 md_wakeup_thread(mddev->thread);
1099                 kfree(plug);
1100                 return;
1101         }
1102
1103         /* we aren't scheduling, so we can do the write-out directly. */
1104         bio = bio_list_get(&plug->pending);
1105         bitmap_unplug(mddev->bitmap);
1106         wake_up(&conf->wait_barrier);
1107
1108         while (bio) { /* submit pending writes */
1109                 struct bio *next = bio->bi_next;
1110                 struct md_rdev *rdev = (void*)bio->bi_disk;
1111                 bio->bi_next = NULL;
1112                 bio_set_dev(bio, rdev->bdev);
1113                 if (test_bit(Faulty, &rdev->flags)) {
1114                         bio_io_error(bio);
1115                 } else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
1116                                     !blk_queue_discard(bio->bi_disk->queue)))
1117                         /* Just ignore it */
1118                         bio_endio(bio);
1119                 else
1120                         generic_make_request(bio);
1121                 bio = next;
1122         }
1123         kfree(plug);
1124 }
1125
1126 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1127                                 struct r10bio *r10_bio)
1128 {
1129         struct r10conf *conf = mddev->private;
1130         struct bio *read_bio;
1131         const int op = bio_op(bio);
1132         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1133         int max_sectors;
1134         sector_t sectors;
1135         struct md_rdev *rdev;
1136         char b[BDEVNAME_SIZE];
1137         int slot = r10_bio->read_slot;
1138         struct md_rdev *err_rdev = NULL;
1139         gfp_t gfp = GFP_NOIO;
1140
1141         if (r10_bio->devs[slot].rdev) {
1142                 /*
1143                  * This is an error retry, but we cannot
1144                  * safely dereference the rdev in the r10_bio,
1145                  * we must use the one in conf.
1146                  * If it has already been disconnected (unlikely)
1147                  * we lose the device name in error messages.
1148                  */
1149                 int disk;
1150                 /*
1151                  * As we are blocking raid10, it is a little safer to
1152                  * use __GFP_HIGH.
1153                  */
1154                 gfp = GFP_NOIO | __GFP_HIGH;
1155
1156                 rcu_read_lock();
1157                 disk = r10_bio->devs[slot].devnum;
1158                 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1159                 if (err_rdev)
1160                         bdevname(err_rdev->bdev, b);
1161                 else {
1162                         strcpy(b, "???");
1163                         /* This never gets dereferenced */
1164                         err_rdev = r10_bio->devs[slot].rdev;
1165                 }
1166                 rcu_read_unlock();
1167         }
1168         /*
1169          * Register the new request and wait if the reconstruction
1170          * thread has put up a bar for new requests.
1171          * Continue immediately if no resync is active currently.
1172          */
1173         wait_barrier(conf);
1174
1175         sectors = r10_bio->sectors;
1176         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1177             bio->bi_iter.bi_sector < conf->reshape_progress &&
1178             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1179                 /*
1180                  * IO spans the reshape position.  Need to wait for reshape to
1181                  * pass
1182                  */
1183                 raid10_log(conf->mddev, "wait reshape");
1184                 allow_barrier(conf);
1185                 wait_event(conf->wait_barrier,
1186                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1187                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1188                            sectors);
1189                 wait_barrier(conf);
1190         }
1191
1192         rdev = read_balance(conf, r10_bio, &max_sectors);
1193         if (!rdev) {
1194                 if (err_rdev) {
1195                         pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1196                                             mdname(mddev), b,
1197                                             (unsigned long long)r10_bio->sector);
1198                 }
1199                 raid_end_bio_io(r10_bio);
1200                 return;
1201         }
1202         if (err_rdev)
1203                 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1204                                    mdname(mddev),
1205                                    bdevname(rdev->bdev, b),
1206                                    (unsigned long long)r10_bio->sector);
1207         if (max_sectors < bio_sectors(bio)) {
1208                 struct bio *split = bio_split(bio, max_sectors,
1209                                               gfp, &conf->bio_split);
1210                 bio_chain(split, bio);
1211                 generic_make_request(bio);
1212                 bio = split;
1213                 r10_bio->master_bio = bio;
1214                 r10_bio->sectors = max_sectors;
1215         }
1216         slot = r10_bio->read_slot;
1217
1218         read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1219
1220         r10_bio->devs[slot].bio = read_bio;
1221         r10_bio->devs[slot].rdev = rdev;
1222
1223         read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1224                 choose_data_offset(r10_bio, rdev);
1225         bio_set_dev(read_bio, rdev->bdev);
1226         read_bio->bi_end_io = raid10_end_read_request;
1227         bio_set_op_attrs(read_bio, op, do_sync);
1228         if (test_bit(FailFast, &rdev->flags) &&
1229             test_bit(R10BIO_FailFast, &r10_bio->state))
1230                 read_bio->bi_opf |= MD_FAILFAST;
1231         read_bio->bi_private = r10_bio;
1232
1233         if (mddev->gendisk)
1234                 trace_block_bio_remap(read_bio->bi_disk->queue,
1235                                       read_bio, disk_devt(mddev->gendisk),
1236                                       r10_bio->sector);
1237         generic_make_request(read_bio);
1238         return;
1239 }
1240
1241 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1242                                   struct bio *bio, bool replacement,
1243                                   int n_copy)
1244 {
1245         const int op = bio_op(bio);
1246         const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1247         const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1248         unsigned long flags;
1249         struct blk_plug_cb *cb;
1250         struct raid10_plug_cb *plug = NULL;
1251         struct r10conf *conf = mddev->private;
1252         struct md_rdev *rdev;
1253         int devnum = r10_bio->devs[n_copy].devnum;
1254         struct bio *mbio;
1255
1256         if (replacement) {
1257                 rdev = conf->mirrors[devnum].replacement;
1258                 if (rdev == NULL) {
1259                         /* Replacement just got moved to main 'rdev' */
1260                         smp_mb();
1261                         rdev = conf->mirrors[devnum].rdev;
1262                 }
1263         } else
1264                 rdev = conf->mirrors[devnum].rdev;
1265
1266         mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1267         if (replacement)
1268                 r10_bio->devs[n_copy].repl_bio = mbio;
1269         else
1270                 r10_bio->devs[n_copy].bio = mbio;
1271
1272         mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1273                                    choose_data_offset(r10_bio, rdev));
1274         bio_set_dev(mbio, rdev->bdev);
1275         mbio->bi_end_io = raid10_end_write_request;
1276         bio_set_op_attrs(mbio, op, do_sync | do_fua);
1277         if (!replacement && test_bit(FailFast,
1278                                      &conf->mirrors[devnum].rdev->flags)
1279                          && enough(conf, devnum))
1280                 mbio->bi_opf |= MD_FAILFAST;
1281         mbio->bi_private = r10_bio;
1282
1283         if (conf->mddev->gendisk)
1284                 trace_block_bio_remap(mbio->bi_disk->queue,
1285                                       mbio, disk_devt(conf->mddev->gendisk),
1286                                       r10_bio->sector);
1287         /* flush_pending_writes() needs access to the rdev so...*/
1288         mbio->bi_disk = (void *)rdev;
1289
1290         atomic_inc(&r10_bio->remaining);
1291
1292         cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1293         if (cb)
1294                 plug = container_of(cb, struct raid10_plug_cb, cb);
1295         else
1296                 plug = NULL;
1297         if (plug) {
1298                 bio_list_add(&plug->pending, mbio);
1299                 plug->pending_cnt++;
1300         } else {
1301                 spin_lock_irqsave(&conf->device_lock, flags);
1302                 bio_list_add(&conf->pending_bio_list, mbio);
1303                 conf->pending_count++;
1304                 spin_unlock_irqrestore(&conf->device_lock, flags);
1305                 md_wakeup_thread(mddev->thread);
1306         }
1307 }
1308
1309 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1310                                  struct r10bio *r10_bio)
1311 {
1312         struct r10conf *conf = mddev->private;
1313         int i;
1314         struct md_rdev *blocked_rdev;
1315         sector_t sectors;
1316         int max_sectors;
1317
1318         if ((mddev_is_clustered(mddev) &&
1319              md_cluster_ops->area_resyncing(mddev, WRITE,
1320                                             bio->bi_iter.bi_sector,
1321                                             bio_end_sector(bio)))) {
1322                 DEFINE_WAIT(w);
1323                 for (;;) {
1324                         prepare_to_wait(&conf->wait_barrier,
1325                                         &w, TASK_IDLE);
1326                         if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1327                                  bio->bi_iter.bi_sector, bio_end_sector(bio)))
1328                                 break;
1329                         schedule();
1330                 }
1331                 finish_wait(&conf->wait_barrier, &w);
1332         }
1333
1334         /*
1335          * Register the new request and wait if the reconstruction
1336          * thread has put up a bar for new requests.
1337          * Continue immediately if no resync is active currently.
1338          */
1339         wait_barrier(conf);
1340
1341         sectors = r10_bio->sectors;
1342         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1343             bio->bi_iter.bi_sector < conf->reshape_progress &&
1344             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1345                 /*
1346                  * IO spans the reshape position.  Need to wait for reshape to
1347                  * pass
1348                  */
1349                 raid10_log(conf->mddev, "wait reshape");
1350                 allow_barrier(conf);
1351                 wait_event(conf->wait_barrier,
1352                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1353                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1354                            sectors);
1355                 wait_barrier(conf);
1356         }
1357
1358         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1359             (mddev->reshape_backwards
1360              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1361                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1362              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1363                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1364                 /* Need to update reshape_position in metadata */
1365                 mddev->reshape_position = conf->reshape_progress;
1366                 set_mask_bits(&mddev->sb_flags, 0,
1367                               BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1368                 md_wakeup_thread(mddev->thread);
1369                 raid10_log(conf->mddev, "wait reshape metadata");
1370                 wait_event(mddev->sb_wait,
1371                            !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1372
1373                 conf->reshape_safe = mddev->reshape_position;
1374         }
1375
1376         if (conf->pending_count >= max_queued_requests) {
1377                 md_wakeup_thread(mddev->thread);
1378                 raid10_log(mddev, "wait queued");
1379                 wait_event(conf->wait_barrier,
1380                            conf->pending_count < max_queued_requests);
1381         }
1382         /* first select target devices under rcu_lock and
1383          * inc refcount on their rdev.  Record them by setting
1384          * bios[x] to bio
1385          * If there are known/acknowledged bad blocks on any device
1386          * on which we have seen a write error, we want to avoid
1387          * writing to those blocks.  This potentially requires several
1388          * writes to write around the bad blocks.  Each set of writes
1389          * gets its own r10_bio with a set of bios attached.
1390          */
1391
1392         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1393         raid10_find_phys(conf, r10_bio);
1394 retry_write:
1395         blocked_rdev = NULL;
1396         rcu_read_lock();
1397         max_sectors = r10_bio->sectors;
1398
1399         for (i = 0;  i < conf->copies; i++) {
1400                 int d = r10_bio->devs[i].devnum;
1401                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1402                 struct md_rdev *rrdev = rcu_dereference(
1403                         conf->mirrors[d].replacement);
1404                 if (rdev == rrdev)
1405                         rrdev = NULL;
1406                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1407                         atomic_inc(&rdev->nr_pending);
1408                         blocked_rdev = rdev;
1409                         break;
1410                 }
1411                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1412                         atomic_inc(&rrdev->nr_pending);
1413                         blocked_rdev = rrdev;
1414                         break;
1415                 }
1416                 if (rdev && (test_bit(Faulty, &rdev->flags)))
1417                         rdev = NULL;
1418                 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1419                         rrdev = NULL;
1420
1421                 r10_bio->devs[i].bio = NULL;
1422                 r10_bio->devs[i].repl_bio = NULL;
1423
1424                 if (!rdev && !rrdev) {
1425                         set_bit(R10BIO_Degraded, &r10_bio->state);
1426                         continue;
1427                 }
1428                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1429                         sector_t first_bad;
1430                         sector_t dev_sector = r10_bio->devs[i].addr;
1431                         int bad_sectors;
1432                         int is_bad;
1433
1434                         is_bad = is_badblock(rdev, dev_sector, max_sectors,
1435                                              &first_bad, &bad_sectors);
1436                         if (is_bad < 0) {
1437                                 /* Mustn't write here until the bad block
1438                                  * is acknowledged
1439                                  */
1440                                 atomic_inc(&rdev->nr_pending);
1441                                 set_bit(BlockedBadBlocks, &rdev->flags);
1442                                 blocked_rdev = rdev;
1443                                 break;
1444                         }
1445                         if (is_bad && first_bad <= dev_sector) {
1446                                 /* Cannot write here at all */
1447                                 bad_sectors -= (dev_sector - first_bad);
1448                                 if (bad_sectors < max_sectors)
1449                                         /* Mustn't write more than bad_sectors
1450                                          * to other devices yet
1451                                          */
1452                                         max_sectors = bad_sectors;
1453                                 /* We don't set R10BIO_Degraded as that
1454                                  * only applies if the disk is missing,
1455                                  * so it might be re-added, and we want to
1456                                  * know to recover this chunk.
1457                                  * In this case the device is here, and the
1458                                  * fact that this chunk is not in-sync is
1459                                  * recorded in the bad block log.
1460                                  */
1461                                 continue;
1462                         }
1463                         if (is_bad) {
1464                                 int good_sectors = first_bad - dev_sector;
1465                                 if (good_sectors < max_sectors)
1466                                         max_sectors = good_sectors;
1467                         }
1468                 }
1469                 if (rdev) {
1470                         r10_bio->devs[i].bio = bio;
1471                         atomic_inc(&rdev->nr_pending);
1472                 }
1473                 if (rrdev) {
1474                         r10_bio->devs[i].repl_bio = bio;
1475                         atomic_inc(&rrdev->nr_pending);
1476                 }
1477         }
1478         rcu_read_unlock();
1479
1480         if (unlikely(blocked_rdev)) {
1481                 /* Have to wait for this device to get unblocked, then retry */
1482                 int j;
1483                 int d;
1484
1485                 for (j = 0; j < i; j++) {
1486                         if (r10_bio->devs[j].bio) {
1487                                 d = r10_bio->devs[j].devnum;
1488                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1489                         }
1490                         if (r10_bio->devs[j].repl_bio) {
1491                                 struct md_rdev *rdev;
1492                                 d = r10_bio->devs[j].devnum;
1493                                 rdev = conf->mirrors[d].replacement;
1494                                 if (!rdev) {
1495                                         /* Race with remove_disk */
1496                                         smp_mb();
1497                                         rdev = conf->mirrors[d].rdev;
1498                                 }
1499                                 rdev_dec_pending(rdev, mddev);
1500                         }
1501                 }
1502                 allow_barrier(conf);
1503                 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1504                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1505                 wait_barrier(conf);
1506                 goto retry_write;
1507         }
1508
1509         if (max_sectors < r10_bio->sectors)
1510                 r10_bio->sectors = max_sectors;
1511
1512         if (r10_bio->sectors < bio_sectors(bio)) {
1513                 struct bio *split = bio_split(bio, r10_bio->sectors,
1514                                               GFP_NOIO, &conf->bio_split);
1515                 bio_chain(split, bio);
1516                 generic_make_request(bio);
1517                 bio = split;
1518                 r10_bio->master_bio = bio;
1519         }
1520
1521         atomic_set(&r10_bio->remaining, 1);
1522         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1523
1524         for (i = 0; i < conf->copies; i++) {
1525                 if (r10_bio->devs[i].bio)
1526                         raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1527                 if (r10_bio->devs[i].repl_bio)
1528                         raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1529         }
1530         one_write_done(r10_bio);
1531 }
1532
1533 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1534 {
1535         struct r10conf *conf = mddev->private;
1536         struct r10bio *r10_bio;
1537
1538         r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1539
1540         r10_bio->master_bio = bio;
1541         r10_bio->sectors = sectors;
1542
1543         r10_bio->mddev = mddev;
1544         r10_bio->sector = bio->bi_iter.bi_sector;
1545         r10_bio->state = 0;
1546         memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1547
1548         if (bio_data_dir(bio) == READ)
1549                 raid10_read_request(mddev, bio, r10_bio);
1550         else
1551                 raid10_write_request(mddev, bio, r10_bio);
1552 }
1553
1554 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1555 {
1556         struct r10conf *conf = mddev->private;
1557         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1558         int chunk_sects = chunk_mask + 1;
1559         int sectors = bio_sectors(bio);
1560
1561         if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1562                 md_flush_request(mddev, bio);
1563                 return true;
1564         }
1565
1566         if (!md_write_start(mddev, bio))
1567                 return false;
1568
1569         /*
1570          * If this request crosses a chunk boundary, we need to split
1571          * it.
1572          */
1573         if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1574                      sectors > chunk_sects
1575                      && (conf->geo.near_copies < conf->geo.raid_disks
1576                          || conf->prev.near_copies <
1577                          conf->prev.raid_disks)))
1578                 sectors = chunk_sects -
1579                         (bio->bi_iter.bi_sector &
1580                          (chunk_sects - 1));
1581         __make_request(mddev, bio, sectors);
1582
1583         /* In case raid10d snuck in to freeze_array */
1584         wake_up(&conf->wait_barrier);
1585         return true;
1586 }
1587
1588 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1589 {
1590         struct r10conf *conf = mddev->private;
1591         int i;
1592
1593         if (conf->geo.near_copies < conf->geo.raid_disks)
1594                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1595         if (conf->geo.near_copies > 1)
1596                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1597         if (conf->geo.far_copies > 1) {
1598                 if (conf->geo.far_offset)
1599                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1600                 else
1601                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1602                 if (conf->geo.far_set_size != conf->geo.raid_disks)
1603                         seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1604         }
1605         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1606                                         conf->geo.raid_disks - mddev->degraded);
1607         rcu_read_lock();
1608         for (i = 0; i < conf->geo.raid_disks; i++) {
1609                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1610                 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1611         }
1612         rcu_read_unlock();
1613         seq_printf(seq, "]");
1614 }
1615
1616 /* check if there are enough drives for
1617  * every block to appear on atleast one.
1618  * Don't consider the device numbered 'ignore'
1619  * as we might be about to remove it.
1620  */
1621 static int _enough(struct r10conf *conf, int previous, int ignore)
1622 {
1623         int first = 0;
1624         int has_enough = 0;
1625         int disks, ncopies;
1626         if (previous) {
1627                 disks = conf->prev.raid_disks;
1628                 ncopies = conf->prev.near_copies;
1629         } else {
1630                 disks = conf->geo.raid_disks;
1631                 ncopies = conf->geo.near_copies;
1632         }
1633
1634         rcu_read_lock();
1635         do {
1636                 int n = conf->copies;
1637                 int cnt = 0;
1638                 int this = first;
1639                 while (n--) {
1640                         struct md_rdev *rdev;
1641                         if (this != ignore &&
1642                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1643                             test_bit(In_sync, &rdev->flags))
1644                                 cnt++;
1645                         this = (this+1) % disks;
1646                 }
1647                 if (cnt == 0)
1648                         goto out;
1649                 first = (first + ncopies) % disks;
1650         } while (first != 0);
1651         has_enough = 1;
1652 out:
1653         rcu_read_unlock();
1654         return has_enough;
1655 }
1656
1657 static int enough(struct r10conf *conf, int ignore)
1658 {
1659         /* when calling 'enough', both 'prev' and 'geo' must
1660          * be stable.
1661          * This is ensured if ->reconfig_mutex or ->device_lock
1662          * is held.
1663          */
1664         return _enough(conf, 0, ignore) &&
1665                 _enough(conf, 1, ignore);
1666 }
1667
1668 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1669 {
1670         char b[BDEVNAME_SIZE];
1671         struct r10conf *conf = mddev->private;
1672         unsigned long flags;
1673
1674         /*
1675          * If it is not operational, then we have already marked it as dead
1676          * else if it is the last working disks, ignore the error, let the
1677          * next level up know.
1678          * else mark the drive as failed
1679          */
1680         spin_lock_irqsave(&conf->device_lock, flags);
1681         if (test_bit(In_sync, &rdev->flags)
1682             && !enough(conf, rdev->raid_disk)) {
1683                 /*
1684                  * Don't fail the drive, just return an IO error.
1685                  */
1686                 spin_unlock_irqrestore(&conf->device_lock, flags);
1687                 return;
1688         }
1689         if (test_and_clear_bit(In_sync, &rdev->flags))
1690                 mddev->degraded++;
1691         /*
1692          * If recovery is running, make sure it aborts.
1693          */
1694         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1695         set_bit(Blocked, &rdev->flags);
1696         set_bit(Faulty, &rdev->flags);
1697         set_mask_bits(&mddev->sb_flags, 0,
1698                       BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1699         spin_unlock_irqrestore(&conf->device_lock, flags);
1700         pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1701                 "md/raid10:%s: Operation continuing on %d devices.\n",
1702                 mdname(mddev), bdevname(rdev->bdev, b),
1703                 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1704 }
1705
1706 static void print_conf(struct r10conf *conf)
1707 {
1708         int i;
1709         struct md_rdev *rdev;
1710
1711         pr_debug("RAID10 conf printout:\n");
1712         if (!conf) {
1713                 pr_debug("(!conf)\n");
1714                 return;
1715         }
1716         pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1717                  conf->geo.raid_disks);
1718
1719         /* This is only called with ->reconfix_mutex held, so
1720          * rcu protection of rdev is not needed */
1721         for (i = 0; i < conf->geo.raid_disks; i++) {
1722                 char b[BDEVNAME_SIZE];
1723                 rdev = conf->mirrors[i].rdev;
1724                 if (rdev)
1725                         pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1726                                  i, !test_bit(In_sync, &rdev->flags),
1727                                  !test_bit(Faulty, &rdev->flags),
1728                                  bdevname(rdev->bdev,b));
1729         }
1730 }
1731
1732 static void close_sync(struct r10conf *conf)
1733 {
1734         wait_barrier(conf);
1735         allow_barrier(conf);
1736
1737         mempool_exit(&conf->r10buf_pool);
1738 }
1739
1740 static int raid10_spare_active(struct mddev *mddev)
1741 {
1742         int i;
1743         struct r10conf *conf = mddev->private;
1744         struct raid10_info *tmp;
1745         int count = 0;
1746         unsigned long flags;
1747
1748         /*
1749          * Find all non-in_sync disks within the RAID10 configuration
1750          * and mark them in_sync
1751          */
1752         for (i = 0; i < conf->geo.raid_disks; i++) {
1753                 tmp = conf->mirrors + i;
1754                 if (tmp->replacement
1755                     && tmp->replacement->recovery_offset == MaxSector
1756                     && !test_bit(Faulty, &tmp->replacement->flags)
1757                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1758                         /* Replacement has just become active */
1759                         if (!tmp->rdev
1760                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1761                                 count++;
1762                         if (tmp->rdev) {
1763                                 /* Replaced device not technically faulty,
1764                                  * but we need to be sure it gets removed
1765                                  * and never re-added.
1766                                  */
1767                                 set_bit(Faulty, &tmp->rdev->flags);
1768                                 sysfs_notify_dirent_safe(
1769                                         tmp->rdev->sysfs_state);
1770                         }
1771                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1772                 } else if (tmp->rdev
1773                            && tmp->rdev->recovery_offset == MaxSector
1774                            && !test_bit(Faulty, &tmp->rdev->flags)
1775                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1776                         count++;
1777                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1778                 }
1779         }
1780         spin_lock_irqsave(&conf->device_lock, flags);
1781         mddev->degraded -= count;
1782         spin_unlock_irqrestore(&conf->device_lock, flags);
1783
1784         print_conf(conf);
1785         return count;
1786 }
1787
1788 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1789 {
1790         struct r10conf *conf = mddev->private;
1791         int err = -EEXIST;
1792         int mirror;
1793         int first = 0;
1794         int last = conf->geo.raid_disks - 1;
1795
1796         if (mddev->recovery_cp < MaxSector)
1797                 /* only hot-add to in-sync arrays, as recovery is
1798                  * very different from resync
1799                  */
1800                 return -EBUSY;
1801         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1802                 return -EINVAL;
1803
1804         if (md_integrity_add_rdev(rdev, mddev))
1805                 return -ENXIO;
1806
1807         if (rdev->raid_disk >= 0)
1808                 first = last = rdev->raid_disk;
1809
1810         if (rdev->saved_raid_disk >= first &&
1811             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1812                 mirror = rdev->saved_raid_disk;
1813         else
1814                 mirror = first;
1815         for ( ; mirror <= last ; mirror++) {
1816                 struct raid10_info *p = &conf->mirrors[mirror];
1817                 if (p->recovery_disabled == mddev->recovery_disabled)
1818                         continue;
1819                 if (p->rdev) {
1820                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1821                             p->replacement != NULL)
1822                                 continue;
1823                         clear_bit(In_sync, &rdev->flags);
1824                         set_bit(Replacement, &rdev->flags);
1825                         rdev->raid_disk = mirror;
1826                         err = 0;
1827                         if (mddev->gendisk)
1828                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1829                                                   rdev->data_offset << 9);
1830                         conf->fullsync = 1;
1831                         rcu_assign_pointer(p->replacement, rdev);
1832                         break;
1833                 }
1834
1835                 if (mddev->gendisk)
1836                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1837                                           rdev->data_offset << 9);
1838
1839                 p->head_position = 0;
1840                 p->recovery_disabled = mddev->recovery_disabled - 1;
1841                 rdev->raid_disk = mirror;
1842                 err = 0;
1843                 if (rdev->saved_raid_disk != mirror)
1844                         conf->fullsync = 1;
1845                 rcu_assign_pointer(p->rdev, rdev);
1846                 break;
1847         }
1848         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1849                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1850
1851         print_conf(conf);
1852         return err;
1853 }
1854
1855 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1856 {
1857         struct r10conf *conf = mddev->private;
1858         int err = 0;
1859         int number = rdev->raid_disk;
1860         struct md_rdev **rdevp;
1861         struct raid10_info *p = conf->mirrors + number;
1862
1863         print_conf(conf);
1864         if (rdev == p->rdev)
1865                 rdevp = &p->rdev;
1866         else if (rdev == p->replacement)
1867                 rdevp = &p->replacement;
1868         else
1869                 return 0;
1870
1871         if (test_bit(In_sync, &rdev->flags) ||
1872             atomic_read(&rdev->nr_pending)) {
1873                 err = -EBUSY;
1874                 goto abort;
1875         }
1876         /* Only remove non-faulty devices if recovery
1877          * is not possible.
1878          */
1879         if (!test_bit(Faulty, &rdev->flags) &&
1880             mddev->recovery_disabled != p->recovery_disabled &&
1881             (!p->replacement || p->replacement == rdev) &&
1882             number < conf->geo.raid_disks &&
1883             enough(conf, -1)) {
1884                 err = -EBUSY;
1885                 goto abort;
1886         }
1887         *rdevp = NULL;
1888         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1889                 synchronize_rcu();
1890                 if (atomic_read(&rdev->nr_pending)) {
1891                         /* lost the race, try later */
1892                         err = -EBUSY;
1893                         *rdevp = rdev;
1894                         goto abort;
1895                 }
1896         }
1897         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         }
1906
1907         clear_bit(WantReplacement, &rdev->flags);
1908         err = md_integrity_register(mddev);
1909
1910 abort:
1911
1912         print_conf(conf);
1913         return err;
1914 }
1915
1916 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1917 {
1918         struct r10conf *conf = r10_bio->mddev->private;
1919
1920         if (!bio->bi_status)
1921                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1922         else
1923                 /* The write handler will notice the lack of
1924                  * R10BIO_Uptodate and record any errors etc
1925                  */
1926                 atomic_add(r10_bio->sectors,
1927                            &conf->mirrors[d].rdev->corrected_errors);
1928
1929         /* for reconstruct, we always reschedule after a read.
1930          * for resync, only after all reads
1931          */
1932         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1933         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1934             atomic_dec_and_test(&r10_bio->remaining)) {
1935                 /* we have read all the blocks,
1936                  * do the comparison in process context in raid10d
1937                  */
1938                 reschedule_retry(r10_bio);
1939         }
1940 }
1941
1942 static void end_sync_read(struct bio *bio)
1943 {
1944         struct r10bio *r10_bio = get_resync_r10bio(bio);
1945         struct r10conf *conf = r10_bio->mddev->private;
1946         int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1947
1948         __end_sync_read(r10_bio, bio, d);
1949 }
1950
1951 static void end_reshape_read(struct bio *bio)
1952 {
1953         /* reshape read bio isn't allocated from r10buf_pool */
1954         struct r10bio *r10_bio = bio->bi_private;
1955
1956         __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1957 }
1958
1959 static void end_sync_request(struct r10bio *r10_bio)
1960 {
1961         struct mddev *mddev = r10_bio->mddev;
1962
1963         while (atomic_dec_and_test(&r10_bio->remaining)) {
1964                 if (r10_bio->master_bio == NULL) {
1965                         /* the primary of several recovery bios */
1966                         sector_t s = r10_bio->sectors;
1967                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1968                             test_bit(R10BIO_WriteError, &r10_bio->state))
1969                                 reschedule_retry(r10_bio);
1970                         else
1971                                 put_buf(r10_bio);
1972                         md_done_sync(mddev, s, 1);
1973                         break;
1974                 } else {
1975                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1976                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1977                             test_bit(R10BIO_WriteError, &r10_bio->state))
1978                                 reschedule_retry(r10_bio);
1979                         else
1980                                 put_buf(r10_bio);
1981                         r10_bio = r10_bio2;
1982                 }
1983         }
1984 }
1985
1986 static void end_sync_write(struct bio *bio)
1987 {
1988         struct r10bio *r10_bio = get_resync_r10bio(bio);
1989         struct mddev *mddev = r10_bio->mddev;
1990         struct r10conf *conf = mddev->private;
1991         int d;
1992         sector_t first_bad;
1993         int bad_sectors;
1994         int slot;
1995         int repl;
1996         struct md_rdev *rdev = NULL;
1997
1998         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1999         if (repl)
2000                 rdev = conf->mirrors[d].replacement;
2001         else
2002                 rdev = conf->mirrors[d].rdev;
2003
2004         if (bio->bi_status) {
2005                 if (repl)
2006                         md_error(mddev, rdev);
2007                 else {
2008                         set_bit(WriteErrorSeen, &rdev->flags);
2009                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
2010                                 set_bit(MD_RECOVERY_NEEDED,
2011                                         &rdev->mddev->recovery);
2012                         set_bit(R10BIO_WriteError, &r10_bio->state);
2013                 }
2014         } else if (is_badblock(rdev,
2015                              r10_bio->devs[slot].addr,
2016                              r10_bio->sectors,
2017                              &first_bad, &bad_sectors))
2018                 set_bit(R10BIO_MadeGood, &r10_bio->state);
2019
2020         rdev_dec_pending(rdev, mddev);
2021
2022         end_sync_request(r10_bio);
2023 }
2024
2025 /*
2026  * Note: sync and recover and handled very differently for raid10
2027  * This code is for resync.
2028  * For resync, we read through virtual addresses and read all blocks.
2029  * If there is any error, we schedule a write.  The lowest numbered
2030  * drive is authoritative.
2031  * However requests come for physical address, so we need to map.
2032  * For every physical address there are raid_disks/copies virtual addresses,
2033  * which is always are least one, but is not necessarly an integer.
2034  * This means that a physical address can span multiple chunks, so we may
2035  * have to submit multiple io requests for a single sync request.
2036  */
2037 /*
2038  * We check if all blocks are in-sync and only write to blocks that
2039  * aren't in sync
2040  */
2041 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2042 {
2043         struct r10conf *conf = mddev->private;
2044         int i, first;
2045         struct bio *tbio, *fbio;
2046         int vcnt;
2047         struct page **tpages, **fpages;
2048
2049         atomic_set(&r10_bio->remaining, 1);
2050
2051         /* find the first device with a block */
2052         for (i=0; i<conf->copies; i++)
2053                 if (!r10_bio->devs[i].bio->bi_status)
2054                         break;
2055
2056         if (i == conf->copies)
2057                 goto done;
2058
2059         first = i;
2060         fbio = r10_bio->devs[i].bio;
2061         fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2062         fbio->bi_iter.bi_idx = 0;
2063         fpages = get_resync_pages(fbio)->pages;
2064
2065         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2066         /* now find blocks with errors */
2067         for (i=0 ; i < conf->copies ; i++) {
2068                 int  j, d;
2069                 struct md_rdev *rdev;
2070                 struct resync_pages *rp;
2071
2072                 tbio = r10_bio->devs[i].bio;
2073
2074                 if (tbio->bi_end_io != end_sync_read)
2075                         continue;
2076                 if (i == first)
2077                         continue;
2078
2079                 tpages = get_resync_pages(tbio)->pages;
2080                 d = r10_bio->devs[i].devnum;
2081                 rdev = conf->mirrors[d].rdev;
2082                 if (!r10_bio->devs[i].bio->bi_status) {
2083                         /* We know that the bi_io_vec layout is the same for
2084                          * both 'first' and 'i', so we just compare them.
2085                          * All vec entries are PAGE_SIZE;
2086                          */
2087                         int sectors = r10_bio->sectors;
2088                         for (j = 0; j < vcnt; j++) {
2089                                 int len = PAGE_SIZE;
2090                                 if (sectors < (len / 512))
2091                                         len = sectors * 512;
2092                                 if (memcmp(page_address(fpages[j]),
2093                                            page_address(tpages[j]),
2094                                            len))
2095                                         break;
2096                                 sectors -= len/512;
2097                         }
2098                         if (j == vcnt)
2099                                 continue;
2100                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2101                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2102                                 /* Don't fix anything. */
2103                                 continue;
2104                 } else if (test_bit(FailFast, &rdev->flags)) {
2105                         /* Just give up on this device */
2106                         md_error(rdev->mddev, rdev);
2107                         continue;
2108                 }
2109                 /* Ok, we need to write this bio, either to correct an
2110                  * inconsistency or to correct an unreadable block.
2111                  * First we need to fixup bv_offset, bv_len and
2112                  * bi_vecs, as the read request might have corrupted these
2113                  */
2114                 rp = get_resync_pages(tbio);
2115                 bio_reset(tbio);
2116
2117                 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2118
2119                 rp->raid_bio = r10_bio;
2120                 tbio->bi_private = rp;
2121                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2122                 tbio->bi_end_io = end_sync_write;
2123                 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2124
2125                 bio_copy_data(tbio, fbio);
2126
2127                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2128                 atomic_inc(&r10_bio->remaining);
2129                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2130
2131                 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2132                         tbio->bi_opf |= MD_FAILFAST;
2133                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2134                 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2135                 generic_make_request(tbio);
2136         }
2137
2138         /* Now write out to any replacement devices
2139          * that are active
2140          */
2141         for (i = 0; i < conf->copies; i++) {
2142                 int d;
2143
2144                 tbio = r10_bio->devs[i].repl_bio;
2145                 if (!tbio || !tbio->bi_end_io)
2146                         continue;
2147                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2148                     && r10_bio->devs[i].bio != fbio)
2149                         bio_copy_data(tbio, fbio);
2150                 d = r10_bio->devs[i].devnum;
2151                 atomic_inc(&r10_bio->remaining);
2152                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2153                              bio_sectors(tbio));
2154                 generic_make_request(tbio);
2155         }
2156
2157 done:
2158         if (atomic_dec_and_test(&r10_bio->remaining)) {
2159                 md_done_sync(mddev, r10_bio->sectors, 1);
2160                 put_buf(r10_bio);
2161         }
2162 }
2163
2164 /*
2165  * Now for the recovery code.
2166  * Recovery happens across physical sectors.
2167  * We recover all non-is_sync drives by finding the virtual address of
2168  * each, and then choose a working drive that also has that virt address.
2169  * There is a separate r10_bio for each non-in_sync drive.
2170  * Only the first two slots are in use. The first for reading,
2171  * The second for writing.
2172  *
2173  */
2174 static void fix_recovery_read_error(struct r10bio *r10_bio)
2175 {
2176         /* We got a read error during recovery.
2177          * We repeat the read in smaller page-sized sections.
2178          * If a read succeeds, write it to the new device or record
2179          * a bad block if we cannot.
2180          * If a read fails, record a bad block on both old and
2181          * new devices.
2182          */
2183         struct mddev *mddev = r10_bio->mddev;
2184         struct r10conf *conf = mddev->private;
2185         struct bio *bio = r10_bio->devs[0].bio;
2186         sector_t sect = 0;
2187         int sectors = r10_bio->sectors;
2188         int idx = 0;
2189         int dr = r10_bio->devs[0].devnum;
2190         int dw = r10_bio->devs[1].devnum;
2191         struct page **pages = get_resync_pages(bio)->pages;
2192
2193         while (sectors) {
2194                 int s = sectors;
2195                 struct md_rdev *rdev;
2196                 sector_t addr;
2197                 int ok;
2198
2199                 if (s > (PAGE_SIZE>>9))
2200                         s = PAGE_SIZE >> 9;
2201
2202                 rdev = conf->mirrors[dr].rdev;
2203                 addr = r10_bio->devs[0].addr + sect,
2204                 ok = sync_page_io(rdev,
2205                                   addr,
2206                                   s << 9,
2207                                   pages[idx],
2208                                   REQ_OP_READ, 0, false);
2209                 if (ok) {
2210                         rdev = conf->mirrors[dw].rdev;
2211                         addr = r10_bio->devs[1].addr + sect;
2212                         ok = sync_page_io(rdev,
2213                                           addr,
2214                                           s << 9,
2215                                           pages[idx],
2216                                           REQ_OP_WRITE, 0, false);
2217                         if (!ok) {
2218                                 set_bit(WriteErrorSeen, &rdev->flags);
2219                                 if (!test_and_set_bit(WantReplacement,
2220                                                       &rdev->flags))
2221                                         set_bit(MD_RECOVERY_NEEDED,
2222                                                 &rdev->mddev->recovery);
2223                         }
2224                 }
2225                 if (!ok) {
2226                         /* We don't worry if we cannot set a bad block -
2227                          * it really is bad so there is no loss in not
2228                          * recording it yet
2229                          */
2230                         rdev_set_badblocks(rdev, addr, s, 0);
2231
2232                         if (rdev != conf->mirrors[dw].rdev) {
2233                                 /* need bad block on destination too */
2234                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2235                                 addr = r10_bio->devs[1].addr + sect;
2236                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2237                                 if (!ok) {
2238                                         /* just abort the recovery */
2239                                         pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2240                                                   mdname(mddev));
2241
2242                                         conf->mirrors[dw].recovery_disabled
2243                                                 = mddev->recovery_disabled;
2244                                         set_bit(MD_RECOVERY_INTR,
2245                                                 &mddev->recovery);
2246                                         break;
2247                                 }
2248                         }
2249                 }
2250
2251                 sectors -= s;
2252                 sect += s;
2253                 idx++;
2254         }
2255 }
2256
2257 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2258 {
2259         struct r10conf *conf = mddev->private;
2260         int d;
2261         struct bio *wbio, *wbio2;
2262
2263         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2264                 fix_recovery_read_error(r10_bio);
2265                 end_sync_request(r10_bio);
2266                 return;
2267         }
2268
2269         /*
2270          * share the pages with the first bio
2271          * and submit the write request
2272          */
2273         d = r10_bio->devs[1].devnum;
2274         wbio = r10_bio->devs[1].bio;
2275         wbio2 = r10_bio->devs[1].repl_bio;
2276         /* Need to test wbio2->bi_end_io before we call
2277          * generic_make_request as if the former is NULL,
2278          * the latter is free to free wbio2.
2279          */
2280         if (wbio2 && !wbio2->bi_end_io)
2281                 wbio2 = NULL;
2282         if (wbio->bi_end_io) {
2283                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2284                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2285                 generic_make_request(wbio);
2286         }
2287         if (wbio2) {
2288                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2289                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2290                              bio_sectors(wbio2));
2291                 generic_make_request(wbio2);
2292         }
2293 }
2294
2295 /*
2296  * Used by fix_read_error() to decay the per rdev read_errors.
2297  * We halve the read error count for every hour that has elapsed
2298  * since the last recorded read error.
2299  *
2300  */
2301 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2302 {
2303         long cur_time_mon;
2304         unsigned long hours_since_last;
2305         unsigned int read_errors = atomic_read(&rdev->read_errors);
2306
2307         cur_time_mon = ktime_get_seconds();
2308
2309         if (rdev->last_read_error == 0) {
2310                 /* first time we've seen a read error */
2311                 rdev->last_read_error = cur_time_mon;
2312                 return;
2313         }
2314
2315         hours_since_last = (long)(cur_time_mon -
2316                             rdev->last_read_error) / 3600;
2317
2318         rdev->last_read_error = cur_time_mon;
2319
2320         /*
2321          * if hours_since_last is > the number of bits in read_errors
2322          * just set read errors to 0. We do this to avoid
2323          * overflowing the shift of read_errors by hours_since_last.
2324          */
2325         if (hours_since_last >= 8 * sizeof(read_errors))
2326                 atomic_set(&rdev->read_errors, 0);
2327         else
2328                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2329 }
2330
2331 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2332                             int sectors, struct page *page, int rw)
2333 {
2334         sector_t first_bad;
2335         int bad_sectors;
2336
2337         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2338             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2339                 return -1;
2340         if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2341                 /* success */
2342                 return 1;
2343         if (rw == WRITE) {
2344                 set_bit(WriteErrorSeen, &rdev->flags);
2345                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2346                         set_bit(MD_RECOVERY_NEEDED,
2347                                 &rdev->mddev->recovery);
2348         }
2349         /* need to record an error - either for the block or the device */
2350         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2351                 md_error(rdev->mddev, rdev);
2352         return 0;
2353 }
2354
2355 /*
2356  * This is a kernel thread which:
2357  *
2358  *      1.      Retries failed read operations on working mirrors.
2359  *      2.      Updates the raid superblock when problems encounter.
2360  *      3.      Performs writes following reads for array synchronising.
2361  */
2362
2363 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2364 {
2365         int sect = 0; /* Offset from r10_bio->sector */
2366         int sectors = r10_bio->sectors;
2367         struct md_rdev *rdev;
2368         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2369         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2370
2371         /* still own a reference to this rdev, so it cannot
2372          * have been cleared recently.
2373          */
2374         rdev = conf->mirrors[d].rdev;
2375
2376         if (test_bit(Faulty, &rdev->flags))
2377                 /* drive has already been failed, just ignore any
2378                    more fix_read_error() attempts */
2379                 return;
2380
2381         check_decay_read_errors(mddev, rdev);
2382         atomic_inc(&rdev->read_errors);
2383         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2384                 char b[BDEVNAME_SIZE];
2385                 bdevname(rdev->bdev, b);
2386
2387                 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2388                           mdname(mddev), b,
2389                           atomic_read(&rdev->read_errors), max_read_errors);
2390                 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2391                           mdname(mddev), b);
2392                 md_error(mddev, rdev);
2393                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2394                 return;
2395         }
2396
2397         while(sectors) {
2398                 int s = sectors;
2399                 int sl = r10_bio->read_slot;
2400                 int success = 0;
2401                 int start;
2402
2403                 if (s > (PAGE_SIZE>>9))
2404                         s = PAGE_SIZE >> 9;
2405
2406                 rcu_read_lock();
2407                 do {
2408                         sector_t first_bad;
2409                         int bad_sectors;
2410
2411                         d = r10_bio->devs[sl].devnum;
2412                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2413                         if (rdev &&
2414                             test_bit(In_sync, &rdev->flags) &&
2415                             !test_bit(Faulty, &rdev->flags) &&
2416                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2417                                         &first_bad, &bad_sectors) == 0) {
2418                                 atomic_inc(&rdev->nr_pending);
2419                                 rcu_read_unlock();
2420                                 success = sync_page_io(rdev,
2421                                                        r10_bio->devs[sl].addr +
2422                                                        sect,
2423                                                        s<<9,
2424                                                        conf->tmppage,
2425                                                        REQ_OP_READ, 0, false);
2426                                 rdev_dec_pending(rdev, mddev);
2427                                 rcu_read_lock();
2428                                 if (success)
2429                                         break;
2430                         }
2431                         sl++;
2432                         if (sl == conf->copies)
2433                                 sl = 0;
2434                 } while (!success && sl != r10_bio->read_slot);
2435                 rcu_read_unlock();
2436
2437                 if (!success) {
2438                         /* Cannot read from anywhere, just mark the block
2439                          * as bad on the first device to discourage future
2440                          * reads.
2441                          */
2442                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2443                         rdev = conf->mirrors[dn].rdev;
2444
2445                         if (!rdev_set_badblocks(
2446                                     rdev,
2447                                     r10_bio->devs[r10_bio->read_slot].addr
2448                                     + sect,
2449                                     s, 0)) {
2450                                 md_error(mddev, rdev);
2451                                 r10_bio->devs[r10_bio->read_slot].bio
2452                                         = IO_BLOCKED;
2453                         }
2454                         break;
2455                 }
2456
2457                 start = sl;
2458                 /* write it back and re-read */
2459                 rcu_read_lock();
2460                 while (sl != r10_bio->read_slot) {
2461                         char b[BDEVNAME_SIZE];
2462
2463                         if (sl==0)
2464                                 sl = conf->copies;
2465                         sl--;
2466                         d = r10_bio->devs[sl].devnum;
2467                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2468                         if (!rdev ||
2469                             test_bit(Faulty, &rdev->flags) ||
2470                             !test_bit(In_sync, &rdev->flags))
2471                                 continue;
2472
2473                         atomic_inc(&rdev->nr_pending);
2474                         rcu_read_unlock();
2475                         if (r10_sync_page_io(rdev,
2476                                              r10_bio->devs[sl].addr +
2477                                              sect,
2478                                              s, conf->tmppage, WRITE)
2479                             == 0) {
2480                                 /* Well, this device is dead */
2481                                 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2482                                           mdname(mddev), s,
2483                                           (unsigned long long)(
2484                                                   sect +
2485                                                   choose_data_offset(r10_bio,
2486                                                                      rdev)),
2487                                           bdevname(rdev->bdev, b));
2488                                 pr_notice("md/raid10:%s: %s: failing drive\n",
2489                                           mdname(mddev),
2490                                           bdevname(rdev->bdev, b));
2491                         }
2492                         rdev_dec_pending(rdev, mddev);
2493                         rcu_read_lock();
2494                 }
2495                 sl = start;
2496                 while (sl != r10_bio->read_slot) {
2497                         char b[BDEVNAME_SIZE];
2498
2499                         if (sl==0)
2500                                 sl = conf->copies;
2501                         sl--;
2502                         d = r10_bio->devs[sl].devnum;
2503                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2504                         if (!rdev ||
2505                             test_bit(Faulty, &rdev->flags) ||
2506                             !test_bit(In_sync, &rdev->flags))
2507                                 continue;
2508
2509                         atomic_inc(&rdev->nr_pending);
2510                         rcu_read_unlock();
2511                         switch (r10_sync_page_io(rdev,
2512                                              r10_bio->devs[sl].addr +
2513                                              sect,
2514                                              s, conf->tmppage,
2515                                                  READ)) {
2516                         case 0:
2517                                 /* Well, this device is dead */
2518                                 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2519                                        mdname(mddev), s,
2520                                        (unsigned long long)(
2521                                                sect +
2522                                                choose_data_offset(r10_bio, rdev)),
2523                                        bdevname(rdev->bdev, b));
2524                                 pr_notice("md/raid10:%s: %s: failing drive\n",
2525                                        mdname(mddev),
2526                                        bdevname(rdev->bdev, b));
2527                                 break;
2528                         case 1:
2529                                 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2530                                        mdname(mddev), s,
2531                                        (unsigned long long)(
2532                                                sect +
2533                                                choose_data_offset(r10_bio, rdev)),
2534                                        bdevname(rdev->bdev, b));
2535                                 atomic_add(s, &rdev->corrected_errors);
2536                         }
2537
2538                         rdev_dec_pending(rdev, mddev);
2539                         rcu_read_lock();
2540                 }
2541                 rcu_read_unlock();
2542
2543                 sectors -= s;
2544                 sect += s;
2545         }
2546 }
2547
2548 static int narrow_write_error(struct r10bio *r10_bio, int i)
2549 {
2550         struct bio *bio = r10_bio->master_bio;
2551         struct mddev *mddev = r10_bio->mddev;
2552         struct r10conf *conf = mddev->private;
2553         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2554         /* bio has the data to be written to slot 'i' where
2555          * we just recently had a write error.
2556          * We repeatedly clone the bio and trim down to one block,
2557          * then try the write.  Where the write fails we record
2558          * a bad block.
2559          * It is conceivable that the bio doesn't exactly align with
2560          * blocks.  We must handle this.
2561          *
2562          * We currently own a reference to the rdev.
2563          */
2564
2565         int block_sectors;
2566         sector_t sector;
2567         int sectors;
2568         int sect_to_write = r10_bio->sectors;
2569         int ok = 1;
2570
2571         if (rdev->badblocks.shift < 0)
2572                 return 0;
2573
2574         block_sectors = roundup(1 << rdev->badblocks.shift,
2575                                 bdev_logical_block_size(rdev->bdev) >> 9);
2576         sector = r10_bio->sector;
2577         sectors = ((r10_bio->sector + block_sectors)
2578                    & ~(sector_t)(block_sectors - 1))
2579                 - sector;
2580
2581         while (sect_to_write) {
2582                 struct bio *wbio;
2583                 sector_t wsector;
2584                 if (sectors > sect_to_write)
2585                         sectors = sect_to_write;
2586                 /* Write at 'sector' for 'sectors' */
2587                 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2588                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2589                 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2590                 wbio->bi_iter.bi_sector = wsector +
2591                                    choose_data_offset(r10_bio, rdev);
2592                 bio_set_dev(wbio, rdev->bdev);
2593                 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2594
2595                 if (submit_bio_wait(wbio) < 0)
2596                         /* Failure! */
2597                         ok = rdev_set_badblocks(rdev, wsector,
2598                                                 sectors, 0)
2599                                 && ok;
2600
2601                 bio_put(wbio);
2602                 sect_to_write -= sectors;
2603                 sector += sectors;
2604                 sectors = block_sectors;
2605         }
2606         return ok;
2607 }
2608
2609 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2610 {
2611         int slot = r10_bio->read_slot;
2612         struct bio *bio;
2613         struct r10conf *conf = mddev->private;
2614         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2615
2616         /* we got a read error. Maybe the drive is bad.  Maybe just
2617          * the block and we can fix it.
2618          * We freeze all other IO, and try reading the block from
2619          * other devices.  When we find one, we re-write
2620          * and check it that fixes the read error.
2621          * This is all done synchronously while the array is
2622          * frozen.
2623          */
2624         bio = r10_bio->devs[slot].bio;
2625         bio_put(bio);
2626         r10_bio->devs[slot].bio = NULL;
2627
2628         if (mddev->ro)
2629                 r10_bio->devs[slot].bio = IO_BLOCKED;
2630         else if (!test_bit(FailFast, &rdev->flags)) {
2631                 freeze_array(conf, 1);
2632                 fix_read_error(conf, mddev, r10_bio);
2633                 unfreeze_array(conf);
2634         } else
2635                 md_error(mddev, rdev);
2636
2637         rdev_dec_pending(rdev, mddev);
2638         allow_barrier(conf);
2639         r10_bio->state = 0;
2640         raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2641 }
2642
2643 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2644 {
2645         /* Some sort of write request has finished and it
2646          * succeeded in writing where we thought there was a
2647          * bad block.  So forget the bad block.
2648          * Or possibly if failed and we need to record
2649          * a bad block.
2650          */
2651         int m;
2652         struct md_rdev *rdev;
2653
2654         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2655             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2656                 for (m = 0; m < conf->copies; m++) {
2657                         int dev = r10_bio->devs[m].devnum;
2658                         rdev = conf->mirrors[dev].rdev;
2659                         if (r10_bio->devs[m].bio == NULL ||
2660                                 r10_bio->devs[m].bio->bi_end_io == NULL)
2661                                 continue;
2662                         if (!r10_bio->devs[m].bio->bi_status) {
2663                                 rdev_clear_badblocks(
2664                                         rdev,
2665                                         r10_bio->devs[m].addr,
2666                                         r10_bio->sectors, 0);
2667                         } else {
2668                                 if (!rdev_set_badblocks(
2669                                             rdev,
2670                                             r10_bio->devs[m].addr,
2671                                             r10_bio->sectors, 0))
2672                                         md_error(conf->mddev, rdev);
2673                         }
2674                         rdev = conf->mirrors[dev].replacement;
2675                         if (r10_bio->devs[m].repl_bio == NULL ||
2676                                 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2677                                 continue;
2678
2679                         if (!r10_bio->devs[m].repl_bio->bi_status) {
2680                                 rdev_clear_badblocks(
2681                                         rdev,
2682                                         r10_bio->devs[m].addr,
2683                                         r10_bio->sectors, 0);
2684                         } else {
2685                                 if (!rdev_set_badblocks(
2686                                             rdev,
2687                                             r10_bio->devs[m].addr,
2688                                             r10_bio->sectors, 0))
2689                                         md_error(conf->mddev, rdev);
2690                         }
2691                 }
2692                 put_buf(r10_bio);
2693         } else {
2694                 bool fail = false;
2695                 for (m = 0; m < conf->copies; m++) {
2696                         int dev = r10_bio->devs[m].devnum;
2697                         struct bio *bio = r10_bio->devs[m].bio;
2698                         rdev = conf->mirrors[dev].rdev;
2699                         if (bio == IO_MADE_GOOD) {
2700                                 rdev_clear_badblocks(
2701                                         rdev,
2702                                         r10_bio->devs[m].addr,
2703                                         r10_bio->sectors, 0);
2704                                 rdev_dec_pending(rdev, conf->mddev);
2705                         } else if (bio != NULL && bio->bi_status) {
2706                                 fail = true;
2707                                 if (!narrow_write_error(r10_bio, m)) {
2708                                         md_error(conf->mddev, rdev);
2709                                         set_bit(R10BIO_Degraded,
2710                                                 &r10_bio->state);
2711                                 }
2712                                 rdev_dec_pending(rdev, conf->mddev);
2713                         }
2714                         bio = r10_bio->devs[m].repl_bio;
2715                         rdev = conf->mirrors[dev].replacement;
2716                         if (rdev && bio == IO_MADE_GOOD) {
2717                                 rdev_clear_badblocks(
2718                                         rdev,
2719                                         r10_bio->devs[m].addr,
2720                                         r10_bio->sectors, 0);
2721                                 rdev_dec_pending(rdev, conf->mddev);
2722                         }
2723                 }
2724                 if (fail) {
2725                         spin_lock_irq(&conf->device_lock);
2726                         list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2727                         conf->nr_queued++;
2728                         spin_unlock_irq(&conf->device_lock);
2729                         /*
2730                          * In case freeze_array() is waiting for condition
2731                          * nr_pending == nr_queued + extra to be true.
2732                          */
2733                         wake_up(&conf->wait_barrier);
2734                         md_wakeup_thread(conf->mddev->thread);
2735                 } else {
2736                         if (test_bit(R10BIO_WriteError,
2737                                      &r10_bio->state))
2738                                 close_write(r10_bio);
2739                         raid_end_bio_io(r10_bio);
2740                 }
2741         }
2742 }
2743
2744 static void raid10d(struct md_thread *thread)
2745 {
2746         struct mddev *mddev = thread->mddev;
2747         struct r10bio *r10_bio;
2748         unsigned long flags;
2749         struct r10conf *conf = mddev->private;
2750         struct list_head *head = &conf->retry_list;
2751         struct blk_plug plug;
2752
2753         md_check_recovery(mddev);
2754
2755         if (!list_empty_careful(&conf->bio_end_io_list) &&
2756             !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2757                 LIST_HEAD(tmp);
2758                 spin_lock_irqsave(&conf->device_lock, flags);
2759                 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2760                         while (!list_empty(&conf->bio_end_io_list)) {
2761                                 list_move(conf->bio_end_io_list.prev, &tmp);
2762                                 conf->nr_queued--;
2763                         }
2764                 }
2765                 spin_unlock_irqrestore(&conf->device_lock, flags);
2766                 while (!list_empty(&tmp)) {
2767                         r10_bio = list_first_entry(&tmp, struct r10bio,
2768                                                    retry_list);
2769                         list_del(&r10_bio->retry_list);
2770                         if (mddev->degraded)
2771                                 set_bit(R10BIO_Degraded, &r10_bio->state);
2772
2773                         if (test_bit(R10BIO_WriteError,
2774                                      &r10_bio->state))
2775                                 close_write(r10_bio);
2776                         raid_end_bio_io(r10_bio);
2777                 }
2778         }
2779
2780         blk_start_plug(&plug);
2781         for (;;) {
2782
2783                 flush_pending_writes(conf);
2784
2785                 spin_lock_irqsave(&conf->device_lock, flags);
2786                 if (list_empty(head)) {
2787                         spin_unlock_irqrestore(&conf->device_lock, flags);
2788                         break;
2789                 }
2790                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2791                 list_del(head->prev);
2792                 conf->nr_queued--;
2793                 spin_unlock_irqrestore(&conf->device_lock, flags);
2794
2795                 mddev = r10_bio->mddev;
2796                 conf = mddev->private;
2797                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2798                     test_bit(R10BIO_WriteError, &r10_bio->state))
2799                         handle_write_completed(conf, r10_bio);
2800                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2801                         reshape_request_write(mddev, r10_bio);
2802                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2803                         sync_request_write(mddev, r10_bio);
2804                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2805                         recovery_request_write(mddev, r10_bio);
2806                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2807                         handle_read_error(mddev, r10_bio);
2808                 else
2809                         WARN_ON_ONCE(1);
2810
2811                 cond_resched();
2812                 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2813                         md_check_recovery(mddev);
2814         }
2815         blk_finish_plug(&plug);
2816 }
2817
2818 static int init_resync(struct r10conf *conf)
2819 {
2820         int ret, buffs, i;
2821
2822         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2823         BUG_ON(mempool_initialized(&conf->r10buf_pool));
2824         conf->have_replacement = 0;
2825         for (i = 0; i < conf->geo.raid_disks; i++)
2826                 if (conf->mirrors[i].replacement)
2827                         conf->have_replacement = 1;
2828         ret = mempool_init(&conf->r10buf_pool, buffs,
2829                            r10buf_pool_alloc, r10buf_pool_free, conf);
2830         if (ret)
2831                 return ret;
2832         conf->next_resync = 0;
2833         return 0;
2834 }
2835
2836 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2837 {
2838         struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2839         struct rsync_pages *rp;
2840         struct bio *bio;
2841         int nalloc;
2842         int i;
2843
2844         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2845             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2846                 nalloc = conf->copies; /* resync */
2847         else
2848                 nalloc = 2; /* recovery */
2849
2850         for (i = 0; i < nalloc; i++) {
2851                 bio = r10bio->devs[i].bio;
2852                 rp = bio->bi_private;
2853                 bio_reset(bio);
2854                 bio->bi_private = rp;
2855                 bio = r10bio->devs[i].repl_bio;
2856                 if (bio) {
2857                         rp = bio->bi_private;
2858                         bio_reset(bio);
2859                         bio->bi_private = rp;
2860                 }
2861         }
2862         return r10bio;
2863 }
2864
2865 /*
2866  * Set cluster_sync_high since we need other nodes to add the
2867  * range [cluster_sync_low, cluster_sync_high] to suspend list.
2868  */
2869 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2870 {
2871         sector_t window_size;
2872         int extra_chunk, chunks;
2873
2874         /*
2875          * First, here we define "stripe" as a unit which across
2876          * all member devices one time, so we get chunks by use
2877          * raid_disks / near_copies. Otherwise, if near_copies is
2878          * close to raid_disks, then resync window could increases
2879          * linearly with the increase of raid_disks, which means
2880          * we will suspend a really large IO window while it is not
2881          * necessary. If raid_disks is not divisible by near_copies,
2882          * an extra chunk is needed to ensure the whole "stripe" is
2883          * covered.
2884          */
2885
2886         chunks = conf->geo.raid_disks / conf->geo.near_copies;
2887         if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2888                 extra_chunk = 0;
2889         else
2890                 extra_chunk = 1;
2891         window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2892
2893         /*
2894          * At least use a 32M window to align with raid1's resync window
2895          */
2896         window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2897                         CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2898
2899         conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2900 }
2901
2902 /*
2903  * perform a "sync" on one "block"
2904  *
2905  * We need to make sure that no normal I/O request - particularly write
2906  * requests - conflict with active sync requests.
2907  *
2908  * This is achieved by tracking pending requests and a 'barrier' concept
2909  * that can be installed to exclude normal IO requests.
2910  *
2911  * Resync and recovery are handled very differently.
2912  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2913  *
2914  * For resync, we iterate over virtual addresses, read all copies,
2915  * and update if there are differences.  If only one copy is live,
2916  * skip it.
2917  * For recovery, we iterate over physical addresses, read a good
2918  * value for each non-in_sync drive, and over-write.
2919  *
2920  * So, for recovery we may have several outstanding complex requests for a
2921  * given address, one for each out-of-sync device.  We model this by allocating
2922  * a number of r10_bio structures, one for each out-of-sync device.
2923  * As we setup these structures, we collect all bio's together into a list
2924  * which we then process collectively to add pages, and then process again
2925  * to pass to generic_make_request.
2926  *
2927  * The r10_bio structures are linked using a borrowed master_bio pointer.
2928  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2929  * has its remaining count decremented to 0, the whole complex operation
2930  * is complete.
2931  *
2932  */
2933
2934 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2935                              int *skipped)
2936 {
2937         struct r10conf *conf = mddev->private;
2938         struct r10bio *r10_bio;
2939         struct bio *biolist = NULL, *bio;
2940         sector_t max_sector, nr_sectors;
2941         int i;
2942         int max_sync;
2943         sector_t sync_blocks;
2944         sector_t sectors_skipped = 0;
2945         int chunks_skipped = 0;
2946         sector_t chunk_mask = conf->geo.chunk_mask;
2947         int page_idx = 0;
2948
2949         if (!mempool_initialized(&conf->r10buf_pool))
2950                 if (init_resync(conf))
2951                         return 0;
2952
2953         /*
2954          * Allow skipping a full rebuild for incremental assembly
2955          * of a clean array, like RAID1 does.
2956          */
2957         if (mddev->bitmap == NULL &&
2958             mddev->recovery_cp == MaxSector &&
2959             mddev->reshape_position == MaxSector &&
2960             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2961             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2962             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2963             conf->fullsync == 0) {
2964                 *skipped = 1;
2965                 return mddev->dev_sectors - sector_nr;
2966         }
2967
2968  skipped:
2969         max_sector = mddev->dev_sectors;
2970         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2971             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2972                 max_sector = mddev->resync_max_sectors;
2973         if (sector_nr >= max_sector) {
2974                 conf->cluster_sync_low = 0;
2975                 conf->cluster_sync_high = 0;
2976
2977                 /* If we aborted, we need to abort the
2978                  * sync on the 'current' bitmap chucks (there can
2979                  * be several when recovering multiple devices).
2980                  * as we may have started syncing it but not finished.
2981                  * We can find the current address in
2982                  * mddev->curr_resync, but for recovery,
2983                  * we need to convert that to several
2984                  * virtual addresses.
2985                  */
2986                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2987                         end_reshape(conf);
2988                         close_sync(conf);
2989                         return 0;
2990                 }
2991
2992                 if (mddev->curr_resync < max_sector) { /* aborted */
2993                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2994                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2995                                                 &sync_blocks, 1);
2996                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2997                                 sector_t sect =
2998                                         raid10_find_virt(conf, mddev->curr_resync, i);
2999                                 bitmap_end_sync(mddev->bitmap, sect,
3000                                                 &sync_blocks, 1);
3001                         }
3002                 } else {
3003                         /* completed sync */
3004                         if ((!mddev->bitmap || conf->fullsync)
3005                             && conf->have_replacement
3006                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3007                                 /* Completed a full sync so the replacements
3008                                  * are now fully recovered.
3009                                  */
3010                                 rcu_read_lock();
3011                                 for (i = 0; i < conf->geo.raid_disks; i++) {
3012                                         struct md_rdev *rdev =
3013                                                 rcu_dereference(conf->mirrors[i].replacement);
3014                                         if (rdev)
3015                                                 rdev->recovery_offset = MaxSector;
3016                                 }
3017                                 rcu_read_unlock();
3018                         }
3019                         conf->fullsync = 0;
3020                 }
3021                 bitmap_close_sync(mddev->bitmap);
3022                 close_sync(conf);
3023                 *skipped = 1;
3024                 return sectors_skipped;
3025         }
3026
3027         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3028                 return reshape_request(mddev, sector_nr, skipped);
3029
3030         if (chunks_skipped >= conf->geo.raid_disks) {
3031                 /* if there has been nothing to do on any drive,
3032                  * then there is nothing to do at all..
3033                  */
3034                 *skipped = 1;
3035                 return (max_sector - sector_nr) + sectors_skipped;
3036         }
3037
3038         if (max_sector > mddev->resync_max)
3039                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3040
3041         /* make sure whole request will fit in a chunk - if chunks
3042          * are meaningful
3043          */
3044         if (conf->geo.near_copies < conf->geo.raid_disks &&
3045             max_sector > (sector_nr | chunk_mask))
3046                 max_sector = (sector_nr | chunk_mask) + 1;
3047
3048         /*
3049          * If there is non-resync activity waiting for a turn, then let it
3050          * though before starting on this new sync request.
3051          */
3052         if (conf->nr_waiting)
3053                 schedule_timeout_uninterruptible(1);
3054
3055         /* Again, very different code for resync and recovery.
3056          * Both must result in an r10bio with a list of bios that
3057          * have bi_end_io, bi_sector, bi_disk set,
3058          * and bi_private set to the r10bio.
3059          * For recovery, we may actually create several r10bios
3060          * with 2 bios in each, that correspond to the bios in the main one.
3061          * In this case, the subordinate r10bios link back through a
3062          * borrowed master_bio pointer, and the counter in the master
3063          * includes a ref from each subordinate.
3064          */
3065         /* First, we decide what to do and set ->bi_end_io
3066          * To end_sync_read if we want to read, and
3067          * end_sync_write if we will want to write.
3068          */
3069
3070         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3071         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3072                 /* recovery... the complicated one */
3073                 int j;
3074                 r10_bio = NULL;
3075
3076                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3077                         int still_degraded;
3078                         struct r10bio *rb2;
3079                         sector_t sect;
3080                         int must_sync;
3081                         int any_working;
3082                         struct raid10_info *mirror = &conf->mirrors[i];
3083                         struct md_rdev *mrdev, *mreplace;
3084
3085                         rcu_read_lock();
3086                         mrdev = rcu_dereference(mirror->rdev);
3087                         mreplace = rcu_dereference(mirror->replacement);
3088
3089                         if ((mrdev == NULL ||
3090                              test_bit(Faulty, &mrdev->flags) ||
3091                              test_bit(In_sync, &mrdev->flags)) &&
3092                             (mreplace == NULL ||
3093                              test_bit(Faulty, &mreplace->flags))) {
3094                                 rcu_read_unlock();
3095                                 continue;
3096                         }
3097
3098                         still_degraded = 0;
3099                         /* want to reconstruct this device */
3100                         rb2 = r10_bio;
3101                         sect = raid10_find_virt(conf, sector_nr, i);
3102                         if (sect >= mddev->resync_max_sectors) {
3103                                 /* last stripe is not complete - don't
3104                                  * try to recover this sector.
3105                                  */
3106                                 rcu_read_unlock();
3107                                 continue;
3108                         }
3109                         if (mreplace && test_bit(Faulty, &mreplace->flags))
3110                                 mreplace = NULL;
3111                         /* Unless we are doing a full sync, or a replacement
3112                          * we only need to recover the block if it is set in
3113                          * the bitmap
3114                          */
3115                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
3116                                                       &sync_blocks, 1);
3117                         if (sync_blocks < max_sync)
3118                                 max_sync = sync_blocks;
3119                         if (!must_sync &&
3120                             mreplace == NULL &&
3121                             !conf->fullsync) {
3122                                 /* yep, skip the sync_blocks here, but don't assume
3123                                  * that there will never be anything to do here
3124                                  */
3125                                 chunks_skipped = -1;
3126                                 rcu_read_unlock();
3127                                 continue;
3128                         }
3129                         atomic_inc(&mrdev->nr_pending);
3130                         if (mreplace)
3131                                 atomic_inc(&mreplace->nr_pending);
3132                         rcu_read_unlock();
3133
3134                         r10_bio = raid10_alloc_init_r10buf(conf);
3135                         r10_bio->state = 0;
3136                         raise_barrier(conf, rb2 != NULL);
3137                         atomic_set(&r10_bio->remaining, 0);
3138
3139                         r10_bio->master_bio = (struct bio*)rb2;
3140                         if (rb2)
3141                                 atomic_inc(&rb2->remaining);
3142                         r10_bio->mddev = mddev;
3143                         set_bit(R10BIO_IsRecover, &r10_bio->state);
3144                         r10_bio->sector = sect;
3145
3146                         raid10_find_phys(conf, r10_bio);
3147
3148                         /* Need to check if the array will still be
3149                          * degraded
3150                          */
3151                         rcu_read_lock();
3152                         for (j = 0; j < conf->geo.raid_disks; j++) {
3153                                 struct md_rdev *rdev = rcu_dereference(
3154                                         conf->mirrors[j].rdev);
3155                                 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3156                                         still_degraded = 1;
3157                                         break;
3158                                 }
3159                         }
3160
3161                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
3162                                                       &sync_blocks, still_degraded);
3163
3164                         any_working = 0;
3165                         for (j=0; j<conf->copies;j++) {
3166                                 int k;
3167                                 int d = r10_bio->devs[j].devnum;
3168                                 sector_t from_addr, to_addr;
3169                                 struct md_rdev *rdev =
3170                                         rcu_dereference(conf->mirrors[d].rdev);
3171                                 sector_t sector, first_bad;
3172                                 int bad_sectors;
3173                                 if (!rdev ||
3174                                     !test_bit(In_sync, &rdev->flags))
3175                                         continue;
3176                                 /* This is where we read from */
3177                                 any_working = 1;
3178                                 sector = r10_bio->devs[j].addr;
3179
3180                                 if (is_badblock(rdev, sector, max_sync,
3181                                                 &first_bad, &bad_sectors)) {
3182                                         if (first_bad > sector)
3183                                                 max_sync = first_bad - sector;
3184                                         else {
3185                                                 bad_sectors -= (sector
3186                                                                 - first_bad);
3187                                                 if (max_sync > bad_sectors)
3188                                                         max_sync = bad_sectors;
3189                                                 continue;
3190                                         }
3191                                 }
3192                                 bio = r10_bio->devs[0].bio;
3193                                 bio->bi_next = biolist;
3194                                 biolist = bio;
3195                                 bio->bi_end_io = end_sync_read;
3196                                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3197                                 if (test_bit(FailFast, &rdev->flags))
3198                                         bio->bi_opf |= MD_FAILFAST;
3199                                 from_addr = r10_bio->devs[j].addr;
3200                                 bio->bi_iter.bi_sector = from_addr +
3201                                         rdev->data_offset;
3202                                 bio_set_dev(bio, rdev->bdev);
3203                                 atomic_inc(&rdev->nr_pending);
3204                                 /* and we write to 'i' (if not in_sync) */
3205
3206                                 for (k=0; k<conf->copies; k++)
3207                                         if (r10_bio->devs[k].devnum == i)
3208                                                 break;
3209                                 BUG_ON(k == conf->copies);
3210                                 to_addr = r10_bio->devs[k].addr;
3211                                 r10_bio->devs[0].devnum = d;
3212                                 r10_bio->devs[0].addr = from_addr;
3213                                 r10_bio->devs[1].devnum = i;
3214                                 r10_bio->devs[1].addr = to_addr;
3215
3216                                 if (!test_bit(In_sync, &mrdev->flags)) {
3217                                         bio = r10_bio->devs[1].bio;
3218                                         bio->bi_next = biolist;
3219                                         biolist = bio;
3220                                         bio->bi_end_io = end_sync_write;
3221                                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3222                                         bio->bi_iter.bi_sector = to_addr
3223                                                 + mrdev->data_offset;
3224                                         bio_set_dev(bio, mrdev->bdev);
3225                                         atomic_inc(&r10_bio->remaining);
3226                                 } else
3227                                         r10_bio->devs[1].bio->bi_end_io = NULL;
3228
3229                                 /* and maybe write to replacement */
3230                                 bio = r10_bio->devs[1].repl_bio;
3231                                 if (bio)
3232                                         bio->bi_end_io = NULL;
3233                                 /* Note: if mreplace != NULL, then bio
3234                                  * cannot be NULL as r10buf_pool_alloc will
3235                                  * have allocated it.
3236                                  * So the second test here is pointless.
3237                                  * But it keeps semantic-checkers happy, and
3238                                  * this comment keeps human reviewers
3239                                  * happy.
3240                                  */
3241                                 if (mreplace == NULL || bio == NULL ||
3242                                     test_bit(Faulty, &mreplace->flags))
3243                                         break;
3244                                 bio->bi_next = biolist;
3245                                 biolist = bio;
3246                                 bio->bi_end_io = end_sync_write;
3247                                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3248                                 bio->bi_iter.bi_sector = to_addr +
3249                                         mreplace->data_offset;
3250                                 bio_set_dev(bio, mreplace->bdev);
3251                                 atomic_inc(&r10_bio->remaining);
3252                                 break;
3253                         }
3254                         rcu_read_unlock();
3255                         if (j == conf->copies) {
3256                                 /* Cannot recover, so abort the recovery or
3257                                  * record a bad block */
3258                                 if (any_working) {
3259                                         /* problem is that there are bad blocks
3260                                          * on other device(s)
3261                                          */
3262                                         int k;
3263                                         for (k = 0; k < conf->copies; k++)
3264                                                 if (r10_bio->devs[k].devnum == i)
3265                                                         break;
3266                                         if (!test_bit(In_sync,
3267                                                       &mrdev->flags)
3268                                             && !rdev_set_badblocks(
3269                                                     mrdev,
3270                                                     r10_bio->devs[k].addr,
3271                                                     max_sync, 0))
3272                                                 any_working = 0;
3273                                         if (mreplace &&
3274                                             !rdev_set_badblocks(
3275                                                     mreplace,
3276                                                     r10_bio->devs[k].addr,
3277                                                     max_sync, 0))
3278                                                 any_working = 0;
3279                                 }
3280                                 if (!any_working)  {
3281                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
3282                                                               &mddev->recovery))
3283                                                 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3284                                                        mdname(mddev));
3285                                         mirror->recovery_disabled
3286                                                 = mddev->recovery_disabled;
3287                                 }
3288                                 put_buf(r10_bio);
3289                                 if (rb2)
3290                                         atomic_dec(&rb2->remaining);
3291                                 r10_bio = rb2;
3292                                 rdev_dec_pending(mrdev, mddev);
3293                                 if (mreplace)
3294                                         rdev_dec_pending(mreplace, mddev);
3295                                 break;
3296                         }
3297                         rdev_dec_pending(mrdev, mddev);
3298                         if (mreplace)
3299                                 rdev_dec_pending(mreplace, mddev);
3300                         if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3301                                 /* Only want this if there is elsewhere to
3302                                  * read from. 'j' is currently the first
3303                                  * readable copy.
3304                                  */
3305                                 int targets = 1;
3306                                 for (; j < conf->copies; j++) {
3307                                         int d = r10_bio->devs[j].devnum;
3308                                         if (conf->mirrors[d].rdev &&
3309                                             test_bit(In_sync,
3310                                                       &conf->mirrors[d].rdev->flags))
3311                                                 targets++;
3312                                 }
3313                                 if (targets == 1)
3314                                         r10_bio->devs[0].bio->bi_opf
3315                                                 &= ~MD_FAILFAST;
3316                         }
3317                 }
3318                 if (biolist == NULL) {
3319                         while (r10_bio) {
3320                                 struct r10bio *rb2 = r10_bio;
3321                                 r10_bio = (struct r10bio*) rb2->master_bio;
3322                                 rb2->master_bio = NULL;
3323                                 put_buf(rb2);
3324                         }
3325                         goto giveup;
3326                 }
3327         } else {
3328                 /* resync. Schedule a read for every block at this virt offset */
3329                 int count = 0;
3330
3331                 /*
3332                  * Since curr_resync_completed could probably not update in
3333                  * time, and we will set cluster_sync_low based on it.
3334                  * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3335                  * safety reason, which ensures curr_resync_completed is
3336                  * updated in bitmap_cond_end_sync.
3337                  */
3338                 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3339                                      mddev_is_clustered(mddev) &&
3340                                      (sector_nr + 2 * RESYNC_SECTORS >
3341                                       conf->cluster_sync_high));
3342
3343                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3344                                        &sync_blocks, mddev->degraded) &&
3345                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3346                                                  &mddev->recovery)) {
3347                         /* We can skip this block */
3348                         *skipped = 1;
3349                         return sync_blocks + sectors_skipped;
3350                 }
3351                 if (sync_blocks < max_sync)
3352                         max_sync = sync_blocks;
3353                 r10_bio = raid10_alloc_init_r10buf(conf);
3354                 r10_bio->state = 0;
3355
3356                 r10_bio->mddev = mddev;
3357                 atomic_set(&r10_bio->remaining, 0);
3358                 raise_barrier(conf, 0);
3359                 conf->next_resync = sector_nr;
3360
3361                 r10_bio->master_bio = NULL;
3362                 r10_bio->sector = sector_nr;
3363                 set_bit(R10BIO_IsSync, &r10_bio->state);
3364                 raid10_find_phys(conf, r10_bio);
3365                 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3366
3367                 for (i = 0; i < conf->copies; i++) {
3368                         int d = r10_bio->devs[i].devnum;
3369                         sector_t first_bad, sector;
3370                         int bad_sectors;
3371                         struct md_rdev *rdev;
3372
3373                         if (r10_bio->devs[i].repl_bio)
3374                                 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3375
3376                         bio = r10_bio->devs[i].bio;
3377                         bio->bi_status = BLK_STS_IOERR;
3378                         rcu_read_lock();
3379                         rdev = rcu_dereference(conf->mirrors[d].rdev);
3380                         if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3381                                 rcu_read_unlock();
3382                                 continue;
3383                         }
3384                         sector = r10_bio->devs[i].addr;
3385                         if (is_badblock(rdev, sector, max_sync,
3386                                         &first_bad, &bad_sectors)) {
3387                                 if (first_bad > sector)
3388                                         max_sync = first_bad - sector;
3389                                 else {
3390                                         bad_sectors -= (sector - first_bad);
3391                                         if (max_sync > bad_sectors)
3392                                                 max_sync = bad_sectors;
3393                                         rcu_read_unlock();
3394                                         continue;
3395                                 }
3396                         }
3397                         atomic_inc(&rdev->nr_pending);
3398                         atomic_inc(&r10_bio->remaining);
3399                         bio->bi_next = biolist;
3400                         biolist = bio;
3401                         bio->bi_end_io = end_sync_read;
3402                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
3403                         if (test_bit(FailFast, &rdev->flags))
3404                                 bio->bi_opf |= MD_FAILFAST;
3405                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3406                         bio_set_dev(bio, rdev->bdev);
3407                         count++;
3408
3409                         rdev = rcu_dereference(conf->mirrors[d].replacement);
3410                         if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3411                                 rcu_read_unlock();
3412                                 continue;
3413                         }
3414                         atomic_inc(&rdev->nr_pending);
3415
3416                         /* Need to set up for writing to the replacement */
3417                         bio = r10_bio->devs[i].repl_bio;
3418                         bio->bi_status = BLK_STS_IOERR;
3419
3420                         sector = r10_bio->devs[i].addr;
3421                         bio->bi_next = biolist;
3422                         biolist = bio;
3423                         bio->bi_end_io = end_sync_write;
3424                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3425                         if (test_bit(FailFast, &rdev->flags))
3426                                 bio->bi_opf |= MD_FAILFAST;
3427                         bio->bi_iter.bi_sector = sector + rdev->data_offset;
3428                         bio_set_dev(bio, rdev->bdev);
3429                         count++;
3430                         rcu_read_unlock();
3431                 }
3432
3433                 if (count < 2) {
3434                         for (i=0; i<conf->copies; i++) {
3435                                 int d = r10_bio->devs[i].devnum;
3436                                 if (r10_bio->devs[i].bio->bi_end_io)
3437                                         rdev_dec_pending(conf->mirrors[d].rdev,
3438                                                          mddev);
3439                                 if (r10_bio->devs[i].repl_bio &&
3440                                     r10_bio->devs[i].repl_bio->bi_end_io)
3441                                         rdev_dec_pending(
3442                                                 conf->mirrors[d].replacement,
3443                                                 mddev);
3444                         }
3445                         put_buf(r10_bio);
3446                         biolist = NULL;
3447                         goto giveup;
3448                 }
3449         }
3450
3451         nr_sectors = 0;
3452         if (sector_nr + max_sync < max_sector)
3453                 max_sector = sector_nr + max_sync;
3454         do {
3455                 struct page *page;
3456                 int len = PAGE_SIZE;
3457                 if (sector_nr + (len>>9) > max_sector)
3458                         len = (max_sector - sector_nr) << 9;
3459                 if (len == 0)
3460                         break;
3461                 for (bio= biolist ; bio ; bio=bio->bi_next) {
3462                         struct resync_pages *rp = get_resync_pages(bio);
3463                         page = resync_fetch_page(rp, page_idx);
3464                         /*
3465                          * won't fail because the vec table is big enough
3466                          * to hold all these pages
3467                          */
3468                         bio_add_page(bio, page, len, 0);
3469                 }
3470                 nr_sectors += len>>9;
3471                 sector_nr += len>>9;
3472         } while (++page_idx < RESYNC_PAGES);
3473         r10_bio->sectors = nr_sectors;
3474
3475         if (mddev_is_clustered(mddev) &&
3476             test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3477                 /* It is resync not recovery */
3478                 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3479                         conf->cluster_sync_low = mddev->curr_resync_completed;
3480                         raid10_set_cluster_sync_high(conf);
3481                         /* Send resync message */
3482                         md_cluster_ops->resync_info_update(mddev,
3483                                                 conf->cluster_sync_low,
3484                                                 conf->cluster_sync_high);
3485                 }
3486         } else if (mddev_is_clustered(mddev)) {
3487                 /* This is recovery not resync */
3488                 sector_t sect_va1, sect_va2;
3489                 bool broadcast_msg = false;
3490
3491                 for (i = 0; i < conf->geo.raid_disks; i++) {
3492                         /*
3493                          * sector_nr is a device address for recovery, so we
3494                          * need translate it to array address before compare
3495                          * with cluster_sync_high.
3496                          */
3497                         sect_va1 = raid10_find_virt(conf, sector_nr, i);
3498
3499                         if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3500                                 broadcast_msg = true;
3501                                 /*
3502                                  * curr_resync_completed is similar as
3503                                  * sector_nr, so make the translation too.
3504                                  */
3505                                 sect_va2 = raid10_find_virt(conf,
3506                                         mddev->curr_resync_completed, i);
3507
3508                                 if (conf->cluster_sync_low == 0 ||
3509                                     conf->cluster_sync_low > sect_va2)
3510                                         conf->cluster_sync_low = sect_va2;
3511                         }
3512                 }
3513                 if (broadcast_msg) {
3514                         raid10_set_cluster_sync_high(conf);
3515                         md_cluster_ops->resync_info_update(mddev,
3516                                                 conf->cluster_sync_low,
3517                                                 conf->cluster_sync_high);
3518                 }
3519         }
3520
3521         while (biolist) {
3522                 bio = biolist;
3523                 biolist = biolist->bi_next;
3524
3525                 bio->bi_next = NULL;
3526                 r10_bio = get_resync_r10bio(bio);
3527                 r10_bio->sectors = nr_sectors;
3528
3529                 if (bio->bi_end_io == end_sync_read) {
3530                         md_sync_acct_bio(bio, nr_sectors);
3531                         bio->bi_status = 0;
3532                         generic_make_request(bio);
3533                 }
3534         }
3535
3536         if (sectors_skipped)
3537                 /* pretend they weren't skipped, it makes
3538                  * no important difference in this case
3539                  */
3540                 md_done_sync(mddev, sectors_skipped, 1);
3541
3542         return sectors_skipped + nr_sectors;
3543  giveup:
3544         /* There is nowhere to write, so all non-sync
3545          * drives must be failed or in resync, all drives
3546          * have a bad block, so try the next chunk...
3547          */
3548         if (sector_nr + max_sync < max_sector)
3549                 max_sector = sector_nr + max_sync;
3550
3551         sectors_skipped += (max_sector - sector_nr);
3552         chunks_skipped ++;
3553         sector_nr = max_sector;
3554         goto skipped;
3555 }
3556
3557 static sector_t
3558 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3559 {
3560         sector_t size;
3561         struct r10conf *conf = mddev->private;
3562
3563         if (!raid_disks)
3564                 raid_disks = min(conf->geo.raid_disks,
3565                                  conf->prev.raid_disks);
3566         if (!sectors)
3567                 sectors = conf->dev_sectors;
3568
3569         size = sectors >> conf->geo.chunk_shift;
3570         sector_div(size, conf->geo.far_copies);
3571         size = size * raid_disks;
3572         sector_div(size, conf->geo.near_copies);
3573
3574         return size << conf->geo.chunk_shift;
3575 }
3576
3577 static void calc_sectors(struct r10conf *conf, sector_t size)
3578 {
3579         /* Calculate the number of sectors-per-device that will
3580          * actually be used, and set conf->dev_sectors and
3581          * conf->stride
3582          */
3583
3584         size = size >> conf->geo.chunk_shift;
3585         sector_div(size, conf->geo.far_copies);
3586         size = size * conf->geo.raid_disks;
3587         sector_div(size, conf->geo.near_copies);
3588         /* 'size' is now the number of chunks in the array */
3589         /* calculate "used chunks per device" */
3590         size = size * conf->copies;
3591
3592         /* We need to round up when dividing by raid_disks to
3593          * get the stride size.
3594          */
3595         size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3596
3597         conf->dev_sectors = size << conf->geo.chunk_shift;
3598
3599         if (conf->geo.far_offset)
3600                 conf->geo.stride = 1 << conf->geo.chunk_shift;
3601         else {
3602                 sector_div(size, conf->geo.far_copies);
3603                 conf->geo.stride = size << conf->geo.chunk_shift;
3604         }
3605 }
3606
3607 enum geo_type {geo_new, geo_old, geo_start};
3608 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3609 {
3610         int nc, fc, fo;
3611         int layout, chunk, disks;
3612         switch (new) {
3613         case geo_old:
3614                 layout = mddev->layout;
3615                 chunk = mddev->chunk_sectors;
3616                 disks = mddev->raid_disks - mddev->delta_disks;
3617                 break;
3618         case geo_new:
3619                 layout = mddev->new_layout;
3620                 chunk = mddev->new_chunk_sectors;
3621                 disks = mddev->raid_disks;
3622                 break;
3623         default: /* avoid 'may be unused' warnings */
3624         case geo_start: /* new when starting reshape - raid_disks not
3625                          * updated yet. */
3626                 layout = mddev->new_layout;
3627                 chunk = mddev->new_chunk_sectors;
3628                 disks = mddev->raid_disks + mddev->delta_disks;
3629                 break;
3630         }
3631         if (layout >> 19)
3632                 return -1;
3633         if (chunk < (PAGE_SIZE >> 9) ||
3634             !is_power_of_2(chunk))
3635                 return -2;
3636         nc = layout & 255;
3637         fc = (layout >> 8) & 255;
3638         fo = layout & (1<<16);
3639         geo->raid_disks = disks;
3640         geo->near_copies = nc;
3641         geo->far_copies = fc;
3642         geo->far_offset = fo;
3643         switch (layout >> 17) {
3644         case 0: /* original layout.  simple but not always optimal */
3645                 geo->far_set_size = disks;
3646                 break;
3647         case 1: /* "improved" layout which was buggy.  Hopefully no-one is
3648                  * actually using this, but leave code here just in case.*/
3649                 geo->far_set_size = disks/fc;
3650                 WARN(geo->far_set_size < fc,
3651                      "This RAID10 layout does not provide data safety - please backup and create new array\n");
3652                 break;
3653         case 2: /* "improved" layout fixed to match documentation */
3654                 geo->far_set_size = fc * nc;
3655                 break;
3656         default: /* Not a valid layout */
3657                 return -1;
3658         }
3659         geo->chunk_mask = chunk - 1;
3660         geo->chunk_shift = ffz(~chunk);
3661         return nc*fc;
3662 }
3663
3664 static struct r10conf *setup_conf(struct mddev *mddev)
3665 {
3666         struct r10conf *conf = NULL;
3667         int err = -EINVAL;
3668         struct geom geo;
3669         int copies;
3670
3671         copies = setup_geo(&geo, mddev, geo_new);
3672
3673         if (copies == -2) {
3674                 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3675                         mdname(mddev), PAGE_SIZE);
3676                 goto out;
3677         }
3678
3679         if (copies < 2 || copies > mddev->raid_disks) {
3680                 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3681                         mdname(mddev), mddev->new_layout);
3682                 goto out;
3683         }
3684
3685         err = -ENOMEM;
3686         conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3687         if (!conf)
3688                 goto out;
3689
3690         /* FIXME calc properly */
3691         conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3692                                 sizeof(struct raid10_info),
3693                                 GFP_KERNEL);
3694         if (!conf->mirrors)
3695                 goto out;
3696
3697         conf->tmppage = alloc_page(GFP_KERNEL);
3698         if (!conf->tmppage)
3699                 goto out;
3700
3701         conf->geo = geo;
3702         conf->copies = copies;
3703         err = mempool_init(&conf->r10bio_pool, NR_RAID10_BIOS, r10bio_pool_alloc,
3704                            r10bio_pool_free, conf);
3705         if (err)
3706                 goto out;
3707
3708         err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3709         if (err)
3710                 goto out;
3711
3712         calc_sectors(conf, mddev->dev_sectors);
3713         if (mddev->reshape_position == MaxSector) {
3714                 conf->prev = conf->geo;
3715                 conf->reshape_progress = MaxSector;
3716         } else {
3717                 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3718                         err = -EINVAL;
3719                         goto out;
3720                 }
3721                 conf->reshape_progress = mddev->reshape_position;
3722                 if (conf->prev.far_offset)
3723                         conf->prev.stride = 1 << conf->prev.chunk_shift;
3724                 else
3725                         /* far_copies must be 1 */
3726                         conf->prev.stride = conf->dev_sectors;
3727         }
3728         conf->reshape_safe = conf->reshape_progress;
3729         spin_lock_init(&conf->device_lock);
3730         INIT_LIST_HEAD(&conf->retry_list);
3731         INIT_LIST_HEAD(&conf->bio_end_io_list);
3732
3733         spin_lock_init(&conf->resync_lock);
3734         init_waitqueue_head(&conf->wait_barrier);
3735         atomic_set(&conf->nr_pending, 0);
3736
3737         err = -ENOMEM;
3738         conf->thread = md_register_thread(raid10d, mddev, "raid10");
3739         if (!conf->thread)
3740                 goto out;
3741
3742         conf->mddev = mddev;
3743         return conf;
3744
3745  out:
3746         if (conf) {
3747                 mempool_exit(&conf->r10bio_pool);
3748                 kfree(conf->mirrors);
3749                 safe_put_page(conf->tmppage);
3750                 bioset_exit(&conf->bio_split);
3751                 kfree(conf);
3752         }
3753         return ERR_PTR(err);
3754 }
3755
3756 static int raid10_run(struct mddev *mddev)
3757 {
3758         struct r10conf *conf;
3759         int i, disk_idx, chunk_size;
3760         struct raid10_info *disk;
3761         struct md_rdev *rdev;
3762         sector_t size;
3763         sector_t min_offset_diff = 0;
3764         int first = 1;
3765         bool discard_supported = false;
3766
3767         if (mddev_init_writes_pending(mddev) < 0)
3768                 return -ENOMEM;
3769
3770         if (mddev->private == NULL) {
3771                 conf = setup_conf(mddev);
3772                 if (IS_ERR(conf))
3773                         return PTR_ERR(conf);
3774                 mddev->private = conf;
3775         }
3776         conf = mddev->private;
3777         if (!conf)
3778                 goto out;
3779
3780         if (mddev_is_clustered(conf->mddev)) {
3781                 int fc, fo;
3782
3783                 fc = (mddev->layout >> 8) & 255;
3784                 fo = mddev->layout & (1<<16);
3785                 if (fc > 1 || fo > 0) {
3786                         pr_err("only near layout is supported by clustered"
3787                                 " raid10\n");
3788                         goto out_free_conf;
3789                 }
3790         }
3791
3792         mddev->thread = conf->thread;
3793         conf->thread = NULL;
3794
3795         chunk_size = mddev->chunk_sectors << 9;
3796         if (mddev->queue) {
3797                 blk_queue_max_discard_sectors(mddev->queue,
3798                                               mddev->chunk_sectors);
3799                 blk_queue_max_write_same_sectors(mddev->queue, 0);
3800                 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3801                 blk_queue_io_min(mddev->queue, chunk_size);
3802                 if (conf->geo.raid_disks % conf->geo.near_copies)
3803                         blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3804                 else
3805                         blk_queue_io_opt(mddev->queue, chunk_size *
3806                                          (conf->geo.raid_disks / conf->geo.near_copies));
3807         }
3808
3809         rdev_for_each(rdev, mddev) {
3810                 long long diff;
3811
3812                 disk_idx = rdev->raid_disk;
3813                 if (disk_idx < 0)
3814                         continue;
3815                 if (disk_idx >= conf->geo.raid_disks &&
3816                     disk_idx >= conf->prev.raid_disks)
3817                         continue;
3818                 disk = conf->mirrors + disk_idx;
3819
3820                 if (test_bit(Replacement, &rdev->flags)) {
3821                         if (disk->replacement)
3822                                 goto out_free_conf;
3823                         disk->replacement = rdev;
3824                 } else {
3825                         if (disk->rdev)
3826                                 goto out_free_conf;
3827                         disk->rdev = rdev;
3828                 }
3829                 diff = (rdev->new_data_offset - rdev->data_offset);
3830                 if (!mddev->reshape_backwards)
3831                         diff = -diff;
3832                 if (diff < 0)
3833                         diff = 0;
3834                 if (first || diff < min_offset_diff)
3835                         min_offset_diff = diff;
3836
3837                 if (mddev->gendisk)
3838                         disk_stack_limits(mddev->gendisk, rdev->bdev,
3839                                           rdev->data_offset << 9);
3840
3841                 disk->head_position = 0;
3842
3843                 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3844                         discard_supported = true;
3845                 first = 0;
3846         }
3847
3848         if (mddev->queue) {
3849                 if (discard_supported)
3850                         blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3851                                                 mddev->queue);
3852                 else
3853                         blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3854                                                   mddev->queue);
3855         }
3856         /* need to check that every block has at least one working mirror */
3857         if (!enough(conf, -1)) {
3858                 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3859                        mdname(mddev));
3860                 goto out_free_conf;
3861         }
3862
3863         if (conf->reshape_progress != MaxSector) {
3864                 /* must ensure that shape change is supported */
3865                 if (conf->geo.far_copies != 1 &&
3866                     conf->geo.far_offset == 0)
3867                         goto out_free_conf;
3868                 if (conf->prev.far_copies != 1 &&
3869                     conf->prev.far_offset == 0)
3870                         goto out_free_conf;
3871         }
3872
3873         mddev->degraded = 0;
3874         for (i = 0;
3875              i < conf->geo.raid_disks
3876                      || i < conf->prev.raid_disks;
3877              i++) {
3878
3879                 disk = conf->mirrors + i;
3880
3881                 if (!disk->rdev && disk->replacement) {
3882                         /* The replacement is all we have - use it */
3883                         disk->rdev = disk->replacement;
3884                         disk->replacement = NULL;
3885                         clear_bit(Replacement, &disk->rdev->flags);
3886                 }
3887
3888                 if (!disk->rdev ||
3889                     !test_bit(In_sync, &disk->rdev->flags)) {
3890                         disk->head_position = 0;
3891                         mddev->degraded++;
3892                         if (disk->rdev &&
3893                             disk->rdev->saved_raid_disk < 0)
3894                                 conf->fullsync = 1;
3895                 }
3896                 disk->recovery_disabled = mddev->recovery_disabled - 1;
3897         }
3898
3899         if (mddev->recovery_cp != MaxSector)
3900                 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3901                           mdname(mddev));
3902         pr_info("md/raid10:%s: active with %d out of %d devices\n",
3903                 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3904                 conf->geo.raid_disks);
3905         /*
3906          * Ok, everything is just fine now
3907          */
3908         mddev->dev_sectors = conf->dev_sectors;
3909         size = raid10_size(mddev, 0, 0);
3910         md_set_array_sectors(mddev, size);
3911         mddev->resync_max_sectors = size;
3912         set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3913
3914         if (mddev->queue) {
3915                 int stripe = conf->geo.raid_disks *
3916                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3917
3918                 /* Calculate max read-ahead size.
3919                  * We need to readahead at least twice a whole stripe....
3920                  * maybe...
3921                  */
3922                 stripe /= conf->geo.near_copies;
3923                 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3924                         mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3925         }
3926
3927         if (md_integrity_register(mddev))
3928                 goto out_free_conf;
3929
3930         if (conf->reshape_progress != MaxSector) {
3931                 unsigned long before_length, after_length;
3932
3933                 before_length = ((1 << conf->prev.chunk_shift) *
3934                                  conf->prev.far_copies);
3935                 after_length = ((1 << conf->geo.chunk_shift) *
3936                                 conf->geo.far_copies);
3937
3938                 if (max(before_length, after_length) > min_offset_diff) {
3939                         /* This cannot work */
3940                         pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3941                         goto out_free_conf;
3942                 }
3943                 conf->offset_diff = min_offset_diff;
3944
3945                 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3946                 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3947                 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3948                 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3949                 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3950                                                         "reshape");
3951         }
3952
3953         return 0;
3954
3955 out_free_conf:
3956         md_unregister_thread(&mddev->thread);
3957         mempool_exit(&conf->r10bio_pool);
3958         safe_put_page(conf->tmppage);
3959         kfree(conf->mirrors);
3960         kfree(conf);
3961         mddev->private = NULL;
3962 out:
3963         return -EIO;
3964 }
3965
3966 static void raid10_free(struct mddev *mddev, void *priv)
3967 {
3968         struct r10conf *conf = priv;
3969
3970         mempool_exit(&conf->r10bio_pool);
3971         safe_put_page(conf->tmppage);
3972         kfree(conf->mirrors);
3973         kfree(conf->mirrors_old);
3974         kfree(conf->mirrors_new);
3975         bioset_exit(&conf->bio_split);
3976         kfree(conf);
3977 }
3978
3979 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3980 {
3981         struct r10conf *conf = mddev->private;
3982
3983         if (quiesce)
3984                 raise_barrier(conf, 0);
3985         else
3986                 lower_barrier(conf);
3987 }
3988
3989 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3990 {
3991         /* Resize of 'far' arrays is not supported.
3992          * For 'near' and 'offset' arrays we can set the
3993          * number of sectors used to be an appropriate multiple
3994          * of the chunk size.
3995          * For 'offset', this is far_copies*chunksize.
3996          * For 'near' the multiplier is the LCM of
3997          * near_copies and raid_disks.
3998          * So if far_copies > 1 && !far_offset, fail.
3999          * Else find LCM(raid_disks, near_copy)*far_copies and
4000          * multiply by chunk_size.  Then round to this number.
4001          * This is mostly done by raid10_size()
4002          */
4003         struct r10conf *conf = mddev->private;
4004         sector_t oldsize, size;
4005
4006         if (mddev->reshape_position != MaxSector)
4007                 return -EBUSY;
4008
4009         if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4010                 return -EINVAL;
4011
4012         oldsize = raid10_size(mddev, 0, 0);
4013         size = raid10_size(mddev, sectors, 0);
4014         if (mddev->external_size &&
4015             mddev->array_sectors > size)
4016                 return -EINVAL;
4017         if (mddev->bitmap) {
4018                 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
4019                 if (ret)
4020                         return ret;
4021         }
4022         md_set_array_sectors(mddev, size);
4023         if (sectors > mddev->dev_sectors &&
4024             mddev->recovery_cp > oldsize) {
4025                 mddev->recovery_cp = oldsize;
4026                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4027         }
4028         calc_sectors(conf, sectors);
4029         mddev->dev_sectors = conf->dev_sectors;
4030         mddev->resync_max_sectors = size;
4031         return 0;
4032 }
4033
4034 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4035 {
4036         struct md_rdev *rdev;
4037         struct r10conf *conf;
4038
4039         if (mddev->degraded > 0) {
4040                 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4041                         mdname(mddev));
4042                 return ERR_PTR(-EINVAL);
4043         }
4044         sector_div(size, devs);
4045
4046         /* Set new parameters */
4047         mddev->new_level = 10;
4048         /* new layout: far_copies = 1, near_copies = 2 */
4049         mddev->new_layout = (1<<8) + 2;
4050         mddev->new_chunk_sectors = mddev->chunk_sectors;
4051         mddev->delta_disks = mddev->raid_disks;
4052         mddev->raid_disks *= 2;
4053         /* make sure it will be not marked as dirty */
4054         mddev->recovery_cp = MaxSector;
4055         mddev->dev_sectors = size;
4056
4057         conf = setup_conf(mddev);
4058         if (!IS_ERR(conf)) {
4059                 rdev_for_each(rdev, mddev)
4060                         if (rdev->raid_disk >= 0) {
4061                                 rdev->new_raid_disk = rdev->raid_disk * 2;
4062                                 rdev->sectors = size;
4063                         }
4064                 conf->barrier = 1;
4065         }
4066
4067         return conf;
4068 }
4069
4070 static void *raid10_takeover(struct mddev *mddev)
4071 {
4072         struct r0conf *raid0_conf;
4073
4074         /* raid10 can take over:
4075          *  raid0 - providing it has only two drives
4076          */
4077         if (mddev->level == 0) {
4078                 /* for raid0 takeover only one zone is supported */
4079                 raid0_conf = mddev->private;
4080                 if (raid0_conf->nr_strip_zones > 1) {
4081                         pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4082                                 mdname(mddev));
4083                         return ERR_PTR(-EINVAL);
4084                 }
4085                 return raid10_takeover_raid0(mddev,
4086                         raid0_conf->strip_zone->zone_end,
4087                         raid0_conf->strip_zone->nb_dev);
4088         }
4089         return ERR_PTR(-EINVAL);
4090 }
4091
4092 static int raid10_check_reshape(struct mddev *mddev)
4093 {
4094         /* Called when there is a request to change
4095          * - layout (to ->new_layout)
4096          * - chunk size (to ->new_chunk_sectors)
4097          * - raid_disks (by delta_disks)
4098          * or when trying to restart a reshape that was ongoing.
4099          *
4100          * We need to validate the request and possibly allocate
4101          * space if that might be an issue later.
4102          *
4103          * Currently we reject any reshape of a 'far' mode array,
4104          * allow chunk size to change if new is generally acceptable,
4105          * allow raid_disks to increase, and allow
4106          * a switch between 'near' mode and 'offset' mode.
4107          */
4108         struct r10conf *conf = mddev->private;
4109         struct geom geo;
4110
4111         if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4112                 return -EINVAL;
4113
4114         if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4115                 /* mustn't change number of copies */
4116                 return -EINVAL;
4117         if (geo.far_copies > 1 && !geo.far_offset)
4118                 /* Cannot switch to 'far' mode */
4119                 return -EINVAL;
4120
4121         if (mddev->array_sectors & geo.chunk_mask)
4122                         /* not factor of array size */
4123                         return -EINVAL;
4124
4125         if (!enough(conf, -1))
4126                 return -EINVAL;
4127
4128         kfree(conf->mirrors_new);
4129         conf->mirrors_new = NULL;
4130         if (mddev->delta_disks > 0) {
4131                 /* allocate new 'mirrors' list */
4132                 conf->mirrors_new =
4133                         kcalloc(mddev->raid_disks + mddev->delta_disks,
4134                                 sizeof(struct raid10_info),
4135                                 GFP_KERNEL);
4136                 if (!conf->mirrors_new)
4137                         return -ENOMEM;
4138         }
4139         return 0;
4140 }
4141
4142 /*
4143  * Need to check if array has failed when deciding whether to:
4144  *  - start an array
4145  *  - remove non-faulty devices
4146  *  - add a spare
4147  *  - allow a reshape
4148  * This determination is simple when no reshape is happening.
4149  * However if there is a reshape, we need to carefully check
4150  * both the before and after sections.
4151  * This is because some failed devices may only affect one
4152  * of the two sections, and some non-in_sync devices may
4153  * be insync in the section most affected by failed devices.
4154  */
4155 static int calc_degraded(struct r10conf *conf)
4156 {
4157         int degraded, degraded2;
4158         int i;
4159
4160         rcu_read_lock();
4161         degraded = 0;
4162         /* 'prev' section first */
4163         for (i = 0; i < conf->prev.raid_disks; i++) {
4164                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4165                 if (!rdev || test_bit(Faulty, &rdev->flags))
4166                         degraded++;
4167                 else if (!test_bit(In_sync, &rdev->flags))
4168                         /* When we can reduce the number of devices in
4169                          * an array, this might not contribute to
4170                          * 'degraded'.  It does now.
4171                          */
4172                         degraded++;
4173         }
4174         rcu_read_unlock();
4175         if (conf->geo.raid_disks == conf->prev.raid_disks)
4176                 return degraded;
4177         rcu_read_lock();
4178         degraded2 = 0;
4179         for (i = 0; i < conf->geo.raid_disks; i++) {
4180                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4181                 if (!rdev || test_bit(Faulty, &rdev->flags))
4182                         degraded2++;
4183                 else if (!test_bit(In_sync, &rdev->flags)) {
4184                         /* If reshape is increasing the number of devices,
4185                          * this section has already been recovered, so
4186                          * it doesn't contribute to degraded.
4187                          * else it does.
4188                          */
4189                         if (conf->geo.raid_disks <= conf->prev.raid_disks)
4190                                 degraded2++;
4191                 }
4192         }
4193         rcu_read_unlock();
4194         if (degraded2 > degraded)
4195                 return degraded2;
4196         return degraded;
4197 }
4198
4199 static int raid10_start_reshape(struct mddev *mddev)
4200 {
4201         /* A 'reshape' has been requested. This commits
4202          * the various 'new' fields and sets MD_RECOVER_RESHAPE
4203          * This also checks if there are enough spares and adds them
4204          * to the array.
4205          * We currently require enough spares to make the final
4206          * array non-degraded.  We also require that the difference
4207          * between old and new data_offset - on each device - is
4208          * enough that we never risk over-writing.
4209          */
4210
4211         unsigned long before_length, after_length;
4212         sector_t min_offset_diff = 0;
4213         int first = 1;
4214         struct geom new;
4215         struct r10conf *conf = mddev->private;
4216         struct md_rdev *rdev;
4217         int spares = 0;
4218         int ret;
4219
4220         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4221                 return -EBUSY;
4222
4223         if (setup_geo(&new, mddev, geo_start) != conf->copies)
4224                 return -EINVAL;
4225
4226         before_length = ((1 << conf->prev.chunk_shift) *
4227                          conf->prev.far_copies);
4228         after_length = ((1 << conf->geo.chunk_shift) *
4229                         conf->geo.far_copies);
4230
4231         rdev_for_each(rdev, mddev) {
4232                 if (!test_bit(In_sync, &rdev->flags)
4233                     && !test_bit(Faulty, &rdev->flags))
4234                         spares++;
4235                 if (rdev->raid_disk >= 0) {
4236                         long long diff = (rdev->new_data_offset
4237                                           - rdev->data_offset);
4238                         if (!mddev->reshape_backwards)
4239                                 diff = -diff;
4240                         if (diff < 0)
4241                                 diff = 0;
4242                         if (first || diff < min_offset_diff)
4243                                 min_offset_diff = diff;
4244                         first = 0;
4245                 }
4246         }
4247
4248         if (max(before_length, after_length) > min_offset_diff)
4249                 return -EINVAL;
4250
4251         if (spares < mddev->delta_disks)
4252                 return -EINVAL;
4253
4254         conf->offset_diff = min_offset_diff;
4255         spin_lock_irq(&conf->device_lock);
4256         if (conf->mirrors_new) {
4257                 memcpy(conf->mirrors_new, conf->mirrors,
4258                        sizeof(struct raid10_info)*conf->prev.raid_disks);
4259                 smp_mb();
4260                 kfree(conf->mirrors_old);
4261                 conf->mirrors_old = conf->mirrors;
4262                 conf->mirrors = conf->mirrors_new;
4263                 conf->mirrors_new = NULL;
4264         }
4265         setup_geo(&conf->geo, mddev, geo_start);
4266         smp_mb();
4267         if (mddev->reshape_backwards) {
4268                 sector_t size = raid10_size(mddev, 0, 0);
4269                 if (size < mddev->array_sectors) {
4270                         spin_unlock_irq(&conf->device_lock);
4271                         pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4272                                 mdname(mddev));
4273                         return -EINVAL;
4274                 }
4275                 mddev->resync_max_sectors = size;
4276                 conf->reshape_progress = size;
4277         } else
4278                 conf->reshape_progress = 0;
4279         conf->reshape_safe = conf->reshape_progress;
4280         spin_unlock_irq(&conf->device_lock);
4281
4282         if (mddev->delta_disks && mddev->bitmap) {
4283                 ret = bitmap_resize(mddev->bitmap,
4284                                     raid10_size(mddev, 0,
4285                                                 conf->geo.raid_disks),
4286                                     0, 0);
4287                 if (ret)
4288                         goto abort;
4289         }
4290         if (mddev->delta_disks > 0) {
4291                 rdev_for_each(rdev, mddev)
4292                         if (rdev->raid_disk < 0 &&
4293                             !test_bit(Faulty, &rdev->flags)) {
4294                                 if (raid10_add_disk(mddev, rdev) == 0) {
4295                                         if (rdev->raid_disk >=
4296                                             conf->prev.raid_disks)
4297                                                 set_bit(In_sync, &rdev->flags);
4298                                         else
4299                                                 rdev->recovery_offset = 0;
4300
4301                                         if (sysfs_link_rdev(mddev, rdev))
4302                                                 /* Failure here  is OK */;
4303                                 }
4304                         } else if (rdev->raid_disk >= conf->prev.raid_disks
4305                                    && !test_bit(Faulty, &rdev->flags)) {
4306                                 /* This is a spare that was manually added */
4307                                 set_bit(In_sync, &rdev->flags);
4308                         }
4309         }
4310         /* When a reshape changes the number of devices,
4311          * ->degraded is measured against the larger of the
4312          * pre and  post numbers.
4313          */
4314         spin_lock_irq(&conf->device_lock);
4315         mddev->degraded = calc_degraded(conf);
4316         spin_unlock_irq(&conf->device_lock);
4317         mddev->raid_disks = conf->geo.raid_disks;
4318         mddev->reshape_position = conf->reshape_progress;
4319         set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4320
4321         clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4322         clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4323         clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4324         set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4325         set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4326
4327         mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4328                                                 "reshape");
4329         if (!mddev->sync_thread) {
4330                 ret = -EAGAIN;
4331                 goto abort;
4332         }
4333         conf->reshape_checkpoint = jiffies;
4334         md_wakeup_thread(mddev->sync_thread);
4335         md_new_event(mddev);
4336         return 0;
4337
4338 abort:
4339         mddev->recovery = 0;
4340         spin_lock_irq(&conf->device_lock);
4341         conf->geo = conf->prev;
4342         mddev->raid_disks = conf->geo.raid_disks;
4343         rdev_for_each(rdev, mddev)
4344                 rdev->new_data_offset = rdev->data_offset;
4345         smp_wmb();
4346         conf->reshape_progress = MaxSector;
4347         conf->reshape_safe = MaxSector;
4348         mddev->reshape_position = MaxSector;
4349         spin_unlock_irq(&conf->device_lock);
4350         return ret;
4351 }
4352
4353 /* Calculate the last device-address that could contain
4354  * any block from the chunk that includes the array-address 's'
4355  * and report the next address.
4356  * i.e. the address returned will be chunk-aligned and after
4357  * any data that is in the chunk containing 's'.
4358  */
4359 static sector_t last_dev_address(sector_t s, struct geom *geo)
4360 {
4361         s = (s | geo->chunk_mask) + 1;
4362         s >>= geo->chunk_shift;
4363         s *= geo->near_copies;
4364         s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4365         s *= geo->far_copies;
4366         s <<= geo->chunk_shift;
4367         return s;
4368 }
4369
4370 /* Calculate the first device-address that could contain
4371  * any block from the chunk that includes the array-address 's'.
4372  * This too will be the start of a chunk
4373  */
4374 static sector_t first_dev_address(sector_t s, struct geom *geo)
4375 {
4376         s >>= geo->chunk_shift;
4377         s *= geo->near_copies;
4378         sector_div(s, geo->raid_disks);
4379         s *= geo->far_copies;
4380         s <<= geo->chunk_shift;
4381         return s;
4382 }
4383
4384 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4385                                 int *skipped)
4386 {
4387         /* We simply copy at most one chunk (smallest of old and new)
4388          * at a time, possibly less if that exceeds RESYNC_PAGES,
4389          * or we hit a bad block or something.
4390          * This might mean we pause for normal IO in the middle of
4391          * a chunk, but that is not a problem as mddev->reshape_position
4392          * can record any location.
4393          *
4394          * If we will want to write to a location that isn't
4395          * yet recorded as 'safe' (i.e. in metadata on disk) then
4396          * we need to flush all reshape requests and update the metadata.
4397          *
4398          * When reshaping forwards (e.g. to more devices), we interpret
4399          * 'safe' as the earliest block which might not have been copied
4400          * down yet.  We divide this by previous stripe size and multiply
4401          * by previous stripe length to get lowest device offset that we
4402          * cannot write to yet.
4403          * We interpret 'sector_nr' as an address that we want to write to.
4404          * From this we use last_device_address() to find where we might
4405          * write to, and first_device_address on the  'safe' position.
4406          * If this 'next' write position is after the 'safe' position,
4407          * we must update the metadata to increase the 'safe' position.
4408          *
4409          * When reshaping backwards, we round in the opposite direction
4410          * and perform the reverse test:  next write position must not be
4411          * less than current safe position.
4412          *
4413          * In all this the minimum difference in data offsets
4414          * (conf->offset_diff - always positive) allows a bit of slack,
4415          * so next can be after 'safe', but not by more than offset_diff
4416          *
4417          * We need to prepare all the bios here before we start any IO
4418          * to ensure the size we choose is acceptable to all devices.
4419          * The means one for each copy for write-out and an extra one for
4420          * read-in.
4421          * We store the read-in bio in ->master_bio and the others in
4422          * ->devs[x].bio and ->devs[x].repl_bio.
4423          */
4424         struct r10conf *conf = mddev->private;
4425         struct r10bio *r10_bio;
4426         sector_t next, safe, last;
4427         int max_sectors;
4428         int nr_sectors;
4429         int s;
4430         struct md_rdev *rdev;
4431         int need_flush = 0;
4432         struct bio *blist;
4433         struct bio *bio, *read_bio;
4434         int sectors_done = 0;
4435         struct page **pages;
4436
4437         if (sector_nr == 0) {
4438                 /* If restarting in the middle, skip the initial sectors */
4439                 if (mddev->reshape_backwards &&
4440                     conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4441                         sector_nr = (raid10_size(mddev, 0, 0)
4442                                      - conf->reshape_progress);
4443                 } else if (!mddev->reshape_backwards &&
4444                            conf->reshape_progress > 0)
4445                         sector_nr = conf->reshape_progress;
4446                 if (sector_nr) {
4447                         mddev->curr_resync_completed = sector_nr;
4448                         sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4449                         *skipped = 1;
4450                         return sector_nr;
4451                 }
4452         }
4453
4454         /* We don't use sector_nr to track where we are up to
4455          * as that doesn't work well for ->reshape_backwards.
4456          * So just use ->reshape_progress.
4457          */
4458         if (mddev->reshape_backwards) {
4459                 /* 'next' is the earliest device address that we might
4460                  * write to for this chunk in the new layout
4461                  */
4462                 next = first_dev_address(conf->reshape_progress - 1,
4463                                          &conf->geo);
4464
4465                 /* 'safe' is the last device address that we might read from
4466                  * in the old layout after a restart
4467                  */
4468                 safe = last_dev_address(conf->reshape_safe - 1,
4469                                         &conf->prev);
4470
4471                 if (next + conf->offset_diff < safe)
4472                         need_flush = 1;
4473
4474                 last = conf->reshape_progress - 1;
4475                 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4476                                                & conf->prev.chunk_mask);
4477                 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4478                         sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4479         } else {
4480                 /* 'next' is after the last device address that we
4481                  * might write to for this chunk in the new layout
4482                  */
4483                 next = last_dev_address(conf->reshape_progress, &conf->geo);
4484
4485                 /* 'safe' is the earliest device address that we might
4486                  * read from in the old layout after a restart
4487                  */
4488                 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4489
4490                 /* Need to update metadata if 'next' might be beyond 'safe'
4491                  * as that would possibly corrupt data
4492                  */
4493                 if (next > safe + conf->offset_diff)
4494                         need_flush = 1;
4495
4496                 sector_nr = conf->reshape_progress;
4497                 last  = sector_nr | (conf->geo.chunk_mask
4498                                      & conf->prev.chunk_mask);
4499
4500                 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4501                         last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4502         }
4503
4504         if (need_flush ||
4505             time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4506                 /* Need to update reshape_position in metadata */
4507                 wait_barrier(conf);
4508                 mddev->reshape_position = conf->reshape_progress;
4509                 if (mddev->reshape_backwards)
4510                         mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4511                                 - conf->reshape_progress;
4512                 else
4513                         mddev->curr_resync_completed = conf->reshape_progress;
4514                 conf->reshape_checkpoint = jiffies;
4515                 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4516                 md_wakeup_thread(mddev->thread);
4517                 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4518                            test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4519                 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4520                         allow_barrier(conf);
4521                         return sectors_done;
4522                 }
4523                 conf->reshape_safe = mddev->reshape_position;
4524                 allow_barrier(conf);
4525         }
4526
4527 read_more:
4528         /* Now schedule reads for blocks from sector_nr to last */
4529         r10_bio = raid10_alloc_init_r10buf(conf);
4530         r10_bio->state = 0;
4531         raise_barrier(conf, sectors_done != 0);
4532         atomic_set(&r10_bio->remaining, 0);
4533         r10_bio->mddev = mddev;
4534         r10_bio->sector = sector_nr;
4535         set_bit(R10BIO_IsReshape, &r10_bio->state);
4536         r10_bio->sectors = last - sector_nr + 1;
4537         rdev = read_balance(conf, r10_bio, &max_sectors);
4538         BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4539
4540         if (!rdev) {
4541                 /* Cannot read from here, so need to record bad blocks
4542                  * on all the target devices.
4543                  */
4544                 // FIXME
4545                 mempool_free(r10_bio, &conf->r10buf_pool);
4546                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4547                 return sectors_done;
4548         }
4549
4550         read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4551
4552         bio_set_dev(read_bio, rdev->bdev);
4553         read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4554                                + rdev->data_offset);
4555         read_bio->bi_private = r10_bio;
4556         read_bio->bi_end_io = end_reshape_read;
4557         bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4558         read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4559         read_bio->bi_status = 0;
4560         read_bio->bi_vcnt = 0;
4561         read_bio->bi_iter.bi_size = 0;
4562         r10_bio->master_bio = read_bio;
4563         r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4564
4565         /* Now find the locations in the new layout */
4566         __raid10_find_phys(&conf->geo, r10_bio);
4567
4568         blist = read_bio;
4569         read_bio->bi_next = NULL;
4570
4571         rcu_read_lock();
4572         for (s = 0; s < conf->copies*2; s++) {
4573                 struct bio *b;
4574                 int d = r10_bio->devs[s/2].devnum;
4575                 struct md_rdev *rdev2;
4576                 if (s&1) {
4577                         rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4578                         b = r10_bio->devs[s/2].repl_bio;
4579                 } else {
4580                         rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4581                         b = r10_bio->devs[s/2].bio;
4582                 }
4583                 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4584                         continue;
4585
4586                 bio_set_dev(b, rdev2->bdev);
4587                 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4588                         rdev2->new_data_offset;
4589                 b->bi_end_io = end_reshape_write;
4590                 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4591                 b->bi_next = blist;
4592                 blist = b;
4593         }
4594
4595         /* Now add as many pages as possible to all of these bios. */
4596
4597         nr_sectors = 0;
4598         pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4599         for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4600                 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4601                 int len = (max_sectors - s) << 9;
4602                 if (len > PAGE_SIZE)
4603                         len = PAGE_SIZE;
4604                 for (bio = blist; bio ; bio = bio->bi_next) {
4605                         /*
4606                          * won't fail because the vec table is big enough
4607                          * to hold all these pages
4608                          */
4609                         bio_add_page(bio, page, len, 0);
4610                 }
4611                 sector_nr += len >> 9;
4612                 nr_sectors += len >> 9;
4613         }
4614         rcu_read_unlock();
4615         r10_bio->sectors = nr_sectors;
4616
4617         /* Now submit the read */
4618         md_sync_acct_bio(read_bio, r10_bio->sectors);
4619         atomic_inc(&r10_bio->remaining);
4620         read_bio->bi_next = NULL;
4621         generic_make_request(read_bio);
4622         sector_nr += nr_sectors;
4623         sectors_done += nr_sectors;
4624         if (sector_nr <= last)
4625                 goto read_more;
4626
4627         /* Now that we have done the whole section we can
4628          * update reshape_progress
4629          */
4630         if (mddev->reshape_backwards)
4631                 conf->reshape_progress -= sectors_done;
4632         else
4633                 conf->reshape_progress += sectors_done;
4634
4635         return sectors_done;
4636 }
4637
4638 static void end_reshape_request(struct r10bio *r10_bio);
4639 static int handle_reshape_read_error(struct mddev *mddev,
4640                                      struct r10bio *r10_bio);
4641 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4642 {
4643         /* Reshape read completed.  Hopefully we have a block
4644          * to write out.
4645          * If we got a read error then we do sync 1-page reads from
4646          * elsewhere until we find the data - or give up.
4647          */
4648         struct r10conf *conf = mddev->private;
4649         int s;
4650
4651         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4652                 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4653                         /* Reshape has been aborted */
4654                         md_done_sync(mddev, r10_bio->sectors, 0);
4655                         return;
4656                 }
4657
4658         /* We definitely have the data in the pages, schedule the
4659          * writes.
4660          */
4661         atomic_set(&r10_bio->remaining, 1);
4662         for (s = 0; s < conf->copies*2; s++) {
4663                 struct bio *b;
4664                 int d = r10_bio->devs[s/2].devnum;
4665                 struct md_rdev *rdev;
4666                 rcu_read_lock();
4667                 if (s&1) {
4668                         rdev = rcu_dereference(conf->mirrors[d].replacement);
4669                         b = r10_bio->devs[s/2].repl_bio;
4670                 } else {
4671                         rdev = rcu_dereference(conf->mirrors[d].rdev);
4672                         b = r10_bio->devs[s/2].bio;
4673                 }
4674                 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4675                         rcu_read_unlock();
4676                         continue;
4677                 }
4678                 atomic_inc(&rdev->nr_pending);
4679                 rcu_read_unlock();
4680                 md_sync_acct_bio(b, r10_bio->sectors);
4681                 atomic_inc(&r10_bio->remaining);
4682                 b->bi_next = NULL;
4683                 generic_make_request(b);
4684         }
4685         end_reshape_request(r10_bio);
4686 }
4687
4688 static void end_reshape(struct r10conf *conf)
4689 {
4690         if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4691                 return;
4692
4693         spin_lock_irq(&conf->device_lock);
4694         conf->prev = conf->geo;
4695         md_finish_reshape(conf->mddev);
4696         smp_wmb();
4697         conf->reshape_progress = MaxSector;
4698         conf->reshape_safe = MaxSector;
4699         spin_unlock_irq(&conf->device_lock);
4700
4701         /* read-ahead size must cover two whole stripes, which is
4702          * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4703          */
4704         if (conf->mddev->queue) {
4705                 int stripe = conf->geo.raid_disks *
4706                         ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4707                 stripe /= conf->geo.near_copies;
4708                 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4709                         conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4710         }
4711         conf->fullsync = 0;
4712 }
4713
4714 static int handle_reshape_read_error(struct mddev *mddev,
4715                                      struct r10bio *r10_bio)
4716 {
4717         /* Use sync reads to get the blocks from somewhere else */
4718         int sectors = r10_bio->sectors;
4719         struct r10conf *conf = mddev->private;
4720         struct r10bio *r10b;
4721         int slot = 0;
4722         int idx = 0;
4723         struct page **pages;
4724
4725         r10b = kmalloc(sizeof(*r10b) +
4726                sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4727         if (!r10b) {
4728                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4729                 return -ENOMEM;
4730         }
4731
4732         /* reshape IOs share pages from .devs[0].bio */
4733         pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4734
4735         r10b->sector = r10_bio->sector;
4736         __raid10_find_phys(&conf->prev, r10b);
4737
4738         while (sectors) {
4739                 int s = sectors;
4740                 int success = 0;
4741                 int first_slot = slot;
4742
4743                 if (s > (PAGE_SIZE >> 9))
4744                         s = PAGE_SIZE >> 9;
4745
4746                 rcu_read_lock();
4747                 while (!success) {
4748                         int d = r10b->devs[slot].devnum;
4749                         struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4750                         sector_t addr;
4751                         if (rdev == NULL ||
4752                             test_bit(Faulty, &rdev->flags) ||
4753                             !test_bit(In_sync, &rdev->flags))
4754                                 goto failed;
4755
4756                         addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4757                         atomic_inc(&rdev->nr_pending);
4758                         rcu_read_unlock();
4759                         success = sync_page_io(rdev,
4760                                                addr,
4761                                                s << 9,
4762                                                pages[idx],
4763                                                REQ_OP_READ, 0, false);
4764                         rdev_dec_pending(rdev, mddev);
4765                         rcu_read_lock();
4766                         if (success)
4767                                 break;
4768                 failed:
4769                         slot++;
4770                         if (slot >= conf->copies)
4771                                 slot = 0;
4772                         if (slot == first_slot)
4773                                 break;
4774                 }
4775                 rcu_read_unlock();
4776                 if (!success) {
4777                         /* couldn't read this block, must give up */
4778                         set_bit(MD_RECOVERY_INTR,
4779                                 &mddev->recovery);
4780                         kfree(r10b);
4781                         return -EIO;
4782                 }
4783                 sectors -= s;
4784                 idx++;
4785         }
4786         kfree(r10b);
4787         return 0;
4788 }
4789
4790 static void end_reshape_write(struct bio *bio)
4791 {
4792         struct r10bio *r10_bio = get_resync_r10bio(bio);
4793         struct mddev *mddev = r10_bio->mddev;
4794         struct r10conf *conf = mddev->private;
4795         int d;
4796         int slot;
4797         int repl;
4798         struct md_rdev *rdev = NULL;
4799
4800         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4801         if (repl)
4802                 rdev = conf->mirrors[d].replacement;
4803         if (!rdev) {
4804                 smp_mb();
4805                 rdev = conf->mirrors[d].rdev;
4806         }
4807
4808         if (bio->bi_status) {
4809                 /* FIXME should record badblock */
4810                 md_error(mddev, rdev);
4811         }
4812
4813         rdev_dec_pending(rdev, mddev);
4814         end_reshape_request(r10_bio);
4815 }
4816
4817 static void end_reshape_request(struct r10bio *r10_bio)
4818 {
4819         if (!atomic_dec_and_test(&r10_bio->remaining))
4820                 return;
4821         md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4822         bio_put(r10_bio->master_bio);
4823         put_buf(r10_bio);
4824 }
4825
4826 static void raid10_finish_reshape(struct mddev *mddev)
4827 {
4828         struct r10conf *conf = mddev->private;
4829
4830         if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4831                 return;
4832
4833         if (mddev->delta_disks > 0) {
4834                 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4835                         mddev->recovery_cp = mddev->resync_max_sectors;
4836                         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4837                 }
4838                 mddev->resync_max_sectors = mddev->array_sectors;
4839         } else {
4840                 int d;
4841                 rcu_read_lock();
4842                 for (d = conf->geo.raid_disks ;
4843                      d < conf->geo.raid_disks - mddev->delta_disks;
4844                      d++) {
4845                         struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4846                         if (rdev)
4847                                 clear_bit(In_sync, &rdev->flags);
4848                         rdev = rcu_dereference(conf->mirrors[d].replacement);
4849                         if (rdev)
4850                                 clear_bit(In_sync, &rdev->flags);
4851                 }
4852                 rcu_read_unlock();
4853         }
4854         mddev->layout = mddev->new_layout;
4855         mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4856         mddev->reshape_position = MaxSector;
4857         mddev->delta_disks = 0;
4858         mddev->reshape_backwards = 0;
4859 }
4860
4861 static struct md_personality raid10_personality =
4862 {
4863         .name           = "raid10",
4864         .level          = 10,
4865         .owner          = THIS_MODULE,
4866         .make_request   = raid10_make_request,
4867         .run            = raid10_run,
4868         .free           = raid10_free,
4869         .status         = raid10_status,
4870         .error_handler  = raid10_error,
4871         .hot_add_disk   = raid10_add_disk,
4872         .hot_remove_disk= raid10_remove_disk,
4873         .spare_active   = raid10_spare_active,
4874         .sync_request   = raid10_sync_request,
4875         .quiesce        = raid10_quiesce,
4876         .size           = raid10_size,
4877         .resize         = raid10_resize,
4878         .takeover       = raid10_takeover,
4879         .check_reshape  = raid10_check_reshape,
4880         .start_reshape  = raid10_start_reshape,
4881         .finish_reshape = raid10_finish_reshape,
4882         .congested      = raid10_congested,
4883 };
4884
4885 static int __init raid_init(void)
4886 {
4887         return register_md_personality(&raid10_personality);
4888 }
4889
4890 static void raid_exit(void)
4891 {
4892         unregister_md_personality(&raid10_personality);
4893 }
4894
4895 module_init(raid_init);
4896 module_exit(raid_exit);
4897 MODULE_LICENSE("GPL");
4898 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4899 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4900 MODULE_ALIAS("md-raid10");
4901 MODULE_ALIAS("md-level-10");
4902
4903 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);