Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid10.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 2000-2004 Neil Brown
6 *
7 * RAID-10 support for md.
8 *
9 * Base on code in raid1.c. See raid1.c for further copyright information.
10 */
11
12#include <linux/slab.h>
13#include <linux/delay.h>
14#include <linux/blkdev.h>
15#include <linux/module.h>
16#include <linux/seq_file.h>
17#include <linux/ratelimit.h>
18#include <linux/kthread.h>
19#include <linux/raid/md_p.h>
20#include <trace/events/block.h>
21#include "md.h"
22
23#define RAID_1_10_NAME "raid10"
24#include "raid10.h"
25#include "raid0.h"
26#include "md-bitmap.h"
27
28/*
29 * RAID10 provides a combination of RAID0 and RAID1 functionality.
30 * The layout of data is defined by
31 * chunk_size
32 * raid_disks
33 * near_copies (stored in low byte of layout)
34 * far_copies (stored in second byte of layout)
35 * far_offset (stored in bit 16 of layout )
36 * use_far_sets (stored in bit 17 of layout )
37 * use_far_sets_bugfixed (stored in bit 18 of layout )
38 *
39 * The data to be stored is divided into chunks using chunksize. Each device
40 * is divided into far_copies sections. In each section, chunks are laid out
41 * in a style similar to raid0, but near_copies copies of each chunk is stored
42 * (each on a different drive). The starting device for each section is offset
43 * near_copies from the starting device of the previous section. Thus there
44 * are (near_copies * far_copies) of each chunk, and each is on a different
45 * drive. near_copies and far_copies must be at least one, and their product
46 * is at most raid_disks.
47 *
48 * If far_offset is true, then the far_copies are handled a bit differently.
49 * The copies are still in different stripes, but instead of being very far
50 * apart on disk, there are adjacent stripes.
51 *
52 * The far and offset algorithms are handled slightly differently if
53 * 'use_far_sets' is true. In this case, the array's devices are grouped into
54 * sets that are (near_copies * far_copies) in size. The far copied stripes
55 * are still shifted by 'near_copies' devices, but this shifting stays confined
56 * to the set rather than the entire array. This is done to improve the number
57 * of device combinations that can fail without causing the array to fail.
58 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
59 * on a device):
60 * A B C D A B C D E
61 * ... ...
62 * D A B C E A B C D
63 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
64 * [A B] [C D] [A B] [C D E]
65 * |...| |...| |...| | ... |
66 * [B A] [D C] [B A] [E C D]
67 */
68
69static void allow_barrier(struct r10conf *conf);
70static void lower_barrier(struct r10conf *conf);
71static int _enough(struct r10conf *conf, int previous, int ignore);
72static int enough(struct r10conf *conf, int ignore);
73static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
74 int *skipped);
75static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
76static void end_reshape_write(struct bio *bio);
77static void end_reshape(struct r10conf *conf);
78
79#include "raid1-10.c"
80
81#define NULL_CMD
82#define cmd_before(conf, cmd) \
83 do { \
84 write_sequnlock_irq(&(conf)->resync_lock); \
85 cmd; \
86 } while (0)
87#define cmd_after(conf) write_seqlock_irq(&(conf)->resync_lock)
88
89#define wait_event_barrier_cmd(conf, cond, cmd) \
90 wait_event_cmd((conf)->wait_barrier, cond, cmd_before(conf, cmd), \
91 cmd_after(conf))
92
93#define wait_event_barrier(conf, cond) \
94 wait_event_barrier_cmd(conf, cond, NULL_CMD)
95
96/*
97 * for resync bio, r10bio pointer can be retrieved from the per-bio
98 * 'struct resync_pages'.
99 */
100static inline struct r10bio *get_resync_r10bio(struct bio *bio)
101{
102 return get_resync_pages(bio)->raid_bio;
103}
104
105static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
106{
107 struct r10conf *conf = data;
108 int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
109
110 /* allocate a r10bio with room for raid_disks entries in the
111 * bios array */
112 return kzalloc(size, gfp_flags);
113}
114
115#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
116/* amount of memory to reserve for resync requests */
117#define RESYNC_WINDOW (1024*1024)
118/* maximum number of concurrent requests, memory permitting */
119#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
120#define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
121#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
122
123/*
124 * When performing a resync, we need to read and compare, so
125 * we need as many pages are there are copies.
126 * When performing a recovery, we need 2 bios, one for read,
127 * one for write (we recover only one drive per r10buf)
128 *
129 */
130static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
131{
132 struct r10conf *conf = data;
133 struct r10bio *r10_bio;
134 struct bio *bio;
135 int j;
136 int nalloc, nalloc_rp;
137 struct resync_pages *rps;
138
139 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
140 if (!r10_bio)
141 return NULL;
142
143 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
144 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
145 nalloc = conf->copies; /* resync */
146 else
147 nalloc = 2; /* recovery */
148
149 /* allocate once for all bios */
150 if (!conf->have_replacement)
151 nalloc_rp = nalloc;
152 else
153 nalloc_rp = nalloc * 2;
154 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
155 if (!rps)
156 goto out_free_r10bio;
157
158 /*
159 * Allocate bios.
160 */
161 for (j = nalloc ; j-- ; ) {
162 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
163 if (!bio)
164 goto out_free_bio;
165 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
166 r10_bio->devs[j].bio = bio;
167 if (!conf->have_replacement)
168 continue;
169 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
170 if (!bio)
171 goto out_free_bio;
172 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
173 r10_bio->devs[j].repl_bio = bio;
174 }
175 /*
176 * Allocate RESYNC_PAGES data pages and attach them
177 * where needed.
178 */
179 for (j = 0; j < nalloc; j++) {
180 struct bio *rbio = r10_bio->devs[j].repl_bio;
181 struct resync_pages *rp, *rp_repl;
182
183 rp = &rps[j];
184 if (rbio)
185 rp_repl = &rps[nalloc + j];
186
187 bio = r10_bio->devs[j].bio;
188
189 if (!j || test_bit(MD_RECOVERY_SYNC,
190 &conf->mddev->recovery)) {
191 if (resync_alloc_pages(rp, gfp_flags))
192 goto out_free_pages;
193 } else {
194 memcpy(rp, &rps[0], sizeof(*rp));
195 resync_get_all_pages(rp);
196 }
197
198 rp->raid_bio = r10_bio;
199 bio->bi_private = rp;
200 if (rbio) {
201 memcpy(rp_repl, rp, sizeof(*rp));
202 rbio->bi_private = rp_repl;
203 }
204 }
205
206 return r10_bio;
207
208out_free_pages:
209 while (--j >= 0)
210 resync_free_pages(&rps[j]);
211
212 j = 0;
213out_free_bio:
214 for ( ; j < nalloc; j++) {
215 if (r10_bio->devs[j].bio)
216 bio_uninit(r10_bio->devs[j].bio);
217 kfree(r10_bio->devs[j].bio);
218 if (r10_bio->devs[j].repl_bio)
219 bio_uninit(r10_bio->devs[j].repl_bio);
220 kfree(r10_bio->devs[j].repl_bio);
221 }
222 kfree(rps);
223out_free_r10bio:
224 rbio_pool_free(r10_bio, conf);
225 return NULL;
226}
227
228static void r10buf_pool_free(void *__r10_bio, void *data)
229{
230 struct r10conf *conf = data;
231 struct r10bio *r10bio = __r10_bio;
232 int j;
233 struct resync_pages *rp = NULL;
234
235 for (j = conf->copies; j--; ) {
236 struct bio *bio = r10bio->devs[j].bio;
237
238 if (bio) {
239 rp = get_resync_pages(bio);
240 resync_free_pages(rp);
241 bio_uninit(bio);
242 kfree(bio);
243 }
244
245 bio = r10bio->devs[j].repl_bio;
246 if (bio) {
247 bio_uninit(bio);
248 kfree(bio);
249 }
250 }
251
252 /* resync pages array stored in the 1st bio's .bi_private */
253 kfree(rp);
254
255 rbio_pool_free(r10bio, conf);
256}
257
258static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
259{
260 int i;
261
262 for (i = 0; i < conf->geo.raid_disks; i++) {
263 struct bio **bio = & r10_bio->devs[i].bio;
264 if (!BIO_SPECIAL(*bio))
265 bio_put(*bio);
266 *bio = NULL;
267 bio = &r10_bio->devs[i].repl_bio;
268 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
269 bio_put(*bio);
270 *bio = NULL;
271 }
272}
273
274static void free_r10bio(struct r10bio *r10_bio)
275{
276 struct r10conf *conf = r10_bio->mddev->private;
277
278 put_all_bios(conf, r10_bio);
279 mempool_free(r10_bio, &conf->r10bio_pool);
280}
281
282static void put_buf(struct r10bio *r10_bio)
283{
284 struct r10conf *conf = r10_bio->mddev->private;
285
286 mempool_free(r10_bio, &conf->r10buf_pool);
287
288 lower_barrier(conf);
289}
290
291static void wake_up_barrier(struct r10conf *conf)
292{
293 if (wq_has_sleeper(&conf->wait_barrier))
294 wake_up(&conf->wait_barrier);
295}
296
297static void reschedule_retry(struct r10bio *r10_bio)
298{
299 unsigned long flags;
300 struct mddev *mddev = r10_bio->mddev;
301 struct r10conf *conf = mddev->private;
302
303 spin_lock_irqsave(&conf->device_lock, flags);
304 list_add(&r10_bio->retry_list, &conf->retry_list);
305 conf->nr_queued ++;
306 spin_unlock_irqrestore(&conf->device_lock, flags);
307
308 /* wake up frozen array... */
309 wake_up(&conf->wait_barrier);
310
311 md_wakeup_thread(mddev->thread);
312}
313
314/*
315 * raid_end_bio_io() is called when we have finished servicing a mirrored
316 * operation and are ready to return a success/failure code to the buffer
317 * cache layer.
318 */
319static void raid_end_bio_io(struct r10bio *r10_bio)
320{
321 struct bio *bio = r10_bio->master_bio;
322 struct r10conf *conf = r10_bio->mddev->private;
323
324 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
325 bio->bi_status = BLK_STS_IOERR;
326
327 bio_endio(bio);
328 /*
329 * Wake up any possible resync thread that waits for the device
330 * to go idle.
331 */
332 allow_barrier(conf);
333
334 free_r10bio(r10_bio);
335}
336
337/*
338 * Update disk head position estimator based on IRQ completion info.
339 */
340static inline void update_head_pos(int slot, struct r10bio *r10_bio)
341{
342 struct r10conf *conf = r10_bio->mddev->private;
343
344 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
345 r10_bio->devs[slot].addr + (r10_bio->sectors);
346}
347
348/*
349 * Find the disk number which triggered given bio
350 */
351static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
352 struct bio *bio, int *slotp, int *replp)
353{
354 int slot;
355 int repl = 0;
356
357 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
358 if (r10_bio->devs[slot].bio == bio)
359 break;
360 if (r10_bio->devs[slot].repl_bio == bio) {
361 repl = 1;
362 break;
363 }
364 }
365
366 update_head_pos(slot, r10_bio);
367
368 if (slotp)
369 *slotp = slot;
370 if (replp)
371 *replp = repl;
372 return r10_bio->devs[slot].devnum;
373}
374
375static void raid10_end_read_request(struct bio *bio)
376{
377 int uptodate = !bio->bi_status;
378 struct r10bio *r10_bio = bio->bi_private;
379 int slot;
380 struct md_rdev *rdev;
381 struct r10conf *conf = r10_bio->mddev->private;
382
383 slot = r10_bio->read_slot;
384 rdev = r10_bio->devs[slot].rdev;
385 /*
386 * this branch is our 'one mirror IO has finished' event handler:
387 */
388 update_head_pos(slot, r10_bio);
389
390 if (uptodate) {
391 /*
392 * Set R10BIO_Uptodate in our master bio, so that
393 * we will return a good error code to the higher
394 * levels even if IO on some other mirrored buffer fails.
395 *
396 * The 'master' represents the composite IO operation to
397 * user-side. So if something waits for IO, then it will
398 * wait for the 'master' bio.
399 */
400 set_bit(R10BIO_Uptodate, &r10_bio->state);
401 } else {
402 /* If all other devices that store this block have
403 * failed, we want to return the error upwards rather
404 * than fail the last device. Here we redefine
405 * "uptodate" to mean "Don't want to retry"
406 */
407 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
408 rdev->raid_disk))
409 uptodate = 1;
410 }
411 if (uptodate) {
412 raid_end_bio_io(r10_bio);
413 rdev_dec_pending(rdev, conf->mddev);
414 } else {
415 /*
416 * oops, read error - keep the refcount on the rdev
417 */
418 pr_err_ratelimited("md/raid10:%s: %pg: rescheduling sector %llu\n",
419 mdname(conf->mddev),
420 rdev->bdev,
421 (unsigned long long)r10_bio->sector);
422 set_bit(R10BIO_ReadError, &r10_bio->state);
423 reschedule_retry(r10_bio);
424 }
425}
426
427static void close_write(struct r10bio *r10_bio)
428{
429 struct mddev *mddev = r10_bio->mddev;
430
431 md_write_end(mddev);
432}
433
434static void one_write_done(struct r10bio *r10_bio)
435{
436 if (atomic_dec_and_test(&r10_bio->remaining)) {
437 if (test_bit(R10BIO_WriteError, &r10_bio->state))
438 reschedule_retry(r10_bio);
439 else {
440 close_write(r10_bio);
441 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
442 reschedule_retry(r10_bio);
443 else
444 raid_end_bio_io(r10_bio);
445 }
446 }
447}
448
449static void raid10_end_write_request(struct bio *bio)
450{
451 struct r10bio *r10_bio = bio->bi_private;
452 int dev;
453 int dec_rdev = 1;
454 struct r10conf *conf = r10_bio->mddev->private;
455 int slot, repl;
456 struct md_rdev *rdev = NULL;
457 struct bio *to_put = NULL;
458 bool discard_error;
459
460 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
461
462 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
463
464 if (repl)
465 rdev = conf->mirrors[dev].replacement;
466 if (!rdev) {
467 smp_rmb();
468 repl = 0;
469 rdev = conf->mirrors[dev].rdev;
470 }
471 /*
472 * this branch is our 'one mirror IO has finished' event handler:
473 */
474 if (bio->bi_status && !discard_error) {
475 if (repl)
476 /* Never record new bad blocks to replacement,
477 * just fail it.
478 */
479 md_error(rdev->mddev, rdev);
480 else {
481 set_bit(WriteErrorSeen, &rdev->flags);
482 if (!test_and_set_bit(WantReplacement, &rdev->flags))
483 set_bit(MD_RECOVERY_NEEDED,
484 &rdev->mddev->recovery);
485
486 dec_rdev = 0;
487 if (test_bit(FailFast, &rdev->flags) &&
488 (bio->bi_opf & MD_FAILFAST)) {
489 md_error(rdev->mddev, rdev);
490 }
491
492 /*
493 * When the device is faulty, it is not necessary to
494 * handle write error.
495 */
496 if (!test_bit(Faulty, &rdev->flags))
497 set_bit(R10BIO_WriteError, &r10_bio->state);
498 else {
499 /* Fail the request */
500 r10_bio->devs[slot].bio = NULL;
501 to_put = bio;
502 dec_rdev = 1;
503 }
504 }
505 } else {
506 /*
507 * Set R10BIO_Uptodate in our master bio, so that
508 * we will return a good error code for to the higher
509 * levels even if IO on some other mirrored buffer fails.
510 *
511 * The 'master' represents the composite IO operation to
512 * user-side. So if something waits for IO, then it will
513 * wait for the 'master' bio.
514 *
515 * Do not set R10BIO_Uptodate if the current device is
516 * rebuilding or Faulty. This is because we cannot use
517 * such device for properly reading the data back (we could
518 * potentially use it, if the current write would have felt
519 * before rdev->recovery_offset, but for simplicity we don't
520 * check this here.
521 */
522 if (test_bit(In_sync, &rdev->flags) &&
523 !test_bit(Faulty, &rdev->flags))
524 set_bit(R10BIO_Uptodate, &r10_bio->state);
525
526 /* Maybe we can clear some bad blocks. */
527 if (rdev_has_badblock(rdev, r10_bio->devs[slot].addr,
528 r10_bio->sectors) &&
529 !discard_error) {
530 bio_put(bio);
531 if (repl)
532 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
533 else
534 r10_bio->devs[slot].bio = IO_MADE_GOOD;
535 dec_rdev = 0;
536 set_bit(R10BIO_MadeGood, &r10_bio->state);
537 }
538 }
539
540 /*
541 *
542 * Let's see if all mirrored write operations have finished
543 * already.
544 */
545 one_write_done(r10_bio);
546 if (dec_rdev)
547 rdev_dec_pending(rdev, conf->mddev);
548 if (to_put)
549 bio_put(to_put);
550}
551
552/*
553 * RAID10 layout manager
554 * As well as the chunksize and raid_disks count, there are two
555 * parameters: near_copies and far_copies.
556 * near_copies * far_copies must be <= raid_disks.
557 * Normally one of these will be 1.
558 * If both are 1, we get raid0.
559 * If near_copies == raid_disks, we get raid1.
560 *
561 * Chunks are laid out in raid0 style with near_copies copies of the
562 * first chunk, followed by near_copies copies of the next chunk and
563 * so on.
564 * If far_copies > 1, then after 1/far_copies of the array has been assigned
565 * as described above, we start again with a device offset of near_copies.
566 * So we effectively have another copy of the whole array further down all
567 * the drives, but with blocks on different drives.
568 * With this layout, and block is never stored twice on the one device.
569 *
570 * raid10_find_phys finds the sector offset of a given virtual sector
571 * on each device that it is on.
572 *
573 * raid10_find_virt does the reverse mapping, from a device and a
574 * sector offset to a virtual address
575 */
576
577static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
578{
579 int n,f;
580 sector_t sector;
581 sector_t chunk;
582 sector_t stripe;
583 int dev;
584 int slot = 0;
585 int last_far_set_start, last_far_set_size;
586
587 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
588 last_far_set_start *= geo->far_set_size;
589
590 last_far_set_size = geo->far_set_size;
591 last_far_set_size += (geo->raid_disks % geo->far_set_size);
592
593 /* now calculate first sector/dev */
594 chunk = r10bio->sector >> geo->chunk_shift;
595 sector = r10bio->sector & geo->chunk_mask;
596
597 chunk *= geo->near_copies;
598 stripe = chunk;
599 dev = sector_div(stripe, geo->raid_disks);
600 if (geo->far_offset)
601 stripe *= geo->far_copies;
602
603 sector += stripe << geo->chunk_shift;
604
605 /* and calculate all the others */
606 for (n = 0; n < geo->near_copies; n++) {
607 int d = dev;
608 int set;
609 sector_t s = sector;
610 r10bio->devs[slot].devnum = d;
611 r10bio->devs[slot].addr = s;
612 slot++;
613
614 for (f = 1; f < geo->far_copies; f++) {
615 set = d / geo->far_set_size;
616 d += geo->near_copies;
617
618 if ((geo->raid_disks % geo->far_set_size) &&
619 (d > last_far_set_start)) {
620 d -= last_far_set_start;
621 d %= last_far_set_size;
622 d += last_far_set_start;
623 } else {
624 d %= geo->far_set_size;
625 d += geo->far_set_size * set;
626 }
627 s += geo->stride;
628 r10bio->devs[slot].devnum = d;
629 r10bio->devs[slot].addr = s;
630 slot++;
631 }
632 dev++;
633 if (dev >= geo->raid_disks) {
634 dev = 0;
635 sector += (geo->chunk_mask + 1);
636 }
637 }
638}
639
640static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
641{
642 struct geom *geo = &conf->geo;
643
644 if (conf->reshape_progress != MaxSector &&
645 ((r10bio->sector >= conf->reshape_progress) !=
646 conf->mddev->reshape_backwards)) {
647 set_bit(R10BIO_Previous, &r10bio->state);
648 geo = &conf->prev;
649 } else
650 clear_bit(R10BIO_Previous, &r10bio->state);
651
652 __raid10_find_phys(geo, r10bio);
653}
654
655static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
656{
657 sector_t offset, chunk, vchunk;
658 /* Never use conf->prev as this is only called during resync
659 * or recovery, so reshape isn't happening
660 */
661 struct geom *geo = &conf->geo;
662 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
663 int far_set_size = geo->far_set_size;
664 int last_far_set_start;
665
666 if (geo->raid_disks % geo->far_set_size) {
667 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
668 last_far_set_start *= geo->far_set_size;
669
670 if (dev >= last_far_set_start) {
671 far_set_size = geo->far_set_size;
672 far_set_size += (geo->raid_disks % geo->far_set_size);
673 far_set_start = last_far_set_start;
674 }
675 }
676
677 offset = sector & geo->chunk_mask;
678 if (geo->far_offset) {
679 int fc;
680 chunk = sector >> geo->chunk_shift;
681 fc = sector_div(chunk, geo->far_copies);
682 dev -= fc * geo->near_copies;
683 if (dev < far_set_start)
684 dev += far_set_size;
685 } else {
686 while (sector >= geo->stride) {
687 sector -= geo->stride;
688 if (dev < (geo->near_copies + far_set_start))
689 dev += far_set_size - geo->near_copies;
690 else
691 dev -= geo->near_copies;
692 }
693 chunk = sector >> geo->chunk_shift;
694 }
695 vchunk = chunk * geo->raid_disks + dev;
696 sector_div(vchunk, geo->near_copies);
697 return (vchunk << geo->chunk_shift) + offset;
698}
699
700/*
701 * This routine returns the disk from which the requested read should
702 * be done. There is a per-array 'next expected sequential IO' sector
703 * number - if this matches on the next IO then we use the last disk.
704 * There is also a per-disk 'last know head position' sector that is
705 * maintained from IRQ contexts, both the normal and the resync IO
706 * completion handlers update this position correctly. If there is no
707 * perfect sequential match then we pick the disk whose head is closest.
708 *
709 * If there are 2 mirrors in the same 2 devices, performance degrades
710 * because position is mirror, not device based.
711 *
712 * The rdev for the device selected will have nr_pending incremented.
713 */
714
715/*
716 * FIXME: possibly should rethink readbalancing and do it differently
717 * depending on near_copies / far_copies geometry.
718 */
719static struct md_rdev *read_balance(struct r10conf *conf,
720 struct r10bio *r10_bio,
721 int *max_sectors)
722{
723 const sector_t this_sector = r10_bio->sector;
724 int disk, slot;
725 int sectors = r10_bio->sectors;
726 int best_good_sectors;
727 sector_t new_distance, best_dist;
728 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
729 int do_balance;
730 int best_dist_slot, best_pending_slot;
731 bool has_nonrot_disk = false;
732 unsigned int min_pending;
733 struct geom *geo = &conf->geo;
734
735 raid10_find_phys(conf, r10_bio);
736 best_dist_slot = -1;
737 min_pending = UINT_MAX;
738 best_dist_rdev = NULL;
739 best_pending_rdev = NULL;
740 best_dist = MaxSector;
741 best_good_sectors = 0;
742 do_balance = 1;
743 clear_bit(R10BIO_FailFast, &r10_bio->state);
744
745 if (raid1_should_read_first(conf->mddev, this_sector, sectors))
746 do_balance = 0;
747
748 for (slot = 0; slot < conf->copies ; slot++) {
749 sector_t first_bad;
750 int bad_sectors;
751 sector_t dev_sector;
752 unsigned int pending;
753 bool nonrot;
754
755 if (r10_bio->devs[slot].bio == IO_BLOCKED)
756 continue;
757 disk = r10_bio->devs[slot].devnum;
758 rdev = conf->mirrors[disk].replacement;
759 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
760 r10_bio->devs[slot].addr + sectors >
761 rdev->recovery_offset)
762 rdev = conf->mirrors[disk].rdev;
763 if (rdev == NULL ||
764 test_bit(Faulty, &rdev->flags))
765 continue;
766 if (!test_bit(In_sync, &rdev->flags) &&
767 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
768 continue;
769
770 dev_sector = r10_bio->devs[slot].addr;
771 if (is_badblock(rdev, dev_sector, sectors,
772 &first_bad, &bad_sectors)) {
773 if (best_dist < MaxSector)
774 /* Already have a better slot */
775 continue;
776 if (first_bad <= dev_sector) {
777 /* Cannot read here. If this is the
778 * 'primary' device, then we must not read
779 * beyond 'bad_sectors' from another device.
780 */
781 bad_sectors -= (dev_sector - first_bad);
782 if (!do_balance && sectors > bad_sectors)
783 sectors = bad_sectors;
784 if (best_good_sectors > sectors)
785 best_good_sectors = sectors;
786 } else {
787 sector_t good_sectors =
788 first_bad - dev_sector;
789 if (good_sectors > best_good_sectors) {
790 best_good_sectors = good_sectors;
791 best_dist_slot = slot;
792 best_dist_rdev = rdev;
793 }
794 if (!do_balance)
795 /* Must read from here */
796 break;
797 }
798 continue;
799 } else
800 best_good_sectors = sectors;
801
802 if (!do_balance)
803 break;
804
805 nonrot = bdev_nonrot(rdev->bdev);
806 has_nonrot_disk |= nonrot;
807 pending = atomic_read(&rdev->nr_pending);
808 if (min_pending > pending && nonrot) {
809 min_pending = pending;
810 best_pending_slot = slot;
811 best_pending_rdev = rdev;
812 }
813
814 if (best_dist_slot >= 0)
815 /* At least 2 disks to choose from so failfast is OK */
816 set_bit(R10BIO_FailFast, &r10_bio->state);
817 /* This optimisation is debatable, and completely destroys
818 * sequential read speed for 'far copies' arrays. So only
819 * keep it for 'near' arrays, and review those later.
820 */
821 if (geo->near_copies > 1 && !pending)
822 new_distance = 0;
823
824 /* for far > 1 always use the lowest address */
825 else if (geo->far_copies > 1)
826 new_distance = r10_bio->devs[slot].addr;
827 else
828 new_distance = abs(r10_bio->devs[slot].addr -
829 conf->mirrors[disk].head_position);
830
831 if (new_distance < best_dist) {
832 best_dist = new_distance;
833 best_dist_slot = slot;
834 best_dist_rdev = rdev;
835 }
836 }
837 if (slot >= conf->copies) {
838 if (has_nonrot_disk) {
839 slot = best_pending_slot;
840 rdev = best_pending_rdev;
841 } else {
842 slot = best_dist_slot;
843 rdev = best_dist_rdev;
844 }
845 }
846
847 if (slot >= 0) {
848 atomic_inc(&rdev->nr_pending);
849 r10_bio->read_slot = slot;
850 } else
851 rdev = NULL;
852 *max_sectors = best_good_sectors;
853
854 return rdev;
855}
856
857static void flush_pending_writes(struct r10conf *conf)
858{
859 /* Any writes that have been queued but are awaiting
860 * bitmap updates get flushed here.
861 */
862 spin_lock_irq(&conf->device_lock);
863
864 if (conf->pending_bio_list.head) {
865 struct blk_plug plug;
866 struct bio *bio;
867
868 bio = bio_list_get(&conf->pending_bio_list);
869 spin_unlock_irq(&conf->device_lock);
870
871 /*
872 * As this is called in a wait_event() loop (see freeze_array),
873 * current->state might be TASK_UNINTERRUPTIBLE which will
874 * cause a warning when we prepare to wait again. As it is
875 * rare that this path is taken, it is perfectly safe to force
876 * us to go around the wait_event() loop again, so the warning
877 * is a false-positive. Silence the warning by resetting
878 * thread state
879 */
880 __set_current_state(TASK_RUNNING);
881
882 blk_start_plug(&plug);
883 raid1_prepare_flush_writes(conf->mddev);
884 wake_up(&conf->wait_barrier);
885
886 while (bio) { /* submit pending writes */
887 struct bio *next = bio->bi_next;
888
889 raid1_submit_write(bio);
890 bio = next;
891 cond_resched();
892 }
893 blk_finish_plug(&plug);
894 } else
895 spin_unlock_irq(&conf->device_lock);
896}
897
898/* Barriers....
899 * Sometimes we need to suspend IO while we do something else,
900 * either some resync/recovery, or reconfigure the array.
901 * To do this we raise a 'barrier'.
902 * The 'barrier' is a counter that can be raised multiple times
903 * to count how many activities are happening which preclude
904 * normal IO.
905 * We can only raise the barrier if there is no pending IO.
906 * i.e. if nr_pending == 0.
907 * We choose only to raise the barrier if no-one is waiting for the
908 * barrier to go down. This means that as soon as an IO request
909 * is ready, no other operations which require a barrier will start
910 * until the IO request has had a chance.
911 *
912 * So: regular IO calls 'wait_barrier'. When that returns there
913 * is no backgroup IO happening, It must arrange to call
914 * allow_barrier when it has finished its IO.
915 * backgroup IO calls must call raise_barrier. Once that returns
916 * there is no normal IO happeing. It must arrange to call
917 * lower_barrier when the particular background IO completes.
918 */
919
920static void raise_barrier(struct r10conf *conf, int force)
921{
922 write_seqlock_irq(&conf->resync_lock);
923
924 if (WARN_ON_ONCE(force && !conf->barrier))
925 force = false;
926
927 /* Wait until no block IO is waiting (unless 'force') */
928 wait_event_barrier(conf, force || !conf->nr_waiting);
929
930 /* block any new IO from starting */
931 WRITE_ONCE(conf->barrier, conf->barrier + 1);
932
933 /* Now wait for all pending IO to complete */
934 wait_event_barrier(conf, !atomic_read(&conf->nr_pending) &&
935 conf->barrier < RESYNC_DEPTH);
936
937 write_sequnlock_irq(&conf->resync_lock);
938}
939
940static void lower_barrier(struct r10conf *conf)
941{
942 unsigned long flags;
943
944 write_seqlock_irqsave(&conf->resync_lock, flags);
945 WRITE_ONCE(conf->barrier, conf->barrier - 1);
946 write_sequnlock_irqrestore(&conf->resync_lock, flags);
947 wake_up(&conf->wait_barrier);
948}
949
950static bool stop_waiting_barrier(struct r10conf *conf)
951{
952 struct bio_list *bio_list = current->bio_list;
953 struct md_thread *thread;
954
955 /* barrier is dropped */
956 if (!conf->barrier)
957 return true;
958
959 /*
960 * If there are already pending requests (preventing the barrier from
961 * rising completely), and the pre-process bio queue isn't empty, then
962 * don't wait, as we need to empty that queue to get the nr_pending
963 * count down.
964 */
965 if (atomic_read(&conf->nr_pending) && bio_list &&
966 (!bio_list_empty(&bio_list[0]) || !bio_list_empty(&bio_list[1])))
967 return true;
968
969 /* daemon thread must exist while handling io */
970 thread = rcu_dereference_protected(conf->mddev->thread, true);
971 /*
972 * move on if io is issued from raid10d(), nr_pending is not released
973 * from original io(see handle_read_error()). All raise barrier is
974 * blocked until this io is done.
975 */
976 if (thread->tsk == current) {
977 WARN_ON_ONCE(atomic_read(&conf->nr_pending) == 0);
978 return true;
979 }
980
981 return false;
982}
983
984static bool wait_barrier_nolock(struct r10conf *conf)
985{
986 unsigned int seq = read_seqbegin(&conf->resync_lock);
987
988 if (READ_ONCE(conf->barrier))
989 return false;
990
991 atomic_inc(&conf->nr_pending);
992 if (!read_seqretry(&conf->resync_lock, seq))
993 return true;
994
995 if (atomic_dec_and_test(&conf->nr_pending))
996 wake_up_barrier(conf);
997
998 return false;
999}
1000
1001static bool wait_barrier(struct r10conf *conf, bool nowait)
1002{
1003 bool ret = true;
1004
1005 if (wait_barrier_nolock(conf))
1006 return true;
1007
1008 write_seqlock_irq(&conf->resync_lock);
1009 if (conf->barrier) {
1010 /* Return false when nowait flag is set */
1011 if (nowait) {
1012 ret = false;
1013 } else {
1014 conf->nr_waiting++;
1015 mddev_add_trace_msg(conf->mddev, "raid10 wait barrier");
1016 wait_event_barrier(conf, stop_waiting_barrier(conf));
1017 conf->nr_waiting--;
1018 }
1019 if (!conf->nr_waiting)
1020 wake_up(&conf->wait_barrier);
1021 }
1022 /* Only increment nr_pending when we wait */
1023 if (ret)
1024 atomic_inc(&conf->nr_pending);
1025 write_sequnlock_irq(&conf->resync_lock);
1026 return ret;
1027}
1028
1029static void allow_barrier(struct r10conf *conf)
1030{
1031 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1032 (conf->array_freeze_pending))
1033 wake_up_barrier(conf);
1034}
1035
1036static void freeze_array(struct r10conf *conf, int extra)
1037{
1038 /* stop syncio and normal IO and wait for everything to
1039 * go quiet.
1040 * We increment barrier and nr_waiting, and then
1041 * wait until nr_pending match nr_queued+extra
1042 * This is called in the context of one normal IO request
1043 * that has failed. Thus any sync request that might be pending
1044 * will be blocked by nr_pending, and we need to wait for
1045 * pending IO requests to complete or be queued for re-try.
1046 * Thus the number queued (nr_queued) plus this request (extra)
1047 * must match the number of pending IOs (nr_pending) before
1048 * we continue.
1049 */
1050 write_seqlock_irq(&conf->resync_lock);
1051 conf->array_freeze_pending++;
1052 WRITE_ONCE(conf->barrier, conf->barrier + 1);
1053 conf->nr_waiting++;
1054 wait_event_barrier_cmd(conf, atomic_read(&conf->nr_pending) ==
1055 conf->nr_queued + extra, flush_pending_writes(conf));
1056 conf->array_freeze_pending--;
1057 write_sequnlock_irq(&conf->resync_lock);
1058}
1059
1060static void unfreeze_array(struct r10conf *conf)
1061{
1062 /* reverse the effect of the freeze */
1063 write_seqlock_irq(&conf->resync_lock);
1064 WRITE_ONCE(conf->barrier, conf->barrier - 1);
1065 conf->nr_waiting--;
1066 wake_up(&conf->wait_barrier);
1067 write_sequnlock_irq(&conf->resync_lock);
1068}
1069
1070static sector_t choose_data_offset(struct r10bio *r10_bio,
1071 struct md_rdev *rdev)
1072{
1073 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1074 test_bit(R10BIO_Previous, &r10_bio->state))
1075 return rdev->data_offset;
1076 else
1077 return rdev->new_data_offset;
1078}
1079
1080static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1081{
1082 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1083 struct mddev *mddev = plug->cb.data;
1084 struct r10conf *conf = mddev->private;
1085 struct bio *bio;
1086
1087 if (from_schedule) {
1088 spin_lock_irq(&conf->device_lock);
1089 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1090 spin_unlock_irq(&conf->device_lock);
1091 wake_up_barrier(conf);
1092 md_wakeup_thread(mddev->thread);
1093 kfree(plug);
1094 return;
1095 }
1096
1097 /* we aren't scheduling, so we can do the write-out directly. */
1098 bio = bio_list_get(&plug->pending);
1099 raid1_prepare_flush_writes(mddev);
1100 wake_up_barrier(conf);
1101
1102 while (bio) { /* submit pending writes */
1103 struct bio *next = bio->bi_next;
1104
1105 raid1_submit_write(bio);
1106 bio = next;
1107 cond_resched();
1108 }
1109 kfree(plug);
1110}
1111
1112/*
1113 * 1. Register the new request and wait if the reconstruction thread has put
1114 * up a bar for new requests. Continue immediately if no resync is active
1115 * currently.
1116 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1117 */
1118static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1119 struct bio *bio, sector_t sectors)
1120{
1121 /* Bail out if REQ_NOWAIT is set for the bio */
1122 if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1123 bio_wouldblock_error(bio);
1124 return false;
1125 }
1126 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1127 bio->bi_iter.bi_sector < conf->reshape_progress &&
1128 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1129 allow_barrier(conf);
1130 if (bio->bi_opf & REQ_NOWAIT) {
1131 bio_wouldblock_error(bio);
1132 return false;
1133 }
1134 mddev_add_trace_msg(conf->mddev, "raid10 wait reshape");
1135 wait_event(conf->wait_barrier,
1136 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1137 conf->reshape_progress >= bio->bi_iter.bi_sector +
1138 sectors);
1139 wait_barrier(conf, false);
1140 }
1141 return true;
1142}
1143
1144static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1145 struct r10bio *r10_bio, bool io_accounting)
1146{
1147 struct r10conf *conf = mddev->private;
1148 struct bio *read_bio;
1149 const enum req_op op = bio_op(bio);
1150 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1151 int max_sectors;
1152 struct md_rdev *rdev;
1153 char b[BDEVNAME_SIZE];
1154 int slot = r10_bio->read_slot;
1155 struct md_rdev *err_rdev = NULL;
1156 gfp_t gfp = GFP_NOIO;
1157 int error;
1158
1159 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1160 /*
1161 * This is an error retry, but we cannot
1162 * safely dereference the rdev in the r10_bio,
1163 * we must use the one in conf.
1164 * If it has already been disconnected (unlikely)
1165 * we lose the device name in error messages.
1166 */
1167 int disk;
1168 /*
1169 * As we are blocking raid10, it is a little safer to
1170 * use __GFP_HIGH.
1171 */
1172 gfp = GFP_NOIO | __GFP_HIGH;
1173
1174 disk = r10_bio->devs[slot].devnum;
1175 err_rdev = conf->mirrors[disk].rdev;
1176 if (err_rdev)
1177 snprintf(b, sizeof(b), "%pg", err_rdev->bdev);
1178 else {
1179 strcpy(b, "???");
1180 /* This never gets dereferenced */
1181 err_rdev = r10_bio->devs[slot].rdev;
1182 }
1183 }
1184
1185 if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1186 return;
1187 rdev = read_balance(conf, r10_bio, &max_sectors);
1188 if (!rdev) {
1189 if (err_rdev) {
1190 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1191 mdname(mddev), b,
1192 (unsigned long long)r10_bio->sector);
1193 }
1194 raid_end_bio_io(r10_bio);
1195 return;
1196 }
1197 if (err_rdev)
1198 pr_err_ratelimited("md/raid10:%s: %pg: redirecting sector %llu to another mirror\n",
1199 mdname(mddev),
1200 rdev->bdev,
1201 (unsigned long long)r10_bio->sector);
1202 if (max_sectors < bio_sectors(bio)) {
1203 struct bio *split = bio_split(bio, max_sectors,
1204 gfp, &conf->bio_split);
1205 if (IS_ERR(split)) {
1206 error = PTR_ERR(split);
1207 goto err_handle;
1208 }
1209 bio_chain(split, bio);
1210 allow_barrier(conf);
1211 submit_bio_noacct(bio);
1212 wait_barrier(conf, false);
1213 bio = split;
1214 r10_bio->master_bio = bio;
1215 r10_bio->sectors = max_sectors;
1216 }
1217 slot = r10_bio->read_slot;
1218
1219 if (io_accounting) {
1220 md_account_bio(mddev, &bio);
1221 r10_bio->master_bio = bio;
1222 }
1223 read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1224
1225 r10_bio->devs[slot].bio = read_bio;
1226 r10_bio->devs[slot].rdev = rdev;
1227
1228 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1229 choose_data_offset(r10_bio, rdev);
1230 read_bio->bi_end_io = raid10_end_read_request;
1231 read_bio->bi_opf = op | do_sync;
1232 if (test_bit(FailFast, &rdev->flags) &&
1233 test_bit(R10BIO_FailFast, &r10_bio->state))
1234 read_bio->bi_opf |= MD_FAILFAST;
1235 read_bio->bi_private = r10_bio;
1236 mddev_trace_remap(mddev, read_bio, r10_bio->sector);
1237 submit_bio_noacct(read_bio);
1238 return;
1239err_handle:
1240 atomic_dec(&rdev->nr_pending);
1241 bio->bi_status = errno_to_blk_status(error);
1242 set_bit(R10BIO_Uptodate, &r10_bio->state);
1243 raid_end_bio_io(r10_bio);
1244}
1245
1246static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1247 struct bio *bio, bool replacement,
1248 int n_copy)
1249{
1250 const enum req_op op = bio_op(bio);
1251 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1252 const blk_opf_t do_fua = bio->bi_opf & REQ_FUA;
1253 const blk_opf_t do_atomic = bio->bi_opf & REQ_ATOMIC;
1254 unsigned long flags;
1255 struct r10conf *conf = mddev->private;
1256 struct md_rdev *rdev;
1257 int devnum = r10_bio->devs[n_copy].devnum;
1258 struct bio *mbio;
1259
1260 rdev = replacement ? conf->mirrors[devnum].replacement :
1261 conf->mirrors[devnum].rdev;
1262
1263 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1264 if (replacement)
1265 r10_bio->devs[n_copy].repl_bio = mbio;
1266 else
1267 r10_bio->devs[n_copy].bio = mbio;
1268
1269 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1270 choose_data_offset(r10_bio, rdev));
1271 mbio->bi_end_io = raid10_end_write_request;
1272 mbio->bi_opf = op | do_sync | do_fua | do_atomic;
1273 if (!replacement && test_bit(FailFast,
1274 &conf->mirrors[devnum].rdev->flags)
1275 && enough(conf, devnum))
1276 mbio->bi_opf |= MD_FAILFAST;
1277 mbio->bi_private = r10_bio;
1278 mddev_trace_remap(mddev, mbio, r10_bio->sector);
1279 /* flush_pending_writes() needs access to the rdev so...*/
1280 mbio->bi_bdev = (void *)rdev;
1281
1282 atomic_inc(&r10_bio->remaining);
1283
1284 if (!raid1_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) {
1285 spin_lock_irqsave(&conf->device_lock, flags);
1286 bio_list_add(&conf->pending_bio_list, mbio);
1287 spin_unlock_irqrestore(&conf->device_lock, flags);
1288 md_wakeup_thread(mddev->thread);
1289 }
1290}
1291
1292static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1293{
1294 struct r10conf *conf = mddev->private;
1295 struct md_rdev *blocked_rdev;
1296 int i;
1297
1298retry_wait:
1299 blocked_rdev = NULL;
1300 for (i = 0; i < conf->copies; i++) {
1301 struct md_rdev *rdev, *rrdev;
1302
1303 rdev = conf->mirrors[i].rdev;
1304 if (rdev) {
1305 sector_t dev_sector = r10_bio->devs[i].addr;
1306
1307 /*
1308 * Discard request doesn't care the write result
1309 * so it doesn't need to wait blocked disk here.
1310 */
1311 if (test_bit(WriteErrorSeen, &rdev->flags) &&
1312 r10_bio->sectors &&
1313 rdev_has_badblock(rdev, dev_sector,
1314 r10_bio->sectors) < 0)
1315 /*
1316 * Mustn't write here until the bad
1317 * block is acknowledged
1318 */
1319 set_bit(BlockedBadBlocks, &rdev->flags);
1320
1321 if (rdev_blocked(rdev)) {
1322 blocked_rdev = rdev;
1323 atomic_inc(&rdev->nr_pending);
1324 break;
1325 }
1326 }
1327
1328 rrdev = conf->mirrors[i].replacement;
1329 if (rrdev && rdev_blocked(rrdev)) {
1330 atomic_inc(&rrdev->nr_pending);
1331 blocked_rdev = rrdev;
1332 break;
1333 }
1334 }
1335
1336 if (unlikely(blocked_rdev)) {
1337 /* Have to wait for this device to get unblocked, then retry */
1338 allow_barrier(conf);
1339 mddev_add_trace_msg(conf->mddev,
1340 "raid10 %s wait rdev %d blocked",
1341 __func__, blocked_rdev->raid_disk);
1342 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1343 wait_barrier(conf, false);
1344 goto retry_wait;
1345 }
1346}
1347
1348static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1349 struct r10bio *r10_bio)
1350{
1351 struct r10conf *conf = mddev->private;
1352 int i, k;
1353 sector_t sectors;
1354 int max_sectors;
1355 int error;
1356
1357 if ((mddev_is_clustered(mddev) &&
1358 md_cluster_ops->area_resyncing(mddev, WRITE,
1359 bio->bi_iter.bi_sector,
1360 bio_end_sector(bio)))) {
1361 DEFINE_WAIT(w);
1362 /* Bail out if REQ_NOWAIT is set for the bio */
1363 if (bio->bi_opf & REQ_NOWAIT) {
1364 bio_wouldblock_error(bio);
1365 return;
1366 }
1367 for (;;) {
1368 prepare_to_wait(&conf->wait_barrier,
1369 &w, TASK_IDLE);
1370 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1371 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1372 break;
1373 schedule();
1374 }
1375 finish_wait(&conf->wait_barrier, &w);
1376 }
1377
1378 sectors = r10_bio->sectors;
1379 if (!regular_request_wait(mddev, conf, bio, sectors))
1380 return;
1381 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1382 (mddev->reshape_backwards
1383 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1384 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1385 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1386 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1387 /* Need to update reshape_position in metadata */
1388 mddev->reshape_position = conf->reshape_progress;
1389 set_mask_bits(&mddev->sb_flags, 0,
1390 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1391 md_wakeup_thread(mddev->thread);
1392 if (bio->bi_opf & REQ_NOWAIT) {
1393 allow_barrier(conf);
1394 bio_wouldblock_error(bio);
1395 return;
1396 }
1397 mddev_add_trace_msg(conf->mddev,
1398 "raid10 wait reshape metadata");
1399 wait_event(mddev->sb_wait,
1400 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1401
1402 conf->reshape_safe = mddev->reshape_position;
1403 }
1404
1405 /* first select target devices under rcu_lock and
1406 * inc refcount on their rdev. Record them by setting
1407 * bios[x] to bio
1408 * If there are known/acknowledged bad blocks on any device
1409 * on which we have seen a write error, we want to avoid
1410 * writing to those blocks. This potentially requires several
1411 * writes to write around the bad blocks. Each set of writes
1412 * gets its own r10_bio with a set of bios attached.
1413 */
1414
1415 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1416 raid10_find_phys(conf, r10_bio);
1417
1418 wait_blocked_dev(mddev, r10_bio);
1419
1420 max_sectors = r10_bio->sectors;
1421
1422 for (i = 0; i < conf->copies; i++) {
1423 int d = r10_bio->devs[i].devnum;
1424 struct md_rdev *rdev, *rrdev;
1425
1426 rdev = conf->mirrors[d].rdev;
1427 rrdev = conf->mirrors[d].replacement;
1428 if (rdev && (test_bit(Faulty, &rdev->flags)))
1429 rdev = NULL;
1430 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1431 rrdev = NULL;
1432
1433 r10_bio->devs[i].bio = NULL;
1434 r10_bio->devs[i].repl_bio = NULL;
1435
1436 if (!rdev && !rrdev)
1437 continue;
1438 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1439 sector_t first_bad;
1440 sector_t dev_sector = r10_bio->devs[i].addr;
1441 int bad_sectors;
1442 int is_bad;
1443
1444 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1445 &first_bad, &bad_sectors);
1446 if (is_bad && first_bad <= dev_sector) {
1447 /* Cannot write here at all */
1448 bad_sectors -= (dev_sector - first_bad);
1449 if (bad_sectors < max_sectors)
1450 /* Mustn't write more than bad_sectors
1451 * to other devices yet
1452 */
1453 max_sectors = bad_sectors;
1454 continue;
1455 }
1456 if (is_bad) {
1457 int good_sectors;
1458
1459 /*
1460 * We cannot atomically write this, so just
1461 * error in that case. It could be possible to
1462 * atomically write other mirrors, but the
1463 * complexity of supporting that is not worth
1464 * the benefit.
1465 */
1466 if (bio->bi_opf & REQ_ATOMIC) {
1467 error = -EIO;
1468 goto err_handle;
1469 }
1470
1471 good_sectors = first_bad - dev_sector;
1472 if (good_sectors < max_sectors)
1473 max_sectors = good_sectors;
1474 }
1475 }
1476 if (rdev) {
1477 r10_bio->devs[i].bio = bio;
1478 atomic_inc(&rdev->nr_pending);
1479 }
1480 if (rrdev) {
1481 r10_bio->devs[i].repl_bio = bio;
1482 atomic_inc(&rrdev->nr_pending);
1483 }
1484 }
1485
1486 if (max_sectors < r10_bio->sectors)
1487 r10_bio->sectors = max_sectors;
1488
1489 if (r10_bio->sectors < bio_sectors(bio)) {
1490 struct bio *split = bio_split(bio, r10_bio->sectors,
1491 GFP_NOIO, &conf->bio_split);
1492 if (IS_ERR(split)) {
1493 error = PTR_ERR(split);
1494 goto err_handle;
1495 }
1496 bio_chain(split, bio);
1497 allow_barrier(conf);
1498 submit_bio_noacct(bio);
1499 wait_barrier(conf, false);
1500 bio = split;
1501 r10_bio->master_bio = bio;
1502 }
1503
1504 md_account_bio(mddev, &bio);
1505 r10_bio->master_bio = bio;
1506 atomic_set(&r10_bio->remaining, 1);
1507
1508 for (i = 0; i < conf->copies; i++) {
1509 if (r10_bio->devs[i].bio)
1510 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1511 if (r10_bio->devs[i].repl_bio)
1512 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1513 }
1514 one_write_done(r10_bio);
1515 return;
1516err_handle:
1517 for (k = 0; k < i; k++) {
1518 int d = r10_bio->devs[k].devnum;
1519 struct md_rdev *rdev = conf->mirrors[d].rdev;
1520 struct md_rdev *rrdev = conf->mirrors[d].replacement;
1521
1522 if (r10_bio->devs[k].bio) {
1523 rdev_dec_pending(rdev, mddev);
1524 r10_bio->devs[k].bio = NULL;
1525 }
1526 if (r10_bio->devs[k].repl_bio) {
1527 rdev_dec_pending(rrdev, mddev);
1528 r10_bio->devs[k].repl_bio = NULL;
1529 }
1530 }
1531
1532 bio->bi_status = errno_to_blk_status(error);
1533 set_bit(R10BIO_Uptodate, &r10_bio->state);
1534 raid_end_bio_io(r10_bio);
1535}
1536
1537static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1538{
1539 struct r10conf *conf = mddev->private;
1540 struct r10bio *r10_bio;
1541
1542 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1543
1544 r10_bio->master_bio = bio;
1545 r10_bio->sectors = sectors;
1546
1547 r10_bio->mddev = mddev;
1548 r10_bio->sector = bio->bi_iter.bi_sector;
1549 r10_bio->state = 0;
1550 r10_bio->read_slot = -1;
1551 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1552 conf->geo.raid_disks);
1553
1554 if (bio_data_dir(bio) == READ)
1555 raid10_read_request(mddev, bio, r10_bio, true);
1556 else
1557 raid10_write_request(mddev, bio, r10_bio);
1558}
1559
1560static void raid_end_discard_bio(struct r10bio *r10bio)
1561{
1562 struct r10conf *conf = r10bio->mddev->private;
1563 struct r10bio *first_r10bio;
1564
1565 while (atomic_dec_and_test(&r10bio->remaining)) {
1566
1567 allow_barrier(conf);
1568
1569 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1570 first_r10bio = (struct r10bio *)r10bio->master_bio;
1571 free_r10bio(r10bio);
1572 r10bio = first_r10bio;
1573 } else {
1574 md_write_end(r10bio->mddev);
1575 bio_endio(r10bio->master_bio);
1576 free_r10bio(r10bio);
1577 break;
1578 }
1579 }
1580}
1581
1582static void raid10_end_discard_request(struct bio *bio)
1583{
1584 struct r10bio *r10_bio = bio->bi_private;
1585 struct r10conf *conf = r10_bio->mddev->private;
1586 struct md_rdev *rdev = NULL;
1587 int dev;
1588 int slot, repl;
1589
1590 /*
1591 * We don't care the return value of discard bio
1592 */
1593 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1594 set_bit(R10BIO_Uptodate, &r10_bio->state);
1595
1596 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1597 rdev = repl ? conf->mirrors[dev].replacement :
1598 conf->mirrors[dev].rdev;
1599
1600 raid_end_discard_bio(r10_bio);
1601 rdev_dec_pending(rdev, conf->mddev);
1602}
1603
1604/*
1605 * There are some limitations to handle discard bio
1606 * 1st, the discard size is bigger than stripe_size*2.
1607 * 2st, if the discard bio spans reshape progress, we use the old way to
1608 * handle discard bio
1609 */
1610static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1611{
1612 struct r10conf *conf = mddev->private;
1613 struct geom *geo = &conf->geo;
1614 int far_copies = geo->far_copies;
1615 bool first_copy = true;
1616 struct r10bio *r10_bio, *first_r10bio;
1617 struct bio *split;
1618 int disk;
1619 sector_t chunk;
1620 unsigned int stripe_size;
1621 unsigned int stripe_data_disks;
1622 sector_t split_size;
1623 sector_t bio_start, bio_end;
1624 sector_t first_stripe_index, last_stripe_index;
1625 sector_t start_disk_offset;
1626 unsigned int start_disk_index;
1627 sector_t end_disk_offset;
1628 unsigned int end_disk_index;
1629 unsigned int remainder;
1630
1631 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1632 return -EAGAIN;
1633
1634 if (WARN_ON_ONCE(bio->bi_opf & REQ_NOWAIT)) {
1635 bio_wouldblock_error(bio);
1636 return 0;
1637 }
1638 wait_barrier(conf, false);
1639
1640 /*
1641 * Check reshape again to avoid reshape happens after checking
1642 * MD_RECOVERY_RESHAPE and before wait_barrier
1643 */
1644 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1645 goto out;
1646
1647 if (geo->near_copies)
1648 stripe_data_disks = geo->raid_disks / geo->near_copies +
1649 geo->raid_disks % geo->near_copies;
1650 else
1651 stripe_data_disks = geo->raid_disks;
1652
1653 stripe_size = stripe_data_disks << geo->chunk_shift;
1654
1655 bio_start = bio->bi_iter.bi_sector;
1656 bio_end = bio_end_sector(bio);
1657
1658 /*
1659 * Maybe one discard bio is smaller than strip size or across one
1660 * stripe and discard region is larger than one stripe size. For far
1661 * offset layout, if the discard region is not aligned with stripe
1662 * size, there is hole when we submit discard bio to member disk.
1663 * For simplicity, we only handle discard bio which discard region
1664 * is bigger than stripe_size * 2
1665 */
1666 if (bio_sectors(bio) < stripe_size*2)
1667 goto out;
1668
1669 /*
1670 * Keep bio aligned with strip size.
1671 */
1672 div_u64_rem(bio_start, stripe_size, &remainder);
1673 if (remainder) {
1674 split_size = stripe_size - remainder;
1675 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1676 if (IS_ERR(split)) {
1677 bio->bi_status = errno_to_blk_status(PTR_ERR(split));
1678 bio_endio(bio);
1679 return 0;
1680 }
1681 bio_chain(split, bio);
1682 allow_barrier(conf);
1683 /* Resend the fist split part */
1684 submit_bio_noacct(split);
1685 wait_barrier(conf, false);
1686 }
1687 div_u64_rem(bio_end, stripe_size, &remainder);
1688 if (remainder) {
1689 split_size = bio_sectors(bio) - remainder;
1690 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1691 if (IS_ERR(split)) {
1692 bio->bi_status = errno_to_blk_status(PTR_ERR(split));
1693 bio_endio(bio);
1694 return 0;
1695 }
1696 bio_chain(split, bio);
1697 allow_barrier(conf);
1698 /* Resend the second split part */
1699 submit_bio_noacct(bio);
1700 bio = split;
1701 wait_barrier(conf, false);
1702 }
1703
1704 bio_start = bio->bi_iter.bi_sector;
1705 bio_end = bio_end_sector(bio);
1706
1707 /*
1708 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1709 * One stripe contains the chunks from all member disk (one chunk from
1710 * one disk at the same HBA address). For layout detail, see 'man md 4'
1711 */
1712 chunk = bio_start >> geo->chunk_shift;
1713 chunk *= geo->near_copies;
1714 first_stripe_index = chunk;
1715 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1716 if (geo->far_offset)
1717 first_stripe_index *= geo->far_copies;
1718 start_disk_offset = (bio_start & geo->chunk_mask) +
1719 (first_stripe_index << geo->chunk_shift);
1720
1721 chunk = bio_end >> geo->chunk_shift;
1722 chunk *= geo->near_copies;
1723 last_stripe_index = chunk;
1724 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1725 if (geo->far_offset)
1726 last_stripe_index *= geo->far_copies;
1727 end_disk_offset = (bio_end & geo->chunk_mask) +
1728 (last_stripe_index << geo->chunk_shift);
1729
1730retry_discard:
1731 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1732 r10_bio->mddev = mddev;
1733 r10_bio->state = 0;
1734 r10_bio->sectors = 0;
1735 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1736 wait_blocked_dev(mddev, r10_bio);
1737
1738 /*
1739 * For far layout it needs more than one r10bio to cover all regions.
1740 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1741 * to record the discard bio. Other r10bio->master_bio record the first
1742 * r10bio. The first r10bio only release after all other r10bios finish.
1743 * The discard bio returns only first r10bio finishes
1744 */
1745 if (first_copy) {
1746 r10_bio->master_bio = bio;
1747 set_bit(R10BIO_Discard, &r10_bio->state);
1748 first_copy = false;
1749 first_r10bio = r10_bio;
1750 } else
1751 r10_bio->master_bio = (struct bio *)first_r10bio;
1752
1753 /*
1754 * first select target devices under rcu_lock and
1755 * inc refcount on their rdev. Record them by setting
1756 * bios[x] to bio
1757 */
1758 for (disk = 0; disk < geo->raid_disks; disk++) {
1759 struct md_rdev *rdev, *rrdev;
1760
1761 rdev = conf->mirrors[disk].rdev;
1762 rrdev = conf->mirrors[disk].replacement;
1763 r10_bio->devs[disk].bio = NULL;
1764 r10_bio->devs[disk].repl_bio = NULL;
1765
1766 if (rdev && (test_bit(Faulty, &rdev->flags)))
1767 rdev = NULL;
1768 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1769 rrdev = NULL;
1770 if (!rdev && !rrdev)
1771 continue;
1772
1773 if (rdev) {
1774 r10_bio->devs[disk].bio = bio;
1775 atomic_inc(&rdev->nr_pending);
1776 }
1777 if (rrdev) {
1778 r10_bio->devs[disk].repl_bio = bio;
1779 atomic_inc(&rrdev->nr_pending);
1780 }
1781 }
1782
1783 atomic_set(&r10_bio->remaining, 1);
1784 for (disk = 0; disk < geo->raid_disks; disk++) {
1785 sector_t dev_start, dev_end;
1786 struct bio *mbio, *rbio = NULL;
1787
1788 /*
1789 * Now start to calculate the start and end address for each disk.
1790 * The space between dev_start and dev_end is the discard region.
1791 *
1792 * For dev_start, it needs to consider three conditions:
1793 * 1st, the disk is before start_disk, you can imagine the disk in
1794 * the next stripe. So the dev_start is the start address of next
1795 * stripe.
1796 * 2st, the disk is after start_disk, it means the disk is at the
1797 * same stripe of first disk
1798 * 3st, the first disk itself, we can use start_disk_offset directly
1799 */
1800 if (disk < start_disk_index)
1801 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1802 else if (disk > start_disk_index)
1803 dev_start = first_stripe_index * mddev->chunk_sectors;
1804 else
1805 dev_start = start_disk_offset;
1806
1807 if (disk < end_disk_index)
1808 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1809 else if (disk > end_disk_index)
1810 dev_end = last_stripe_index * mddev->chunk_sectors;
1811 else
1812 dev_end = end_disk_offset;
1813
1814 /*
1815 * It only handles discard bio which size is >= stripe size, so
1816 * dev_end > dev_start all the time.
1817 * It doesn't need to use rcu lock to get rdev here. We already
1818 * add rdev->nr_pending in the first loop.
1819 */
1820 if (r10_bio->devs[disk].bio) {
1821 struct md_rdev *rdev = conf->mirrors[disk].rdev;
1822 mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1823 &mddev->bio_set);
1824 mbio->bi_end_io = raid10_end_discard_request;
1825 mbio->bi_private = r10_bio;
1826 r10_bio->devs[disk].bio = mbio;
1827 r10_bio->devs[disk].devnum = disk;
1828 atomic_inc(&r10_bio->remaining);
1829 md_submit_discard_bio(mddev, rdev, mbio,
1830 dev_start + choose_data_offset(r10_bio, rdev),
1831 dev_end - dev_start);
1832 bio_endio(mbio);
1833 }
1834 if (r10_bio->devs[disk].repl_bio) {
1835 struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1836 rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1837 &mddev->bio_set);
1838 rbio->bi_end_io = raid10_end_discard_request;
1839 rbio->bi_private = r10_bio;
1840 r10_bio->devs[disk].repl_bio = rbio;
1841 r10_bio->devs[disk].devnum = disk;
1842 atomic_inc(&r10_bio->remaining);
1843 md_submit_discard_bio(mddev, rrdev, rbio,
1844 dev_start + choose_data_offset(r10_bio, rrdev),
1845 dev_end - dev_start);
1846 bio_endio(rbio);
1847 }
1848 }
1849
1850 if (!geo->far_offset && --far_copies) {
1851 first_stripe_index += geo->stride >> geo->chunk_shift;
1852 start_disk_offset += geo->stride;
1853 last_stripe_index += geo->stride >> geo->chunk_shift;
1854 end_disk_offset += geo->stride;
1855 atomic_inc(&first_r10bio->remaining);
1856 raid_end_discard_bio(r10_bio);
1857 wait_barrier(conf, false);
1858 goto retry_discard;
1859 }
1860
1861 raid_end_discard_bio(r10_bio);
1862
1863 return 0;
1864out:
1865 allow_barrier(conf);
1866 return -EAGAIN;
1867}
1868
1869static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1870{
1871 struct r10conf *conf = mddev->private;
1872 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1873 int chunk_sects = chunk_mask + 1;
1874 int sectors = bio_sectors(bio);
1875
1876 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1877 && md_flush_request(mddev, bio))
1878 return true;
1879
1880 md_write_start(mddev, bio);
1881
1882 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1883 if (!raid10_handle_discard(mddev, bio))
1884 return true;
1885
1886 /*
1887 * If this request crosses a chunk boundary, we need to split
1888 * it.
1889 */
1890 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1891 sectors > chunk_sects
1892 && (conf->geo.near_copies < conf->geo.raid_disks
1893 || conf->prev.near_copies <
1894 conf->prev.raid_disks)))
1895 sectors = chunk_sects -
1896 (bio->bi_iter.bi_sector &
1897 (chunk_sects - 1));
1898 __make_request(mddev, bio, sectors);
1899
1900 /* In case raid10d snuck in to freeze_array */
1901 wake_up_barrier(conf);
1902 return true;
1903}
1904
1905static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1906{
1907 struct r10conf *conf = mddev->private;
1908 int i;
1909
1910 lockdep_assert_held(&mddev->lock);
1911
1912 if (conf->geo.near_copies < conf->geo.raid_disks)
1913 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1914 if (conf->geo.near_copies > 1)
1915 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1916 if (conf->geo.far_copies > 1) {
1917 if (conf->geo.far_offset)
1918 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1919 else
1920 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1921 if (conf->geo.far_set_size != conf->geo.raid_disks)
1922 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1923 }
1924 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1925 conf->geo.raid_disks - mddev->degraded);
1926 for (i = 0; i < conf->geo.raid_disks; i++) {
1927 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1928
1929 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1930 }
1931 seq_printf(seq, "]");
1932}
1933
1934/* check if there are enough drives for
1935 * every block to appear on atleast one.
1936 * Don't consider the device numbered 'ignore'
1937 * as we might be about to remove it.
1938 */
1939static int _enough(struct r10conf *conf, int previous, int ignore)
1940{
1941 int first = 0;
1942 int has_enough = 0;
1943 int disks, ncopies;
1944 if (previous) {
1945 disks = conf->prev.raid_disks;
1946 ncopies = conf->prev.near_copies;
1947 } else {
1948 disks = conf->geo.raid_disks;
1949 ncopies = conf->geo.near_copies;
1950 }
1951
1952 do {
1953 int n = conf->copies;
1954 int cnt = 0;
1955 int this = first;
1956 while (n--) {
1957 struct md_rdev *rdev;
1958 if (this != ignore &&
1959 (rdev = conf->mirrors[this].rdev) &&
1960 test_bit(In_sync, &rdev->flags))
1961 cnt++;
1962 this = (this+1) % disks;
1963 }
1964 if (cnt == 0)
1965 goto out;
1966 first = (first + ncopies) % disks;
1967 } while (first != 0);
1968 has_enough = 1;
1969out:
1970 return has_enough;
1971}
1972
1973static int enough(struct r10conf *conf, int ignore)
1974{
1975 /* when calling 'enough', both 'prev' and 'geo' must
1976 * be stable.
1977 * This is ensured if ->reconfig_mutex or ->device_lock
1978 * is held.
1979 */
1980 return _enough(conf, 0, ignore) &&
1981 _enough(conf, 1, ignore);
1982}
1983
1984/**
1985 * raid10_error() - RAID10 error handler.
1986 * @mddev: affected md device.
1987 * @rdev: member device to fail.
1988 *
1989 * The routine acknowledges &rdev failure and determines new @mddev state.
1990 * If it failed, then:
1991 * - &MD_BROKEN flag is set in &mddev->flags.
1992 * Otherwise, it must be degraded:
1993 * - recovery is interrupted.
1994 * - &mddev->degraded is bumped.
1995 *
1996 * @rdev is marked as &Faulty excluding case when array is failed and
1997 * &mddev->fail_last_dev is off.
1998 */
1999static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
2000{
2001 struct r10conf *conf = mddev->private;
2002 unsigned long flags;
2003
2004 spin_lock_irqsave(&conf->device_lock, flags);
2005
2006 if (test_bit(In_sync, &rdev->flags) && !enough(conf, rdev->raid_disk)) {
2007 set_bit(MD_BROKEN, &mddev->flags);
2008
2009 if (!mddev->fail_last_dev) {
2010 spin_unlock_irqrestore(&conf->device_lock, flags);
2011 return;
2012 }
2013 }
2014 if (test_and_clear_bit(In_sync, &rdev->flags))
2015 mddev->degraded++;
2016
2017 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2018 set_bit(Blocked, &rdev->flags);
2019 set_bit(Faulty, &rdev->flags);
2020 set_mask_bits(&mddev->sb_flags, 0,
2021 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2022 spin_unlock_irqrestore(&conf->device_lock, flags);
2023 pr_crit("md/raid10:%s: Disk failure on %pg, disabling device.\n"
2024 "md/raid10:%s: Operation continuing on %d devices.\n",
2025 mdname(mddev), rdev->bdev,
2026 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
2027}
2028
2029static void print_conf(struct r10conf *conf)
2030{
2031 int i;
2032 struct md_rdev *rdev;
2033
2034 pr_debug("RAID10 conf printout:\n");
2035 if (!conf) {
2036 pr_debug("(!conf)\n");
2037 return;
2038 }
2039 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2040 conf->geo.raid_disks);
2041
2042 lockdep_assert_held(&conf->mddev->reconfig_mutex);
2043 for (i = 0; i < conf->geo.raid_disks; i++) {
2044 rdev = conf->mirrors[i].rdev;
2045 if (rdev)
2046 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
2047 i, !test_bit(In_sync, &rdev->flags),
2048 !test_bit(Faulty, &rdev->flags),
2049 rdev->bdev);
2050 }
2051}
2052
2053static void close_sync(struct r10conf *conf)
2054{
2055 wait_barrier(conf, false);
2056 allow_barrier(conf);
2057
2058 mempool_exit(&conf->r10buf_pool);
2059}
2060
2061static int raid10_spare_active(struct mddev *mddev)
2062{
2063 int i;
2064 struct r10conf *conf = mddev->private;
2065 struct raid10_info *tmp;
2066 int count = 0;
2067 unsigned long flags;
2068
2069 /*
2070 * Find all non-in_sync disks within the RAID10 configuration
2071 * and mark them in_sync
2072 */
2073 for (i = 0; i < conf->geo.raid_disks; i++) {
2074 tmp = conf->mirrors + i;
2075 if (tmp->replacement
2076 && tmp->replacement->recovery_offset == MaxSector
2077 && !test_bit(Faulty, &tmp->replacement->flags)
2078 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2079 /* Replacement has just become active */
2080 if (!tmp->rdev
2081 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2082 count++;
2083 if (tmp->rdev) {
2084 /* Replaced device not technically faulty,
2085 * but we need to be sure it gets removed
2086 * and never re-added.
2087 */
2088 set_bit(Faulty, &tmp->rdev->flags);
2089 sysfs_notify_dirent_safe(
2090 tmp->rdev->sysfs_state);
2091 }
2092 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2093 } else if (tmp->rdev
2094 && tmp->rdev->recovery_offset == MaxSector
2095 && !test_bit(Faulty, &tmp->rdev->flags)
2096 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2097 count++;
2098 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2099 }
2100 }
2101 spin_lock_irqsave(&conf->device_lock, flags);
2102 mddev->degraded -= count;
2103 spin_unlock_irqrestore(&conf->device_lock, flags);
2104
2105 print_conf(conf);
2106 return count;
2107}
2108
2109static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2110{
2111 struct r10conf *conf = mddev->private;
2112 int err = -EEXIST;
2113 int mirror, repl_slot = -1;
2114 int first = 0;
2115 int last = conf->geo.raid_disks - 1;
2116 struct raid10_info *p;
2117
2118 if (mddev->recovery_cp < MaxSector)
2119 /* only hot-add to in-sync arrays, as recovery is
2120 * very different from resync
2121 */
2122 return -EBUSY;
2123 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2124 return -EINVAL;
2125
2126 if (rdev->raid_disk >= 0)
2127 first = last = rdev->raid_disk;
2128
2129 if (rdev->saved_raid_disk >= first &&
2130 rdev->saved_raid_disk < conf->geo.raid_disks &&
2131 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2132 mirror = rdev->saved_raid_disk;
2133 else
2134 mirror = first;
2135 for ( ; mirror <= last ; mirror++) {
2136 p = &conf->mirrors[mirror];
2137 if (p->recovery_disabled == mddev->recovery_disabled)
2138 continue;
2139 if (p->rdev) {
2140 if (test_bit(WantReplacement, &p->rdev->flags) &&
2141 p->replacement == NULL && repl_slot < 0)
2142 repl_slot = mirror;
2143 continue;
2144 }
2145
2146 err = mddev_stack_new_rdev(mddev, rdev);
2147 if (err)
2148 return err;
2149 p->head_position = 0;
2150 p->recovery_disabled = mddev->recovery_disabled - 1;
2151 rdev->raid_disk = mirror;
2152 err = 0;
2153 if (rdev->saved_raid_disk != mirror)
2154 conf->fullsync = 1;
2155 WRITE_ONCE(p->rdev, rdev);
2156 break;
2157 }
2158
2159 if (err && repl_slot >= 0) {
2160 p = &conf->mirrors[repl_slot];
2161 clear_bit(In_sync, &rdev->flags);
2162 set_bit(Replacement, &rdev->flags);
2163 rdev->raid_disk = repl_slot;
2164 err = mddev_stack_new_rdev(mddev, rdev);
2165 if (err)
2166 return err;
2167 conf->fullsync = 1;
2168 WRITE_ONCE(p->replacement, rdev);
2169 }
2170
2171 print_conf(conf);
2172 return err;
2173}
2174
2175static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2176{
2177 struct r10conf *conf = mddev->private;
2178 int err = 0;
2179 int number = rdev->raid_disk;
2180 struct md_rdev **rdevp;
2181 struct raid10_info *p;
2182
2183 print_conf(conf);
2184 if (unlikely(number >= mddev->raid_disks))
2185 return 0;
2186 p = conf->mirrors + number;
2187 if (rdev == p->rdev)
2188 rdevp = &p->rdev;
2189 else if (rdev == p->replacement)
2190 rdevp = &p->replacement;
2191 else
2192 return 0;
2193
2194 if (test_bit(In_sync, &rdev->flags) ||
2195 atomic_read(&rdev->nr_pending)) {
2196 err = -EBUSY;
2197 goto abort;
2198 }
2199 /* Only remove non-faulty devices if recovery
2200 * is not possible.
2201 */
2202 if (!test_bit(Faulty, &rdev->flags) &&
2203 mddev->recovery_disabled != p->recovery_disabled &&
2204 (!p->replacement || p->replacement == rdev) &&
2205 number < conf->geo.raid_disks &&
2206 enough(conf, -1)) {
2207 err = -EBUSY;
2208 goto abort;
2209 }
2210 WRITE_ONCE(*rdevp, NULL);
2211 if (p->replacement) {
2212 /* We must have just cleared 'rdev' */
2213 WRITE_ONCE(p->rdev, p->replacement);
2214 clear_bit(Replacement, &p->replacement->flags);
2215 WRITE_ONCE(p->replacement, NULL);
2216 }
2217
2218 clear_bit(WantReplacement, &rdev->flags);
2219 err = md_integrity_register(mddev);
2220
2221abort:
2222
2223 print_conf(conf);
2224 return err;
2225}
2226
2227static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2228{
2229 struct r10conf *conf = r10_bio->mddev->private;
2230
2231 if (!bio->bi_status)
2232 set_bit(R10BIO_Uptodate, &r10_bio->state);
2233 else
2234 /* The write handler will notice the lack of
2235 * R10BIO_Uptodate and record any errors etc
2236 */
2237 atomic_add(r10_bio->sectors,
2238 &conf->mirrors[d].rdev->corrected_errors);
2239
2240 /* for reconstruct, we always reschedule after a read.
2241 * for resync, only after all reads
2242 */
2243 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2244 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2245 atomic_dec_and_test(&r10_bio->remaining)) {
2246 /* we have read all the blocks,
2247 * do the comparison in process context in raid10d
2248 */
2249 reschedule_retry(r10_bio);
2250 }
2251}
2252
2253static void end_sync_read(struct bio *bio)
2254{
2255 struct r10bio *r10_bio = get_resync_r10bio(bio);
2256 struct r10conf *conf = r10_bio->mddev->private;
2257 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2258
2259 __end_sync_read(r10_bio, bio, d);
2260}
2261
2262static void end_reshape_read(struct bio *bio)
2263{
2264 /* reshape read bio isn't allocated from r10buf_pool */
2265 struct r10bio *r10_bio = bio->bi_private;
2266
2267 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2268}
2269
2270static void end_sync_request(struct r10bio *r10_bio)
2271{
2272 struct mddev *mddev = r10_bio->mddev;
2273
2274 while (atomic_dec_and_test(&r10_bio->remaining)) {
2275 if (r10_bio->master_bio == NULL) {
2276 /* the primary of several recovery bios */
2277 sector_t s = r10_bio->sectors;
2278 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2279 test_bit(R10BIO_WriteError, &r10_bio->state))
2280 reschedule_retry(r10_bio);
2281 else
2282 put_buf(r10_bio);
2283 md_done_sync(mddev, s, 1);
2284 break;
2285 } else {
2286 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2287 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2288 test_bit(R10BIO_WriteError, &r10_bio->state))
2289 reschedule_retry(r10_bio);
2290 else
2291 put_buf(r10_bio);
2292 r10_bio = r10_bio2;
2293 }
2294 }
2295}
2296
2297static void end_sync_write(struct bio *bio)
2298{
2299 struct r10bio *r10_bio = get_resync_r10bio(bio);
2300 struct mddev *mddev = r10_bio->mddev;
2301 struct r10conf *conf = mddev->private;
2302 int d;
2303 int slot;
2304 int repl;
2305 struct md_rdev *rdev = NULL;
2306
2307 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2308 if (repl)
2309 rdev = conf->mirrors[d].replacement;
2310 else
2311 rdev = conf->mirrors[d].rdev;
2312
2313 if (bio->bi_status) {
2314 if (repl)
2315 md_error(mddev, rdev);
2316 else {
2317 set_bit(WriteErrorSeen, &rdev->flags);
2318 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2319 set_bit(MD_RECOVERY_NEEDED,
2320 &rdev->mddev->recovery);
2321 set_bit(R10BIO_WriteError, &r10_bio->state);
2322 }
2323 } else if (rdev_has_badblock(rdev, r10_bio->devs[slot].addr,
2324 r10_bio->sectors)) {
2325 set_bit(R10BIO_MadeGood, &r10_bio->state);
2326 }
2327
2328 rdev_dec_pending(rdev, mddev);
2329
2330 end_sync_request(r10_bio);
2331}
2332
2333/*
2334 * Note: sync and recover and handled very differently for raid10
2335 * This code is for resync.
2336 * For resync, we read through virtual addresses and read all blocks.
2337 * If there is any error, we schedule a write. The lowest numbered
2338 * drive is authoritative.
2339 * However requests come for physical address, so we need to map.
2340 * For every physical address there are raid_disks/copies virtual addresses,
2341 * which is always are least one, but is not necessarly an integer.
2342 * This means that a physical address can span multiple chunks, so we may
2343 * have to submit multiple io requests for a single sync request.
2344 */
2345/*
2346 * We check if all blocks are in-sync and only write to blocks that
2347 * aren't in sync
2348 */
2349static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2350{
2351 struct r10conf *conf = mddev->private;
2352 int i, first;
2353 struct bio *tbio, *fbio;
2354 int vcnt;
2355 struct page **tpages, **fpages;
2356
2357 atomic_set(&r10_bio->remaining, 1);
2358
2359 /* find the first device with a block */
2360 for (i=0; i<conf->copies; i++)
2361 if (!r10_bio->devs[i].bio->bi_status)
2362 break;
2363
2364 if (i == conf->copies)
2365 goto done;
2366
2367 first = i;
2368 fbio = r10_bio->devs[i].bio;
2369 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2370 fbio->bi_iter.bi_idx = 0;
2371 fpages = get_resync_pages(fbio)->pages;
2372
2373 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2374 /* now find blocks with errors */
2375 for (i=0 ; i < conf->copies ; i++) {
2376 int j, d;
2377 struct md_rdev *rdev;
2378 struct resync_pages *rp;
2379
2380 tbio = r10_bio->devs[i].bio;
2381
2382 if (tbio->bi_end_io != end_sync_read)
2383 continue;
2384 if (i == first)
2385 continue;
2386
2387 tpages = get_resync_pages(tbio)->pages;
2388 d = r10_bio->devs[i].devnum;
2389 rdev = conf->mirrors[d].rdev;
2390 if (!r10_bio->devs[i].bio->bi_status) {
2391 /* We know that the bi_io_vec layout is the same for
2392 * both 'first' and 'i', so we just compare them.
2393 * All vec entries are PAGE_SIZE;
2394 */
2395 int sectors = r10_bio->sectors;
2396 for (j = 0; j < vcnt; j++) {
2397 int len = PAGE_SIZE;
2398 if (sectors < (len / 512))
2399 len = sectors * 512;
2400 if (memcmp(page_address(fpages[j]),
2401 page_address(tpages[j]),
2402 len))
2403 break;
2404 sectors -= len/512;
2405 }
2406 if (j == vcnt)
2407 continue;
2408 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2409 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2410 /* Don't fix anything. */
2411 continue;
2412 } else if (test_bit(FailFast, &rdev->flags)) {
2413 /* Just give up on this device */
2414 md_error(rdev->mddev, rdev);
2415 continue;
2416 }
2417 /* Ok, we need to write this bio, either to correct an
2418 * inconsistency or to correct an unreadable block.
2419 * First we need to fixup bv_offset, bv_len and
2420 * bi_vecs, as the read request might have corrupted these
2421 */
2422 rp = get_resync_pages(tbio);
2423 bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2424
2425 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2426
2427 rp->raid_bio = r10_bio;
2428 tbio->bi_private = rp;
2429 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2430 tbio->bi_end_io = end_sync_write;
2431
2432 bio_copy_data(tbio, fbio);
2433
2434 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2435 atomic_inc(&r10_bio->remaining);
2436 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2437
2438 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2439 tbio->bi_opf |= MD_FAILFAST;
2440 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2441 submit_bio_noacct(tbio);
2442 }
2443
2444 /* Now write out to any replacement devices
2445 * that are active
2446 */
2447 for (i = 0; i < conf->copies; i++) {
2448 int d;
2449
2450 tbio = r10_bio->devs[i].repl_bio;
2451 if (!tbio || !tbio->bi_end_io)
2452 continue;
2453 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2454 && r10_bio->devs[i].bio != fbio)
2455 bio_copy_data(tbio, fbio);
2456 d = r10_bio->devs[i].devnum;
2457 atomic_inc(&r10_bio->remaining);
2458 md_sync_acct(conf->mirrors[d].replacement->bdev,
2459 bio_sectors(tbio));
2460 submit_bio_noacct(tbio);
2461 }
2462
2463done:
2464 if (atomic_dec_and_test(&r10_bio->remaining)) {
2465 md_done_sync(mddev, r10_bio->sectors, 1);
2466 put_buf(r10_bio);
2467 }
2468}
2469
2470/*
2471 * Now for the recovery code.
2472 * Recovery happens across physical sectors.
2473 * We recover all non-is_sync drives by finding the virtual address of
2474 * each, and then choose a working drive that also has that virt address.
2475 * There is a separate r10_bio for each non-in_sync drive.
2476 * Only the first two slots are in use. The first for reading,
2477 * The second for writing.
2478 *
2479 */
2480static void fix_recovery_read_error(struct r10bio *r10_bio)
2481{
2482 /* We got a read error during recovery.
2483 * We repeat the read in smaller page-sized sections.
2484 * If a read succeeds, write it to the new device or record
2485 * a bad block if we cannot.
2486 * If a read fails, record a bad block on both old and
2487 * new devices.
2488 */
2489 struct mddev *mddev = r10_bio->mddev;
2490 struct r10conf *conf = mddev->private;
2491 struct bio *bio = r10_bio->devs[0].bio;
2492 sector_t sect = 0;
2493 int sectors = r10_bio->sectors;
2494 int idx = 0;
2495 int dr = r10_bio->devs[0].devnum;
2496 int dw = r10_bio->devs[1].devnum;
2497 struct page **pages = get_resync_pages(bio)->pages;
2498
2499 while (sectors) {
2500 int s = sectors;
2501 struct md_rdev *rdev;
2502 sector_t addr;
2503 int ok;
2504
2505 if (s > (PAGE_SIZE>>9))
2506 s = PAGE_SIZE >> 9;
2507
2508 rdev = conf->mirrors[dr].rdev;
2509 addr = r10_bio->devs[0].addr + sect;
2510 ok = sync_page_io(rdev,
2511 addr,
2512 s << 9,
2513 pages[idx],
2514 REQ_OP_READ, false);
2515 if (ok) {
2516 rdev = conf->mirrors[dw].rdev;
2517 addr = r10_bio->devs[1].addr + sect;
2518 ok = sync_page_io(rdev,
2519 addr,
2520 s << 9,
2521 pages[idx],
2522 REQ_OP_WRITE, false);
2523 if (!ok) {
2524 set_bit(WriteErrorSeen, &rdev->flags);
2525 if (!test_and_set_bit(WantReplacement,
2526 &rdev->flags))
2527 set_bit(MD_RECOVERY_NEEDED,
2528 &rdev->mddev->recovery);
2529 }
2530 }
2531 if (!ok) {
2532 /* We don't worry if we cannot set a bad block -
2533 * it really is bad so there is no loss in not
2534 * recording it yet
2535 */
2536 rdev_set_badblocks(rdev, addr, s, 0);
2537
2538 if (rdev != conf->mirrors[dw].rdev) {
2539 /* need bad block on destination too */
2540 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2541 addr = r10_bio->devs[1].addr + sect;
2542 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2543 if (!ok) {
2544 /* just abort the recovery */
2545 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2546 mdname(mddev));
2547
2548 conf->mirrors[dw].recovery_disabled
2549 = mddev->recovery_disabled;
2550 set_bit(MD_RECOVERY_INTR,
2551 &mddev->recovery);
2552 break;
2553 }
2554 }
2555 }
2556
2557 sectors -= s;
2558 sect += s;
2559 idx++;
2560 }
2561}
2562
2563static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2564{
2565 struct r10conf *conf = mddev->private;
2566 int d;
2567 struct bio *wbio = r10_bio->devs[1].bio;
2568 struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2569
2570 /* Need to test wbio2->bi_end_io before we call
2571 * submit_bio_noacct as if the former is NULL,
2572 * the latter is free to free wbio2.
2573 */
2574 if (wbio2 && !wbio2->bi_end_io)
2575 wbio2 = NULL;
2576
2577 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2578 fix_recovery_read_error(r10_bio);
2579 if (wbio->bi_end_io)
2580 end_sync_request(r10_bio);
2581 if (wbio2)
2582 end_sync_request(r10_bio);
2583 return;
2584 }
2585
2586 /*
2587 * share the pages with the first bio
2588 * and submit the write request
2589 */
2590 d = r10_bio->devs[1].devnum;
2591 if (wbio->bi_end_io) {
2592 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2593 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2594 submit_bio_noacct(wbio);
2595 }
2596 if (wbio2) {
2597 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2598 md_sync_acct(conf->mirrors[d].replacement->bdev,
2599 bio_sectors(wbio2));
2600 submit_bio_noacct(wbio2);
2601 }
2602}
2603
2604static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2605 int sectors, struct page *page, enum req_op op)
2606{
2607 if (rdev_has_badblock(rdev, sector, sectors) &&
2608 (op == REQ_OP_READ || test_bit(WriteErrorSeen, &rdev->flags)))
2609 return -1;
2610 if (sync_page_io(rdev, sector, sectors << 9, page, op, false))
2611 /* success */
2612 return 1;
2613 if (op == REQ_OP_WRITE) {
2614 set_bit(WriteErrorSeen, &rdev->flags);
2615 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2616 set_bit(MD_RECOVERY_NEEDED,
2617 &rdev->mddev->recovery);
2618 }
2619 /* need to record an error - either for the block or the device */
2620 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2621 md_error(rdev->mddev, rdev);
2622 return 0;
2623}
2624
2625/*
2626 * This is a kernel thread which:
2627 *
2628 * 1. Retries failed read operations on working mirrors.
2629 * 2. Updates the raid superblock when problems encounter.
2630 * 3. Performs writes following reads for array synchronising.
2631 */
2632
2633static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2634{
2635 int sect = 0; /* Offset from r10_bio->sector */
2636 int sectors = r10_bio->sectors, slot = r10_bio->read_slot;
2637 struct md_rdev *rdev;
2638 int d = r10_bio->devs[slot].devnum;
2639
2640 /* still own a reference to this rdev, so it cannot
2641 * have been cleared recently.
2642 */
2643 rdev = conf->mirrors[d].rdev;
2644
2645 if (test_bit(Faulty, &rdev->flags))
2646 /* drive has already been failed, just ignore any
2647 more fix_read_error() attempts */
2648 return;
2649
2650 if (exceed_read_errors(mddev, rdev)) {
2651 r10_bio->devs[slot].bio = IO_BLOCKED;
2652 return;
2653 }
2654
2655 while(sectors) {
2656 int s = sectors;
2657 int sl = slot;
2658 int success = 0;
2659 int start;
2660
2661 if (s > (PAGE_SIZE>>9))
2662 s = PAGE_SIZE >> 9;
2663
2664 do {
2665 d = r10_bio->devs[sl].devnum;
2666 rdev = conf->mirrors[d].rdev;
2667 if (rdev &&
2668 test_bit(In_sync, &rdev->flags) &&
2669 !test_bit(Faulty, &rdev->flags) &&
2670 rdev_has_badblock(rdev,
2671 r10_bio->devs[sl].addr + sect,
2672 s) == 0) {
2673 atomic_inc(&rdev->nr_pending);
2674 success = sync_page_io(rdev,
2675 r10_bio->devs[sl].addr +
2676 sect,
2677 s<<9,
2678 conf->tmppage,
2679 REQ_OP_READ, false);
2680 rdev_dec_pending(rdev, mddev);
2681 if (success)
2682 break;
2683 }
2684 sl++;
2685 if (sl == conf->copies)
2686 sl = 0;
2687 } while (sl != slot);
2688
2689 if (!success) {
2690 /* Cannot read from anywhere, just mark the block
2691 * as bad on the first device to discourage future
2692 * reads.
2693 */
2694 int dn = r10_bio->devs[slot].devnum;
2695 rdev = conf->mirrors[dn].rdev;
2696
2697 if (!rdev_set_badblocks(
2698 rdev,
2699 r10_bio->devs[slot].addr
2700 + sect,
2701 s, 0)) {
2702 md_error(mddev, rdev);
2703 r10_bio->devs[slot].bio
2704 = IO_BLOCKED;
2705 }
2706 break;
2707 }
2708
2709 start = sl;
2710 /* write it back and re-read */
2711 while (sl != slot) {
2712 if (sl==0)
2713 sl = conf->copies;
2714 sl--;
2715 d = r10_bio->devs[sl].devnum;
2716 rdev = conf->mirrors[d].rdev;
2717 if (!rdev ||
2718 test_bit(Faulty, &rdev->flags) ||
2719 !test_bit(In_sync, &rdev->flags))
2720 continue;
2721
2722 atomic_inc(&rdev->nr_pending);
2723 if (r10_sync_page_io(rdev,
2724 r10_bio->devs[sl].addr +
2725 sect,
2726 s, conf->tmppage, REQ_OP_WRITE)
2727 == 0) {
2728 /* Well, this device is dead */
2729 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %pg)\n",
2730 mdname(mddev), s,
2731 (unsigned long long)(
2732 sect +
2733 choose_data_offset(r10_bio,
2734 rdev)),
2735 rdev->bdev);
2736 pr_notice("md/raid10:%s: %pg: failing drive\n",
2737 mdname(mddev),
2738 rdev->bdev);
2739 }
2740 rdev_dec_pending(rdev, mddev);
2741 }
2742 sl = start;
2743 while (sl != slot) {
2744 if (sl==0)
2745 sl = conf->copies;
2746 sl--;
2747 d = r10_bio->devs[sl].devnum;
2748 rdev = conf->mirrors[d].rdev;
2749 if (!rdev ||
2750 test_bit(Faulty, &rdev->flags) ||
2751 !test_bit(In_sync, &rdev->flags))
2752 continue;
2753
2754 atomic_inc(&rdev->nr_pending);
2755 switch (r10_sync_page_io(rdev,
2756 r10_bio->devs[sl].addr +
2757 sect,
2758 s, conf->tmppage, REQ_OP_READ)) {
2759 case 0:
2760 /* Well, this device is dead */
2761 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %pg)\n",
2762 mdname(mddev), s,
2763 (unsigned long long)(
2764 sect +
2765 choose_data_offset(r10_bio, rdev)),
2766 rdev->bdev);
2767 pr_notice("md/raid10:%s: %pg: failing drive\n",
2768 mdname(mddev),
2769 rdev->bdev);
2770 break;
2771 case 1:
2772 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %pg)\n",
2773 mdname(mddev), s,
2774 (unsigned long long)(
2775 sect +
2776 choose_data_offset(r10_bio, rdev)),
2777 rdev->bdev);
2778 atomic_add(s, &rdev->corrected_errors);
2779 }
2780
2781 rdev_dec_pending(rdev, mddev);
2782 }
2783
2784 sectors -= s;
2785 sect += s;
2786 }
2787}
2788
2789static int narrow_write_error(struct r10bio *r10_bio, int i)
2790{
2791 struct bio *bio = r10_bio->master_bio;
2792 struct mddev *mddev = r10_bio->mddev;
2793 struct r10conf *conf = mddev->private;
2794 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2795 /* bio has the data to be written to slot 'i' where
2796 * we just recently had a write error.
2797 * We repeatedly clone the bio and trim down to one block,
2798 * then try the write. Where the write fails we record
2799 * a bad block.
2800 * It is conceivable that the bio doesn't exactly align with
2801 * blocks. We must handle this.
2802 *
2803 * We currently own a reference to the rdev.
2804 */
2805
2806 int block_sectors;
2807 sector_t sector;
2808 int sectors;
2809 int sect_to_write = r10_bio->sectors;
2810 int ok = 1;
2811
2812 if (rdev->badblocks.shift < 0)
2813 return 0;
2814
2815 block_sectors = roundup(1 << rdev->badblocks.shift,
2816 bdev_logical_block_size(rdev->bdev) >> 9);
2817 sector = r10_bio->sector;
2818 sectors = ((r10_bio->sector + block_sectors)
2819 & ~(sector_t)(block_sectors - 1))
2820 - sector;
2821
2822 while (sect_to_write) {
2823 struct bio *wbio;
2824 sector_t wsector;
2825 if (sectors > sect_to_write)
2826 sectors = sect_to_write;
2827 /* Write at 'sector' for 'sectors' */
2828 wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2829 &mddev->bio_set);
2830 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2831 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2832 wbio->bi_iter.bi_sector = wsector +
2833 choose_data_offset(r10_bio, rdev);
2834 wbio->bi_opf = REQ_OP_WRITE;
2835
2836 if (submit_bio_wait(wbio) < 0)
2837 /* Failure! */
2838 ok = rdev_set_badblocks(rdev, wsector,
2839 sectors, 0)
2840 && ok;
2841
2842 bio_put(wbio);
2843 sect_to_write -= sectors;
2844 sector += sectors;
2845 sectors = block_sectors;
2846 }
2847 return ok;
2848}
2849
2850static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2851{
2852 int slot = r10_bio->read_slot;
2853 struct bio *bio;
2854 struct r10conf *conf = mddev->private;
2855 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2856
2857 /* we got a read error. Maybe the drive is bad. Maybe just
2858 * the block and we can fix it.
2859 * We freeze all other IO, and try reading the block from
2860 * other devices. When we find one, we re-write
2861 * and check it that fixes the read error.
2862 * This is all done synchronously while the array is
2863 * frozen.
2864 */
2865 bio = r10_bio->devs[slot].bio;
2866 bio_put(bio);
2867 r10_bio->devs[slot].bio = NULL;
2868
2869 if (mddev->ro)
2870 r10_bio->devs[slot].bio = IO_BLOCKED;
2871 else if (!test_bit(FailFast, &rdev->flags)) {
2872 freeze_array(conf, 1);
2873 fix_read_error(conf, mddev, r10_bio);
2874 unfreeze_array(conf);
2875 } else
2876 md_error(mddev, rdev);
2877
2878 rdev_dec_pending(rdev, mddev);
2879 r10_bio->state = 0;
2880 raid10_read_request(mddev, r10_bio->master_bio, r10_bio, false);
2881 /*
2882 * allow_barrier after re-submit to ensure no sync io
2883 * can be issued while regular io pending.
2884 */
2885 allow_barrier(conf);
2886}
2887
2888static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2889{
2890 /* Some sort of write request has finished and it
2891 * succeeded in writing where we thought there was a
2892 * bad block. So forget the bad block.
2893 * Or possibly if failed and we need to record
2894 * a bad block.
2895 */
2896 int m;
2897 struct md_rdev *rdev;
2898
2899 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2900 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2901 for (m = 0; m < conf->copies; m++) {
2902 int dev = r10_bio->devs[m].devnum;
2903 rdev = conf->mirrors[dev].rdev;
2904 if (r10_bio->devs[m].bio == NULL ||
2905 r10_bio->devs[m].bio->bi_end_io == NULL)
2906 continue;
2907 if (!r10_bio->devs[m].bio->bi_status) {
2908 rdev_clear_badblocks(
2909 rdev,
2910 r10_bio->devs[m].addr,
2911 r10_bio->sectors, 0);
2912 } else {
2913 if (!rdev_set_badblocks(
2914 rdev,
2915 r10_bio->devs[m].addr,
2916 r10_bio->sectors, 0))
2917 md_error(conf->mddev, rdev);
2918 }
2919 rdev = conf->mirrors[dev].replacement;
2920 if (r10_bio->devs[m].repl_bio == NULL ||
2921 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2922 continue;
2923
2924 if (!r10_bio->devs[m].repl_bio->bi_status) {
2925 rdev_clear_badblocks(
2926 rdev,
2927 r10_bio->devs[m].addr,
2928 r10_bio->sectors, 0);
2929 } else {
2930 if (!rdev_set_badblocks(
2931 rdev,
2932 r10_bio->devs[m].addr,
2933 r10_bio->sectors, 0))
2934 md_error(conf->mddev, rdev);
2935 }
2936 }
2937 put_buf(r10_bio);
2938 } else {
2939 bool fail = false;
2940 for (m = 0; m < conf->copies; m++) {
2941 int dev = r10_bio->devs[m].devnum;
2942 struct bio *bio = r10_bio->devs[m].bio;
2943 rdev = conf->mirrors[dev].rdev;
2944 if (bio == IO_MADE_GOOD) {
2945 rdev_clear_badblocks(
2946 rdev,
2947 r10_bio->devs[m].addr,
2948 r10_bio->sectors, 0);
2949 rdev_dec_pending(rdev, conf->mddev);
2950 } else if (bio != NULL && bio->bi_status) {
2951 fail = true;
2952 if (!narrow_write_error(r10_bio, m))
2953 md_error(conf->mddev, rdev);
2954 rdev_dec_pending(rdev, conf->mddev);
2955 }
2956 bio = r10_bio->devs[m].repl_bio;
2957 rdev = conf->mirrors[dev].replacement;
2958 if (rdev && bio == IO_MADE_GOOD) {
2959 rdev_clear_badblocks(
2960 rdev,
2961 r10_bio->devs[m].addr,
2962 r10_bio->sectors, 0);
2963 rdev_dec_pending(rdev, conf->mddev);
2964 }
2965 }
2966 if (fail) {
2967 spin_lock_irq(&conf->device_lock);
2968 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2969 conf->nr_queued++;
2970 spin_unlock_irq(&conf->device_lock);
2971 /*
2972 * In case freeze_array() is waiting for condition
2973 * nr_pending == nr_queued + extra to be true.
2974 */
2975 wake_up(&conf->wait_barrier);
2976 md_wakeup_thread(conf->mddev->thread);
2977 } else {
2978 if (test_bit(R10BIO_WriteError,
2979 &r10_bio->state))
2980 close_write(r10_bio);
2981 raid_end_bio_io(r10_bio);
2982 }
2983 }
2984}
2985
2986static void raid10d(struct md_thread *thread)
2987{
2988 struct mddev *mddev = thread->mddev;
2989 struct r10bio *r10_bio;
2990 unsigned long flags;
2991 struct r10conf *conf = mddev->private;
2992 struct list_head *head = &conf->retry_list;
2993 struct blk_plug plug;
2994
2995 md_check_recovery(mddev);
2996
2997 if (!list_empty_careful(&conf->bio_end_io_list) &&
2998 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2999 LIST_HEAD(tmp);
3000 spin_lock_irqsave(&conf->device_lock, flags);
3001 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3002 while (!list_empty(&conf->bio_end_io_list)) {
3003 list_move(conf->bio_end_io_list.prev, &tmp);
3004 conf->nr_queued--;
3005 }
3006 }
3007 spin_unlock_irqrestore(&conf->device_lock, flags);
3008 while (!list_empty(&tmp)) {
3009 r10_bio = list_first_entry(&tmp, struct r10bio,
3010 retry_list);
3011 list_del(&r10_bio->retry_list);
3012
3013 if (test_bit(R10BIO_WriteError,
3014 &r10_bio->state))
3015 close_write(r10_bio);
3016 raid_end_bio_io(r10_bio);
3017 }
3018 }
3019
3020 blk_start_plug(&plug);
3021 for (;;) {
3022
3023 flush_pending_writes(conf);
3024
3025 spin_lock_irqsave(&conf->device_lock, flags);
3026 if (list_empty(head)) {
3027 spin_unlock_irqrestore(&conf->device_lock, flags);
3028 break;
3029 }
3030 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3031 list_del(head->prev);
3032 conf->nr_queued--;
3033 spin_unlock_irqrestore(&conf->device_lock, flags);
3034
3035 mddev = r10_bio->mddev;
3036 conf = mddev->private;
3037 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3038 test_bit(R10BIO_WriteError, &r10_bio->state))
3039 handle_write_completed(conf, r10_bio);
3040 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3041 reshape_request_write(mddev, r10_bio);
3042 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3043 sync_request_write(mddev, r10_bio);
3044 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3045 recovery_request_write(mddev, r10_bio);
3046 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3047 handle_read_error(mddev, r10_bio);
3048 else
3049 WARN_ON_ONCE(1);
3050
3051 cond_resched();
3052 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3053 md_check_recovery(mddev);
3054 }
3055 blk_finish_plug(&plug);
3056}
3057
3058static int init_resync(struct r10conf *conf)
3059{
3060 int ret, buffs, i;
3061
3062 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3063 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3064 conf->have_replacement = 0;
3065 for (i = 0; i < conf->geo.raid_disks; i++)
3066 if (conf->mirrors[i].replacement)
3067 conf->have_replacement = 1;
3068 ret = mempool_init(&conf->r10buf_pool, buffs,
3069 r10buf_pool_alloc, r10buf_pool_free, conf);
3070 if (ret)
3071 return ret;
3072 conf->next_resync = 0;
3073 return 0;
3074}
3075
3076static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3077{
3078 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3079 struct rsync_pages *rp;
3080 struct bio *bio;
3081 int nalloc;
3082 int i;
3083
3084 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3085 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3086 nalloc = conf->copies; /* resync */
3087 else
3088 nalloc = 2; /* recovery */
3089
3090 for (i = 0; i < nalloc; i++) {
3091 bio = r10bio->devs[i].bio;
3092 rp = bio->bi_private;
3093 bio_reset(bio, NULL, 0);
3094 bio->bi_private = rp;
3095 bio = r10bio->devs[i].repl_bio;
3096 if (bio) {
3097 rp = bio->bi_private;
3098 bio_reset(bio, NULL, 0);
3099 bio->bi_private = rp;
3100 }
3101 }
3102 return r10bio;
3103}
3104
3105/*
3106 * Set cluster_sync_high since we need other nodes to add the
3107 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3108 */
3109static void raid10_set_cluster_sync_high(struct r10conf *conf)
3110{
3111 sector_t window_size;
3112 int extra_chunk, chunks;
3113
3114 /*
3115 * First, here we define "stripe" as a unit which across
3116 * all member devices one time, so we get chunks by use
3117 * raid_disks / near_copies. Otherwise, if near_copies is
3118 * close to raid_disks, then resync window could increases
3119 * linearly with the increase of raid_disks, which means
3120 * we will suspend a really large IO window while it is not
3121 * necessary. If raid_disks is not divisible by near_copies,
3122 * an extra chunk is needed to ensure the whole "stripe" is
3123 * covered.
3124 */
3125
3126 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3127 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3128 extra_chunk = 0;
3129 else
3130 extra_chunk = 1;
3131 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3132
3133 /*
3134 * At least use a 32M window to align with raid1's resync window
3135 */
3136 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3137 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3138
3139 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3140}
3141
3142/*
3143 * perform a "sync" on one "block"
3144 *
3145 * We need to make sure that no normal I/O request - particularly write
3146 * requests - conflict with active sync requests.
3147 *
3148 * This is achieved by tracking pending requests and a 'barrier' concept
3149 * that can be installed to exclude normal IO requests.
3150 *
3151 * Resync and recovery are handled very differently.
3152 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3153 *
3154 * For resync, we iterate over virtual addresses, read all copies,
3155 * and update if there are differences. If only one copy is live,
3156 * skip it.
3157 * For recovery, we iterate over physical addresses, read a good
3158 * value for each non-in_sync drive, and over-write.
3159 *
3160 * So, for recovery we may have several outstanding complex requests for a
3161 * given address, one for each out-of-sync device. We model this by allocating
3162 * a number of r10_bio structures, one for each out-of-sync device.
3163 * As we setup these structures, we collect all bio's together into a list
3164 * which we then process collectively to add pages, and then process again
3165 * to pass to submit_bio_noacct.
3166 *
3167 * The r10_bio structures are linked using a borrowed master_bio pointer.
3168 * This link is counted in ->remaining. When the r10_bio that points to NULL
3169 * has its remaining count decremented to 0, the whole complex operation
3170 * is complete.
3171 *
3172 */
3173
3174static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3175 sector_t max_sector, int *skipped)
3176{
3177 struct r10conf *conf = mddev->private;
3178 struct r10bio *r10_bio;
3179 struct bio *biolist = NULL, *bio;
3180 sector_t nr_sectors;
3181 int i;
3182 int max_sync;
3183 sector_t sync_blocks;
3184 sector_t sectors_skipped = 0;
3185 int chunks_skipped = 0;
3186 sector_t chunk_mask = conf->geo.chunk_mask;
3187 int page_idx = 0;
3188 int error_disk = -1;
3189
3190 /*
3191 * Allow skipping a full rebuild for incremental assembly
3192 * of a clean array, like RAID1 does.
3193 */
3194 if (mddev->bitmap == NULL &&
3195 mddev->recovery_cp == MaxSector &&
3196 mddev->reshape_position == MaxSector &&
3197 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3198 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3199 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3200 conf->fullsync == 0) {
3201 *skipped = 1;
3202 return mddev->dev_sectors - sector_nr;
3203 }
3204
3205 if (!mempool_initialized(&conf->r10buf_pool))
3206 if (init_resync(conf))
3207 return 0;
3208
3209 skipped:
3210 if (sector_nr >= max_sector) {
3211 conf->cluster_sync_low = 0;
3212 conf->cluster_sync_high = 0;
3213
3214 /* If we aborted, we need to abort the
3215 * sync on the 'current' bitmap chucks (there can
3216 * be several when recovering multiple devices).
3217 * as we may have started syncing it but not finished.
3218 * We can find the current address in
3219 * mddev->curr_resync, but for recovery,
3220 * we need to convert that to several
3221 * virtual addresses.
3222 */
3223 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3224 end_reshape(conf);
3225 close_sync(conf);
3226 return 0;
3227 }
3228
3229 if (mddev->curr_resync < max_sector) { /* aborted */
3230 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3231 mddev->bitmap_ops->end_sync(mddev,
3232 mddev->curr_resync,
3233 &sync_blocks);
3234 else for (i = 0; i < conf->geo.raid_disks; i++) {
3235 sector_t sect =
3236 raid10_find_virt(conf, mddev->curr_resync, i);
3237
3238 mddev->bitmap_ops->end_sync(mddev, sect,
3239 &sync_blocks);
3240 }
3241 } else {
3242 /* completed sync */
3243 if ((!mddev->bitmap || conf->fullsync)
3244 && conf->have_replacement
3245 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3246 /* Completed a full sync so the replacements
3247 * are now fully recovered.
3248 */
3249 for (i = 0; i < conf->geo.raid_disks; i++) {
3250 struct md_rdev *rdev =
3251 conf->mirrors[i].replacement;
3252
3253 if (rdev)
3254 rdev->recovery_offset = MaxSector;
3255 }
3256 }
3257 conf->fullsync = 0;
3258 }
3259 mddev->bitmap_ops->close_sync(mddev);
3260 close_sync(conf);
3261 *skipped = 1;
3262 return sectors_skipped;
3263 }
3264
3265 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3266 return reshape_request(mddev, sector_nr, skipped);
3267
3268 if (chunks_skipped >= conf->geo.raid_disks) {
3269 pr_err("md/raid10:%s: %s fails\n", mdname(mddev),
3270 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ? "resync" : "recovery");
3271 if (error_disk >= 0 &&
3272 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3273 /*
3274 * recovery fails, set mirrors.recovery_disabled,
3275 * device shouldn't be added to there.
3276 */
3277 conf->mirrors[error_disk].recovery_disabled =
3278 mddev->recovery_disabled;
3279 return 0;
3280 }
3281 /*
3282 * if there has been nothing to do on any drive,
3283 * then there is nothing to do at all.
3284 */
3285 *skipped = 1;
3286 return (max_sector - sector_nr) + sectors_skipped;
3287 }
3288
3289 if (max_sector > mddev->resync_max)
3290 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3291
3292 /* make sure whole request will fit in a chunk - if chunks
3293 * are meaningful
3294 */
3295 if (conf->geo.near_copies < conf->geo.raid_disks &&
3296 max_sector > (sector_nr | chunk_mask))
3297 max_sector = (sector_nr | chunk_mask) + 1;
3298
3299 /*
3300 * If there is non-resync activity waiting for a turn, then let it
3301 * though before starting on this new sync request.
3302 */
3303 if (conf->nr_waiting)
3304 schedule_timeout_uninterruptible(1);
3305
3306 /* Again, very different code for resync and recovery.
3307 * Both must result in an r10bio with a list of bios that
3308 * have bi_end_io, bi_sector, bi_bdev set,
3309 * and bi_private set to the r10bio.
3310 * For recovery, we may actually create several r10bios
3311 * with 2 bios in each, that correspond to the bios in the main one.
3312 * In this case, the subordinate r10bios link back through a
3313 * borrowed master_bio pointer, and the counter in the master
3314 * includes a ref from each subordinate.
3315 */
3316 /* First, we decide what to do and set ->bi_end_io
3317 * To end_sync_read if we want to read, and
3318 * end_sync_write if we will want to write.
3319 */
3320
3321 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3322 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3323 /* recovery... the complicated one */
3324 int j;
3325 r10_bio = NULL;
3326
3327 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3328 bool still_degraded;
3329 struct r10bio *rb2;
3330 sector_t sect;
3331 bool must_sync;
3332 int any_working;
3333 struct raid10_info *mirror = &conf->mirrors[i];
3334 struct md_rdev *mrdev, *mreplace;
3335
3336 mrdev = mirror->rdev;
3337 mreplace = mirror->replacement;
3338
3339 if (mrdev && (test_bit(Faulty, &mrdev->flags) ||
3340 test_bit(In_sync, &mrdev->flags)))
3341 mrdev = NULL;
3342 if (mreplace && test_bit(Faulty, &mreplace->flags))
3343 mreplace = NULL;
3344
3345 if (!mrdev && !mreplace)
3346 continue;
3347
3348 still_degraded = false;
3349 /* want to reconstruct this device */
3350 rb2 = r10_bio;
3351 sect = raid10_find_virt(conf, sector_nr, i);
3352 if (sect >= mddev->resync_max_sectors)
3353 /* last stripe is not complete - don't
3354 * try to recover this sector.
3355 */
3356 continue;
3357 /* Unless we are doing a full sync, or a replacement
3358 * we only need to recover the block if it is set in
3359 * the bitmap
3360 */
3361 must_sync = mddev->bitmap_ops->start_sync(mddev, sect,
3362 &sync_blocks,
3363 true);
3364 if (sync_blocks < max_sync)
3365 max_sync = sync_blocks;
3366 if (!must_sync &&
3367 mreplace == NULL &&
3368 !conf->fullsync) {
3369 /* yep, skip the sync_blocks here, but don't assume
3370 * that there will never be anything to do here
3371 */
3372 chunks_skipped = -1;
3373 continue;
3374 }
3375 if (mrdev)
3376 atomic_inc(&mrdev->nr_pending);
3377 if (mreplace)
3378 atomic_inc(&mreplace->nr_pending);
3379
3380 r10_bio = raid10_alloc_init_r10buf(conf);
3381 r10_bio->state = 0;
3382 raise_barrier(conf, rb2 != NULL);
3383 atomic_set(&r10_bio->remaining, 0);
3384
3385 r10_bio->master_bio = (struct bio*)rb2;
3386 if (rb2)
3387 atomic_inc(&rb2->remaining);
3388 r10_bio->mddev = mddev;
3389 set_bit(R10BIO_IsRecover, &r10_bio->state);
3390 r10_bio->sector = sect;
3391
3392 raid10_find_phys(conf, r10_bio);
3393
3394 /* Need to check if the array will still be
3395 * degraded
3396 */
3397 for (j = 0; j < conf->geo.raid_disks; j++) {
3398 struct md_rdev *rdev = conf->mirrors[j].rdev;
3399
3400 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3401 still_degraded = false;
3402 break;
3403 }
3404 }
3405
3406 must_sync = mddev->bitmap_ops->start_sync(mddev, sect,
3407 &sync_blocks, still_degraded);
3408
3409 any_working = 0;
3410 for (j=0; j<conf->copies;j++) {
3411 int k;
3412 int d = r10_bio->devs[j].devnum;
3413 sector_t from_addr, to_addr;
3414 struct md_rdev *rdev = conf->mirrors[d].rdev;
3415 sector_t sector, first_bad;
3416 int bad_sectors;
3417 if (!rdev ||
3418 !test_bit(In_sync, &rdev->flags))
3419 continue;
3420 /* This is where we read from */
3421 any_working = 1;
3422 sector = r10_bio->devs[j].addr;
3423
3424 if (is_badblock(rdev, sector, max_sync,
3425 &first_bad, &bad_sectors)) {
3426 if (first_bad > sector)
3427 max_sync = first_bad - sector;
3428 else {
3429 bad_sectors -= (sector
3430 - first_bad);
3431 if (max_sync > bad_sectors)
3432 max_sync = bad_sectors;
3433 continue;
3434 }
3435 }
3436 bio = r10_bio->devs[0].bio;
3437 bio->bi_next = biolist;
3438 biolist = bio;
3439 bio->bi_end_io = end_sync_read;
3440 bio->bi_opf = REQ_OP_READ;
3441 if (test_bit(FailFast, &rdev->flags))
3442 bio->bi_opf |= MD_FAILFAST;
3443 from_addr = r10_bio->devs[j].addr;
3444 bio->bi_iter.bi_sector = from_addr +
3445 rdev->data_offset;
3446 bio_set_dev(bio, rdev->bdev);
3447 atomic_inc(&rdev->nr_pending);
3448 /* and we write to 'i' (if not in_sync) */
3449
3450 for (k=0; k<conf->copies; k++)
3451 if (r10_bio->devs[k].devnum == i)
3452 break;
3453 BUG_ON(k == conf->copies);
3454 to_addr = r10_bio->devs[k].addr;
3455 r10_bio->devs[0].devnum = d;
3456 r10_bio->devs[0].addr = from_addr;
3457 r10_bio->devs[1].devnum = i;
3458 r10_bio->devs[1].addr = to_addr;
3459
3460 if (mrdev) {
3461 bio = r10_bio->devs[1].bio;
3462 bio->bi_next = biolist;
3463 biolist = bio;
3464 bio->bi_end_io = end_sync_write;
3465 bio->bi_opf = REQ_OP_WRITE;
3466 bio->bi_iter.bi_sector = to_addr
3467 + mrdev->data_offset;
3468 bio_set_dev(bio, mrdev->bdev);
3469 atomic_inc(&r10_bio->remaining);
3470 } else
3471 r10_bio->devs[1].bio->bi_end_io = NULL;
3472
3473 /* and maybe write to replacement */
3474 bio = r10_bio->devs[1].repl_bio;
3475 if (bio)
3476 bio->bi_end_io = NULL;
3477 /* Note: if replace is not NULL, then bio
3478 * cannot be NULL as r10buf_pool_alloc will
3479 * have allocated it.
3480 */
3481 if (!mreplace)
3482 break;
3483 bio->bi_next = biolist;
3484 biolist = bio;
3485 bio->bi_end_io = end_sync_write;
3486 bio->bi_opf = REQ_OP_WRITE;
3487 bio->bi_iter.bi_sector = to_addr +
3488 mreplace->data_offset;
3489 bio_set_dev(bio, mreplace->bdev);
3490 atomic_inc(&r10_bio->remaining);
3491 break;
3492 }
3493 if (j == conf->copies) {
3494 /* Cannot recover, so abort the recovery or
3495 * record a bad block */
3496 if (any_working) {
3497 /* problem is that there are bad blocks
3498 * on other device(s)
3499 */
3500 int k;
3501 for (k = 0; k < conf->copies; k++)
3502 if (r10_bio->devs[k].devnum == i)
3503 break;
3504 if (mrdev && !test_bit(In_sync,
3505 &mrdev->flags)
3506 && !rdev_set_badblocks(
3507 mrdev,
3508 r10_bio->devs[k].addr,
3509 max_sync, 0))
3510 any_working = 0;
3511 if (mreplace &&
3512 !rdev_set_badblocks(
3513 mreplace,
3514 r10_bio->devs[k].addr,
3515 max_sync, 0))
3516 any_working = 0;
3517 }
3518 if (!any_working) {
3519 if (!test_and_set_bit(MD_RECOVERY_INTR,
3520 &mddev->recovery))
3521 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3522 mdname(mddev));
3523 mirror->recovery_disabled
3524 = mddev->recovery_disabled;
3525 } else {
3526 error_disk = i;
3527 }
3528 put_buf(r10_bio);
3529 if (rb2)
3530 atomic_dec(&rb2->remaining);
3531 r10_bio = rb2;
3532 if (mrdev)
3533 rdev_dec_pending(mrdev, mddev);
3534 if (mreplace)
3535 rdev_dec_pending(mreplace, mddev);
3536 break;
3537 }
3538 if (mrdev)
3539 rdev_dec_pending(mrdev, mddev);
3540 if (mreplace)
3541 rdev_dec_pending(mreplace, mddev);
3542 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3543 /* Only want this if there is elsewhere to
3544 * read from. 'j' is currently the first
3545 * readable copy.
3546 */
3547 int targets = 1;
3548 for (; j < conf->copies; j++) {
3549 int d = r10_bio->devs[j].devnum;
3550 if (conf->mirrors[d].rdev &&
3551 test_bit(In_sync,
3552 &conf->mirrors[d].rdev->flags))
3553 targets++;
3554 }
3555 if (targets == 1)
3556 r10_bio->devs[0].bio->bi_opf
3557 &= ~MD_FAILFAST;
3558 }
3559 }
3560 if (biolist == NULL) {
3561 while (r10_bio) {
3562 struct r10bio *rb2 = r10_bio;
3563 r10_bio = (struct r10bio*) rb2->master_bio;
3564 rb2->master_bio = NULL;
3565 put_buf(rb2);
3566 }
3567 goto giveup;
3568 }
3569 } else {
3570 /* resync. Schedule a read for every block at this virt offset */
3571 int count = 0;
3572
3573 /*
3574 * Since curr_resync_completed could probably not update in
3575 * time, and we will set cluster_sync_low based on it.
3576 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3577 * safety reason, which ensures curr_resync_completed is
3578 * updated in bitmap_cond_end_sync.
3579 */
3580 mddev->bitmap_ops->cond_end_sync(mddev, sector_nr,
3581 mddev_is_clustered(mddev) &&
3582 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3583
3584 if (!mddev->bitmap_ops->start_sync(mddev, sector_nr,
3585 &sync_blocks,
3586 mddev->degraded) &&
3587 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3588 &mddev->recovery)) {
3589 /* We can skip this block */
3590 *skipped = 1;
3591 return sync_blocks + sectors_skipped;
3592 }
3593 if (sync_blocks < max_sync)
3594 max_sync = sync_blocks;
3595 r10_bio = raid10_alloc_init_r10buf(conf);
3596 r10_bio->state = 0;
3597
3598 r10_bio->mddev = mddev;
3599 atomic_set(&r10_bio->remaining, 0);
3600 raise_barrier(conf, 0);
3601 conf->next_resync = sector_nr;
3602
3603 r10_bio->master_bio = NULL;
3604 r10_bio->sector = sector_nr;
3605 set_bit(R10BIO_IsSync, &r10_bio->state);
3606 raid10_find_phys(conf, r10_bio);
3607 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3608
3609 for (i = 0; i < conf->copies; i++) {
3610 int d = r10_bio->devs[i].devnum;
3611 sector_t first_bad, sector;
3612 int bad_sectors;
3613 struct md_rdev *rdev;
3614
3615 if (r10_bio->devs[i].repl_bio)
3616 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3617
3618 bio = r10_bio->devs[i].bio;
3619 bio->bi_status = BLK_STS_IOERR;
3620 rdev = conf->mirrors[d].rdev;
3621 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3622 continue;
3623
3624 sector = r10_bio->devs[i].addr;
3625 if (is_badblock(rdev, sector, max_sync,
3626 &first_bad, &bad_sectors)) {
3627 if (first_bad > sector)
3628 max_sync = first_bad - sector;
3629 else {
3630 bad_sectors -= (sector - first_bad);
3631 if (max_sync > bad_sectors)
3632 max_sync = bad_sectors;
3633 continue;
3634 }
3635 }
3636 atomic_inc(&rdev->nr_pending);
3637 atomic_inc(&r10_bio->remaining);
3638 bio->bi_next = biolist;
3639 biolist = bio;
3640 bio->bi_end_io = end_sync_read;
3641 bio->bi_opf = REQ_OP_READ;
3642 if (test_bit(FailFast, &rdev->flags))
3643 bio->bi_opf |= MD_FAILFAST;
3644 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3645 bio_set_dev(bio, rdev->bdev);
3646 count++;
3647
3648 rdev = conf->mirrors[d].replacement;
3649 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3650 continue;
3651
3652 atomic_inc(&rdev->nr_pending);
3653
3654 /* Need to set up for writing to the replacement */
3655 bio = r10_bio->devs[i].repl_bio;
3656 bio->bi_status = BLK_STS_IOERR;
3657
3658 sector = r10_bio->devs[i].addr;
3659 bio->bi_next = biolist;
3660 biolist = bio;
3661 bio->bi_end_io = end_sync_write;
3662 bio->bi_opf = REQ_OP_WRITE;
3663 if (test_bit(FailFast, &rdev->flags))
3664 bio->bi_opf |= MD_FAILFAST;
3665 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3666 bio_set_dev(bio, rdev->bdev);
3667 count++;
3668 }
3669
3670 if (count < 2) {
3671 for (i=0; i<conf->copies; i++) {
3672 int d = r10_bio->devs[i].devnum;
3673 if (r10_bio->devs[i].bio->bi_end_io)
3674 rdev_dec_pending(conf->mirrors[d].rdev,
3675 mddev);
3676 if (r10_bio->devs[i].repl_bio &&
3677 r10_bio->devs[i].repl_bio->bi_end_io)
3678 rdev_dec_pending(
3679 conf->mirrors[d].replacement,
3680 mddev);
3681 }
3682 put_buf(r10_bio);
3683 biolist = NULL;
3684 goto giveup;
3685 }
3686 }
3687
3688 nr_sectors = 0;
3689 if (sector_nr + max_sync < max_sector)
3690 max_sector = sector_nr + max_sync;
3691 do {
3692 struct page *page;
3693 int len = PAGE_SIZE;
3694 if (sector_nr + (len>>9) > max_sector)
3695 len = (max_sector - sector_nr) << 9;
3696 if (len == 0)
3697 break;
3698 for (bio= biolist ; bio ; bio=bio->bi_next) {
3699 struct resync_pages *rp = get_resync_pages(bio);
3700 page = resync_fetch_page(rp, page_idx);
3701 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
3702 bio->bi_status = BLK_STS_RESOURCE;
3703 bio_endio(bio);
3704 goto giveup;
3705 }
3706 }
3707 nr_sectors += len>>9;
3708 sector_nr += len>>9;
3709 } while (++page_idx < RESYNC_PAGES);
3710 r10_bio->sectors = nr_sectors;
3711
3712 if (mddev_is_clustered(mddev) &&
3713 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3714 /* It is resync not recovery */
3715 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3716 conf->cluster_sync_low = mddev->curr_resync_completed;
3717 raid10_set_cluster_sync_high(conf);
3718 /* Send resync message */
3719 md_cluster_ops->resync_info_update(mddev,
3720 conf->cluster_sync_low,
3721 conf->cluster_sync_high);
3722 }
3723 } else if (mddev_is_clustered(mddev)) {
3724 /* This is recovery not resync */
3725 sector_t sect_va1, sect_va2;
3726 bool broadcast_msg = false;
3727
3728 for (i = 0; i < conf->geo.raid_disks; i++) {
3729 /*
3730 * sector_nr is a device address for recovery, so we
3731 * need translate it to array address before compare
3732 * with cluster_sync_high.
3733 */
3734 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3735
3736 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3737 broadcast_msg = true;
3738 /*
3739 * curr_resync_completed is similar as
3740 * sector_nr, so make the translation too.
3741 */
3742 sect_va2 = raid10_find_virt(conf,
3743 mddev->curr_resync_completed, i);
3744
3745 if (conf->cluster_sync_low == 0 ||
3746 conf->cluster_sync_low > sect_va2)
3747 conf->cluster_sync_low = sect_va2;
3748 }
3749 }
3750 if (broadcast_msg) {
3751 raid10_set_cluster_sync_high(conf);
3752 md_cluster_ops->resync_info_update(mddev,
3753 conf->cluster_sync_low,
3754 conf->cluster_sync_high);
3755 }
3756 }
3757
3758 while (biolist) {
3759 bio = biolist;
3760 biolist = biolist->bi_next;
3761
3762 bio->bi_next = NULL;
3763 r10_bio = get_resync_r10bio(bio);
3764 r10_bio->sectors = nr_sectors;
3765
3766 if (bio->bi_end_io == end_sync_read) {
3767 md_sync_acct_bio(bio, nr_sectors);
3768 bio->bi_status = 0;
3769 submit_bio_noacct(bio);
3770 }
3771 }
3772
3773 if (sectors_skipped)
3774 /* pretend they weren't skipped, it makes
3775 * no important difference in this case
3776 */
3777 md_done_sync(mddev, sectors_skipped, 1);
3778
3779 return sectors_skipped + nr_sectors;
3780 giveup:
3781 /* There is nowhere to write, so all non-sync
3782 * drives must be failed or in resync, all drives
3783 * have a bad block, so try the next chunk...
3784 */
3785 if (sector_nr + max_sync < max_sector)
3786 max_sector = sector_nr + max_sync;
3787
3788 sectors_skipped += (max_sector - sector_nr);
3789 chunks_skipped ++;
3790 sector_nr = max_sector;
3791 goto skipped;
3792}
3793
3794static sector_t
3795raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3796{
3797 sector_t size;
3798 struct r10conf *conf = mddev->private;
3799
3800 if (!raid_disks)
3801 raid_disks = min(conf->geo.raid_disks,
3802 conf->prev.raid_disks);
3803 if (!sectors)
3804 sectors = conf->dev_sectors;
3805
3806 size = sectors >> conf->geo.chunk_shift;
3807 sector_div(size, conf->geo.far_copies);
3808 size = size * raid_disks;
3809 sector_div(size, conf->geo.near_copies);
3810
3811 return size << conf->geo.chunk_shift;
3812}
3813
3814static void calc_sectors(struct r10conf *conf, sector_t size)
3815{
3816 /* Calculate the number of sectors-per-device that will
3817 * actually be used, and set conf->dev_sectors and
3818 * conf->stride
3819 */
3820
3821 size = size >> conf->geo.chunk_shift;
3822 sector_div(size, conf->geo.far_copies);
3823 size = size * conf->geo.raid_disks;
3824 sector_div(size, conf->geo.near_copies);
3825 /* 'size' is now the number of chunks in the array */
3826 /* calculate "used chunks per device" */
3827 size = size * conf->copies;
3828
3829 /* We need to round up when dividing by raid_disks to
3830 * get the stride size.
3831 */
3832 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3833
3834 conf->dev_sectors = size << conf->geo.chunk_shift;
3835
3836 if (conf->geo.far_offset)
3837 conf->geo.stride = 1 << conf->geo.chunk_shift;
3838 else {
3839 sector_div(size, conf->geo.far_copies);
3840 conf->geo.stride = size << conf->geo.chunk_shift;
3841 }
3842}
3843
3844enum geo_type {geo_new, geo_old, geo_start};
3845static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3846{
3847 int nc, fc, fo;
3848 int layout, chunk, disks;
3849 switch (new) {
3850 case geo_old:
3851 layout = mddev->layout;
3852 chunk = mddev->chunk_sectors;
3853 disks = mddev->raid_disks - mddev->delta_disks;
3854 break;
3855 case geo_new:
3856 layout = mddev->new_layout;
3857 chunk = mddev->new_chunk_sectors;
3858 disks = mddev->raid_disks;
3859 break;
3860 default: /* avoid 'may be unused' warnings */
3861 case geo_start: /* new when starting reshape - raid_disks not
3862 * updated yet. */
3863 layout = mddev->new_layout;
3864 chunk = mddev->new_chunk_sectors;
3865 disks = mddev->raid_disks + mddev->delta_disks;
3866 break;
3867 }
3868 if (layout >> 19)
3869 return -1;
3870 if (chunk < (PAGE_SIZE >> 9) ||
3871 !is_power_of_2(chunk))
3872 return -2;
3873 nc = layout & 255;
3874 fc = (layout >> 8) & 255;
3875 fo = layout & (1<<16);
3876 geo->raid_disks = disks;
3877 geo->near_copies = nc;
3878 geo->far_copies = fc;
3879 geo->far_offset = fo;
3880 switch (layout >> 17) {
3881 case 0: /* original layout. simple but not always optimal */
3882 geo->far_set_size = disks;
3883 break;
3884 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3885 * actually using this, but leave code here just in case.*/
3886 geo->far_set_size = disks/fc;
3887 WARN(geo->far_set_size < fc,
3888 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3889 break;
3890 case 2: /* "improved" layout fixed to match documentation */
3891 geo->far_set_size = fc * nc;
3892 break;
3893 default: /* Not a valid layout */
3894 return -1;
3895 }
3896 geo->chunk_mask = chunk - 1;
3897 geo->chunk_shift = ffz(~chunk);
3898 return nc*fc;
3899}
3900
3901static void raid10_free_conf(struct r10conf *conf)
3902{
3903 if (!conf)
3904 return;
3905
3906 mempool_exit(&conf->r10bio_pool);
3907 kfree(conf->mirrors);
3908 kfree(conf->mirrors_old);
3909 kfree(conf->mirrors_new);
3910 safe_put_page(conf->tmppage);
3911 bioset_exit(&conf->bio_split);
3912 kfree(conf);
3913}
3914
3915static struct r10conf *setup_conf(struct mddev *mddev)
3916{
3917 struct r10conf *conf = NULL;
3918 int err = -EINVAL;
3919 struct geom geo;
3920 int copies;
3921
3922 copies = setup_geo(&geo, mddev, geo_new);
3923
3924 if (copies == -2) {
3925 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3926 mdname(mddev), PAGE_SIZE);
3927 goto out;
3928 }
3929
3930 if (copies < 2 || copies > mddev->raid_disks) {
3931 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3932 mdname(mddev), mddev->new_layout);
3933 goto out;
3934 }
3935
3936 err = -ENOMEM;
3937 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3938 if (!conf)
3939 goto out;
3940
3941 /* FIXME calc properly */
3942 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3943 sizeof(struct raid10_info),
3944 GFP_KERNEL);
3945 if (!conf->mirrors)
3946 goto out;
3947
3948 conf->tmppage = alloc_page(GFP_KERNEL);
3949 if (!conf->tmppage)
3950 goto out;
3951
3952 conf->geo = geo;
3953 conf->copies = copies;
3954 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3955 rbio_pool_free, conf);
3956 if (err)
3957 goto out;
3958
3959 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3960 if (err)
3961 goto out;
3962
3963 calc_sectors(conf, mddev->dev_sectors);
3964 if (mddev->reshape_position == MaxSector) {
3965 conf->prev = conf->geo;
3966 conf->reshape_progress = MaxSector;
3967 } else {
3968 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3969 err = -EINVAL;
3970 goto out;
3971 }
3972 conf->reshape_progress = mddev->reshape_position;
3973 if (conf->prev.far_offset)
3974 conf->prev.stride = 1 << conf->prev.chunk_shift;
3975 else
3976 /* far_copies must be 1 */
3977 conf->prev.stride = conf->dev_sectors;
3978 }
3979 conf->reshape_safe = conf->reshape_progress;
3980 spin_lock_init(&conf->device_lock);
3981 INIT_LIST_HEAD(&conf->retry_list);
3982 INIT_LIST_HEAD(&conf->bio_end_io_list);
3983
3984 seqlock_init(&conf->resync_lock);
3985 init_waitqueue_head(&conf->wait_barrier);
3986 atomic_set(&conf->nr_pending, 0);
3987
3988 err = -ENOMEM;
3989 rcu_assign_pointer(conf->thread,
3990 md_register_thread(raid10d, mddev, "raid10"));
3991 if (!conf->thread)
3992 goto out;
3993
3994 conf->mddev = mddev;
3995 return conf;
3996
3997 out:
3998 raid10_free_conf(conf);
3999 return ERR_PTR(err);
4000}
4001
4002static unsigned int raid10_nr_stripes(struct r10conf *conf)
4003{
4004 unsigned int raid_disks = conf->geo.raid_disks;
4005
4006 if (conf->geo.raid_disks % conf->geo.near_copies)
4007 return raid_disks;
4008 return raid_disks / conf->geo.near_copies;
4009}
4010
4011static int raid10_set_queue_limits(struct mddev *mddev)
4012{
4013 struct r10conf *conf = mddev->private;
4014 struct queue_limits lim;
4015 int err;
4016
4017 md_init_stacking_limits(&lim);
4018 lim.max_write_zeroes_sectors = 0;
4019 lim.io_min = mddev->chunk_sectors << 9;
4020 lim.io_opt = lim.io_min * raid10_nr_stripes(conf);
4021 lim.features |= BLK_FEAT_ATOMIC_WRITES_STACKED;
4022 err = mddev_stack_rdev_limits(mddev, &lim, MDDEV_STACK_INTEGRITY);
4023 if (err)
4024 return err;
4025 return queue_limits_set(mddev->gendisk->queue, &lim);
4026}
4027
4028static int raid10_run(struct mddev *mddev)
4029{
4030 struct r10conf *conf;
4031 int i, disk_idx;
4032 struct raid10_info *disk;
4033 struct md_rdev *rdev;
4034 sector_t size;
4035 sector_t min_offset_diff = 0;
4036 int first = 1;
4037 int ret = -EIO;
4038
4039 if (mddev->private == NULL) {
4040 conf = setup_conf(mddev);
4041 if (IS_ERR(conf))
4042 return PTR_ERR(conf);
4043 mddev->private = conf;
4044 }
4045 conf = mddev->private;
4046 if (!conf)
4047 goto out;
4048
4049 rcu_assign_pointer(mddev->thread, conf->thread);
4050 rcu_assign_pointer(conf->thread, NULL);
4051
4052 if (mddev_is_clustered(conf->mddev)) {
4053 int fc, fo;
4054
4055 fc = (mddev->layout >> 8) & 255;
4056 fo = mddev->layout & (1<<16);
4057 if (fc > 1 || fo > 0) {
4058 pr_err("only near layout is supported by clustered"
4059 " raid10\n");
4060 goto out_free_conf;
4061 }
4062 }
4063
4064 rdev_for_each(rdev, mddev) {
4065 long long diff;
4066
4067 disk_idx = rdev->raid_disk;
4068 if (disk_idx < 0)
4069 continue;
4070 if (disk_idx >= conf->geo.raid_disks &&
4071 disk_idx >= conf->prev.raid_disks)
4072 continue;
4073 disk = conf->mirrors + disk_idx;
4074
4075 if (test_bit(Replacement, &rdev->flags)) {
4076 if (disk->replacement)
4077 goto out_free_conf;
4078 disk->replacement = rdev;
4079 } else {
4080 if (disk->rdev)
4081 goto out_free_conf;
4082 disk->rdev = rdev;
4083 }
4084 diff = (rdev->new_data_offset - rdev->data_offset);
4085 if (!mddev->reshape_backwards)
4086 diff = -diff;
4087 if (diff < 0)
4088 diff = 0;
4089 if (first || diff < min_offset_diff)
4090 min_offset_diff = diff;
4091
4092 disk->head_position = 0;
4093 first = 0;
4094 }
4095
4096 if (!mddev_is_dm(conf->mddev)) {
4097 int err = raid10_set_queue_limits(mddev);
4098
4099 if (err) {
4100 ret = err;
4101 goto out_free_conf;
4102 }
4103 }
4104
4105 /* need to check that every block has at least one working mirror */
4106 if (!enough(conf, -1)) {
4107 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4108 mdname(mddev));
4109 goto out_free_conf;
4110 }
4111
4112 if (conf->reshape_progress != MaxSector) {
4113 /* must ensure that shape change is supported */
4114 if (conf->geo.far_copies != 1 &&
4115 conf->geo.far_offset == 0)
4116 goto out_free_conf;
4117 if (conf->prev.far_copies != 1 &&
4118 conf->prev.far_offset == 0)
4119 goto out_free_conf;
4120 }
4121
4122 mddev->degraded = 0;
4123 for (i = 0;
4124 i < conf->geo.raid_disks
4125 || i < conf->prev.raid_disks;
4126 i++) {
4127
4128 disk = conf->mirrors + i;
4129
4130 if (!disk->rdev && disk->replacement) {
4131 /* The replacement is all we have - use it */
4132 disk->rdev = disk->replacement;
4133 disk->replacement = NULL;
4134 clear_bit(Replacement, &disk->rdev->flags);
4135 }
4136
4137 if (!disk->rdev ||
4138 !test_bit(In_sync, &disk->rdev->flags)) {
4139 disk->head_position = 0;
4140 mddev->degraded++;
4141 if (disk->rdev &&
4142 disk->rdev->saved_raid_disk < 0)
4143 conf->fullsync = 1;
4144 }
4145
4146 if (disk->replacement &&
4147 !test_bit(In_sync, &disk->replacement->flags) &&
4148 disk->replacement->saved_raid_disk < 0) {
4149 conf->fullsync = 1;
4150 }
4151
4152 disk->recovery_disabled = mddev->recovery_disabled - 1;
4153 }
4154
4155 if (mddev->recovery_cp != MaxSector)
4156 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4157 mdname(mddev));
4158 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4159 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4160 conf->geo.raid_disks);
4161 /*
4162 * Ok, everything is just fine now
4163 */
4164 mddev->dev_sectors = conf->dev_sectors;
4165 size = raid10_size(mddev, 0, 0);
4166 md_set_array_sectors(mddev, size);
4167 mddev->resync_max_sectors = size;
4168 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4169
4170 if (md_integrity_register(mddev))
4171 goto out_free_conf;
4172
4173 if (conf->reshape_progress != MaxSector) {
4174 unsigned long before_length, after_length;
4175
4176 before_length = ((1 << conf->prev.chunk_shift) *
4177 conf->prev.far_copies);
4178 after_length = ((1 << conf->geo.chunk_shift) *
4179 conf->geo.far_copies);
4180
4181 if (max(before_length, after_length) > min_offset_diff) {
4182 /* This cannot work */
4183 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4184 goto out_free_conf;
4185 }
4186 conf->offset_diff = min_offset_diff;
4187
4188 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4189 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4190 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4191 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4192 }
4193
4194 return 0;
4195
4196out_free_conf:
4197 md_unregister_thread(mddev, &mddev->thread);
4198 raid10_free_conf(conf);
4199 mddev->private = NULL;
4200out:
4201 return ret;
4202}
4203
4204static void raid10_free(struct mddev *mddev, void *priv)
4205{
4206 raid10_free_conf(priv);
4207}
4208
4209static void raid10_quiesce(struct mddev *mddev, int quiesce)
4210{
4211 struct r10conf *conf = mddev->private;
4212
4213 if (quiesce)
4214 raise_barrier(conf, 0);
4215 else
4216 lower_barrier(conf);
4217}
4218
4219static int raid10_resize(struct mddev *mddev, sector_t sectors)
4220{
4221 /* Resize of 'far' arrays is not supported.
4222 * For 'near' and 'offset' arrays we can set the
4223 * number of sectors used to be an appropriate multiple
4224 * of the chunk size.
4225 * For 'offset', this is far_copies*chunksize.
4226 * For 'near' the multiplier is the LCM of
4227 * near_copies and raid_disks.
4228 * So if far_copies > 1 && !far_offset, fail.
4229 * Else find LCM(raid_disks, near_copy)*far_copies and
4230 * multiply by chunk_size. Then round to this number.
4231 * This is mostly done by raid10_size()
4232 */
4233 struct r10conf *conf = mddev->private;
4234 sector_t oldsize, size;
4235 int ret;
4236
4237 if (mddev->reshape_position != MaxSector)
4238 return -EBUSY;
4239
4240 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4241 return -EINVAL;
4242
4243 oldsize = raid10_size(mddev, 0, 0);
4244 size = raid10_size(mddev, sectors, 0);
4245 if (mddev->external_size &&
4246 mddev->array_sectors > size)
4247 return -EINVAL;
4248
4249 ret = mddev->bitmap_ops->resize(mddev, size, 0, false);
4250 if (ret)
4251 return ret;
4252
4253 md_set_array_sectors(mddev, size);
4254 if (sectors > mddev->dev_sectors &&
4255 mddev->recovery_cp > oldsize) {
4256 mddev->recovery_cp = oldsize;
4257 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4258 }
4259 calc_sectors(conf, sectors);
4260 mddev->dev_sectors = conf->dev_sectors;
4261 mddev->resync_max_sectors = size;
4262 return 0;
4263}
4264
4265static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4266{
4267 struct md_rdev *rdev;
4268 struct r10conf *conf;
4269
4270 if (mddev->degraded > 0) {
4271 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4272 mdname(mddev));
4273 return ERR_PTR(-EINVAL);
4274 }
4275 sector_div(size, devs);
4276
4277 /* Set new parameters */
4278 mddev->new_level = 10;
4279 /* new layout: far_copies = 1, near_copies = 2 */
4280 mddev->new_layout = (1<<8) + 2;
4281 mddev->new_chunk_sectors = mddev->chunk_sectors;
4282 mddev->delta_disks = mddev->raid_disks;
4283 mddev->raid_disks *= 2;
4284 /* make sure it will be not marked as dirty */
4285 mddev->recovery_cp = MaxSector;
4286 mddev->dev_sectors = size;
4287
4288 conf = setup_conf(mddev);
4289 if (!IS_ERR(conf)) {
4290 rdev_for_each(rdev, mddev)
4291 if (rdev->raid_disk >= 0) {
4292 rdev->new_raid_disk = rdev->raid_disk * 2;
4293 rdev->sectors = size;
4294 }
4295 }
4296
4297 return conf;
4298}
4299
4300static void *raid10_takeover(struct mddev *mddev)
4301{
4302 struct r0conf *raid0_conf;
4303
4304 /* raid10 can take over:
4305 * raid0 - providing it has only two drives
4306 */
4307 if (mddev->level == 0) {
4308 /* for raid0 takeover only one zone is supported */
4309 raid0_conf = mddev->private;
4310 if (raid0_conf->nr_strip_zones > 1) {
4311 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4312 mdname(mddev));
4313 return ERR_PTR(-EINVAL);
4314 }
4315 return raid10_takeover_raid0(mddev,
4316 raid0_conf->strip_zone->zone_end,
4317 raid0_conf->strip_zone->nb_dev);
4318 }
4319 return ERR_PTR(-EINVAL);
4320}
4321
4322static int raid10_check_reshape(struct mddev *mddev)
4323{
4324 /* Called when there is a request to change
4325 * - layout (to ->new_layout)
4326 * - chunk size (to ->new_chunk_sectors)
4327 * - raid_disks (by delta_disks)
4328 * or when trying to restart a reshape that was ongoing.
4329 *
4330 * We need to validate the request and possibly allocate
4331 * space if that might be an issue later.
4332 *
4333 * Currently we reject any reshape of a 'far' mode array,
4334 * allow chunk size to change if new is generally acceptable,
4335 * allow raid_disks to increase, and allow
4336 * a switch between 'near' mode and 'offset' mode.
4337 */
4338 struct r10conf *conf = mddev->private;
4339 struct geom geo;
4340
4341 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4342 return -EINVAL;
4343
4344 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4345 /* mustn't change number of copies */
4346 return -EINVAL;
4347 if (geo.far_copies > 1 && !geo.far_offset)
4348 /* Cannot switch to 'far' mode */
4349 return -EINVAL;
4350
4351 if (mddev->array_sectors & geo.chunk_mask)
4352 /* not factor of array size */
4353 return -EINVAL;
4354
4355 if (!enough(conf, -1))
4356 return -EINVAL;
4357
4358 kfree(conf->mirrors_new);
4359 conf->mirrors_new = NULL;
4360 if (mddev->delta_disks > 0) {
4361 /* allocate new 'mirrors' list */
4362 conf->mirrors_new =
4363 kcalloc(mddev->raid_disks + mddev->delta_disks,
4364 sizeof(struct raid10_info),
4365 GFP_KERNEL);
4366 if (!conf->mirrors_new)
4367 return -ENOMEM;
4368 }
4369 return 0;
4370}
4371
4372/*
4373 * Need to check if array has failed when deciding whether to:
4374 * - start an array
4375 * - remove non-faulty devices
4376 * - add a spare
4377 * - allow a reshape
4378 * This determination is simple when no reshape is happening.
4379 * However if there is a reshape, we need to carefully check
4380 * both the before and after sections.
4381 * This is because some failed devices may only affect one
4382 * of the two sections, and some non-in_sync devices may
4383 * be insync in the section most affected by failed devices.
4384 */
4385static int calc_degraded(struct r10conf *conf)
4386{
4387 int degraded, degraded2;
4388 int i;
4389
4390 degraded = 0;
4391 /* 'prev' section first */
4392 for (i = 0; i < conf->prev.raid_disks; i++) {
4393 struct md_rdev *rdev = conf->mirrors[i].rdev;
4394
4395 if (!rdev || test_bit(Faulty, &rdev->flags))
4396 degraded++;
4397 else if (!test_bit(In_sync, &rdev->flags))
4398 /* When we can reduce the number of devices in
4399 * an array, this might not contribute to
4400 * 'degraded'. It does now.
4401 */
4402 degraded++;
4403 }
4404 if (conf->geo.raid_disks == conf->prev.raid_disks)
4405 return degraded;
4406 degraded2 = 0;
4407 for (i = 0; i < conf->geo.raid_disks; i++) {
4408 struct md_rdev *rdev = conf->mirrors[i].rdev;
4409
4410 if (!rdev || test_bit(Faulty, &rdev->flags))
4411 degraded2++;
4412 else if (!test_bit(In_sync, &rdev->flags)) {
4413 /* If reshape is increasing the number of devices,
4414 * this section has already been recovered, so
4415 * it doesn't contribute to degraded.
4416 * else it does.
4417 */
4418 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4419 degraded2++;
4420 }
4421 }
4422 if (degraded2 > degraded)
4423 return degraded2;
4424 return degraded;
4425}
4426
4427static int raid10_start_reshape(struct mddev *mddev)
4428{
4429 /* A 'reshape' has been requested. This commits
4430 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4431 * This also checks if there are enough spares and adds them
4432 * to the array.
4433 * We currently require enough spares to make the final
4434 * array non-degraded. We also require that the difference
4435 * between old and new data_offset - on each device - is
4436 * enough that we never risk over-writing.
4437 */
4438
4439 unsigned long before_length, after_length;
4440 sector_t min_offset_diff = 0;
4441 int first = 1;
4442 struct geom new;
4443 struct r10conf *conf = mddev->private;
4444 struct md_rdev *rdev;
4445 int spares = 0;
4446 int ret;
4447
4448 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4449 return -EBUSY;
4450
4451 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4452 return -EINVAL;
4453
4454 before_length = ((1 << conf->prev.chunk_shift) *
4455 conf->prev.far_copies);
4456 after_length = ((1 << conf->geo.chunk_shift) *
4457 conf->geo.far_copies);
4458
4459 rdev_for_each(rdev, mddev) {
4460 if (!test_bit(In_sync, &rdev->flags)
4461 && !test_bit(Faulty, &rdev->flags))
4462 spares++;
4463 if (rdev->raid_disk >= 0) {
4464 long long diff = (rdev->new_data_offset
4465 - rdev->data_offset);
4466 if (!mddev->reshape_backwards)
4467 diff = -diff;
4468 if (diff < 0)
4469 diff = 0;
4470 if (first || diff < min_offset_diff)
4471 min_offset_diff = diff;
4472 first = 0;
4473 }
4474 }
4475
4476 if (max(before_length, after_length) > min_offset_diff)
4477 return -EINVAL;
4478
4479 if (spares < mddev->delta_disks)
4480 return -EINVAL;
4481
4482 conf->offset_diff = min_offset_diff;
4483 spin_lock_irq(&conf->device_lock);
4484 if (conf->mirrors_new) {
4485 memcpy(conf->mirrors_new, conf->mirrors,
4486 sizeof(struct raid10_info)*conf->prev.raid_disks);
4487 smp_mb();
4488 kfree(conf->mirrors_old);
4489 conf->mirrors_old = conf->mirrors;
4490 conf->mirrors = conf->mirrors_new;
4491 conf->mirrors_new = NULL;
4492 }
4493 setup_geo(&conf->geo, mddev, geo_start);
4494 smp_mb();
4495 if (mddev->reshape_backwards) {
4496 sector_t size = raid10_size(mddev, 0, 0);
4497 if (size < mddev->array_sectors) {
4498 spin_unlock_irq(&conf->device_lock);
4499 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4500 mdname(mddev));
4501 return -EINVAL;
4502 }
4503 mddev->resync_max_sectors = size;
4504 conf->reshape_progress = size;
4505 } else
4506 conf->reshape_progress = 0;
4507 conf->reshape_safe = conf->reshape_progress;
4508 spin_unlock_irq(&conf->device_lock);
4509
4510 if (mddev->delta_disks && mddev->bitmap) {
4511 struct mdp_superblock_1 *sb = NULL;
4512 sector_t oldsize, newsize;
4513
4514 oldsize = raid10_size(mddev, 0, 0);
4515 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4516
4517 if (!mddev_is_clustered(mddev)) {
4518 ret = mddev->bitmap_ops->resize(mddev, newsize, 0, false);
4519 if (ret)
4520 goto abort;
4521 else
4522 goto out;
4523 }
4524
4525 rdev_for_each(rdev, mddev) {
4526 if (rdev->raid_disk > -1 &&
4527 !test_bit(Faulty, &rdev->flags))
4528 sb = page_address(rdev->sb_page);
4529 }
4530
4531 /*
4532 * some node is already performing reshape, and no need to
4533 * call bitmap_ops->resize again since it should be called when
4534 * receiving BITMAP_RESIZE msg
4535 */
4536 if ((sb && (le32_to_cpu(sb->feature_map) &
4537 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4538 goto out;
4539
4540 ret = mddev->bitmap_ops->resize(mddev, newsize, 0, false);
4541 if (ret)
4542 goto abort;
4543
4544 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4545 if (ret) {
4546 mddev->bitmap_ops->resize(mddev, oldsize, 0, false);
4547 goto abort;
4548 }
4549 }
4550out:
4551 if (mddev->delta_disks > 0) {
4552 rdev_for_each(rdev, mddev)
4553 if (rdev->raid_disk < 0 &&
4554 !test_bit(Faulty, &rdev->flags)) {
4555 if (raid10_add_disk(mddev, rdev) == 0) {
4556 if (rdev->raid_disk >=
4557 conf->prev.raid_disks)
4558 set_bit(In_sync, &rdev->flags);
4559 else
4560 rdev->recovery_offset = 0;
4561
4562 /* Failure here is OK */
4563 sysfs_link_rdev(mddev, rdev);
4564 }
4565 } else if (rdev->raid_disk >= conf->prev.raid_disks
4566 && !test_bit(Faulty, &rdev->flags)) {
4567 /* This is a spare that was manually added */
4568 set_bit(In_sync, &rdev->flags);
4569 }
4570 }
4571 /* When a reshape changes the number of devices,
4572 * ->degraded is measured against the larger of the
4573 * pre and post numbers.
4574 */
4575 spin_lock_irq(&conf->device_lock);
4576 mddev->degraded = calc_degraded(conf);
4577 spin_unlock_irq(&conf->device_lock);
4578 mddev->raid_disks = conf->geo.raid_disks;
4579 mddev->reshape_position = conf->reshape_progress;
4580 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4581
4582 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4583 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4584 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4585 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4586 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4587 conf->reshape_checkpoint = jiffies;
4588 md_new_event();
4589 return 0;
4590
4591abort:
4592 mddev->recovery = 0;
4593 spin_lock_irq(&conf->device_lock);
4594 conf->geo = conf->prev;
4595 mddev->raid_disks = conf->geo.raid_disks;
4596 rdev_for_each(rdev, mddev)
4597 rdev->new_data_offset = rdev->data_offset;
4598 smp_wmb();
4599 conf->reshape_progress = MaxSector;
4600 conf->reshape_safe = MaxSector;
4601 mddev->reshape_position = MaxSector;
4602 spin_unlock_irq(&conf->device_lock);
4603 return ret;
4604}
4605
4606/* Calculate the last device-address that could contain
4607 * any block from the chunk that includes the array-address 's'
4608 * and report the next address.
4609 * i.e. the address returned will be chunk-aligned and after
4610 * any data that is in the chunk containing 's'.
4611 */
4612static sector_t last_dev_address(sector_t s, struct geom *geo)
4613{
4614 s = (s | geo->chunk_mask) + 1;
4615 s >>= geo->chunk_shift;
4616 s *= geo->near_copies;
4617 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4618 s *= geo->far_copies;
4619 s <<= geo->chunk_shift;
4620 return s;
4621}
4622
4623/* Calculate the first device-address that could contain
4624 * any block from the chunk that includes the array-address 's'.
4625 * This too will be the start of a chunk
4626 */
4627static sector_t first_dev_address(sector_t s, struct geom *geo)
4628{
4629 s >>= geo->chunk_shift;
4630 s *= geo->near_copies;
4631 sector_div(s, geo->raid_disks);
4632 s *= geo->far_copies;
4633 s <<= geo->chunk_shift;
4634 return s;
4635}
4636
4637static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4638 int *skipped)
4639{
4640 /* We simply copy at most one chunk (smallest of old and new)
4641 * at a time, possibly less if that exceeds RESYNC_PAGES,
4642 * or we hit a bad block or something.
4643 * This might mean we pause for normal IO in the middle of
4644 * a chunk, but that is not a problem as mddev->reshape_position
4645 * can record any location.
4646 *
4647 * If we will want to write to a location that isn't
4648 * yet recorded as 'safe' (i.e. in metadata on disk) then
4649 * we need to flush all reshape requests and update the metadata.
4650 *
4651 * When reshaping forwards (e.g. to more devices), we interpret
4652 * 'safe' as the earliest block which might not have been copied
4653 * down yet. We divide this by previous stripe size and multiply
4654 * by previous stripe length to get lowest device offset that we
4655 * cannot write to yet.
4656 * We interpret 'sector_nr' as an address that we want to write to.
4657 * From this we use last_device_address() to find where we might
4658 * write to, and first_device_address on the 'safe' position.
4659 * If this 'next' write position is after the 'safe' position,
4660 * we must update the metadata to increase the 'safe' position.
4661 *
4662 * When reshaping backwards, we round in the opposite direction
4663 * and perform the reverse test: next write position must not be
4664 * less than current safe position.
4665 *
4666 * In all this the minimum difference in data offsets
4667 * (conf->offset_diff - always positive) allows a bit of slack,
4668 * so next can be after 'safe', but not by more than offset_diff
4669 *
4670 * We need to prepare all the bios here before we start any IO
4671 * to ensure the size we choose is acceptable to all devices.
4672 * The means one for each copy for write-out and an extra one for
4673 * read-in.
4674 * We store the read-in bio in ->master_bio and the others in
4675 * ->devs[x].bio and ->devs[x].repl_bio.
4676 */
4677 struct r10conf *conf = mddev->private;
4678 struct r10bio *r10_bio;
4679 sector_t next, safe, last;
4680 int max_sectors;
4681 int nr_sectors;
4682 int s;
4683 struct md_rdev *rdev;
4684 int need_flush = 0;
4685 struct bio *blist;
4686 struct bio *bio, *read_bio;
4687 int sectors_done = 0;
4688 struct page **pages;
4689
4690 if (sector_nr == 0) {
4691 /* If restarting in the middle, skip the initial sectors */
4692 if (mddev->reshape_backwards &&
4693 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4694 sector_nr = (raid10_size(mddev, 0, 0)
4695 - conf->reshape_progress);
4696 } else if (!mddev->reshape_backwards &&
4697 conf->reshape_progress > 0)
4698 sector_nr = conf->reshape_progress;
4699 if (sector_nr) {
4700 mddev->curr_resync_completed = sector_nr;
4701 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4702 *skipped = 1;
4703 return sector_nr;
4704 }
4705 }
4706
4707 /* We don't use sector_nr to track where we are up to
4708 * as that doesn't work well for ->reshape_backwards.
4709 * So just use ->reshape_progress.
4710 */
4711 if (mddev->reshape_backwards) {
4712 /* 'next' is the earliest device address that we might
4713 * write to for this chunk in the new layout
4714 */
4715 next = first_dev_address(conf->reshape_progress - 1,
4716 &conf->geo);
4717
4718 /* 'safe' is the last device address that we might read from
4719 * in the old layout after a restart
4720 */
4721 safe = last_dev_address(conf->reshape_safe - 1,
4722 &conf->prev);
4723
4724 if (next + conf->offset_diff < safe)
4725 need_flush = 1;
4726
4727 last = conf->reshape_progress - 1;
4728 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4729 & conf->prev.chunk_mask);
4730 if (sector_nr + RESYNC_SECTORS < last)
4731 sector_nr = last + 1 - RESYNC_SECTORS;
4732 } else {
4733 /* 'next' is after the last device address that we
4734 * might write to for this chunk in the new layout
4735 */
4736 next = last_dev_address(conf->reshape_progress, &conf->geo);
4737
4738 /* 'safe' is the earliest device address that we might
4739 * read from in the old layout after a restart
4740 */
4741 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4742
4743 /* Need to update metadata if 'next' might be beyond 'safe'
4744 * as that would possibly corrupt data
4745 */
4746 if (next > safe + conf->offset_diff)
4747 need_flush = 1;
4748
4749 sector_nr = conf->reshape_progress;
4750 last = sector_nr | (conf->geo.chunk_mask
4751 & conf->prev.chunk_mask);
4752
4753 if (sector_nr + RESYNC_SECTORS <= last)
4754 last = sector_nr + RESYNC_SECTORS - 1;
4755 }
4756
4757 if (need_flush ||
4758 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4759 /* Need to update reshape_position in metadata */
4760 wait_barrier(conf, false);
4761 mddev->reshape_position = conf->reshape_progress;
4762 if (mddev->reshape_backwards)
4763 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4764 - conf->reshape_progress;
4765 else
4766 mddev->curr_resync_completed = conf->reshape_progress;
4767 conf->reshape_checkpoint = jiffies;
4768 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4769 md_wakeup_thread(mddev->thread);
4770 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4771 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4772 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4773 allow_barrier(conf);
4774 return sectors_done;
4775 }
4776 conf->reshape_safe = mddev->reshape_position;
4777 allow_barrier(conf);
4778 }
4779
4780 raise_barrier(conf, 0);
4781read_more:
4782 /* Now schedule reads for blocks from sector_nr to last */
4783 r10_bio = raid10_alloc_init_r10buf(conf);
4784 r10_bio->state = 0;
4785 raise_barrier(conf, 1);
4786 atomic_set(&r10_bio->remaining, 0);
4787 r10_bio->mddev = mddev;
4788 r10_bio->sector = sector_nr;
4789 set_bit(R10BIO_IsReshape, &r10_bio->state);
4790 r10_bio->sectors = last - sector_nr + 1;
4791 rdev = read_balance(conf, r10_bio, &max_sectors);
4792 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4793
4794 if (!rdev) {
4795 /* Cannot read from here, so need to record bad blocks
4796 * on all the target devices.
4797 */
4798 // FIXME
4799 mempool_free(r10_bio, &conf->r10buf_pool);
4800 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4801 return sectors_done;
4802 }
4803
4804 read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4805 GFP_KERNEL, &mddev->bio_set);
4806 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4807 + rdev->data_offset);
4808 read_bio->bi_private = r10_bio;
4809 read_bio->bi_end_io = end_reshape_read;
4810 r10_bio->master_bio = read_bio;
4811 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4812
4813 /*
4814 * Broadcast RESYNC message to other nodes, so all nodes would not
4815 * write to the region to avoid conflict.
4816 */
4817 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4818 struct mdp_superblock_1 *sb = NULL;
4819 int sb_reshape_pos = 0;
4820
4821 conf->cluster_sync_low = sector_nr;
4822 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4823 sb = page_address(rdev->sb_page);
4824 if (sb) {
4825 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4826 /*
4827 * Set cluster_sync_low again if next address for array
4828 * reshape is less than cluster_sync_low. Since we can't
4829 * update cluster_sync_low until it has finished reshape.
4830 */
4831 if (sb_reshape_pos < conf->cluster_sync_low)
4832 conf->cluster_sync_low = sb_reshape_pos;
4833 }
4834
4835 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4836 conf->cluster_sync_high);
4837 }
4838
4839 /* Now find the locations in the new layout */
4840 __raid10_find_phys(&conf->geo, r10_bio);
4841
4842 blist = read_bio;
4843 read_bio->bi_next = NULL;
4844
4845 for (s = 0; s < conf->copies*2; s++) {
4846 struct bio *b;
4847 int d = r10_bio->devs[s/2].devnum;
4848 struct md_rdev *rdev2;
4849 if (s&1) {
4850 rdev2 = conf->mirrors[d].replacement;
4851 b = r10_bio->devs[s/2].repl_bio;
4852 } else {
4853 rdev2 = conf->mirrors[d].rdev;
4854 b = r10_bio->devs[s/2].bio;
4855 }
4856 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4857 continue;
4858
4859 bio_set_dev(b, rdev2->bdev);
4860 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4861 rdev2->new_data_offset;
4862 b->bi_end_io = end_reshape_write;
4863 b->bi_opf = REQ_OP_WRITE;
4864 b->bi_next = blist;
4865 blist = b;
4866 }
4867
4868 /* Now add as many pages as possible to all of these bios. */
4869
4870 nr_sectors = 0;
4871 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4872 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4873 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4874 int len = (max_sectors - s) << 9;
4875 if (len > PAGE_SIZE)
4876 len = PAGE_SIZE;
4877 for (bio = blist; bio ; bio = bio->bi_next) {
4878 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
4879 bio->bi_status = BLK_STS_RESOURCE;
4880 bio_endio(bio);
4881 return sectors_done;
4882 }
4883 }
4884 sector_nr += len >> 9;
4885 nr_sectors += len >> 9;
4886 }
4887 r10_bio->sectors = nr_sectors;
4888
4889 /* Now submit the read */
4890 md_sync_acct_bio(read_bio, r10_bio->sectors);
4891 atomic_inc(&r10_bio->remaining);
4892 read_bio->bi_next = NULL;
4893 submit_bio_noacct(read_bio);
4894 sectors_done += nr_sectors;
4895 if (sector_nr <= last)
4896 goto read_more;
4897
4898 lower_barrier(conf);
4899
4900 /* Now that we have done the whole section we can
4901 * update reshape_progress
4902 */
4903 if (mddev->reshape_backwards)
4904 conf->reshape_progress -= sectors_done;
4905 else
4906 conf->reshape_progress += sectors_done;
4907
4908 return sectors_done;
4909}
4910
4911static void end_reshape_request(struct r10bio *r10_bio);
4912static int handle_reshape_read_error(struct mddev *mddev,
4913 struct r10bio *r10_bio);
4914static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4915{
4916 /* Reshape read completed. Hopefully we have a block
4917 * to write out.
4918 * If we got a read error then we do sync 1-page reads from
4919 * elsewhere until we find the data - or give up.
4920 */
4921 struct r10conf *conf = mddev->private;
4922 int s;
4923
4924 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4925 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4926 /* Reshape has been aborted */
4927 md_done_sync(mddev, r10_bio->sectors, 0);
4928 return;
4929 }
4930
4931 /* We definitely have the data in the pages, schedule the
4932 * writes.
4933 */
4934 atomic_set(&r10_bio->remaining, 1);
4935 for (s = 0; s < conf->copies*2; s++) {
4936 struct bio *b;
4937 int d = r10_bio->devs[s/2].devnum;
4938 struct md_rdev *rdev;
4939 if (s&1) {
4940 rdev = conf->mirrors[d].replacement;
4941 b = r10_bio->devs[s/2].repl_bio;
4942 } else {
4943 rdev = conf->mirrors[d].rdev;
4944 b = r10_bio->devs[s/2].bio;
4945 }
4946 if (!rdev || test_bit(Faulty, &rdev->flags))
4947 continue;
4948
4949 atomic_inc(&rdev->nr_pending);
4950 md_sync_acct_bio(b, r10_bio->sectors);
4951 atomic_inc(&r10_bio->remaining);
4952 b->bi_next = NULL;
4953 submit_bio_noacct(b);
4954 }
4955 end_reshape_request(r10_bio);
4956}
4957
4958static void end_reshape(struct r10conf *conf)
4959{
4960 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4961 return;
4962
4963 spin_lock_irq(&conf->device_lock);
4964 conf->prev = conf->geo;
4965 md_finish_reshape(conf->mddev);
4966 smp_wmb();
4967 conf->reshape_progress = MaxSector;
4968 conf->reshape_safe = MaxSector;
4969 spin_unlock_irq(&conf->device_lock);
4970
4971 mddev_update_io_opt(conf->mddev, raid10_nr_stripes(conf));
4972 conf->fullsync = 0;
4973}
4974
4975static void raid10_update_reshape_pos(struct mddev *mddev)
4976{
4977 struct r10conf *conf = mddev->private;
4978 sector_t lo, hi;
4979
4980 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4981 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4982 || mddev->reshape_position == MaxSector)
4983 conf->reshape_progress = mddev->reshape_position;
4984 else
4985 WARN_ON_ONCE(1);
4986}
4987
4988static int handle_reshape_read_error(struct mddev *mddev,
4989 struct r10bio *r10_bio)
4990{
4991 /* Use sync reads to get the blocks from somewhere else */
4992 int sectors = r10_bio->sectors;
4993 struct r10conf *conf = mddev->private;
4994 struct r10bio *r10b;
4995 int slot = 0;
4996 int idx = 0;
4997 struct page **pages;
4998
4999 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
5000 if (!r10b) {
5001 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
5002 return -ENOMEM;
5003 }
5004
5005 /* reshape IOs share pages from .devs[0].bio */
5006 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
5007
5008 r10b->sector = r10_bio->sector;
5009 __raid10_find_phys(&conf->prev, r10b);
5010
5011 while (sectors) {
5012 int s = sectors;
5013 int success = 0;
5014 int first_slot = slot;
5015
5016 if (s > (PAGE_SIZE >> 9))
5017 s = PAGE_SIZE >> 9;
5018
5019 while (!success) {
5020 int d = r10b->devs[slot].devnum;
5021 struct md_rdev *rdev = conf->mirrors[d].rdev;
5022 sector_t addr;
5023 if (rdev == NULL ||
5024 test_bit(Faulty, &rdev->flags) ||
5025 !test_bit(In_sync, &rdev->flags))
5026 goto failed;
5027
5028 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5029 atomic_inc(&rdev->nr_pending);
5030 success = sync_page_io(rdev,
5031 addr,
5032 s << 9,
5033 pages[idx],
5034 REQ_OP_READ, false);
5035 rdev_dec_pending(rdev, mddev);
5036 if (success)
5037 break;
5038 failed:
5039 slot++;
5040 if (slot >= conf->copies)
5041 slot = 0;
5042 if (slot == first_slot)
5043 break;
5044 }
5045 if (!success) {
5046 /* couldn't read this block, must give up */
5047 set_bit(MD_RECOVERY_INTR,
5048 &mddev->recovery);
5049 kfree(r10b);
5050 return -EIO;
5051 }
5052 sectors -= s;
5053 idx++;
5054 }
5055 kfree(r10b);
5056 return 0;
5057}
5058
5059static void end_reshape_write(struct bio *bio)
5060{
5061 struct r10bio *r10_bio = get_resync_r10bio(bio);
5062 struct mddev *mddev = r10_bio->mddev;
5063 struct r10conf *conf = mddev->private;
5064 int d;
5065 int slot;
5066 int repl;
5067 struct md_rdev *rdev = NULL;
5068
5069 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5070 rdev = repl ? conf->mirrors[d].replacement :
5071 conf->mirrors[d].rdev;
5072
5073 if (bio->bi_status) {
5074 /* FIXME should record badblock */
5075 md_error(mddev, rdev);
5076 }
5077
5078 rdev_dec_pending(rdev, mddev);
5079 end_reshape_request(r10_bio);
5080}
5081
5082static void end_reshape_request(struct r10bio *r10_bio)
5083{
5084 if (!atomic_dec_and_test(&r10_bio->remaining))
5085 return;
5086 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5087 bio_put(r10_bio->master_bio);
5088 put_buf(r10_bio);
5089}
5090
5091static void raid10_finish_reshape(struct mddev *mddev)
5092{
5093 struct r10conf *conf = mddev->private;
5094
5095 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5096 return;
5097
5098 if (mddev->delta_disks > 0) {
5099 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5100 mddev->recovery_cp = mddev->resync_max_sectors;
5101 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5102 }
5103 mddev->resync_max_sectors = mddev->array_sectors;
5104 } else {
5105 int d;
5106 for (d = conf->geo.raid_disks ;
5107 d < conf->geo.raid_disks - mddev->delta_disks;
5108 d++) {
5109 struct md_rdev *rdev = conf->mirrors[d].rdev;
5110 if (rdev)
5111 clear_bit(In_sync, &rdev->flags);
5112 rdev = conf->mirrors[d].replacement;
5113 if (rdev)
5114 clear_bit(In_sync, &rdev->flags);
5115 }
5116 }
5117 mddev->layout = mddev->new_layout;
5118 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5119 mddev->reshape_position = MaxSector;
5120 mddev->delta_disks = 0;
5121 mddev->reshape_backwards = 0;
5122}
5123
5124static struct md_personality raid10_personality =
5125{
5126 .name = "raid10",
5127 .level = 10,
5128 .owner = THIS_MODULE,
5129 .make_request = raid10_make_request,
5130 .run = raid10_run,
5131 .free = raid10_free,
5132 .status = raid10_status,
5133 .error_handler = raid10_error,
5134 .hot_add_disk = raid10_add_disk,
5135 .hot_remove_disk= raid10_remove_disk,
5136 .spare_active = raid10_spare_active,
5137 .sync_request = raid10_sync_request,
5138 .quiesce = raid10_quiesce,
5139 .size = raid10_size,
5140 .resize = raid10_resize,
5141 .takeover = raid10_takeover,
5142 .check_reshape = raid10_check_reshape,
5143 .start_reshape = raid10_start_reshape,
5144 .finish_reshape = raid10_finish_reshape,
5145 .update_reshape_pos = raid10_update_reshape_pos,
5146};
5147
5148static int __init raid_init(void)
5149{
5150 return register_md_personality(&raid10_personality);
5151}
5152
5153static void raid_exit(void)
5154{
5155 unregister_md_personality(&raid10_personality);
5156}
5157
5158module_init(raid_init);
5159module_exit(raid_exit);
5160MODULE_LICENSE("GPL");
5161MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5162MODULE_ALIAS("md-personality-9"); /* RAID10 */
5163MODULE_ALIAS("md-raid10");
5164MODULE_ALIAS("md-level-10");
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/seq_file.h>
25#include <linux/ratelimit.h>
26#include "md.h"
27#include "raid10.h"
28#include "raid0.h"
29#include "bitmap.h"
30
31/*
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
34 * chunk_size
35 * raid_disks
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
39 *
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47 * drive.
48 * near_copies and far_copies must be at least one, and their product is at most
49 * raid_disks.
50 *
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
54 */
55
56/*
57 * Number of guaranteed r10bios in case of extreme VM load:
58 */
59#define NR_RAID10_BIOS 256
60
61static void allow_barrier(conf_t *conf);
62static void lower_barrier(conf_t *conf);
63
64static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
65{
66 conf_t *conf = data;
67 int size = offsetof(struct r10bio_s, devs[conf->copies]);
68
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size, gfp_flags);
71}
72
73static void r10bio_pool_free(void *r10_bio, void *data)
74{
75 kfree(r10_bio);
76}
77
78/* Maximum size of each resync request */
79#define RESYNC_BLOCK_SIZE (64*1024)
80#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81/* amount of memory to reserve for resync requests */
82#define RESYNC_WINDOW (1024*1024)
83/* maximum number of concurrent requests, memory permitting */
84#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
85
86/*
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
91 *
92 */
93static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
94{
95 conf_t *conf = data;
96 struct page *page;
97 r10bio_t *r10_bio;
98 struct bio *bio;
99 int i, j;
100 int nalloc;
101
102 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
103 if (!r10_bio)
104 return NULL;
105
106 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107 nalloc = conf->copies; /* resync */
108 else
109 nalloc = 2; /* recovery */
110
111 /*
112 * Allocate bios.
113 */
114 for (j = nalloc ; j-- ; ) {
115 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
116 if (!bio)
117 goto out_free_bio;
118 r10_bio->devs[j].bio = bio;
119 }
120 /*
121 * Allocate RESYNC_PAGES data pages and attach them
122 * where needed.
123 */
124 for (j = 0 ; j < nalloc; j++) {
125 bio = r10_bio->devs[j].bio;
126 for (i = 0; i < RESYNC_PAGES; i++) {
127 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128 &conf->mddev->recovery)) {
129 /* we can share bv_page's during recovery */
130 struct bio *rbio = r10_bio->devs[0].bio;
131 page = rbio->bi_io_vec[i].bv_page;
132 get_page(page);
133 } else
134 page = alloc_page(gfp_flags);
135 if (unlikely(!page))
136 goto out_free_pages;
137
138 bio->bi_io_vec[i].bv_page = page;
139 }
140 }
141
142 return r10_bio;
143
144out_free_pages:
145 for ( ; i > 0 ; i--)
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
147 while (j--)
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
150 j = -1;
151out_free_bio:
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
155 return NULL;
156}
157
158static void r10buf_pool_free(void *__r10_bio, void *data)
159{
160 int i;
161 conf_t *conf = data;
162 r10bio_t *r10bio = __r10_bio;
163 int j;
164
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
167 if (bio) {
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
171 }
172 bio_put(bio);
173 }
174 }
175 r10bio_pool_free(r10bio, conf);
176}
177
178static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
179{
180 int i;
181
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (!BIO_SPECIAL(*bio))
185 bio_put(*bio);
186 *bio = NULL;
187 }
188}
189
190static void free_r10bio(r10bio_t *r10_bio)
191{
192 conf_t *conf = r10_bio->mddev->private;
193
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
196}
197
198static void put_buf(r10bio_t *r10_bio)
199{
200 conf_t *conf = r10_bio->mddev->private;
201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
204 lower_barrier(conf);
205}
206
207static void reschedule_retry(r10bio_t *r10_bio)
208{
209 unsigned long flags;
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev->private;
212
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
215 conf->nr_queued ++;
216 spin_unlock_irqrestore(&conf->device_lock, flags);
217
218 /* wake up frozen array... */
219 wake_up(&conf->wait_barrier);
220
221 md_wakeup_thread(mddev->thread);
222}
223
224/*
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
227 * cache layer.
228 */
229static void raid_end_bio_io(r10bio_t *r10_bio)
230{
231 struct bio *bio = r10_bio->master_bio;
232 int done;
233 conf_t *conf = r10_bio->mddev->private;
234
235 if (bio->bi_phys_segments) {
236 unsigned long flags;
237 spin_lock_irqsave(&conf->device_lock, flags);
238 bio->bi_phys_segments--;
239 done = (bio->bi_phys_segments == 0);
240 spin_unlock_irqrestore(&conf->device_lock, flags);
241 } else
242 done = 1;
243 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
244 clear_bit(BIO_UPTODATE, &bio->bi_flags);
245 if (done) {
246 bio_endio(bio, 0);
247 /*
248 * Wake up any possible resync thread that waits for the device
249 * to go idle.
250 */
251 allow_barrier(conf);
252 }
253 free_r10bio(r10_bio);
254}
255
256/*
257 * Update disk head position estimator based on IRQ completion info.
258 */
259static inline void update_head_pos(int slot, r10bio_t *r10_bio)
260{
261 conf_t *conf = r10_bio->mddev->private;
262
263 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
264 r10_bio->devs[slot].addr + (r10_bio->sectors);
265}
266
267/*
268 * Find the disk number which triggered given bio
269 */
270static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio,
271 struct bio *bio, int *slotp)
272{
273 int slot;
274
275 for (slot = 0; slot < conf->copies; slot++)
276 if (r10_bio->devs[slot].bio == bio)
277 break;
278
279 BUG_ON(slot == conf->copies);
280 update_head_pos(slot, r10_bio);
281
282 if (slotp)
283 *slotp = slot;
284 return r10_bio->devs[slot].devnum;
285}
286
287static void raid10_end_read_request(struct bio *bio, int error)
288{
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t *r10_bio = bio->bi_private;
291 int slot, dev;
292 conf_t *conf = r10_bio->mddev->private;
293
294
295 slot = r10_bio->read_slot;
296 dev = r10_bio->devs[slot].devnum;
297 /*
298 * this branch is our 'one mirror IO has finished' event handler:
299 */
300 update_head_pos(slot, r10_bio);
301
302 if (uptodate) {
303 /*
304 * Set R10BIO_Uptodate in our master bio, so that
305 * we will return a good error code to the higher
306 * levels even if IO on some other mirrored buffer fails.
307 *
308 * The 'master' represents the composite IO operation to
309 * user-side. So if something waits for IO, then it will
310 * wait for the 'master' bio.
311 */
312 set_bit(R10BIO_Uptodate, &r10_bio->state);
313 raid_end_bio_io(r10_bio);
314 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
315 } else {
316 /*
317 * oops, read error - keep the refcount on the rdev
318 */
319 char b[BDEVNAME_SIZE];
320 printk_ratelimited(KERN_ERR
321 "md/raid10:%s: %s: rescheduling sector %llu\n",
322 mdname(conf->mddev),
323 bdevname(conf->mirrors[dev].rdev->bdev, b),
324 (unsigned long long)r10_bio->sector);
325 set_bit(R10BIO_ReadError, &r10_bio->state);
326 reschedule_retry(r10_bio);
327 }
328}
329
330static void close_write(r10bio_t *r10_bio)
331{
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
334 r10_bio->sectors,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
336 0);
337 md_write_end(r10_bio->mddev);
338}
339
340static void one_write_done(r10bio_t *r10_bio)
341{
342 if (atomic_dec_and_test(&r10_bio->remaining)) {
343 if (test_bit(R10BIO_WriteError, &r10_bio->state))
344 reschedule_retry(r10_bio);
345 else {
346 close_write(r10_bio);
347 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
348 reschedule_retry(r10_bio);
349 else
350 raid_end_bio_io(r10_bio);
351 }
352 }
353}
354
355static void raid10_end_write_request(struct bio *bio, int error)
356{
357 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
358 r10bio_t *r10_bio = bio->bi_private;
359 int dev;
360 int dec_rdev = 1;
361 conf_t *conf = r10_bio->mddev->private;
362 int slot;
363
364 dev = find_bio_disk(conf, r10_bio, bio, &slot);
365
366 /*
367 * this branch is our 'one mirror IO has finished' event handler:
368 */
369 if (!uptodate) {
370 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
371 set_bit(R10BIO_WriteError, &r10_bio->state);
372 dec_rdev = 0;
373 } else {
374 /*
375 * Set R10BIO_Uptodate in our master bio, so that
376 * we will return a good error code for to the higher
377 * levels even if IO on some other mirrored buffer fails.
378 *
379 * The 'master' represents the composite IO operation to
380 * user-side. So if something waits for IO, then it will
381 * wait for the 'master' bio.
382 */
383 sector_t first_bad;
384 int bad_sectors;
385
386 set_bit(R10BIO_Uptodate, &r10_bio->state);
387
388 /* Maybe we can clear some bad blocks. */
389 if (is_badblock(conf->mirrors[dev].rdev,
390 r10_bio->devs[slot].addr,
391 r10_bio->sectors,
392 &first_bad, &bad_sectors)) {
393 bio_put(bio);
394 r10_bio->devs[slot].bio = IO_MADE_GOOD;
395 dec_rdev = 0;
396 set_bit(R10BIO_MadeGood, &r10_bio->state);
397 }
398 }
399
400 /*
401 *
402 * Let's see if all mirrored write operations have finished
403 * already.
404 */
405 one_write_done(r10_bio);
406 if (dec_rdev)
407 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
408}
409
410
411/*
412 * RAID10 layout manager
413 * As well as the chunksize and raid_disks count, there are two
414 * parameters: near_copies and far_copies.
415 * near_copies * far_copies must be <= raid_disks.
416 * Normally one of these will be 1.
417 * If both are 1, we get raid0.
418 * If near_copies == raid_disks, we get raid1.
419 *
420 * Chunks are laid out in raid0 style with near_copies copies of the
421 * first chunk, followed by near_copies copies of the next chunk and
422 * so on.
423 * If far_copies > 1, then after 1/far_copies of the array has been assigned
424 * as described above, we start again with a device offset of near_copies.
425 * So we effectively have another copy of the whole array further down all
426 * the drives, but with blocks on different drives.
427 * With this layout, and block is never stored twice on the one device.
428 *
429 * raid10_find_phys finds the sector offset of a given virtual sector
430 * on each device that it is on.
431 *
432 * raid10_find_virt does the reverse mapping, from a device and a
433 * sector offset to a virtual address
434 */
435
436static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
437{
438 int n,f;
439 sector_t sector;
440 sector_t chunk;
441 sector_t stripe;
442 int dev;
443
444 int slot = 0;
445
446 /* now calculate first sector/dev */
447 chunk = r10bio->sector >> conf->chunk_shift;
448 sector = r10bio->sector & conf->chunk_mask;
449
450 chunk *= conf->near_copies;
451 stripe = chunk;
452 dev = sector_div(stripe, conf->raid_disks);
453 if (conf->far_offset)
454 stripe *= conf->far_copies;
455
456 sector += stripe << conf->chunk_shift;
457
458 /* and calculate all the others */
459 for (n=0; n < conf->near_copies; n++) {
460 int d = dev;
461 sector_t s = sector;
462 r10bio->devs[slot].addr = sector;
463 r10bio->devs[slot].devnum = d;
464 slot++;
465
466 for (f = 1; f < conf->far_copies; f++) {
467 d += conf->near_copies;
468 if (d >= conf->raid_disks)
469 d -= conf->raid_disks;
470 s += conf->stride;
471 r10bio->devs[slot].devnum = d;
472 r10bio->devs[slot].addr = s;
473 slot++;
474 }
475 dev++;
476 if (dev >= conf->raid_disks) {
477 dev = 0;
478 sector += (conf->chunk_mask + 1);
479 }
480 }
481 BUG_ON(slot != conf->copies);
482}
483
484static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
485{
486 sector_t offset, chunk, vchunk;
487
488 offset = sector & conf->chunk_mask;
489 if (conf->far_offset) {
490 int fc;
491 chunk = sector >> conf->chunk_shift;
492 fc = sector_div(chunk, conf->far_copies);
493 dev -= fc * conf->near_copies;
494 if (dev < 0)
495 dev += conf->raid_disks;
496 } else {
497 while (sector >= conf->stride) {
498 sector -= conf->stride;
499 if (dev < conf->near_copies)
500 dev += conf->raid_disks - conf->near_copies;
501 else
502 dev -= conf->near_copies;
503 }
504 chunk = sector >> conf->chunk_shift;
505 }
506 vchunk = chunk * conf->raid_disks + dev;
507 sector_div(vchunk, conf->near_copies);
508 return (vchunk << conf->chunk_shift) + offset;
509}
510
511/**
512 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
513 * @q: request queue
514 * @bvm: properties of new bio
515 * @biovec: the request that could be merged to it.
516 *
517 * Return amount of bytes we can accept at this offset
518 * If near_copies == raid_disk, there are no striping issues,
519 * but in that case, the function isn't called at all.
520 */
521static int raid10_mergeable_bvec(struct request_queue *q,
522 struct bvec_merge_data *bvm,
523 struct bio_vec *biovec)
524{
525 mddev_t *mddev = q->queuedata;
526 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
527 int max;
528 unsigned int chunk_sectors = mddev->chunk_sectors;
529 unsigned int bio_sectors = bvm->bi_size >> 9;
530
531 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
532 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
533 if (max <= biovec->bv_len && bio_sectors == 0)
534 return biovec->bv_len;
535 else
536 return max;
537}
538
539/*
540 * This routine returns the disk from which the requested read should
541 * be done. There is a per-array 'next expected sequential IO' sector
542 * number - if this matches on the next IO then we use the last disk.
543 * There is also a per-disk 'last know head position' sector that is
544 * maintained from IRQ contexts, both the normal and the resync IO
545 * completion handlers update this position correctly. If there is no
546 * perfect sequential match then we pick the disk whose head is closest.
547 *
548 * If there are 2 mirrors in the same 2 devices, performance degrades
549 * because position is mirror, not device based.
550 *
551 * The rdev for the device selected will have nr_pending incremented.
552 */
553
554/*
555 * FIXME: possibly should rethink readbalancing and do it differently
556 * depending on near_copies / far_copies geometry.
557 */
558static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors)
559{
560 const sector_t this_sector = r10_bio->sector;
561 int disk, slot;
562 int sectors = r10_bio->sectors;
563 int best_good_sectors;
564 sector_t new_distance, best_dist;
565 mdk_rdev_t *rdev;
566 int do_balance;
567 int best_slot;
568
569 raid10_find_phys(conf, r10_bio);
570 rcu_read_lock();
571retry:
572 sectors = r10_bio->sectors;
573 best_slot = -1;
574 best_dist = MaxSector;
575 best_good_sectors = 0;
576 do_balance = 1;
577 /*
578 * Check if we can balance. We can balance on the whole
579 * device if no resync is going on (recovery is ok), or below
580 * the resync window. We take the first readable disk when
581 * above the resync window.
582 */
583 if (conf->mddev->recovery_cp < MaxSector
584 && (this_sector + sectors >= conf->next_resync))
585 do_balance = 0;
586
587 for (slot = 0; slot < conf->copies ; slot++) {
588 sector_t first_bad;
589 int bad_sectors;
590 sector_t dev_sector;
591
592 if (r10_bio->devs[slot].bio == IO_BLOCKED)
593 continue;
594 disk = r10_bio->devs[slot].devnum;
595 rdev = rcu_dereference(conf->mirrors[disk].rdev);
596 if (rdev == NULL)
597 continue;
598 if (!test_bit(In_sync, &rdev->flags))
599 continue;
600
601 dev_sector = r10_bio->devs[slot].addr;
602 if (is_badblock(rdev, dev_sector, sectors,
603 &first_bad, &bad_sectors)) {
604 if (best_dist < MaxSector)
605 /* Already have a better slot */
606 continue;
607 if (first_bad <= dev_sector) {
608 /* Cannot read here. If this is the
609 * 'primary' device, then we must not read
610 * beyond 'bad_sectors' from another device.
611 */
612 bad_sectors -= (dev_sector - first_bad);
613 if (!do_balance && sectors > bad_sectors)
614 sectors = bad_sectors;
615 if (best_good_sectors > sectors)
616 best_good_sectors = sectors;
617 } else {
618 sector_t good_sectors =
619 first_bad - dev_sector;
620 if (good_sectors > best_good_sectors) {
621 best_good_sectors = good_sectors;
622 best_slot = slot;
623 }
624 if (!do_balance)
625 /* Must read from here */
626 break;
627 }
628 continue;
629 } else
630 best_good_sectors = sectors;
631
632 if (!do_balance)
633 break;
634
635 /* This optimisation is debatable, and completely destroys
636 * sequential read speed for 'far copies' arrays. So only
637 * keep it for 'near' arrays, and review those later.
638 */
639 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
640 break;
641
642 /* for far > 1 always use the lowest address */
643 if (conf->far_copies > 1)
644 new_distance = r10_bio->devs[slot].addr;
645 else
646 new_distance = abs(r10_bio->devs[slot].addr -
647 conf->mirrors[disk].head_position);
648 if (new_distance < best_dist) {
649 best_dist = new_distance;
650 best_slot = slot;
651 }
652 }
653 if (slot == conf->copies)
654 slot = best_slot;
655
656 if (slot >= 0) {
657 disk = r10_bio->devs[slot].devnum;
658 rdev = rcu_dereference(conf->mirrors[disk].rdev);
659 if (!rdev)
660 goto retry;
661 atomic_inc(&rdev->nr_pending);
662 if (test_bit(Faulty, &rdev->flags)) {
663 /* Cannot risk returning a device that failed
664 * before we inc'ed nr_pending
665 */
666 rdev_dec_pending(rdev, conf->mddev);
667 goto retry;
668 }
669 r10_bio->read_slot = slot;
670 } else
671 disk = -1;
672 rcu_read_unlock();
673 *max_sectors = best_good_sectors;
674
675 return disk;
676}
677
678static int raid10_congested(void *data, int bits)
679{
680 mddev_t *mddev = data;
681 conf_t *conf = mddev->private;
682 int i, ret = 0;
683
684 if (mddev_congested(mddev, bits))
685 return 1;
686 rcu_read_lock();
687 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
688 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
689 if (rdev && !test_bit(Faulty, &rdev->flags)) {
690 struct request_queue *q = bdev_get_queue(rdev->bdev);
691
692 ret |= bdi_congested(&q->backing_dev_info, bits);
693 }
694 }
695 rcu_read_unlock();
696 return ret;
697}
698
699static void flush_pending_writes(conf_t *conf)
700{
701 /* Any writes that have been queued but are awaiting
702 * bitmap updates get flushed here.
703 */
704 spin_lock_irq(&conf->device_lock);
705
706 if (conf->pending_bio_list.head) {
707 struct bio *bio;
708 bio = bio_list_get(&conf->pending_bio_list);
709 spin_unlock_irq(&conf->device_lock);
710 /* flush any pending bitmap writes to disk
711 * before proceeding w/ I/O */
712 bitmap_unplug(conf->mddev->bitmap);
713
714 while (bio) { /* submit pending writes */
715 struct bio *next = bio->bi_next;
716 bio->bi_next = NULL;
717 generic_make_request(bio);
718 bio = next;
719 }
720 } else
721 spin_unlock_irq(&conf->device_lock);
722}
723
724/* Barriers....
725 * Sometimes we need to suspend IO while we do something else,
726 * either some resync/recovery, or reconfigure the array.
727 * To do this we raise a 'barrier'.
728 * The 'barrier' is a counter that can be raised multiple times
729 * to count how many activities are happening which preclude
730 * normal IO.
731 * We can only raise the barrier if there is no pending IO.
732 * i.e. if nr_pending == 0.
733 * We choose only to raise the barrier if no-one is waiting for the
734 * barrier to go down. This means that as soon as an IO request
735 * is ready, no other operations which require a barrier will start
736 * until the IO request has had a chance.
737 *
738 * So: regular IO calls 'wait_barrier'. When that returns there
739 * is no backgroup IO happening, It must arrange to call
740 * allow_barrier when it has finished its IO.
741 * backgroup IO calls must call raise_barrier. Once that returns
742 * there is no normal IO happeing. It must arrange to call
743 * lower_barrier when the particular background IO completes.
744 */
745
746static void raise_barrier(conf_t *conf, int force)
747{
748 BUG_ON(force && !conf->barrier);
749 spin_lock_irq(&conf->resync_lock);
750
751 /* Wait until no block IO is waiting (unless 'force') */
752 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
753 conf->resync_lock, );
754
755 /* block any new IO from starting */
756 conf->barrier++;
757
758 /* Now wait for all pending IO to complete */
759 wait_event_lock_irq(conf->wait_barrier,
760 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
761 conf->resync_lock, );
762
763 spin_unlock_irq(&conf->resync_lock);
764}
765
766static void lower_barrier(conf_t *conf)
767{
768 unsigned long flags;
769 spin_lock_irqsave(&conf->resync_lock, flags);
770 conf->barrier--;
771 spin_unlock_irqrestore(&conf->resync_lock, flags);
772 wake_up(&conf->wait_barrier);
773}
774
775static void wait_barrier(conf_t *conf)
776{
777 spin_lock_irq(&conf->resync_lock);
778 if (conf->barrier) {
779 conf->nr_waiting++;
780 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
781 conf->resync_lock,
782 );
783 conf->nr_waiting--;
784 }
785 conf->nr_pending++;
786 spin_unlock_irq(&conf->resync_lock);
787}
788
789static void allow_barrier(conf_t *conf)
790{
791 unsigned long flags;
792 spin_lock_irqsave(&conf->resync_lock, flags);
793 conf->nr_pending--;
794 spin_unlock_irqrestore(&conf->resync_lock, flags);
795 wake_up(&conf->wait_barrier);
796}
797
798static void freeze_array(conf_t *conf)
799{
800 /* stop syncio and normal IO and wait for everything to
801 * go quiet.
802 * We increment barrier and nr_waiting, and then
803 * wait until nr_pending match nr_queued+1
804 * This is called in the context of one normal IO request
805 * that has failed. Thus any sync request that might be pending
806 * will be blocked by nr_pending, and we need to wait for
807 * pending IO requests to complete or be queued for re-try.
808 * Thus the number queued (nr_queued) plus this request (1)
809 * must match the number of pending IOs (nr_pending) before
810 * we continue.
811 */
812 spin_lock_irq(&conf->resync_lock);
813 conf->barrier++;
814 conf->nr_waiting++;
815 wait_event_lock_irq(conf->wait_barrier,
816 conf->nr_pending == conf->nr_queued+1,
817 conf->resync_lock,
818 flush_pending_writes(conf));
819
820 spin_unlock_irq(&conf->resync_lock);
821}
822
823static void unfreeze_array(conf_t *conf)
824{
825 /* reverse the effect of the freeze */
826 spin_lock_irq(&conf->resync_lock);
827 conf->barrier--;
828 conf->nr_waiting--;
829 wake_up(&conf->wait_barrier);
830 spin_unlock_irq(&conf->resync_lock);
831}
832
833static int make_request(mddev_t *mddev, struct bio * bio)
834{
835 conf_t *conf = mddev->private;
836 mirror_info_t *mirror;
837 r10bio_t *r10_bio;
838 struct bio *read_bio;
839 int i;
840 int chunk_sects = conf->chunk_mask + 1;
841 const int rw = bio_data_dir(bio);
842 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
843 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
844 unsigned long flags;
845 mdk_rdev_t *blocked_rdev;
846 int plugged;
847 int sectors_handled;
848 int max_sectors;
849
850 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
851 md_flush_request(mddev, bio);
852 return 0;
853 }
854
855 /* If this request crosses a chunk boundary, we need to
856 * split it. This will only happen for 1 PAGE (or less) requests.
857 */
858 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
859 > chunk_sects &&
860 conf->near_copies < conf->raid_disks)) {
861 struct bio_pair *bp;
862 /* Sanity check -- queue functions should prevent this happening */
863 if (bio->bi_vcnt != 1 ||
864 bio->bi_idx != 0)
865 goto bad_map;
866 /* This is a one page bio that upper layers
867 * refuse to split for us, so we need to split it.
868 */
869 bp = bio_split(bio,
870 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
871
872 /* Each of these 'make_request' calls will call 'wait_barrier'.
873 * If the first succeeds but the second blocks due to the resync
874 * thread raising the barrier, we will deadlock because the
875 * IO to the underlying device will be queued in generic_make_request
876 * and will never complete, so will never reduce nr_pending.
877 * So increment nr_waiting here so no new raise_barriers will
878 * succeed, and so the second wait_barrier cannot block.
879 */
880 spin_lock_irq(&conf->resync_lock);
881 conf->nr_waiting++;
882 spin_unlock_irq(&conf->resync_lock);
883
884 if (make_request(mddev, &bp->bio1))
885 generic_make_request(&bp->bio1);
886 if (make_request(mddev, &bp->bio2))
887 generic_make_request(&bp->bio2);
888
889 spin_lock_irq(&conf->resync_lock);
890 conf->nr_waiting--;
891 wake_up(&conf->wait_barrier);
892 spin_unlock_irq(&conf->resync_lock);
893
894 bio_pair_release(bp);
895 return 0;
896 bad_map:
897 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
898 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
899 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
900
901 bio_io_error(bio);
902 return 0;
903 }
904
905 md_write_start(mddev, bio);
906
907 /*
908 * Register the new request and wait if the reconstruction
909 * thread has put up a bar for new requests.
910 * Continue immediately if no resync is active currently.
911 */
912 wait_barrier(conf);
913
914 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
915
916 r10_bio->master_bio = bio;
917 r10_bio->sectors = bio->bi_size >> 9;
918
919 r10_bio->mddev = mddev;
920 r10_bio->sector = bio->bi_sector;
921 r10_bio->state = 0;
922
923 /* We might need to issue multiple reads to different
924 * devices if there are bad blocks around, so we keep
925 * track of the number of reads in bio->bi_phys_segments.
926 * If this is 0, there is only one r10_bio and no locking
927 * will be needed when the request completes. If it is
928 * non-zero, then it is the number of not-completed requests.
929 */
930 bio->bi_phys_segments = 0;
931 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
932
933 if (rw == READ) {
934 /*
935 * read balancing logic:
936 */
937 int disk;
938 int slot;
939
940read_again:
941 disk = read_balance(conf, r10_bio, &max_sectors);
942 slot = r10_bio->read_slot;
943 if (disk < 0) {
944 raid_end_bio_io(r10_bio);
945 return 0;
946 }
947 mirror = conf->mirrors + disk;
948
949 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
950 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
951 max_sectors);
952
953 r10_bio->devs[slot].bio = read_bio;
954
955 read_bio->bi_sector = r10_bio->devs[slot].addr +
956 mirror->rdev->data_offset;
957 read_bio->bi_bdev = mirror->rdev->bdev;
958 read_bio->bi_end_io = raid10_end_read_request;
959 read_bio->bi_rw = READ | do_sync;
960 read_bio->bi_private = r10_bio;
961
962 if (max_sectors < r10_bio->sectors) {
963 /* Could not read all from this device, so we will
964 * need another r10_bio.
965 */
966 sectors_handled = (r10_bio->sectors + max_sectors
967 - bio->bi_sector);
968 r10_bio->sectors = max_sectors;
969 spin_lock_irq(&conf->device_lock);
970 if (bio->bi_phys_segments == 0)
971 bio->bi_phys_segments = 2;
972 else
973 bio->bi_phys_segments++;
974 spin_unlock(&conf->device_lock);
975 /* Cannot call generic_make_request directly
976 * as that will be queued in __generic_make_request
977 * and subsequent mempool_alloc might block
978 * waiting for it. so hand bio over to raid10d.
979 */
980 reschedule_retry(r10_bio);
981
982 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
983
984 r10_bio->master_bio = bio;
985 r10_bio->sectors = ((bio->bi_size >> 9)
986 - sectors_handled);
987 r10_bio->state = 0;
988 r10_bio->mddev = mddev;
989 r10_bio->sector = bio->bi_sector + sectors_handled;
990 goto read_again;
991 } else
992 generic_make_request(read_bio);
993 return 0;
994 }
995
996 /*
997 * WRITE:
998 */
999 /* first select target devices under rcu_lock and
1000 * inc refcount on their rdev. Record them by setting
1001 * bios[x] to bio
1002 * If there are known/acknowledged bad blocks on any device
1003 * on which we have seen a write error, we want to avoid
1004 * writing to those blocks. This potentially requires several
1005 * writes to write around the bad blocks. Each set of writes
1006 * gets its own r10_bio with a set of bios attached. The number
1007 * of r10_bios is recored in bio->bi_phys_segments just as with
1008 * the read case.
1009 */
1010 plugged = mddev_check_plugged(mddev);
1011
1012 raid10_find_phys(conf, r10_bio);
1013retry_write:
1014 blocked_rdev = NULL;
1015 rcu_read_lock();
1016 max_sectors = r10_bio->sectors;
1017
1018 for (i = 0; i < conf->copies; i++) {
1019 int d = r10_bio->devs[i].devnum;
1020 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
1021 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1022 atomic_inc(&rdev->nr_pending);
1023 blocked_rdev = rdev;
1024 break;
1025 }
1026 r10_bio->devs[i].bio = NULL;
1027 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1028 set_bit(R10BIO_Degraded, &r10_bio->state);
1029 continue;
1030 }
1031 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1032 sector_t first_bad;
1033 sector_t dev_sector = r10_bio->devs[i].addr;
1034 int bad_sectors;
1035 int is_bad;
1036
1037 is_bad = is_badblock(rdev, dev_sector,
1038 max_sectors,
1039 &first_bad, &bad_sectors);
1040 if (is_bad < 0) {
1041 /* Mustn't write here until the bad block
1042 * is acknowledged
1043 */
1044 atomic_inc(&rdev->nr_pending);
1045 set_bit(BlockedBadBlocks, &rdev->flags);
1046 blocked_rdev = rdev;
1047 break;
1048 }
1049 if (is_bad && first_bad <= dev_sector) {
1050 /* Cannot write here at all */
1051 bad_sectors -= (dev_sector - first_bad);
1052 if (bad_sectors < max_sectors)
1053 /* Mustn't write more than bad_sectors
1054 * to other devices yet
1055 */
1056 max_sectors = bad_sectors;
1057 /* We don't set R10BIO_Degraded as that
1058 * only applies if the disk is missing,
1059 * so it might be re-added, and we want to
1060 * know to recover this chunk.
1061 * In this case the device is here, and the
1062 * fact that this chunk is not in-sync is
1063 * recorded in the bad block log.
1064 */
1065 continue;
1066 }
1067 if (is_bad) {
1068 int good_sectors = first_bad - dev_sector;
1069 if (good_sectors < max_sectors)
1070 max_sectors = good_sectors;
1071 }
1072 }
1073 r10_bio->devs[i].bio = bio;
1074 atomic_inc(&rdev->nr_pending);
1075 }
1076 rcu_read_unlock();
1077
1078 if (unlikely(blocked_rdev)) {
1079 /* Have to wait for this device to get unblocked, then retry */
1080 int j;
1081 int d;
1082
1083 for (j = 0; j < i; j++)
1084 if (r10_bio->devs[j].bio) {
1085 d = r10_bio->devs[j].devnum;
1086 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1087 }
1088 allow_barrier(conf);
1089 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1090 wait_barrier(conf);
1091 goto retry_write;
1092 }
1093
1094 if (max_sectors < r10_bio->sectors) {
1095 /* We are splitting this into multiple parts, so
1096 * we need to prepare for allocating another r10_bio.
1097 */
1098 r10_bio->sectors = max_sectors;
1099 spin_lock_irq(&conf->device_lock);
1100 if (bio->bi_phys_segments == 0)
1101 bio->bi_phys_segments = 2;
1102 else
1103 bio->bi_phys_segments++;
1104 spin_unlock_irq(&conf->device_lock);
1105 }
1106 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1107
1108 atomic_set(&r10_bio->remaining, 1);
1109 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1110
1111 for (i = 0; i < conf->copies; i++) {
1112 struct bio *mbio;
1113 int d = r10_bio->devs[i].devnum;
1114 if (!r10_bio->devs[i].bio)
1115 continue;
1116
1117 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1118 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1119 max_sectors);
1120 r10_bio->devs[i].bio = mbio;
1121
1122 mbio->bi_sector = (r10_bio->devs[i].addr+
1123 conf->mirrors[d].rdev->data_offset);
1124 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1125 mbio->bi_end_io = raid10_end_write_request;
1126 mbio->bi_rw = WRITE | do_sync | do_fua;
1127 mbio->bi_private = r10_bio;
1128
1129 atomic_inc(&r10_bio->remaining);
1130 spin_lock_irqsave(&conf->device_lock, flags);
1131 bio_list_add(&conf->pending_bio_list, mbio);
1132 spin_unlock_irqrestore(&conf->device_lock, flags);
1133 }
1134
1135 /* Don't remove the bias on 'remaining' (one_write_done) until
1136 * after checking if we need to go around again.
1137 */
1138
1139 if (sectors_handled < (bio->bi_size >> 9)) {
1140 one_write_done(r10_bio);
1141 /* We need another r10_bio. It has already been counted
1142 * in bio->bi_phys_segments.
1143 */
1144 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1145
1146 r10_bio->master_bio = bio;
1147 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1148
1149 r10_bio->mddev = mddev;
1150 r10_bio->sector = bio->bi_sector + sectors_handled;
1151 r10_bio->state = 0;
1152 goto retry_write;
1153 }
1154 one_write_done(r10_bio);
1155
1156 /* In case raid10d snuck in to freeze_array */
1157 wake_up(&conf->wait_barrier);
1158
1159 if (do_sync || !mddev->bitmap || !plugged)
1160 md_wakeup_thread(mddev->thread);
1161 return 0;
1162}
1163
1164static void status(struct seq_file *seq, mddev_t *mddev)
1165{
1166 conf_t *conf = mddev->private;
1167 int i;
1168
1169 if (conf->near_copies < conf->raid_disks)
1170 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1171 if (conf->near_copies > 1)
1172 seq_printf(seq, " %d near-copies", conf->near_copies);
1173 if (conf->far_copies > 1) {
1174 if (conf->far_offset)
1175 seq_printf(seq, " %d offset-copies", conf->far_copies);
1176 else
1177 seq_printf(seq, " %d far-copies", conf->far_copies);
1178 }
1179 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1180 conf->raid_disks - mddev->degraded);
1181 for (i = 0; i < conf->raid_disks; i++)
1182 seq_printf(seq, "%s",
1183 conf->mirrors[i].rdev &&
1184 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1185 seq_printf(seq, "]");
1186}
1187
1188/* check if there are enough drives for
1189 * every block to appear on atleast one.
1190 * Don't consider the device numbered 'ignore'
1191 * as we might be about to remove it.
1192 */
1193static int enough(conf_t *conf, int ignore)
1194{
1195 int first = 0;
1196
1197 do {
1198 int n = conf->copies;
1199 int cnt = 0;
1200 while (n--) {
1201 if (conf->mirrors[first].rdev &&
1202 first != ignore)
1203 cnt++;
1204 first = (first+1) % conf->raid_disks;
1205 }
1206 if (cnt == 0)
1207 return 0;
1208 } while (first != 0);
1209 return 1;
1210}
1211
1212static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1213{
1214 char b[BDEVNAME_SIZE];
1215 conf_t *conf = mddev->private;
1216
1217 /*
1218 * If it is not operational, then we have already marked it as dead
1219 * else if it is the last working disks, ignore the error, let the
1220 * next level up know.
1221 * else mark the drive as failed
1222 */
1223 if (test_bit(In_sync, &rdev->flags)
1224 && !enough(conf, rdev->raid_disk))
1225 /*
1226 * Don't fail the drive, just return an IO error.
1227 */
1228 return;
1229 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1230 unsigned long flags;
1231 spin_lock_irqsave(&conf->device_lock, flags);
1232 mddev->degraded++;
1233 spin_unlock_irqrestore(&conf->device_lock, flags);
1234 /*
1235 * if recovery is running, make sure it aborts.
1236 */
1237 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1238 }
1239 set_bit(Blocked, &rdev->flags);
1240 set_bit(Faulty, &rdev->flags);
1241 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1242 printk(KERN_ALERT
1243 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1244 "md/raid10:%s: Operation continuing on %d devices.\n",
1245 mdname(mddev), bdevname(rdev->bdev, b),
1246 mdname(mddev), conf->raid_disks - mddev->degraded);
1247}
1248
1249static void print_conf(conf_t *conf)
1250{
1251 int i;
1252 mirror_info_t *tmp;
1253
1254 printk(KERN_DEBUG "RAID10 conf printout:\n");
1255 if (!conf) {
1256 printk(KERN_DEBUG "(!conf)\n");
1257 return;
1258 }
1259 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1260 conf->raid_disks);
1261
1262 for (i = 0; i < conf->raid_disks; i++) {
1263 char b[BDEVNAME_SIZE];
1264 tmp = conf->mirrors + i;
1265 if (tmp->rdev)
1266 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1267 i, !test_bit(In_sync, &tmp->rdev->flags),
1268 !test_bit(Faulty, &tmp->rdev->flags),
1269 bdevname(tmp->rdev->bdev,b));
1270 }
1271}
1272
1273static void close_sync(conf_t *conf)
1274{
1275 wait_barrier(conf);
1276 allow_barrier(conf);
1277
1278 mempool_destroy(conf->r10buf_pool);
1279 conf->r10buf_pool = NULL;
1280}
1281
1282static int raid10_spare_active(mddev_t *mddev)
1283{
1284 int i;
1285 conf_t *conf = mddev->private;
1286 mirror_info_t *tmp;
1287 int count = 0;
1288 unsigned long flags;
1289
1290 /*
1291 * Find all non-in_sync disks within the RAID10 configuration
1292 * and mark them in_sync
1293 */
1294 for (i = 0; i < conf->raid_disks; i++) {
1295 tmp = conf->mirrors + i;
1296 if (tmp->rdev
1297 && !test_bit(Faulty, &tmp->rdev->flags)
1298 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1299 count++;
1300 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1301 }
1302 }
1303 spin_lock_irqsave(&conf->device_lock, flags);
1304 mddev->degraded -= count;
1305 spin_unlock_irqrestore(&conf->device_lock, flags);
1306
1307 print_conf(conf);
1308 return count;
1309}
1310
1311
1312static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1313{
1314 conf_t *conf = mddev->private;
1315 int err = -EEXIST;
1316 int mirror;
1317 int first = 0;
1318 int last = conf->raid_disks - 1;
1319
1320 if (mddev->recovery_cp < MaxSector)
1321 /* only hot-add to in-sync arrays, as recovery is
1322 * very different from resync
1323 */
1324 return -EBUSY;
1325 if (!enough(conf, -1))
1326 return -EINVAL;
1327
1328 if (rdev->raid_disk >= 0)
1329 first = last = rdev->raid_disk;
1330
1331 if (rdev->saved_raid_disk >= first &&
1332 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1333 mirror = rdev->saved_raid_disk;
1334 else
1335 mirror = first;
1336 for ( ; mirror <= last ; mirror++) {
1337 mirror_info_t *p = &conf->mirrors[mirror];
1338 if (p->recovery_disabled == mddev->recovery_disabled)
1339 continue;
1340 if (!p->rdev)
1341 continue;
1342
1343 disk_stack_limits(mddev->gendisk, rdev->bdev,
1344 rdev->data_offset << 9);
1345 /* as we don't honour merge_bvec_fn, we must
1346 * never risk violating it, so limit
1347 * ->max_segments to one lying with a single
1348 * page, as a one page request is never in
1349 * violation.
1350 */
1351 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1352 blk_queue_max_segments(mddev->queue, 1);
1353 blk_queue_segment_boundary(mddev->queue,
1354 PAGE_CACHE_SIZE - 1);
1355 }
1356
1357 p->head_position = 0;
1358 rdev->raid_disk = mirror;
1359 err = 0;
1360 if (rdev->saved_raid_disk != mirror)
1361 conf->fullsync = 1;
1362 rcu_assign_pointer(p->rdev, rdev);
1363 break;
1364 }
1365
1366 md_integrity_add_rdev(rdev, mddev);
1367 print_conf(conf);
1368 return err;
1369}
1370
1371static int raid10_remove_disk(mddev_t *mddev, int number)
1372{
1373 conf_t *conf = mddev->private;
1374 int err = 0;
1375 mdk_rdev_t *rdev;
1376 mirror_info_t *p = conf->mirrors+ number;
1377
1378 print_conf(conf);
1379 rdev = p->rdev;
1380 if (rdev) {
1381 if (test_bit(In_sync, &rdev->flags) ||
1382 atomic_read(&rdev->nr_pending)) {
1383 err = -EBUSY;
1384 goto abort;
1385 }
1386 /* Only remove faulty devices in recovery
1387 * is not possible.
1388 */
1389 if (!test_bit(Faulty, &rdev->flags) &&
1390 mddev->recovery_disabled != p->recovery_disabled &&
1391 enough(conf, -1)) {
1392 err = -EBUSY;
1393 goto abort;
1394 }
1395 p->rdev = NULL;
1396 synchronize_rcu();
1397 if (atomic_read(&rdev->nr_pending)) {
1398 /* lost the race, try later */
1399 err = -EBUSY;
1400 p->rdev = rdev;
1401 goto abort;
1402 }
1403 err = md_integrity_register(mddev);
1404 }
1405abort:
1406
1407 print_conf(conf);
1408 return err;
1409}
1410
1411
1412static void end_sync_read(struct bio *bio, int error)
1413{
1414 r10bio_t *r10_bio = bio->bi_private;
1415 conf_t *conf = r10_bio->mddev->private;
1416 int d;
1417
1418 d = find_bio_disk(conf, r10_bio, bio, NULL);
1419
1420 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1421 set_bit(R10BIO_Uptodate, &r10_bio->state);
1422 else
1423 /* The write handler will notice the lack of
1424 * R10BIO_Uptodate and record any errors etc
1425 */
1426 atomic_add(r10_bio->sectors,
1427 &conf->mirrors[d].rdev->corrected_errors);
1428
1429 /* for reconstruct, we always reschedule after a read.
1430 * for resync, only after all reads
1431 */
1432 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1433 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1434 atomic_dec_and_test(&r10_bio->remaining)) {
1435 /* we have read all the blocks,
1436 * do the comparison in process context in raid10d
1437 */
1438 reschedule_retry(r10_bio);
1439 }
1440}
1441
1442static void end_sync_request(r10bio_t *r10_bio)
1443{
1444 mddev_t *mddev = r10_bio->mddev;
1445
1446 while (atomic_dec_and_test(&r10_bio->remaining)) {
1447 if (r10_bio->master_bio == NULL) {
1448 /* the primary of several recovery bios */
1449 sector_t s = r10_bio->sectors;
1450 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1451 test_bit(R10BIO_WriteError, &r10_bio->state))
1452 reschedule_retry(r10_bio);
1453 else
1454 put_buf(r10_bio);
1455 md_done_sync(mddev, s, 1);
1456 break;
1457 } else {
1458 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1459 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1460 test_bit(R10BIO_WriteError, &r10_bio->state))
1461 reschedule_retry(r10_bio);
1462 else
1463 put_buf(r10_bio);
1464 r10_bio = r10_bio2;
1465 }
1466 }
1467}
1468
1469static void end_sync_write(struct bio *bio, int error)
1470{
1471 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1472 r10bio_t *r10_bio = bio->bi_private;
1473 mddev_t *mddev = r10_bio->mddev;
1474 conf_t *conf = mddev->private;
1475 int d;
1476 sector_t first_bad;
1477 int bad_sectors;
1478 int slot;
1479
1480 d = find_bio_disk(conf, r10_bio, bio, &slot);
1481
1482 if (!uptodate) {
1483 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1484 set_bit(R10BIO_WriteError, &r10_bio->state);
1485 } else if (is_badblock(conf->mirrors[d].rdev,
1486 r10_bio->devs[slot].addr,
1487 r10_bio->sectors,
1488 &first_bad, &bad_sectors))
1489 set_bit(R10BIO_MadeGood, &r10_bio->state);
1490
1491 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1492
1493 end_sync_request(r10_bio);
1494}
1495
1496/*
1497 * Note: sync and recover and handled very differently for raid10
1498 * This code is for resync.
1499 * For resync, we read through virtual addresses and read all blocks.
1500 * If there is any error, we schedule a write. The lowest numbered
1501 * drive is authoritative.
1502 * However requests come for physical address, so we need to map.
1503 * For every physical address there are raid_disks/copies virtual addresses,
1504 * which is always are least one, but is not necessarly an integer.
1505 * This means that a physical address can span multiple chunks, so we may
1506 * have to submit multiple io requests for a single sync request.
1507 */
1508/*
1509 * We check if all blocks are in-sync and only write to blocks that
1510 * aren't in sync
1511 */
1512static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1513{
1514 conf_t *conf = mddev->private;
1515 int i, first;
1516 struct bio *tbio, *fbio;
1517
1518 atomic_set(&r10_bio->remaining, 1);
1519
1520 /* find the first device with a block */
1521 for (i=0; i<conf->copies; i++)
1522 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1523 break;
1524
1525 if (i == conf->copies)
1526 goto done;
1527
1528 first = i;
1529 fbio = r10_bio->devs[i].bio;
1530
1531 /* now find blocks with errors */
1532 for (i=0 ; i < conf->copies ; i++) {
1533 int j, d;
1534 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1535
1536 tbio = r10_bio->devs[i].bio;
1537
1538 if (tbio->bi_end_io != end_sync_read)
1539 continue;
1540 if (i == first)
1541 continue;
1542 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1543 /* We know that the bi_io_vec layout is the same for
1544 * both 'first' and 'i', so we just compare them.
1545 * All vec entries are PAGE_SIZE;
1546 */
1547 for (j = 0; j < vcnt; j++)
1548 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1549 page_address(tbio->bi_io_vec[j].bv_page),
1550 PAGE_SIZE))
1551 break;
1552 if (j == vcnt)
1553 continue;
1554 mddev->resync_mismatches += r10_bio->sectors;
1555 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1556 /* Don't fix anything. */
1557 continue;
1558 }
1559 /* Ok, we need to write this bio, either to correct an
1560 * inconsistency or to correct an unreadable block.
1561 * First we need to fixup bv_offset, bv_len and
1562 * bi_vecs, as the read request might have corrupted these
1563 */
1564 tbio->bi_vcnt = vcnt;
1565 tbio->bi_size = r10_bio->sectors << 9;
1566 tbio->bi_idx = 0;
1567 tbio->bi_phys_segments = 0;
1568 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1569 tbio->bi_flags |= 1 << BIO_UPTODATE;
1570 tbio->bi_next = NULL;
1571 tbio->bi_rw = WRITE;
1572 tbio->bi_private = r10_bio;
1573 tbio->bi_sector = r10_bio->devs[i].addr;
1574
1575 for (j=0; j < vcnt ; j++) {
1576 tbio->bi_io_vec[j].bv_offset = 0;
1577 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1578
1579 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1580 page_address(fbio->bi_io_vec[j].bv_page),
1581 PAGE_SIZE);
1582 }
1583 tbio->bi_end_io = end_sync_write;
1584
1585 d = r10_bio->devs[i].devnum;
1586 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1587 atomic_inc(&r10_bio->remaining);
1588 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1589
1590 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1591 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1592 generic_make_request(tbio);
1593 }
1594
1595done:
1596 if (atomic_dec_and_test(&r10_bio->remaining)) {
1597 md_done_sync(mddev, r10_bio->sectors, 1);
1598 put_buf(r10_bio);
1599 }
1600}
1601
1602/*
1603 * Now for the recovery code.
1604 * Recovery happens across physical sectors.
1605 * We recover all non-is_sync drives by finding the virtual address of
1606 * each, and then choose a working drive that also has that virt address.
1607 * There is a separate r10_bio for each non-in_sync drive.
1608 * Only the first two slots are in use. The first for reading,
1609 * The second for writing.
1610 *
1611 */
1612static void fix_recovery_read_error(r10bio_t *r10_bio)
1613{
1614 /* We got a read error during recovery.
1615 * We repeat the read in smaller page-sized sections.
1616 * If a read succeeds, write it to the new device or record
1617 * a bad block if we cannot.
1618 * If a read fails, record a bad block on both old and
1619 * new devices.
1620 */
1621 mddev_t *mddev = r10_bio->mddev;
1622 conf_t *conf = mddev->private;
1623 struct bio *bio = r10_bio->devs[0].bio;
1624 sector_t sect = 0;
1625 int sectors = r10_bio->sectors;
1626 int idx = 0;
1627 int dr = r10_bio->devs[0].devnum;
1628 int dw = r10_bio->devs[1].devnum;
1629
1630 while (sectors) {
1631 int s = sectors;
1632 mdk_rdev_t *rdev;
1633 sector_t addr;
1634 int ok;
1635
1636 if (s > (PAGE_SIZE>>9))
1637 s = PAGE_SIZE >> 9;
1638
1639 rdev = conf->mirrors[dr].rdev;
1640 addr = r10_bio->devs[0].addr + sect,
1641 ok = sync_page_io(rdev,
1642 addr,
1643 s << 9,
1644 bio->bi_io_vec[idx].bv_page,
1645 READ, false);
1646 if (ok) {
1647 rdev = conf->mirrors[dw].rdev;
1648 addr = r10_bio->devs[1].addr + sect;
1649 ok = sync_page_io(rdev,
1650 addr,
1651 s << 9,
1652 bio->bi_io_vec[idx].bv_page,
1653 WRITE, false);
1654 if (!ok)
1655 set_bit(WriteErrorSeen, &rdev->flags);
1656 }
1657 if (!ok) {
1658 /* We don't worry if we cannot set a bad block -
1659 * it really is bad so there is no loss in not
1660 * recording it yet
1661 */
1662 rdev_set_badblocks(rdev, addr, s, 0);
1663
1664 if (rdev != conf->mirrors[dw].rdev) {
1665 /* need bad block on destination too */
1666 mdk_rdev_t *rdev2 = conf->mirrors[dw].rdev;
1667 addr = r10_bio->devs[1].addr + sect;
1668 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1669 if (!ok) {
1670 /* just abort the recovery */
1671 printk(KERN_NOTICE
1672 "md/raid10:%s: recovery aborted"
1673 " due to read error\n",
1674 mdname(mddev));
1675
1676 conf->mirrors[dw].recovery_disabled
1677 = mddev->recovery_disabled;
1678 set_bit(MD_RECOVERY_INTR,
1679 &mddev->recovery);
1680 break;
1681 }
1682 }
1683 }
1684
1685 sectors -= s;
1686 sect += s;
1687 idx++;
1688 }
1689}
1690
1691static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1692{
1693 conf_t *conf = mddev->private;
1694 int d;
1695 struct bio *wbio;
1696
1697 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1698 fix_recovery_read_error(r10_bio);
1699 end_sync_request(r10_bio);
1700 return;
1701 }
1702
1703 /*
1704 * share the pages with the first bio
1705 * and submit the write request
1706 */
1707 wbio = r10_bio->devs[1].bio;
1708 d = r10_bio->devs[1].devnum;
1709
1710 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1711 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1712 generic_make_request(wbio);
1713}
1714
1715
1716/*
1717 * Used by fix_read_error() to decay the per rdev read_errors.
1718 * We halve the read error count for every hour that has elapsed
1719 * since the last recorded read error.
1720 *
1721 */
1722static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1723{
1724 struct timespec cur_time_mon;
1725 unsigned long hours_since_last;
1726 unsigned int read_errors = atomic_read(&rdev->read_errors);
1727
1728 ktime_get_ts(&cur_time_mon);
1729
1730 if (rdev->last_read_error.tv_sec == 0 &&
1731 rdev->last_read_error.tv_nsec == 0) {
1732 /* first time we've seen a read error */
1733 rdev->last_read_error = cur_time_mon;
1734 return;
1735 }
1736
1737 hours_since_last = (cur_time_mon.tv_sec -
1738 rdev->last_read_error.tv_sec) / 3600;
1739
1740 rdev->last_read_error = cur_time_mon;
1741
1742 /*
1743 * if hours_since_last is > the number of bits in read_errors
1744 * just set read errors to 0. We do this to avoid
1745 * overflowing the shift of read_errors by hours_since_last.
1746 */
1747 if (hours_since_last >= 8 * sizeof(read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1749 else
1750 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1751}
1752
1753static int r10_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1754 int sectors, struct page *page, int rw)
1755{
1756 sector_t first_bad;
1757 int bad_sectors;
1758
1759 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1760 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1761 return -1;
1762 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1763 /* success */
1764 return 1;
1765 if (rw == WRITE)
1766 set_bit(WriteErrorSeen, &rdev->flags);
1767 /* need to record an error - either for the block or the device */
1768 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1769 md_error(rdev->mddev, rdev);
1770 return 0;
1771}
1772
1773/*
1774 * This is a kernel thread which:
1775 *
1776 * 1. Retries failed read operations on working mirrors.
1777 * 2. Updates the raid superblock when problems encounter.
1778 * 3. Performs writes following reads for array synchronising.
1779 */
1780
1781static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1782{
1783 int sect = 0; /* Offset from r10_bio->sector */
1784 int sectors = r10_bio->sectors;
1785 mdk_rdev_t*rdev;
1786 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1787 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1788
1789 /* still own a reference to this rdev, so it cannot
1790 * have been cleared recently.
1791 */
1792 rdev = conf->mirrors[d].rdev;
1793
1794 if (test_bit(Faulty, &rdev->flags))
1795 /* drive has already been failed, just ignore any
1796 more fix_read_error() attempts */
1797 return;
1798
1799 check_decay_read_errors(mddev, rdev);
1800 atomic_inc(&rdev->read_errors);
1801 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1802 char b[BDEVNAME_SIZE];
1803 bdevname(rdev->bdev, b);
1804
1805 printk(KERN_NOTICE
1806 "md/raid10:%s: %s: Raid device exceeded "
1807 "read_error threshold [cur %d:max %d]\n",
1808 mdname(mddev), b,
1809 atomic_read(&rdev->read_errors), max_read_errors);
1810 printk(KERN_NOTICE
1811 "md/raid10:%s: %s: Failing raid device\n",
1812 mdname(mddev), b);
1813 md_error(mddev, conf->mirrors[d].rdev);
1814 return;
1815 }
1816
1817 while(sectors) {
1818 int s = sectors;
1819 int sl = r10_bio->read_slot;
1820 int success = 0;
1821 int start;
1822
1823 if (s > (PAGE_SIZE>>9))
1824 s = PAGE_SIZE >> 9;
1825
1826 rcu_read_lock();
1827 do {
1828 sector_t first_bad;
1829 int bad_sectors;
1830
1831 d = r10_bio->devs[sl].devnum;
1832 rdev = rcu_dereference(conf->mirrors[d].rdev);
1833 if (rdev &&
1834 test_bit(In_sync, &rdev->flags) &&
1835 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1836 &first_bad, &bad_sectors) == 0) {
1837 atomic_inc(&rdev->nr_pending);
1838 rcu_read_unlock();
1839 success = sync_page_io(rdev,
1840 r10_bio->devs[sl].addr +
1841 sect,
1842 s<<9,
1843 conf->tmppage, READ, false);
1844 rdev_dec_pending(rdev, mddev);
1845 rcu_read_lock();
1846 if (success)
1847 break;
1848 }
1849 sl++;
1850 if (sl == conf->copies)
1851 sl = 0;
1852 } while (!success && sl != r10_bio->read_slot);
1853 rcu_read_unlock();
1854
1855 if (!success) {
1856 /* Cannot read from anywhere, just mark the block
1857 * as bad on the first device to discourage future
1858 * reads.
1859 */
1860 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1861 rdev = conf->mirrors[dn].rdev;
1862
1863 if (!rdev_set_badblocks(
1864 rdev,
1865 r10_bio->devs[r10_bio->read_slot].addr
1866 + sect,
1867 s, 0))
1868 md_error(mddev, rdev);
1869 break;
1870 }
1871
1872 start = sl;
1873 /* write it back and re-read */
1874 rcu_read_lock();
1875 while (sl != r10_bio->read_slot) {
1876 char b[BDEVNAME_SIZE];
1877
1878 if (sl==0)
1879 sl = conf->copies;
1880 sl--;
1881 d = r10_bio->devs[sl].devnum;
1882 rdev = rcu_dereference(conf->mirrors[d].rdev);
1883 if (!rdev ||
1884 !test_bit(In_sync, &rdev->flags))
1885 continue;
1886
1887 atomic_inc(&rdev->nr_pending);
1888 rcu_read_unlock();
1889 if (r10_sync_page_io(rdev,
1890 r10_bio->devs[sl].addr +
1891 sect,
1892 s<<9, conf->tmppage, WRITE)
1893 == 0) {
1894 /* Well, this device is dead */
1895 printk(KERN_NOTICE
1896 "md/raid10:%s: read correction "
1897 "write failed"
1898 " (%d sectors at %llu on %s)\n",
1899 mdname(mddev), s,
1900 (unsigned long long)(
1901 sect + rdev->data_offset),
1902 bdevname(rdev->bdev, b));
1903 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1904 "drive\n",
1905 mdname(mddev),
1906 bdevname(rdev->bdev, b));
1907 }
1908 rdev_dec_pending(rdev, mddev);
1909 rcu_read_lock();
1910 }
1911 sl = start;
1912 while (sl != r10_bio->read_slot) {
1913 char b[BDEVNAME_SIZE];
1914
1915 if (sl==0)
1916 sl = conf->copies;
1917 sl--;
1918 d = r10_bio->devs[sl].devnum;
1919 rdev = rcu_dereference(conf->mirrors[d].rdev);
1920 if (!rdev ||
1921 !test_bit(In_sync, &rdev->flags))
1922 continue;
1923
1924 atomic_inc(&rdev->nr_pending);
1925 rcu_read_unlock();
1926 switch (r10_sync_page_io(rdev,
1927 r10_bio->devs[sl].addr +
1928 sect,
1929 s<<9, conf->tmppage,
1930 READ)) {
1931 case 0:
1932 /* Well, this device is dead */
1933 printk(KERN_NOTICE
1934 "md/raid10:%s: unable to read back "
1935 "corrected sectors"
1936 " (%d sectors at %llu on %s)\n",
1937 mdname(mddev), s,
1938 (unsigned long long)(
1939 sect + rdev->data_offset),
1940 bdevname(rdev->bdev, b));
1941 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1942 "drive\n",
1943 mdname(mddev),
1944 bdevname(rdev->bdev, b));
1945 break;
1946 case 1:
1947 printk(KERN_INFO
1948 "md/raid10:%s: read error corrected"
1949 " (%d sectors at %llu on %s)\n",
1950 mdname(mddev), s,
1951 (unsigned long long)(
1952 sect + rdev->data_offset),
1953 bdevname(rdev->bdev, b));
1954 atomic_add(s, &rdev->corrected_errors);
1955 }
1956
1957 rdev_dec_pending(rdev, mddev);
1958 rcu_read_lock();
1959 }
1960 rcu_read_unlock();
1961
1962 sectors -= s;
1963 sect += s;
1964 }
1965}
1966
1967static void bi_complete(struct bio *bio, int error)
1968{
1969 complete((struct completion *)bio->bi_private);
1970}
1971
1972static int submit_bio_wait(int rw, struct bio *bio)
1973{
1974 struct completion event;
1975 rw |= REQ_SYNC;
1976
1977 init_completion(&event);
1978 bio->bi_private = &event;
1979 bio->bi_end_io = bi_complete;
1980 submit_bio(rw, bio);
1981 wait_for_completion(&event);
1982
1983 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1984}
1985
1986static int narrow_write_error(r10bio_t *r10_bio, int i)
1987{
1988 struct bio *bio = r10_bio->master_bio;
1989 mddev_t *mddev = r10_bio->mddev;
1990 conf_t *conf = mddev->private;
1991 mdk_rdev_t *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
1992 /* bio has the data to be written to slot 'i' where
1993 * we just recently had a write error.
1994 * We repeatedly clone the bio and trim down to one block,
1995 * then try the write. Where the write fails we record
1996 * a bad block.
1997 * It is conceivable that the bio doesn't exactly align with
1998 * blocks. We must handle this.
1999 *
2000 * We currently own a reference to the rdev.
2001 */
2002
2003 int block_sectors;
2004 sector_t sector;
2005 int sectors;
2006 int sect_to_write = r10_bio->sectors;
2007 int ok = 1;
2008
2009 if (rdev->badblocks.shift < 0)
2010 return 0;
2011
2012 block_sectors = 1 << rdev->badblocks.shift;
2013 sector = r10_bio->sector;
2014 sectors = ((r10_bio->sector + block_sectors)
2015 & ~(sector_t)(block_sectors - 1))
2016 - sector;
2017
2018 while (sect_to_write) {
2019 struct bio *wbio;
2020 if (sectors > sect_to_write)
2021 sectors = sect_to_write;
2022 /* Write at 'sector' for 'sectors' */
2023 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2024 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2025 wbio->bi_sector = (r10_bio->devs[i].addr+
2026 rdev->data_offset+
2027 (sector - r10_bio->sector));
2028 wbio->bi_bdev = rdev->bdev;
2029 if (submit_bio_wait(WRITE, wbio) == 0)
2030 /* Failure! */
2031 ok = rdev_set_badblocks(rdev, sector,
2032 sectors, 0)
2033 && ok;
2034
2035 bio_put(wbio);
2036 sect_to_write -= sectors;
2037 sector += sectors;
2038 sectors = block_sectors;
2039 }
2040 return ok;
2041}
2042
2043static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio)
2044{
2045 int slot = r10_bio->read_slot;
2046 int mirror = r10_bio->devs[slot].devnum;
2047 struct bio *bio;
2048 conf_t *conf = mddev->private;
2049 mdk_rdev_t *rdev;
2050 char b[BDEVNAME_SIZE];
2051 unsigned long do_sync;
2052 int max_sectors;
2053
2054 /* we got a read error. Maybe the drive is bad. Maybe just
2055 * the block and we can fix it.
2056 * We freeze all other IO, and try reading the block from
2057 * other devices. When we find one, we re-write
2058 * and check it that fixes the read error.
2059 * This is all done synchronously while the array is
2060 * frozen.
2061 */
2062 if (mddev->ro == 0) {
2063 freeze_array(conf);
2064 fix_read_error(conf, mddev, r10_bio);
2065 unfreeze_array(conf);
2066 }
2067 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2068
2069 bio = r10_bio->devs[slot].bio;
2070 bdevname(bio->bi_bdev, b);
2071 r10_bio->devs[slot].bio =
2072 mddev->ro ? IO_BLOCKED : NULL;
2073read_more:
2074 mirror = read_balance(conf, r10_bio, &max_sectors);
2075 if (mirror == -1) {
2076 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2077 " read error for block %llu\n",
2078 mdname(mddev), b,
2079 (unsigned long long)r10_bio->sector);
2080 raid_end_bio_io(r10_bio);
2081 bio_put(bio);
2082 return;
2083 }
2084
2085 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2086 if (bio)
2087 bio_put(bio);
2088 slot = r10_bio->read_slot;
2089 rdev = conf->mirrors[mirror].rdev;
2090 printk_ratelimited(
2091 KERN_ERR
2092 "md/raid10:%s: %s: redirecting"
2093 "sector %llu to another mirror\n",
2094 mdname(mddev),
2095 bdevname(rdev->bdev, b),
2096 (unsigned long long)r10_bio->sector);
2097 bio = bio_clone_mddev(r10_bio->master_bio,
2098 GFP_NOIO, mddev);
2099 md_trim_bio(bio,
2100 r10_bio->sector - bio->bi_sector,
2101 max_sectors);
2102 r10_bio->devs[slot].bio = bio;
2103 bio->bi_sector = r10_bio->devs[slot].addr
2104 + rdev->data_offset;
2105 bio->bi_bdev = rdev->bdev;
2106 bio->bi_rw = READ | do_sync;
2107 bio->bi_private = r10_bio;
2108 bio->bi_end_io = raid10_end_read_request;
2109 if (max_sectors < r10_bio->sectors) {
2110 /* Drat - have to split this up more */
2111 struct bio *mbio = r10_bio->master_bio;
2112 int sectors_handled =
2113 r10_bio->sector + max_sectors
2114 - mbio->bi_sector;
2115 r10_bio->sectors = max_sectors;
2116 spin_lock_irq(&conf->device_lock);
2117 if (mbio->bi_phys_segments == 0)
2118 mbio->bi_phys_segments = 2;
2119 else
2120 mbio->bi_phys_segments++;
2121 spin_unlock_irq(&conf->device_lock);
2122 generic_make_request(bio);
2123 bio = NULL;
2124
2125 r10_bio = mempool_alloc(conf->r10bio_pool,
2126 GFP_NOIO);
2127 r10_bio->master_bio = mbio;
2128 r10_bio->sectors = (mbio->bi_size >> 9)
2129 - sectors_handled;
2130 r10_bio->state = 0;
2131 set_bit(R10BIO_ReadError,
2132 &r10_bio->state);
2133 r10_bio->mddev = mddev;
2134 r10_bio->sector = mbio->bi_sector
2135 + sectors_handled;
2136
2137 goto read_more;
2138 } else
2139 generic_make_request(bio);
2140}
2141
2142static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio)
2143{
2144 /* Some sort of write request has finished and it
2145 * succeeded in writing where we thought there was a
2146 * bad block. So forget the bad block.
2147 * Or possibly if failed and we need to record
2148 * a bad block.
2149 */
2150 int m;
2151 mdk_rdev_t *rdev;
2152
2153 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2154 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2155 for (m = 0; m < conf->copies; m++) {
2156 int dev = r10_bio->devs[m].devnum;
2157 rdev = conf->mirrors[dev].rdev;
2158 if (r10_bio->devs[m].bio == NULL)
2159 continue;
2160 if (test_bit(BIO_UPTODATE,
2161 &r10_bio->devs[m].bio->bi_flags)) {
2162 rdev_clear_badblocks(
2163 rdev,
2164 r10_bio->devs[m].addr,
2165 r10_bio->sectors);
2166 } else {
2167 if (!rdev_set_badblocks(
2168 rdev,
2169 r10_bio->devs[m].addr,
2170 r10_bio->sectors, 0))
2171 md_error(conf->mddev, rdev);
2172 }
2173 }
2174 put_buf(r10_bio);
2175 } else {
2176 for (m = 0; m < conf->copies; m++) {
2177 int dev = r10_bio->devs[m].devnum;
2178 struct bio *bio = r10_bio->devs[m].bio;
2179 rdev = conf->mirrors[dev].rdev;
2180 if (bio == IO_MADE_GOOD) {
2181 rdev_clear_badblocks(
2182 rdev,
2183 r10_bio->devs[m].addr,
2184 r10_bio->sectors);
2185 rdev_dec_pending(rdev, conf->mddev);
2186 } else if (bio != NULL &&
2187 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2188 if (!narrow_write_error(r10_bio, m)) {
2189 md_error(conf->mddev, rdev);
2190 set_bit(R10BIO_Degraded,
2191 &r10_bio->state);
2192 }
2193 rdev_dec_pending(rdev, conf->mddev);
2194 }
2195 }
2196 if (test_bit(R10BIO_WriteError,
2197 &r10_bio->state))
2198 close_write(r10_bio);
2199 raid_end_bio_io(r10_bio);
2200 }
2201}
2202
2203static void raid10d(mddev_t *mddev)
2204{
2205 r10bio_t *r10_bio;
2206 unsigned long flags;
2207 conf_t *conf = mddev->private;
2208 struct list_head *head = &conf->retry_list;
2209 struct blk_plug plug;
2210
2211 md_check_recovery(mddev);
2212
2213 blk_start_plug(&plug);
2214 for (;;) {
2215
2216 flush_pending_writes(conf);
2217
2218 spin_lock_irqsave(&conf->device_lock, flags);
2219 if (list_empty(head)) {
2220 spin_unlock_irqrestore(&conf->device_lock, flags);
2221 break;
2222 }
2223 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
2224 list_del(head->prev);
2225 conf->nr_queued--;
2226 spin_unlock_irqrestore(&conf->device_lock, flags);
2227
2228 mddev = r10_bio->mddev;
2229 conf = mddev->private;
2230 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2231 test_bit(R10BIO_WriteError, &r10_bio->state))
2232 handle_write_completed(conf, r10_bio);
2233 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2234 sync_request_write(mddev, r10_bio);
2235 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2236 recovery_request_write(mddev, r10_bio);
2237 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2238 handle_read_error(mddev, r10_bio);
2239 else {
2240 /* just a partial read to be scheduled from a
2241 * separate context
2242 */
2243 int slot = r10_bio->read_slot;
2244 generic_make_request(r10_bio->devs[slot].bio);
2245 }
2246
2247 cond_resched();
2248 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2249 md_check_recovery(mddev);
2250 }
2251 blk_finish_plug(&plug);
2252}
2253
2254
2255static int init_resync(conf_t *conf)
2256{
2257 int buffs;
2258
2259 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2260 BUG_ON(conf->r10buf_pool);
2261 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2262 if (!conf->r10buf_pool)
2263 return -ENOMEM;
2264 conf->next_resync = 0;
2265 return 0;
2266}
2267
2268/*
2269 * perform a "sync" on one "block"
2270 *
2271 * We need to make sure that no normal I/O request - particularly write
2272 * requests - conflict with active sync requests.
2273 *
2274 * This is achieved by tracking pending requests and a 'barrier' concept
2275 * that can be installed to exclude normal IO requests.
2276 *
2277 * Resync and recovery are handled very differently.
2278 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2279 *
2280 * For resync, we iterate over virtual addresses, read all copies,
2281 * and update if there are differences. If only one copy is live,
2282 * skip it.
2283 * For recovery, we iterate over physical addresses, read a good
2284 * value for each non-in_sync drive, and over-write.
2285 *
2286 * So, for recovery we may have several outstanding complex requests for a
2287 * given address, one for each out-of-sync device. We model this by allocating
2288 * a number of r10_bio structures, one for each out-of-sync device.
2289 * As we setup these structures, we collect all bio's together into a list
2290 * which we then process collectively to add pages, and then process again
2291 * to pass to generic_make_request.
2292 *
2293 * The r10_bio structures are linked using a borrowed master_bio pointer.
2294 * This link is counted in ->remaining. When the r10_bio that points to NULL
2295 * has its remaining count decremented to 0, the whole complex operation
2296 * is complete.
2297 *
2298 */
2299
2300static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
2301 int *skipped, int go_faster)
2302{
2303 conf_t *conf = mddev->private;
2304 r10bio_t *r10_bio;
2305 struct bio *biolist = NULL, *bio;
2306 sector_t max_sector, nr_sectors;
2307 int i;
2308 int max_sync;
2309 sector_t sync_blocks;
2310 sector_t sectors_skipped = 0;
2311 int chunks_skipped = 0;
2312
2313 if (!conf->r10buf_pool)
2314 if (init_resync(conf))
2315 return 0;
2316
2317 skipped:
2318 max_sector = mddev->dev_sectors;
2319 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2320 max_sector = mddev->resync_max_sectors;
2321 if (sector_nr >= max_sector) {
2322 /* If we aborted, we need to abort the
2323 * sync on the 'current' bitmap chucks (there can
2324 * be several when recovering multiple devices).
2325 * as we may have started syncing it but not finished.
2326 * We can find the current address in
2327 * mddev->curr_resync, but for recovery,
2328 * we need to convert that to several
2329 * virtual addresses.
2330 */
2331 if (mddev->curr_resync < max_sector) { /* aborted */
2332 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2333 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2334 &sync_blocks, 1);
2335 else for (i=0; i<conf->raid_disks; i++) {
2336 sector_t sect =
2337 raid10_find_virt(conf, mddev->curr_resync, i);
2338 bitmap_end_sync(mddev->bitmap, sect,
2339 &sync_blocks, 1);
2340 }
2341 } else /* completed sync */
2342 conf->fullsync = 0;
2343
2344 bitmap_close_sync(mddev->bitmap);
2345 close_sync(conf);
2346 *skipped = 1;
2347 return sectors_skipped;
2348 }
2349 if (chunks_skipped >= conf->raid_disks) {
2350 /* if there has been nothing to do on any drive,
2351 * then there is nothing to do at all..
2352 */
2353 *skipped = 1;
2354 return (max_sector - sector_nr) + sectors_skipped;
2355 }
2356
2357 if (max_sector > mddev->resync_max)
2358 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2359
2360 /* make sure whole request will fit in a chunk - if chunks
2361 * are meaningful
2362 */
2363 if (conf->near_copies < conf->raid_disks &&
2364 max_sector > (sector_nr | conf->chunk_mask))
2365 max_sector = (sector_nr | conf->chunk_mask) + 1;
2366 /*
2367 * If there is non-resync activity waiting for us then
2368 * put in a delay to throttle resync.
2369 */
2370 if (!go_faster && conf->nr_waiting)
2371 msleep_interruptible(1000);
2372
2373 /* Again, very different code for resync and recovery.
2374 * Both must result in an r10bio with a list of bios that
2375 * have bi_end_io, bi_sector, bi_bdev set,
2376 * and bi_private set to the r10bio.
2377 * For recovery, we may actually create several r10bios
2378 * with 2 bios in each, that correspond to the bios in the main one.
2379 * In this case, the subordinate r10bios link back through a
2380 * borrowed master_bio pointer, and the counter in the master
2381 * includes a ref from each subordinate.
2382 */
2383 /* First, we decide what to do and set ->bi_end_io
2384 * To end_sync_read if we want to read, and
2385 * end_sync_write if we will want to write.
2386 */
2387
2388 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2389 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2390 /* recovery... the complicated one */
2391 int j;
2392 r10_bio = NULL;
2393
2394 for (i=0 ; i<conf->raid_disks; i++) {
2395 int still_degraded;
2396 r10bio_t *rb2;
2397 sector_t sect;
2398 int must_sync;
2399 int any_working;
2400
2401 if (conf->mirrors[i].rdev == NULL ||
2402 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
2403 continue;
2404
2405 still_degraded = 0;
2406 /* want to reconstruct this device */
2407 rb2 = r10_bio;
2408 sect = raid10_find_virt(conf, sector_nr, i);
2409 /* Unless we are doing a full sync, we only need
2410 * to recover the block if it is set in the bitmap
2411 */
2412 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2413 &sync_blocks, 1);
2414 if (sync_blocks < max_sync)
2415 max_sync = sync_blocks;
2416 if (!must_sync &&
2417 !conf->fullsync) {
2418 /* yep, skip the sync_blocks here, but don't assume
2419 * that there will never be anything to do here
2420 */
2421 chunks_skipped = -1;
2422 continue;
2423 }
2424
2425 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2426 raise_barrier(conf, rb2 != NULL);
2427 atomic_set(&r10_bio->remaining, 0);
2428
2429 r10_bio->master_bio = (struct bio*)rb2;
2430 if (rb2)
2431 atomic_inc(&rb2->remaining);
2432 r10_bio->mddev = mddev;
2433 set_bit(R10BIO_IsRecover, &r10_bio->state);
2434 r10_bio->sector = sect;
2435
2436 raid10_find_phys(conf, r10_bio);
2437
2438 /* Need to check if the array will still be
2439 * degraded
2440 */
2441 for (j=0; j<conf->raid_disks; j++)
2442 if (conf->mirrors[j].rdev == NULL ||
2443 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2444 still_degraded = 1;
2445 break;
2446 }
2447
2448 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2449 &sync_blocks, still_degraded);
2450
2451 any_working = 0;
2452 for (j=0; j<conf->copies;j++) {
2453 int k;
2454 int d = r10_bio->devs[j].devnum;
2455 sector_t from_addr, to_addr;
2456 mdk_rdev_t *rdev;
2457 sector_t sector, first_bad;
2458 int bad_sectors;
2459 if (!conf->mirrors[d].rdev ||
2460 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2461 continue;
2462 /* This is where we read from */
2463 any_working = 1;
2464 rdev = conf->mirrors[d].rdev;
2465 sector = r10_bio->devs[j].addr;
2466
2467 if (is_badblock(rdev, sector, max_sync,
2468 &first_bad, &bad_sectors)) {
2469 if (first_bad > sector)
2470 max_sync = first_bad - sector;
2471 else {
2472 bad_sectors -= (sector
2473 - first_bad);
2474 if (max_sync > bad_sectors)
2475 max_sync = bad_sectors;
2476 continue;
2477 }
2478 }
2479 bio = r10_bio->devs[0].bio;
2480 bio->bi_next = biolist;
2481 biolist = bio;
2482 bio->bi_private = r10_bio;
2483 bio->bi_end_io = end_sync_read;
2484 bio->bi_rw = READ;
2485 from_addr = r10_bio->devs[j].addr;
2486 bio->bi_sector = from_addr +
2487 conf->mirrors[d].rdev->data_offset;
2488 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2489 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2490 atomic_inc(&r10_bio->remaining);
2491 /* and we write to 'i' */
2492
2493 for (k=0; k<conf->copies; k++)
2494 if (r10_bio->devs[k].devnum == i)
2495 break;
2496 BUG_ON(k == conf->copies);
2497 bio = r10_bio->devs[1].bio;
2498 bio->bi_next = biolist;
2499 biolist = bio;
2500 bio->bi_private = r10_bio;
2501 bio->bi_end_io = end_sync_write;
2502 bio->bi_rw = WRITE;
2503 to_addr = r10_bio->devs[k].addr;
2504 bio->bi_sector = to_addr +
2505 conf->mirrors[i].rdev->data_offset;
2506 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2507
2508 r10_bio->devs[0].devnum = d;
2509 r10_bio->devs[0].addr = from_addr;
2510 r10_bio->devs[1].devnum = i;
2511 r10_bio->devs[1].addr = to_addr;
2512
2513 break;
2514 }
2515 if (j == conf->copies) {
2516 /* Cannot recover, so abort the recovery or
2517 * record a bad block */
2518 put_buf(r10_bio);
2519 if (rb2)
2520 atomic_dec(&rb2->remaining);
2521 r10_bio = rb2;
2522 if (any_working) {
2523 /* problem is that there are bad blocks
2524 * on other device(s)
2525 */
2526 int k;
2527 for (k = 0; k < conf->copies; k++)
2528 if (r10_bio->devs[k].devnum == i)
2529 break;
2530 if (!rdev_set_badblocks(
2531 conf->mirrors[i].rdev,
2532 r10_bio->devs[k].addr,
2533 max_sync, 0))
2534 any_working = 0;
2535 }
2536 if (!any_working) {
2537 if (!test_and_set_bit(MD_RECOVERY_INTR,
2538 &mddev->recovery))
2539 printk(KERN_INFO "md/raid10:%s: insufficient "
2540 "working devices for recovery.\n",
2541 mdname(mddev));
2542 conf->mirrors[i].recovery_disabled
2543 = mddev->recovery_disabled;
2544 }
2545 break;
2546 }
2547 }
2548 if (biolist == NULL) {
2549 while (r10_bio) {
2550 r10bio_t *rb2 = r10_bio;
2551 r10_bio = (r10bio_t*) rb2->master_bio;
2552 rb2->master_bio = NULL;
2553 put_buf(rb2);
2554 }
2555 goto giveup;
2556 }
2557 } else {
2558 /* resync. Schedule a read for every block at this virt offset */
2559 int count = 0;
2560
2561 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2562
2563 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2564 &sync_blocks, mddev->degraded) &&
2565 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2566 &mddev->recovery)) {
2567 /* We can skip this block */
2568 *skipped = 1;
2569 return sync_blocks + sectors_skipped;
2570 }
2571 if (sync_blocks < max_sync)
2572 max_sync = sync_blocks;
2573 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2574
2575 r10_bio->mddev = mddev;
2576 atomic_set(&r10_bio->remaining, 0);
2577 raise_barrier(conf, 0);
2578 conf->next_resync = sector_nr;
2579
2580 r10_bio->master_bio = NULL;
2581 r10_bio->sector = sector_nr;
2582 set_bit(R10BIO_IsSync, &r10_bio->state);
2583 raid10_find_phys(conf, r10_bio);
2584 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2585
2586 for (i=0; i<conf->copies; i++) {
2587 int d = r10_bio->devs[i].devnum;
2588 sector_t first_bad, sector;
2589 int bad_sectors;
2590
2591 bio = r10_bio->devs[i].bio;
2592 bio->bi_end_io = NULL;
2593 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2594 if (conf->mirrors[d].rdev == NULL ||
2595 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2596 continue;
2597 sector = r10_bio->devs[i].addr;
2598 if (is_badblock(conf->mirrors[d].rdev,
2599 sector, max_sync,
2600 &first_bad, &bad_sectors)) {
2601 if (first_bad > sector)
2602 max_sync = first_bad - sector;
2603 else {
2604 bad_sectors -= (sector - first_bad);
2605 if (max_sync > bad_sectors)
2606 max_sync = max_sync;
2607 continue;
2608 }
2609 }
2610 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2611 atomic_inc(&r10_bio->remaining);
2612 bio->bi_next = biolist;
2613 biolist = bio;
2614 bio->bi_private = r10_bio;
2615 bio->bi_end_io = end_sync_read;
2616 bio->bi_rw = READ;
2617 bio->bi_sector = sector +
2618 conf->mirrors[d].rdev->data_offset;
2619 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2620 count++;
2621 }
2622
2623 if (count < 2) {
2624 for (i=0; i<conf->copies; i++) {
2625 int d = r10_bio->devs[i].devnum;
2626 if (r10_bio->devs[i].bio->bi_end_io)
2627 rdev_dec_pending(conf->mirrors[d].rdev,
2628 mddev);
2629 }
2630 put_buf(r10_bio);
2631 biolist = NULL;
2632 goto giveup;
2633 }
2634 }
2635
2636 for (bio = biolist; bio ; bio=bio->bi_next) {
2637
2638 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2639 if (bio->bi_end_io)
2640 bio->bi_flags |= 1 << BIO_UPTODATE;
2641 bio->bi_vcnt = 0;
2642 bio->bi_idx = 0;
2643 bio->bi_phys_segments = 0;
2644 bio->bi_size = 0;
2645 }
2646
2647 nr_sectors = 0;
2648 if (sector_nr + max_sync < max_sector)
2649 max_sector = sector_nr + max_sync;
2650 do {
2651 struct page *page;
2652 int len = PAGE_SIZE;
2653 if (sector_nr + (len>>9) > max_sector)
2654 len = (max_sector - sector_nr) << 9;
2655 if (len == 0)
2656 break;
2657 for (bio= biolist ; bio ; bio=bio->bi_next) {
2658 struct bio *bio2;
2659 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2660 if (bio_add_page(bio, page, len, 0))
2661 continue;
2662
2663 /* stop here */
2664 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2665 for (bio2 = biolist;
2666 bio2 && bio2 != bio;
2667 bio2 = bio2->bi_next) {
2668 /* remove last page from this bio */
2669 bio2->bi_vcnt--;
2670 bio2->bi_size -= len;
2671 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2672 }
2673 goto bio_full;
2674 }
2675 nr_sectors += len>>9;
2676 sector_nr += len>>9;
2677 } while (biolist->bi_vcnt < RESYNC_PAGES);
2678 bio_full:
2679 r10_bio->sectors = nr_sectors;
2680
2681 while (biolist) {
2682 bio = biolist;
2683 biolist = biolist->bi_next;
2684
2685 bio->bi_next = NULL;
2686 r10_bio = bio->bi_private;
2687 r10_bio->sectors = nr_sectors;
2688
2689 if (bio->bi_end_io == end_sync_read) {
2690 md_sync_acct(bio->bi_bdev, nr_sectors);
2691 generic_make_request(bio);
2692 }
2693 }
2694
2695 if (sectors_skipped)
2696 /* pretend they weren't skipped, it makes
2697 * no important difference in this case
2698 */
2699 md_done_sync(mddev, sectors_skipped, 1);
2700
2701 return sectors_skipped + nr_sectors;
2702 giveup:
2703 /* There is nowhere to write, so all non-sync
2704 * drives must be failed or in resync, all drives
2705 * have a bad block, so try the next chunk...
2706 */
2707 if (sector_nr + max_sync < max_sector)
2708 max_sector = sector_nr + max_sync;
2709
2710 sectors_skipped += (max_sector - sector_nr);
2711 chunks_skipped ++;
2712 sector_nr = max_sector;
2713 goto skipped;
2714}
2715
2716static sector_t
2717raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2718{
2719 sector_t size;
2720 conf_t *conf = mddev->private;
2721
2722 if (!raid_disks)
2723 raid_disks = conf->raid_disks;
2724 if (!sectors)
2725 sectors = conf->dev_sectors;
2726
2727 size = sectors >> conf->chunk_shift;
2728 sector_div(size, conf->far_copies);
2729 size = size * raid_disks;
2730 sector_div(size, conf->near_copies);
2731
2732 return size << conf->chunk_shift;
2733}
2734
2735
2736static conf_t *setup_conf(mddev_t *mddev)
2737{
2738 conf_t *conf = NULL;
2739 int nc, fc, fo;
2740 sector_t stride, size;
2741 int err = -EINVAL;
2742
2743 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2744 !is_power_of_2(mddev->new_chunk_sectors)) {
2745 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2746 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2747 mdname(mddev), PAGE_SIZE);
2748 goto out;
2749 }
2750
2751 nc = mddev->new_layout & 255;
2752 fc = (mddev->new_layout >> 8) & 255;
2753 fo = mddev->new_layout & (1<<16);
2754
2755 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2756 (mddev->new_layout >> 17)) {
2757 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2758 mdname(mddev), mddev->new_layout);
2759 goto out;
2760 }
2761
2762 err = -ENOMEM;
2763 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2764 if (!conf)
2765 goto out;
2766
2767 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2768 GFP_KERNEL);
2769 if (!conf->mirrors)
2770 goto out;
2771
2772 conf->tmppage = alloc_page(GFP_KERNEL);
2773 if (!conf->tmppage)
2774 goto out;
2775
2776
2777 conf->raid_disks = mddev->raid_disks;
2778 conf->near_copies = nc;
2779 conf->far_copies = fc;
2780 conf->copies = nc*fc;
2781 conf->far_offset = fo;
2782 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2783 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2784
2785 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2786 r10bio_pool_free, conf);
2787 if (!conf->r10bio_pool)
2788 goto out;
2789
2790 size = mddev->dev_sectors >> conf->chunk_shift;
2791 sector_div(size, fc);
2792 size = size * conf->raid_disks;
2793 sector_div(size, nc);
2794 /* 'size' is now the number of chunks in the array */
2795 /* calculate "used chunks per device" in 'stride' */
2796 stride = size * conf->copies;
2797
2798 /* We need to round up when dividing by raid_disks to
2799 * get the stride size.
2800 */
2801 stride += conf->raid_disks - 1;
2802 sector_div(stride, conf->raid_disks);
2803
2804 conf->dev_sectors = stride << conf->chunk_shift;
2805
2806 if (fo)
2807 stride = 1;
2808 else
2809 sector_div(stride, fc);
2810 conf->stride = stride << conf->chunk_shift;
2811
2812
2813 spin_lock_init(&conf->device_lock);
2814 INIT_LIST_HEAD(&conf->retry_list);
2815
2816 spin_lock_init(&conf->resync_lock);
2817 init_waitqueue_head(&conf->wait_barrier);
2818
2819 conf->thread = md_register_thread(raid10d, mddev, NULL);
2820 if (!conf->thread)
2821 goto out;
2822
2823 conf->mddev = mddev;
2824 return conf;
2825
2826 out:
2827 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2828 mdname(mddev));
2829 if (conf) {
2830 if (conf->r10bio_pool)
2831 mempool_destroy(conf->r10bio_pool);
2832 kfree(conf->mirrors);
2833 safe_put_page(conf->tmppage);
2834 kfree(conf);
2835 }
2836 return ERR_PTR(err);
2837}
2838
2839static int run(mddev_t *mddev)
2840{
2841 conf_t *conf;
2842 int i, disk_idx, chunk_size;
2843 mirror_info_t *disk;
2844 mdk_rdev_t *rdev;
2845 sector_t size;
2846
2847 /*
2848 * copy the already verified devices into our private RAID10
2849 * bookkeeping area. [whatever we allocate in run(),
2850 * should be freed in stop()]
2851 */
2852
2853 if (mddev->private == NULL) {
2854 conf = setup_conf(mddev);
2855 if (IS_ERR(conf))
2856 return PTR_ERR(conf);
2857 mddev->private = conf;
2858 }
2859 conf = mddev->private;
2860 if (!conf)
2861 goto out;
2862
2863 mddev->thread = conf->thread;
2864 conf->thread = NULL;
2865
2866 chunk_size = mddev->chunk_sectors << 9;
2867 blk_queue_io_min(mddev->queue, chunk_size);
2868 if (conf->raid_disks % conf->near_copies)
2869 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2870 else
2871 blk_queue_io_opt(mddev->queue, chunk_size *
2872 (conf->raid_disks / conf->near_copies));
2873
2874 list_for_each_entry(rdev, &mddev->disks, same_set) {
2875
2876 disk_idx = rdev->raid_disk;
2877 if (disk_idx >= conf->raid_disks
2878 || disk_idx < 0)
2879 continue;
2880 disk = conf->mirrors + disk_idx;
2881
2882 disk->rdev = rdev;
2883 disk_stack_limits(mddev->gendisk, rdev->bdev,
2884 rdev->data_offset << 9);
2885 /* as we don't honour merge_bvec_fn, we must never risk
2886 * violating it, so limit max_segments to 1 lying
2887 * within a single page.
2888 */
2889 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2890 blk_queue_max_segments(mddev->queue, 1);
2891 blk_queue_segment_boundary(mddev->queue,
2892 PAGE_CACHE_SIZE - 1);
2893 }
2894
2895 disk->head_position = 0;
2896 }
2897 /* need to check that every block has at least one working mirror */
2898 if (!enough(conf, -1)) {
2899 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2900 mdname(mddev));
2901 goto out_free_conf;
2902 }
2903
2904 mddev->degraded = 0;
2905 for (i = 0; i < conf->raid_disks; i++) {
2906
2907 disk = conf->mirrors + i;
2908
2909 if (!disk->rdev ||
2910 !test_bit(In_sync, &disk->rdev->flags)) {
2911 disk->head_position = 0;
2912 mddev->degraded++;
2913 if (disk->rdev)
2914 conf->fullsync = 1;
2915 }
2916 }
2917
2918 if (mddev->recovery_cp != MaxSector)
2919 printk(KERN_NOTICE "md/raid10:%s: not clean"
2920 " -- starting background reconstruction\n",
2921 mdname(mddev));
2922 printk(KERN_INFO
2923 "md/raid10:%s: active with %d out of %d devices\n",
2924 mdname(mddev), conf->raid_disks - mddev->degraded,
2925 conf->raid_disks);
2926 /*
2927 * Ok, everything is just fine now
2928 */
2929 mddev->dev_sectors = conf->dev_sectors;
2930 size = raid10_size(mddev, 0, 0);
2931 md_set_array_sectors(mddev, size);
2932 mddev->resync_max_sectors = size;
2933
2934 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2935 mddev->queue->backing_dev_info.congested_data = mddev;
2936
2937 /* Calculate max read-ahead size.
2938 * We need to readahead at least twice a whole stripe....
2939 * maybe...
2940 */
2941 {
2942 int stripe = conf->raid_disks *
2943 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2944 stripe /= conf->near_copies;
2945 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2946 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2947 }
2948
2949 if (conf->near_copies < conf->raid_disks)
2950 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2951
2952 if (md_integrity_register(mddev))
2953 goto out_free_conf;
2954
2955 return 0;
2956
2957out_free_conf:
2958 md_unregister_thread(&mddev->thread);
2959 if (conf->r10bio_pool)
2960 mempool_destroy(conf->r10bio_pool);
2961 safe_put_page(conf->tmppage);
2962 kfree(conf->mirrors);
2963 kfree(conf);
2964 mddev->private = NULL;
2965out:
2966 return -EIO;
2967}
2968
2969static int stop(mddev_t *mddev)
2970{
2971 conf_t *conf = mddev->private;
2972
2973 raise_barrier(conf, 0);
2974 lower_barrier(conf);
2975
2976 md_unregister_thread(&mddev->thread);
2977 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2978 if (conf->r10bio_pool)
2979 mempool_destroy(conf->r10bio_pool);
2980 kfree(conf->mirrors);
2981 kfree(conf);
2982 mddev->private = NULL;
2983 return 0;
2984}
2985
2986static void raid10_quiesce(mddev_t *mddev, int state)
2987{
2988 conf_t *conf = mddev->private;
2989
2990 switch(state) {
2991 case 1:
2992 raise_barrier(conf, 0);
2993 break;
2994 case 0:
2995 lower_barrier(conf);
2996 break;
2997 }
2998}
2999
3000static void *raid10_takeover_raid0(mddev_t *mddev)
3001{
3002 mdk_rdev_t *rdev;
3003 conf_t *conf;
3004
3005 if (mddev->degraded > 0) {
3006 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3007 mdname(mddev));
3008 return ERR_PTR(-EINVAL);
3009 }
3010
3011 /* Set new parameters */
3012 mddev->new_level = 10;
3013 /* new layout: far_copies = 1, near_copies = 2 */
3014 mddev->new_layout = (1<<8) + 2;
3015 mddev->new_chunk_sectors = mddev->chunk_sectors;
3016 mddev->delta_disks = mddev->raid_disks;
3017 mddev->raid_disks *= 2;
3018 /* make sure it will be not marked as dirty */
3019 mddev->recovery_cp = MaxSector;
3020
3021 conf = setup_conf(mddev);
3022 if (!IS_ERR(conf)) {
3023 list_for_each_entry(rdev, &mddev->disks, same_set)
3024 if (rdev->raid_disk >= 0)
3025 rdev->new_raid_disk = rdev->raid_disk * 2;
3026 conf->barrier = 1;
3027 }
3028
3029 return conf;
3030}
3031
3032static void *raid10_takeover(mddev_t *mddev)
3033{
3034 struct raid0_private_data *raid0_priv;
3035
3036 /* raid10 can take over:
3037 * raid0 - providing it has only two drives
3038 */
3039 if (mddev->level == 0) {
3040 /* for raid0 takeover only one zone is supported */
3041 raid0_priv = mddev->private;
3042 if (raid0_priv->nr_strip_zones > 1) {
3043 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3044 " with more than one zone.\n",
3045 mdname(mddev));
3046 return ERR_PTR(-EINVAL);
3047 }
3048 return raid10_takeover_raid0(mddev);
3049 }
3050 return ERR_PTR(-EINVAL);
3051}
3052
3053static struct mdk_personality raid10_personality =
3054{
3055 .name = "raid10",
3056 .level = 10,
3057 .owner = THIS_MODULE,
3058 .make_request = make_request,
3059 .run = run,
3060 .stop = stop,
3061 .status = status,
3062 .error_handler = error,
3063 .hot_add_disk = raid10_add_disk,
3064 .hot_remove_disk= raid10_remove_disk,
3065 .spare_active = raid10_spare_active,
3066 .sync_request = sync_request,
3067 .quiesce = raid10_quiesce,
3068 .size = raid10_size,
3069 .takeover = raid10_takeover,
3070};
3071
3072static int __init raid_init(void)
3073{
3074 return register_md_personality(&raid10_personality);
3075}
3076
3077static void raid_exit(void)
3078{
3079 unregister_md_personality(&raid10_personality);
3080}
3081
3082module_init(raid_init);
3083module_exit(raid_exit);
3084MODULE_LICENSE("GPL");
3085MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3086MODULE_ALIAS("md-personality-9"); /* RAID10 */
3087MODULE_ALIAS("md-raid10");
3088MODULE_ALIAS("md-level-10");