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#define raid10_log(md, fmt, args...) \
80 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
81
82#include "raid1-10.c"
83
84#define NULL_CMD
85#define cmd_before(conf, cmd) \
86 do { \
87 write_sequnlock_irq(&(conf)->resync_lock); \
88 cmd; \
89 } while (0)
90#define cmd_after(conf) write_seqlock_irq(&(conf)->resync_lock)
91
92#define wait_event_barrier_cmd(conf, cond, cmd) \
93 wait_event_cmd((conf)->wait_barrier, cond, cmd_before(conf, cmd), \
94 cmd_after(conf))
95
96#define wait_event_barrier(conf, cond) \
97 wait_event_barrier_cmd(conf, cond, NULL_CMD)
98
99/*
100 * for resync bio, r10bio pointer can be retrieved from the per-bio
101 * 'struct resync_pages'.
102 */
103static inline struct r10bio *get_resync_r10bio(struct bio *bio)
104{
105 return get_resync_pages(bio)->raid_bio;
106}
107
108static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
109{
110 struct r10conf *conf = data;
111 int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
112
113 /* allocate a r10bio with room for raid_disks entries in the
114 * bios array */
115 return kzalloc(size, gfp_flags);
116}
117
118#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
119/* amount of memory to reserve for resync requests */
120#define RESYNC_WINDOW (1024*1024)
121/* maximum number of concurrent requests, memory permitting */
122#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
123#define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
124#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
125
126/*
127 * When performing a resync, we need to read and compare, so
128 * we need as many pages are there are copies.
129 * When performing a recovery, we need 2 bios, one for read,
130 * one for write (we recover only one drive per r10buf)
131 *
132 */
133static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
134{
135 struct r10conf *conf = data;
136 struct r10bio *r10_bio;
137 struct bio *bio;
138 int j;
139 int nalloc, nalloc_rp;
140 struct resync_pages *rps;
141
142 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
143 if (!r10_bio)
144 return NULL;
145
146 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
147 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
148 nalloc = conf->copies; /* resync */
149 else
150 nalloc = 2; /* recovery */
151
152 /* allocate once for all bios */
153 if (!conf->have_replacement)
154 nalloc_rp = nalloc;
155 else
156 nalloc_rp = nalloc * 2;
157 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
158 if (!rps)
159 goto out_free_r10bio;
160
161 /*
162 * Allocate bios.
163 */
164 for (j = nalloc ; j-- ; ) {
165 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
166 if (!bio)
167 goto out_free_bio;
168 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
169 r10_bio->devs[j].bio = bio;
170 if (!conf->have_replacement)
171 continue;
172 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
173 if (!bio)
174 goto out_free_bio;
175 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
176 r10_bio->devs[j].repl_bio = bio;
177 }
178 /*
179 * Allocate RESYNC_PAGES data pages and attach them
180 * where needed.
181 */
182 for (j = 0; j < nalloc; j++) {
183 struct bio *rbio = r10_bio->devs[j].repl_bio;
184 struct resync_pages *rp, *rp_repl;
185
186 rp = &rps[j];
187 if (rbio)
188 rp_repl = &rps[nalloc + j];
189
190 bio = r10_bio->devs[j].bio;
191
192 if (!j || test_bit(MD_RECOVERY_SYNC,
193 &conf->mddev->recovery)) {
194 if (resync_alloc_pages(rp, gfp_flags))
195 goto out_free_pages;
196 } else {
197 memcpy(rp, &rps[0], sizeof(*rp));
198 resync_get_all_pages(rp);
199 }
200
201 rp->raid_bio = r10_bio;
202 bio->bi_private = rp;
203 if (rbio) {
204 memcpy(rp_repl, rp, sizeof(*rp));
205 rbio->bi_private = rp_repl;
206 }
207 }
208
209 return r10_bio;
210
211out_free_pages:
212 while (--j >= 0)
213 resync_free_pages(&rps[j]);
214
215 j = 0;
216out_free_bio:
217 for ( ; j < nalloc; j++) {
218 if (r10_bio->devs[j].bio)
219 bio_uninit(r10_bio->devs[j].bio);
220 kfree(r10_bio->devs[j].bio);
221 if (r10_bio->devs[j].repl_bio)
222 bio_uninit(r10_bio->devs[j].repl_bio);
223 kfree(r10_bio->devs[j].repl_bio);
224 }
225 kfree(rps);
226out_free_r10bio:
227 rbio_pool_free(r10_bio, conf);
228 return NULL;
229}
230
231static void r10buf_pool_free(void *__r10_bio, void *data)
232{
233 struct r10conf *conf = data;
234 struct r10bio *r10bio = __r10_bio;
235 int j;
236 struct resync_pages *rp = NULL;
237
238 for (j = conf->copies; j--; ) {
239 struct bio *bio = r10bio->devs[j].bio;
240
241 if (bio) {
242 rp = get_resync_pages(bio);
243 resync_free_pages(rp);
244 bio_uninit(bio);
245 kfree(bio);
246 }
247
248 bio = r10bio->devs[j].repl_bio;
249 if (bio) {
250 bio_uninit(bio);
251 kfree(bio);
252 }
253 }
254
255 /* resync pages array stored in the 1st bio's .bi_private */
256 kfree(rp);
257
258 rbio_pool_free(r10bio, conf);
259}
260
261static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
262{
263 int i;
264
265 for (i = 0; i < conf->geo.raid_disks; i++) {
266 struct bio **bio = & r10_bio->devs[i].bio;
267 if (!BIO_SPECIAL(*bio))
268 bio_put(*bio);
269 *bio = NULL;
270 bio = &r10_bio->devs[i].repl_bio;
271 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
272 bio_put(*bio);
273 *bio = NULL;
274 }
275}
276
277static void free_r10bio(struct r10bio *r10_bio)
278{
279 struct r10conf *conf = r10_bio->mddev->private;
280
281 put_all_bios(conf, r10_bio);
282 mempool_free(r10_bio, &conf->r10bio_pool);
283}
284
285static void put_buf(struct r10bio *r10_bio)
286{
287 struct r10conf *conf = r10_bio->mddev->private;
288
289 mempool_free(r10_bio, &conf->r10buf_pool);
290
291 lower_barrier(conf);
292}
293
294static void wake_up_barrier(struct r10conf *conf)
295{
296 if (wq_has_sleeper(&conf->wait_barrier))
297 wake_up(&conf->wait_barrier);
298}
299
300static void reschedule_retry(struct r10bio *r10_bio)
301{
302 unsigned long flags;
303 struct mddev *mddev = r10_bio->mddev;
304 struct r10conf *conf = mddev->private;
305
306 spin_lock_irqsave(&conf->device_lock, flags);
307 list_add(&r10_bio->retry_list, &conf->retry_list);
308 conf->nr_queued ++;
309 spin_unlock_irqrestore(&conf->device_lock, flags);
310
311 /* wake up frozen array... */
312 wake_up(&conf->wait_barrier);
313
314 md_wakeup_thread(mddev->thread);
315}
316
317/*
318 * raid_end_bio_io() is called when we have finished servicing a mirrored
319 * operation and are ready to return a success/failure code to the buffer
320 * cache layer.
321 */
322static void raid_end_bio_io(struct r10bio *r10_bio)
323{
324 struct bio *bio = r10_bio->master_bio;
325 struct r10conf *conf = r10_bio->mddev->private;
326
327 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
328 bio->bi_status = BLK_STS_IOERR;
329
330 bio_endio(bio);
331 /*
332 * Wake up any possible resync thread that waits for the device
333 * to go idle.
334 */
335 allow_barrier(conf);
336
337 free_r10bio(r10_bio);
338}
339
340/*
341 * Update disk head position estimator based on IRQ completion info.
342 */
343static inline void update_head_pos(int slot, struct r10bio *r10_bio)
344{
345 struct r10conf *conf = r10_bio->mddev->private;
346
347 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
348 r10_bio->devs[slot].addr + (r10_bio->sectors);
349}
350
351/*
352 * Find the disk number which triggered given bio
353 */
354static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
355 struct bio *bio, int *slotp, int *replp)
356{
357 int slot;
358 int repl = 0;
359
360 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
361 if (r10_bio->devs[slot].bio == bio)
362 break;
363 if (r10_bio->devs[slot].repl_bio == bio) {
364 repl = 1;
365 break;
366 }
367 }
368
369 update_head_pos(slot, r10_bio);
370
371 if (slotp)
372 *slotp = slot;
373 if (replp)
374 *replp = repl;
375 return r10_bio->devs[slot].devnum;
376}
377
378static void raid10_end_read_request(struct bio *bio)
379{
380 int uptodate = !bio->bi_status;
381 struct r10bio *r10_bio = bio->bi_private;
382 int slot;
383 struct md_rdev *rdev;
384 struct r10conf *conf = r10_bio->mddev->private;
385
386 slot = r10_bio->read_slot;
387 rdev = r10_bio->devs[slot].rdev;
388 /*
389 * this branch is our 'one mirror IO has finished' event handler:
390 */
391 update_head_pos(slot, r10_bio);
392
393 if (uptodate) {
394 /*
395 * Set R10BIO_Uptodate in our master bio, so that
396 * we will return a good error code to the higher
397 * levels even if IO on some other mirrored buffer fails.
398 *
399 * The 'master' represents the composite IO operation to
400 * user-side. So if something waits for IO, then it will
401 * wait for the 'master' bio.
402 */
403 set_bit(R10BIO_Uptodate, &r10_bio->state);
404 } else {
405 /* If all other devices that store this block have
406 * failed, we want to return the error upwards rather
407 * than fail the last device. Here we redefine
408 * "uptodate" to mean "Don't want to retry"
409 */
410 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
411 rdev->raid_disk))
412 uptodate = 1;
413 }
414 if (uptodate) {
415 raid_end_bio_io(r10_bio);
416 rdev_dec_pending(rdev, conf->mddev);
417 } else {
418 /*
419 * oops, read error - keep the refcount on the rdev
420 */
421 pr_err_ratelimited("md/raid10:%s: %pg: rescheduling sector %llu\n",
422 mdname(conf->mddev),
423 rdev->bdev,
424 (unsigned long long)r10_bio->sector);
425 set_bit(R10BIO_ReadError, &r10_bio->state);
426 reschedule_retry(r10_bio);
427 }
428}
429
430static void close_write(struct r10bio *r10_bio)
431{
432 /* clear the bitmap if all writes complete successfully */
433 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
434 r10_bio->sectors,
435 !test_bit(R10BIO_Degraded, &r10_bio->state),
436 0);
437 md_write_end(r10_bio->mddev);
438}
439
440static void one_write_done(struct r10bio *r10_bio)
441{
442 if (atomic_dec_and_test(&r10_bio->remaining)) {
443 if (test_bit(R10BIO_WriteError, &r10_bio->state))
444 reschedule_retry(r10_bio);
445 else {
446 close_write(r10_bio);
447 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
448 reschedule_retry(r10_bio);
449 else
450 raid_end_bio_io(r10_bio);
451 }
452 }
453}
454
455static void raid10_end_write_request(struct bio *bio)
456{
457 struct r10bio *r10_bio = bio->bi_private;
458 int dev;
459 int dec_rdev = 1;
460 struct r10conf *conf = r10_bio->mddev->private;
461 int slot, repl;
462 struct md_rdev *rdev = NULL;
463 struct bio *to_put = NULL;
464 bool discard_error;
465
466 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
467
468 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
469
470 if (repl)
471 rdev = conf->mirrors[dev].replacement;
472 if (!rdev) {
473 smp_rmb();
474 repl = 0;
475 rdev = conf->mirrors[dev].rdev;
476 }
477 /*
478 * this branch is our 'one mirror IO has finished' event handler:
479 */
480 if (bio->bi_status && !discard_error) {
481 if (repl)
482 /* Never record new bad blocks to replacement,
483 * just fail it.
484 */
485 md_error(rdev->mddev, rdev);
486 else {
487 set_bit(WriteErrorSeen, &rdev->flags);
488 if (!test_and_set_bit(WantReplacement, &rdev->flags))
489 set_bit(MD_RECOVERY_NEEDED,
490 &rdev->mddev->recovery);
491
492 dec_rdev = 0;
493 if (test_bit(FailFast, &rdev->flags) &&
494 (bio->bi_opf & MD_FAILFAST)) {
495 md_error(rdev->mddev, rdev);
496 }
497
498 /*
499 * When the device is faulty, it is not necessary to
500 * handle write error.
501 */
502 if (!test_bit(Faulty, &rdev->flags))
503 set_bit(R10BIO_WriteError, &r10_bio->state);
504 else {
505 /* Fail the request */
506 set_bit(R10BIO_Degraded, &r10_bio->state);
507 r10_bio->devs[slot].bio = NULL;
508 to_put = bio;
509 dec_rdev = 1;
510 }
511 }
512 } else {
513 /*
514 * Set R10BIO_Uptodate in our master bio, so that
515 * we will return a good error code for to the higher
516 * levels even if IO on some other mirrored buffer fails.
517 *
518 * The 'master' represents the composite IO operation to
519 * user-side. So if something waits for IO, then it will
520 * wait for the 'master' bio.
521 */
522 sector_t first_bad;
523 int bad_sectors;
524
525 /*
526 * Do not set R10BIO_Uptodate if the current device is
527 * rebuilding or Faulty. This is because we cannot use
528 * such device for properly reading the data back (we could
529 * potentially use it, if the current write would have felt
530 * before rdev->recovery_offset, but for simplicity we don't
531 * check this here.
532 */
533 if (test_bit(In_sync, &rdev->flags) &&
534 !test_bit(Faulty, &rdev->flags))
535 set_bit(R10BIO_Uptodate, &r10_bio->state);
536
537 /* Maybe we can clear some bad blocks. */
538 if (is_badblock(rdev,
539 r10_bio->devs[slot].addr,
540 r10_bio->sectors,
541 &first_bad, &bad_sectors) && !discard_error) {
542 bio_put(bio);
543 if (repl)
544 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
545 else
546 r10_bio->devs[slot].bio = IO_MADE_GOOD;
547 dec_rdev = 0;
548 set_bit(R10BIO_MadeGood, &r10_bio->state);
549 }
550 }
551
552 /*
553 *
554 * Let's see if all mirrored write operations have finished
555 * already.
556 */
557 one_write_done(r10_bio);
558 if (dec_rdev)
559 rdev_dec_pending(rdev, conf->mddev);
560 if (to_put)
561 bio_put(to_put);
562}
563
564/*
565 * RAID10 layout manager
566 * As well as the chunksize and raid_disks count, there are two
567 * parameters: near_copies and far_copies.
568 * near_copies * far_copies must be <= raid_disks.
569 * Normally one of these will be 1.
570 * If both are 1, we get raid0.
571 * If near_copies == raid_disks, we get raid1.
572 *
573 * Chunks are laid out in raid0 style with near_copies copies of the
574 * first chunk, followed by near_copies copies of the next chunk and
575 * so on.
576 * If far_copies > 1, then after 1/far_copies of the array has been assigned
577 * as described above, we start again with a device offset of near_copies.
578 * So we effectively have another copy of the whole array further down all
579 * the drives, but with blocks on different drives.
580 * With this layout, and block is never stored twice on the one device.
581 *
582 * raid10_find_phys finds the sector offset of a given virtual sector
583 * on each device that it is on.
584 *
585 * raid10_find_virt does the reverse mapping, from a device and a
586 * sector offset to a virtual address
587 */
588
589static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
590{
591 int n,f;
592 sector_t sector;
593 sector_t chunk;
594 sector_t stripe;
595 int dev;
596 int slot = 0;
597 int last_far_set_start, last_far_set_size;
598
599 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
600 last_far_set_start *= geo->far_set_size;
601
602 last_far_set_size = geo->far_set_size;
603 last_far_set_size += (geo->raid_disks % geo->far_set_size);
604
605 /* now calculate first sector/dev */
606 chunk = r10bio->sector >> geo->chunk_shift;
607 sector = r10bio->sector & geo->chunk_mask;
608
609 chunk *= geo->near_copies;
610 stripe = chunk;
611 dev = sector_div(stripe, geo->raid_disks);
612 if (geo->far_offset)
613 stripe *= geo->far_copies;
614
615 sector += stripe << geo->chunk_shift;
616
617 /* and calculate all the others */
618 for (n = 0; n < geo->near_copies; n++) {
619 int d = dev;
620 int set;
621 sector_t s = sector;
622 r10bio->devs[slot].devnum = d;
623 r10bio->devs[slot].addr = s;
624 slot++;
625
626 for (f = 1; f < geo->far_copies; f++) {
627 set = d / geo->far_set_size;
628 d += geo->near_copies;
629
630 if ((geo->raid_disks % geo->far_set_size) &&
631 (d > last_far_set_start)) {
632 d -= last_far_set_start;
633 d %= last_far_set_size;
634 d += last_far_set_start;
635 } else {
636 d %= geo->far_set_size;
637 d += geo->far_set_size * set;
638 }
639 s += geo->stride;
640 r10bio->devs[slot].devnum = d;
641 r10bio->devs[slot].addr = s;
642 slot++;
643 }
644 dev++;
645 if (dev >= geo->raid_disks) {
646 dev = 0;
647 sector += (geo->chunk_mask + 1);
648 }
649 }
650}
651
652static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
653{
654 struct geom *geo = &conf->geo;
655
656 if (conf->reshape_progress != MaxSector &&
657 ((r10bio->sector >= conf->reshape_progress) !=
658 conf->mddev->reshape_backwards)) {
659 set_bit(R10BIO_Previous, &r10bio->state);
660 geo = &conf->prev;
661 } else
662 clear_bit(R10BIO_Previous, &r10bio->state);
663
664 __raid10_find_phys(geo, r10bio);
665}
666
667static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
668{
669 sector_t offset, chunk, vchunk;
670 /* Never use conf->prev as this is only called during resync
671 * or recovery, so reshape isn't happening
672 */
673 struct geom *geo = &conf->geo;
674 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
675 int far_set_size = geo->far_set_size;
676 int last_far_set_start;
677
678 if (geo->raid_disks % geo->far_set_size) {
679 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
680 last_far_set_start *= geo->far_set_size;
681
682 if (dev >= last_far_set_start) {
683 far_set_size = geo->far_set_size;
684 far_set_size += (geo->raid_disks % geo->far_set_size);
685 far_set_start = last_far_set_start;
686 }
687 }
688
689 offset = sector & geo->chunk_mask;
690 if (geo->far_offset) {
691 int fc;
692 chunk = sector >> geo->chunk_shift;
693 fc = sector_div(chunk, geo->far_copies);
694 dev -= fc * geo->near_copies;
695 if (dev < far_set_start)
696 dev += far_set_size;
697 } else {
698 while (sector >= geo->stride) {
699 sector -= geo->stride;
700 if (dev < (geo->near_copies + far_set_start))
701 dev += far_set_size - geo->near_copies;
702 else
703 dev -= geo->near_copies;
704 }
705 chunk = sector >> geo->chunk_shift;
706 }
707 vchunk = chunk * geo->raid_disks + dev;
708 sector_div(vchunk, geo->near_copies);
709 return (vchunk << geo->chunk_shift) + offset;
710}
711
712/*
713 * This routine returns the disk from which the requested read should
714 * be done. There is a per-array 'next expected sequential IO' sector
715 * number - if this matches on the next IO then we use the last disk.
716 * There is also a per-disk 'last know head position' sector that is
717 * maintained from IRQ contexts, both the normal and the resync IO
718 * completion handlers update this position correctly. If there is no
719 * perfect sequential match then we pick the disk whose head is closest.
720 *
721 * If there are 2 mirrors in the same 2 devices, performance degrades
722 * because position is mirror, not device based.
723 *
724 * The rdev for the device selected will have nr_pending incremented.
725 */
726
727/*
728 * FIXME: possibly should rethink readbalancing and do it differently
729 * depending on near_copies / far_copies geometry.
730 */
731static struct md_rdev *read_balance(struct r10conf *conf,
732 struct r10bio *r10_bio,
733 int *max_sectors)
734{
735 const sector_t this_sector = r10_bio->sector;
736 int disk, slot;
737 int sectors = r10_bio->sectors;
738 int best_good_sectors;
739 sector_t new_distance, best_dist;
740 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
741 int do_balance;
742 int best_dist_slot, best_pending_slot;
743 bool has_nonrot_disk = false;
744 unsigned int min_pending;
745 struct geom *geo = &conf->geo;
746
747 raid10_find_phys(conf, r10_bio);
748 best_dist_slot = -1;
749 min_pending = UINT_MAX;
750 best_dist_rdev = NULL;
751 best_pending_rdev = NULL;
752 best_dist = MaxSector;
753 best_good_sectors = 0;
754 do_balance = 1;
755 clear_bit(R10BIO_FailFast, &r10_bio->state);
756 /*
757 * Check if we can balance. We can balance on the whole
758 * device if no resync is going on (recovery is ok), or below
759 * the resync window. We take the first readable disk when
760 * above the resync window.
761 */
762 if ((conf->mddev->recovery_cp < MaxSector
763 && (this_sector + sectors >= conf->next_resync)) ||
764 (mddev_is_clustered(conf->mddev) &&
765 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
766 this_sector + sectors)))
767 do_balance = 0;
768
769 for (slot = 0; slot < conf->copies ; slot++) {
770 sector_t first_bad;
771 int bad_sectors;
772 sector_t dev_sector;
773 unsigned int pending;
774 bool nonrot;
775
776 if (r10_bio->devs[slot].bio == IO_BLOCKED)
777 continue;
778 disk = r10_bio->devs[slot].devnum;
779 rdev = conf->mirrors[disk].replacement;
780 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
781 r10_bio->devs[slot].addr + sectors >
782 rdev->recovery_offset)
783 rdev = conf->mirrors[disk].rdev;
784 if (rdev == NULL ||
785 test_bit(Faulty, &rdev->flags))
786 continue;
787 if (!test_bit(In_sync, &rdev->flags) &&
788 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
789 continue;
790
791 dev_sector = r10_bio->devs[slot].addr;
792 if (is_badblock(rdev, dev_sector, sectors,
793 &first_bad, &bad_sectors)) {
794 if (best_dist < MaxSector)
795 /* Already have a better slot */
796 continue;
797 if (first_bad <= dev_sector) {
798 /* Cannot read here. If this is the
799 * 'primary' device, then we must not read
800 * beyond 'bad_sectors' from another device.
801 */
802 bad_sectors -= (dev_sector - first_bad);
803 if (!do_balance && sectors > bad_sectors)
804 sectors = bad_sectors;
805 if (best_good_sectors > sectors)
806 best_good_sectors = sectors;
807 } else {
808 sector_t good_sectors =
809 first_bad - dev_sector;
810 if (good_sectors > best_good_sectors) {
811 best_good_sectors = good_sectors;
812 best_dist_slot = slot;
813 best_dist_rdev = rdev;
814 }
815 if (!do_balance)
816 /* Must read from here */
817 break;
818 }
819 continue;
820 } else
821 best_good_sectors = sectors;
822
823 if (!do_balance)
824 break;
825
826 nonrot = bdev_nonrot(rdev->bdev);
827 has_nonrot_disk |= nonrot;
828 pending = atomic_read(&rdev->nr_pending);
829 if (min_pending > pending && nonrot) {
830 min_pending = pending;
831 best_pending_slot = slot;
832 best_pending_rdev = rdev;
833 }
834
835 if (best_dist_slot >= 0)
836 /* At least 2 disks to choose from so failfast is OK */
837 set_bit(R10BIO_FailFast, &r10_bio->state);
838 /* This optimisation is debatable, and completely destroys
839 * sequential read speed for 'far copies' arrays. So only
840 * keep it for 'near' arrays, and review those later.
841 */
842 if (geo->near_copies > 1 && !pending)
843 new_distance = 0;
844
845 /* for far > 1 always use the lowest address */
846 else if (geo->far_copies > 1)
847 new_distance = r10_bio->devs[slot].addr;
848 else
849 new_distance = abs(r10_bio->devs[slot].addr -
850 conf->mirrors[disk].head_position);
851
852 if (new_distance < best_dist) {
853 best_dist = new_distance;
854 best_dist_slot = slot;
855 best_dist_rdev = rdev;
856 }
857 }
858 if (slot >= conf->copies) {
859 if (has_nonrot_disk) {
860 slot = best_pending_slot;
861 rdev = best_pending_rdev;
862 } else {
863 slot = best_dist_slot;
864 rdev = best_dist_rdev;
865 }
866 }
867
868 if (slot >= 0) {
869 atomic_inc(&rdev->nr_pending);
870 r10_bio->read_slot = slot;
871 } else
872 rdev = NULL;
873 *max_sectors = best_good_sectors;
874
875 return rdev;
876}
877
878static void flush_pending_writes(struct r10conf *conf)
879{
880 /* Any writes that have been queued but are awaiting
881 * bitmap updates get flushed here.
882 */
883 spin_lock_irq(&conf->device_lock);
884
885 if (conf->pending_bio_list.head) {
886 struct blk_plug plug;
887 struct bio *bio;
888
889 bio = bio_list_get(&conf->pending_bio_list);
890 spin_unlock_irq(&conf->device_lock);
891
892 /*
893 * As this is called in a wait_event() loop (see freeze_array),
894 * current->state might be TASK_UNINTERRUPTIBLE which will
895 * cause a warning when we prepare to wait again. As it is
896 * rare that this path is taken, it is perfectly safe to force
897 * us to go around the wait_event() loop again, so the warning
898 * is a false-positive. Silence the warning by resetting
899 * thread state
900 */
901 __set_current_state(TASK_RUNNING);
902
903 blk_start_plug(&plug);
904 raid1_prepare_flush_writes(conf->mddev->bitmap);
905 wake_up(&conf->wait_barrier);
906
907 while (bio) { /* submit pending writes */
908 struct bio *next = bio->bi_next;
909
910 raid1_submit_write(bio);
911 bio = next;
912 cond_resched();
913 }
914 blk_finish_plug(&plug);
915 } else
916 spin_unlock_irq(&conf->device_lock);
917}
918
919/* Barriers....
920 * Sometimes we need to suspend IO while we do something else,
921 * either some resync/recovery, or reconfigure the array.
922 * To do this we raise a 'barrier'.
923 * The 'barrier' is a counter that can be raised multiple times
924 * to count how many activities are happening which preclude
925 * normal IO.
926 * We can only raise the barrier if there is no pending IO.
927 * i.e. if nr_pending == 0.
928 * We choose only to raise the barrier if no-one is waiting for the
929 * barrier to go down. This means that as soon as an IO request
930 * is ready, no other operations which require a barrier will start
931 * until the IO request has had a chance.
932 *
933 * So: regular IO calls 'wait_barrier'. When that returns there
934 * is no backgroup IO happening, It must arrange to call
935 * allow_barrier when it has finished its IO.
936 * backgroup IO calls must call raise_barrier. Once that returns
937 * there is no normal IO happeing. It must arrange to call
938 * lower_barrier when the particular background IO completes.
939 */
940
941static void raise_barrier(struct r10conf *conf, int force)
942{
943 write_seqlock_irq(&conf->resync_lock);
944
945 if (WARN_ON_ONCE(force && !conf->barrier))
946 force = false;
947
948 /* Wait until no block IO is waiting (unless 'force') */
949 wait_event_barrier(conf, force || !conf->nr_waiting);
950
951 /* block any new IO from starting */
952 WRITE_ONCE(conf->barrier, conf->barrier + 1);
953
954 /* Now wait for all pending IO to complete */
955 wait_event_barrier(conf, !atomic_read(&conf->nr_pending) &&
956 conf->barrier < RESYNC_DEPTH);
957
958 write_sequnlock_irq(&conf->resync_lock);
959}
960
961static void lower_barrier(struct r10conf *conf)
962{
963 unsigned long flags;
964
965 write_seqlock_irqsave(&conf->resync_lock, flags);
966 WRITE_ONCE(conf->barrier, conf->barrier - 1);
967 write_sequnlock_irqrestore(&conf->resync_lock, flags);
968 wake_up(&conf->wait_barrier);
969}
970
971static bool stop_waiting_barrier(struct r10conf *conf)
972{
973 struct bio_list *bio_list = current->bio_list;
974 struct md_thread *thread;
975
976 /* barrier is dropped */
977 if (!conf->barrier)
978 return true;
979
980 /*
981 * If there are already pending requests (preventing the barrier from
982 * rising completely), and the pre-process bio queue isn't empty, then
983 * don't wait, as we need to empty that queue to get the nr_pending
984 * count down.
985 */
986 if (atomic_read(&conf->nr_pending) && bio_list &&
987 (!bio_list_empty(&bio_list[0]) || !bio_list_empty(&bio_list[1])))
988 return true;
989
990 /* daemon thread must exist while handling io */
991 thread = rcu_dereference_protected(conf->mddev->thread, true);
992 /*
993 * move on if io is issued from raid10d(), nr_pending is not released
994 * from original io(see handle_read_error()). All raise barrier is
995 * blocked until this io is done.
996 */
997 if (thread->tsk == current) {
998 WARN_ON_ONCE(atomic_read(&conf->nr_pending) == 0);
999 return true;
1000 }
1001
1002 return false;
1003}
1004
1005static bool wait_barrier_nolock(struct r10conf *conf)
1006{
1007 unsigned int seq = read_seqbegin(&conf->resync_lock);
1008
1009 if (READ_ONCE(conf->barrier))
1010 return false;
1011
1012 atomic_inc(&conf->nr_pending);
1013 if (!read_seqretry(&conf->resync_lock, seq))
1014 return true;
1015
1016 if (atomic_dec_and_test(&conf->nr_pending))
1017 wake_up_barrier(conf);
1018
1019 return false;
1020}
1021
1022static bool wait_barrier(struct r10conf *conf, bool nowait)
1023{
1024 bool ret = true;
1025
1026 if (wait_barrier_nolock(conf))
1027 return true;
1028
1029 write_seqlock_irq(&conf->resync_lock);
1030 if (conf->barrier) {
1031 /* Return false when nowait flag is set */
1032 if (nowait) {
1033 ret = false;
1034 } else {
1035 conf->nr_waiting++;
1036 raid10_log(conf->mddev, "wait barrier");
1037 wait_event_barrier(conf, stop_waiting_barrier(conf));
1038 conf->nr_waiting--;
1039 }
1040 if (!conf->nr_waiting)
1041 wake_up(&conf->wait_barrier);
1042 }
1043 /* Only increment nr_pending when we wait */
1044 if (ret)
1045 atomic_inc(&conf->nr_pending);
1046 write_sequnlock_irq(&conf->resync_lock);
1047 return ret;
1048}
1049
1050static void allow_barrier(struct r10conf *conf)
1051{
1052 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1053 (conf->array_freeze_pending))
1054 wake_up_barrier(conf);
1055}
1056
1057static void freeze_array(struct r10conf *conf, int extra)
1058{
1059 /* stop syncio and normal IO and wait for everything to
1060 * go quiet.
1061 * We increment barrier and nr_waiting, and then
1062 * wait until nr_pending match nr_queued+extra
1063 * This is called in the context of one normal IO request
1064 * that has failed. Thus any sync request that might be pending
1065 * will be blocked by nr_pending, and we need to wait for
1066 * pending IO requests to complete or be queued for re-try.
1067 * Thus the number queued (nr_queued) plus this request (extra)
1068 * must match the number of pending IOs (nr_pending) before
1069 * we continue.
1070 */
1071 write_seqlock_irq(&conf->resync_lock);
1072 conf->array_freeze_pending++;
1073 WRITE_ONCE(conf->barrier, conf->barrier + 1);
1074 conf->nr_waiting++;
1075 wait_event_barrier_cmd(conf, atomic_read(&conf->nr_pending) ==
1076 conf->nr_queued + extra, flush_pending_writes(conf));
1077 conf->array_freeze_pending--;
1078 write_sequnlock_irq(&conf->resync_lock);
1079}
1080
1081static void unfreeze_array(struct r10conf *conf)
1082{
1083 /* reverse the effect of the freeze */
1084 write_seqlock_irq(&conf->resync_lock);
1085 WRITE_ONCE(conf->barrier, conf->barrier - 1);
1086 conf->nr_waiting--;
1087 wake_up(&conf->wait_barrier);
1088 write_sequnlock_irq(&conf->resync_lock);
1089}
1090
1091static sector_t choose_data_offset(struct r10bio *r10_bio,
1092 struct md_rdev *rdev)
1093{
1094 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1095 test_bit(R10BIO_Previous, &r10_bio->state))
1096 return rdev->data_offset;
1097 else
1098 return rdev->new_data_offset;
1099}
1100
1101static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1102{
1103 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1104 struct mddev *mddev = plug->cb.data;
1105 struct r10conf *conf = mddev->private;
1106 struct bio *bio;
1107
1108 if (from_schedule) {
1109 spin_lock_irq(&conf->device_lock);
1110 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1111 spin_unlock_irq(&conf->device_lock);
1112 wake_up_barrier(conf);
1113 md_wakeup_thread(mddev->thread);
1114 kfree(plug);
1115 return;
1116 }
1117
1118 /* we aren't scheduling, so we can do the write-out directly. */
1119 bio = bio_list_get(&plug->pending);
1120 raid1_prepare_flush_writes(mddev->bitmap);
1121 wake_up_barrier(conf);
1122
1123 while (bio) { /* submit pending writes */
1124 struct bio *next = bio->bi_next;
1125
1126 raid1_submit_write(bio);
1127 bio = next;
1128 cond_resched();
1129 }
1130 kfree(plug);
1131}
1132
1133/*
1134 * 1. Register the new request and wait if the reconstruction thread has put
1135 * up a bar for new requests. Continue immediately if no resync is active
1136 * currently.
1137 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1138 */
1139static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1140 struct bio *bio, sector_t sectors)
1141{
1142 /* Bail out if REQ_NOWAIT is set for the bio */
1143 if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1144 bio_wouldblock_error(bio);
1145 return false;
1146 }
1147 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1148 bio->bi_iter.bi_sector < conf->reshape_progress &&
1149 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1150 allow_barrier(conf);
1151 if (bio->bi_opf & REQ_NOWAIT) {
1152 bio_wouldblock_error(bio);
1153 return false;
1154 }
1155 raid10_log(conf->mddev, "wait reshape");
1156 wait_event(conf->wait_barrier,
1157 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1158 conf->reshape_progress >= bio->bi_iter.bi_sector +
1159 sectors);
1160 wait_barrier(conf, false);
1161 }
1162 return true;
1163}
1164
1165static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1166 struct r10bio *r10_bio, bool io_accounting)
1167{
1168 struct r10conf *conf = mddev->private;
1169 struct bio *read_bio;
1170 const enum req_op op = bio_op(bio);
1171 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1172 int max_sectors;
1173 struct md_rdev *rdev;
1174 char b[BDEVNAME_SIZE];
1175 int slot = r10_bio->read_slot;
1176 struct md_rdev *err_rdev = NULL;
1177 gfp_t gfp = GFP_NOIO;
1178
1179 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1180 /*
1181 * This is an error retry, but we cannot
1182 * safely dereference the rdev in the r10_bio,
1183 * we must use the one in conf.
1184 * If it has already been disconnected (unlikely)
1185 * we lose the device name in error messages.
1186 */
1187 int disk;
1188 /*
1189 * As we are blocking raid10, it is a little safer to
1190 * use __GFP_HIGH.
1191 */
1192 gfp = GFP_NOIO | __GFP_HIGH;
1193
1194 disk = r10_bio->devs[slot].devnum;
1195 err_rdev = conf->mirrors[disk].rdev;
1196 if (err_rdev)
1197 snprintf(b, sizeof(b), "%pg", err_rdev->bdev);
1198 else {
1199 strcpy(b, "???");
1200 /* This never gets dereferenced */
1201 err_rdev = r10_bio->devs[slot].rdev;
1202 }
1203 }
1204
1205 if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1206 return;
1207 rdev = read_balance(conf, r10_bio, &max_sectors);
1208 if (!rdev) {
1209 if (err_rdev) {
1210 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1211 mdname(mddev), b,
1212 (unsigned long long)r10_bio->sector);
1213 }
1214 raid_end_bio_io(r10_bio);
1215 return;
1216 }
1217 if (err_rdev)
1218 pr_err_ratelimited("md/raid10:%s: %pg: redirecting sector %llu to another mirror\n",
1219 mdname(mddev),
1220 rdev->bdev,
1221 (unsigned long long)r10_bio->sector);
1222 if (max_sectors < bio_sectors(bio)) {
1223 struct bio *split = bio_split(bio, max_sectors,
1224 gfp, &conf->bio_split);
1225 bio_chain(split, bio);
1226 allow_barrier(conf);
1227 submit_bio_noacct(bio);
1228 wait_barrier(conf, false);
1229 bio = split;
1230 r10_bio->master_bio = bio;
1231 r10_bio->sectors = max_sectors;
1232 }
1233 slot = r10_bio->read_slot;
1234
1235 if (io_accounting) {
1236 md_account_bio(mddev, &bio);
1237 r10_bio->master_bio = bio;
1238 }
1239 read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1240
1241 r10_bio->devs[slot].bio = read_bio;
1242 r10_bio->devs[slot].rdev = rdev;
1243
1244 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1245 choose_data_offset(r10_bio, rdev);
1246 read_bio->bi_end_io = raid10_end_read_request;
1247 read_bio->bi_opf = op | do_sync;
1248 if (test_bit(FailFast, &rdev->flags) &&
1249 test_bit(R10BIO_FailFast, &r10_bio->state))
1250 read_bio->bi_opf |= MD_FAILFAST;
1251 read_bio->bi_private = r10_bio;
1252
1253 if (mddev->gendisk)
1254 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1255 r10_bio->sector);
1256 submit_bio_noacct(read_bio);
1257 return;
1258}
1259
1260static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1261 struct bio *bio, bool replacement,
1262 int n_copy)
1263{
1264 const enum req_op op = bio_op(bio);
1265 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1266 const blk_opf_t do_fua = bio->bi_opf & REQ_FUA;
1267 unsigned long flags;
1268 struct r10conf *conf = mddev->private;
1269 struct md_rdev *rdev;
1270 int devnum = r10_bio->devs[n_copy].devnum;
1271 struct bio *mbio;
1272
1273 rdev = replacement ? conf->mirrors[devnum].replacement :
1274 conf->mirrors[devnum].rdev;
1275
1276 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1277 if (replacement)
1278 r10_bio->devs[n_copy].repl_bio = mbio;
1279 else
1280 r10_bio->devs[n_copy].bio = mbio;
1281
1282 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1283 choose_data_offset(r10_bio, rdev));
1284 mbio->bi_end_io = raid10_end_write_request;
1285 mbio->bi_opf = op | do_sync | do_fua;
1286 if (!replacement && test_bit(FailFast,
1287 &conf->mirrors[devnum].rdev->flags)
1288 && enough(conf, devnum))
1289 mbio->bi_opf |= MD_FAILFAST;
1290 mbio->bi_private = r10_bio;
1291
1292 if (conf->mddev->gendisk)
1293 trace_block_bio_remap(mbio, disk_devt(conf->mddev->gendisk),
1294 r10_bio->sector);
1295 /* flush_pending_writes() needs access to the rdev so...*/
1296 mbio->bi_bdev = (void *)rdev;
1297
1298 atomic_inc(&r10_bio->remaining);
1299
1300 if (!raid1_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) {
1301 spin_lock_irqsave(&conf->device_lock, flags);
1302 bio_list_add(&conf->pending_bio_list, mbio);
1303 spin_unlock_irqrestore(&conf->device_lock, flags);
1304 md_wakeup_thread(mddev->thread);
1305 }
1306}
1307
1308static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1309{
1310 int i;
1311 struct r10conf *conf = mddev->private;
1312 struct md_rdev *blocked_rdev;
1313
1314retry_wait:
1315 blocked_rdev = NULL;
1316 for (i = 0; i < conf->copies; i++) {
1317 struct md_rdev *rdev, *rrdev;
1318
1319 rdev = conf->mirrors[i].rdev;
1320 rrdev = conf->mirrors[i].replacement;
1321 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1322 atomic_inc(&rdev->nr_pending);
1323 blocked_rdev = rdev;
1324 break;
1325 }
1326 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1327 atomic_inc(&rrdev->nr_pending);
1328 blocked_rdev = rrdev;
1329 break;
1330 }
1331
1332 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1333 sector_t first_bad;
1334 sector_t dev_sector = r10_bio->devs[i].addr;
1335 int bad_sectors;
1336 int is_bad;
1337
1338 /*
1339 * Discard request doesn't care the write result
1340 * so it doesn't need to wait blocked disk here.
1341 */
1342 if (!r10_bio->sectors)
1343 continue;
1344
1345 is_bad = is_badblock(rdev, dev_sector, r10_bio->sectors,
1346 &first_bad, &bad_sectors);
1347 if (is_bad < 0) {
1348 /*
1349 * Mustn't write here until the bad block
1350 * is acknowledged
1351 */
1352 atomic_inc(&rdev->nr_pending);
1353 set_bit(BlockedBadBlocks, &rdev->flags);
1354 blocked_rdev = rdev;
1355 break;
1356 }
1357 }
1358 }
1359
1360 if (unlikely(blocked_rdev)) {
1361 /* Have to wait for this device to get unblocked, then retry */
1362 allow_barrier(conf);
1363 raid10_log(conf->mddev, "%s wait rdev %d blocked",
1364 __func__, blocked_rdev->raid_disk);
1365 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1366 wait_barrier(conf, false);
1367 goto retry_wait;
1368 }
1369}
1370
1371static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1372 struct r10bio *r10_bio)
1373{
1374 struct r10conf *conf = mddev->private;
1375 int i;
1376 sector_t sectors;
1377 int max_sectors;
1378
1379 if ((mddev_is_clustered(mddev) &&
1380 md_cluster_ops->area_resyncing(mddev, WRITE,
1381 bio->bi_iter.bi_sector,
1382 bio_end_sector(bio)))) {
1383 DEFINE_WAIT(w);
1384 /* Bail out if REQ_NOWAIT is set for the bio */
1385 if (bio->bi_opf & REQ_NOWAIT) {
1386 bio_wouldblock_error(bio);
1387 return;
1388 }
1389 for (;;) {
1390 prepare_to_wait(&conf->wait_barrier,
1391 &w, TASK_IDLE);
1392 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1393 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1394 break;
1395 schedule();
1396 }
1397 finish_wait(&conf->wait_barrier, &w);
1398 }
1399
1400 sectors = r10_bio->sectors;
1401 if (!regular_request_wait(mddev, conf, bio, sectors))
1402 return;
1403 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1404 (mddev->reshape_backwards
1405 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1406 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1407 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1408 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1409 /* Need to update reshape_position in metadata */
1410 mddev->reshape_position = conf->reshape_progress;
1411 set_mask_bits(&mddev->sb_flags, 0,
1412 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1413 md_wakeup_thread(mddev->thread);
1414 if (bio->bi_opf & REQ_NOWAIT) {
1415 allow_barrier(conf);
1416 bio_wouldblock_error(bio);
1417 return;
1418 }
1419 raid10_log(conf->mddev, "wait reshape metadata");
1420 wait_event(mddev->sb_wait,
1421 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1422
1423 conf->reshape_safe = mddev->reshape_position;
1424 }
1425
1426 /* first select target devices under rcu_lock and
1427 * inc refcount on their rdev. Record them by setting
1428 * bios[x] to bio
1429 * If there are known/acknowledged bad blocks on any device
1430 * on which we have seen a write error, we want to avoid
1431 * writing to those blocks. This potentially requires several
1432 * writes to write around the bad blocks. Each set of writes
1433 * gets its own r10_bio with a set of bios attached.
1434 */
1435
1436 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1437 raid10_find_phys(conf, r10_bio);
1438
1439 wait_blocked_dev(mddev, r10_bio);
1440
1441 max_sectors = r10_bio->sectors;
1442
1443 for (i = 0; i < conf->copies; i++) {
1444 int d = r10_bio->devs[i].devnum;
1445 struct md_rdev *rdev, *rrdev;
1446
1447 rdev = conf->mirrors[d].rdev;
1448 rrdev = conf->mirrors[d].replacement;
1449 if (rdev && (test_bit(Faulty, &rdev->flags)))
1450 rdev = NULL;
1451 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1452 rrdev = NULL;
1453
1454 r10_bio->devs[i].bio = NULL;
1455 r10_bio->devs[i].repl_bio = NULL;
1456
1457 if (!rdev && !rrdev) {
1458 set_bit(R10BIO_Degraded, &r10_bio->state);
1459 continue;
1460 }
1461 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1462 sector_t first_bad;
1463 sector_t dev_sector = r10_bio->devs[i].addr;
1464 int bad_sectors;
1465 int is_bad;
1466
1467 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1468 &first_bad, &bad_sectors);
1469 if (is_bad && first_bad <= dev_sector) {
1470 /* Cannot write here at all */
1471 bad_sectors -= (dev_sector - first_bad);
1472 if (bad_sectors < max_sectors)
1473 /* Mustn't write more than bad_sectors
1474 * to other devices yet
1475 */
1476 max_sectors = bad_sectors;
1477 /* We don't set R10BIO_Degraded as that
1478 * only applies if the disk is missing,
1479 * so it might be re-added, and we want to
1480 * know to recover this chunk.
1481 * In this case the device is here, and the
1482 * fact that this chunk is not in-sync is
1483 * recorded in the bad block log.
1484 */
1485 continue;
1486 }
1487 if (is_bad) {
1488 int good_sectors = first_bad - dev_sector;
1489 if (good_sectors < max_sectors)
1490 max_sectors = good_sectors;
1491 }
1492 }
1493 if (rdev) {
1494 r10_bio->devs[i].bio = bio;
1495 atomic_inc(&rdev->nr_pending);
1496 }
1497 if (rrdev) {
1498 r10_bio->devs[i].repl_bio = bio;
1499 atomic_inc(&rrdev->nr_pending);
1500 }
1501 }
1502
1503 if (max_sectors < r10_bio->sectors)
1504 r10_bio->sectors = max_sectors;
1505
1506 if (r10_bio->sectors < bio_sectors(bio)) {
1507 struct bio *split = bio_split(bio, r10_bio->sectors,
1508 GFP_NOIO, &conf->bio_split);
1509 bio_chain(split, bio);
1510 allow_barrier(conf);
1511 submit_bio_noacct(bio);
1512 wait_barrier(conf, false);
1513 bio = split;
1514 r10_bio->master_bio = bio;
1515 }
1516
1517 md_account_bio(mddev, &bio);
1518 r10_bio->master_bio = bio;
1519 atomic_set(&r10_bio->remaining, 1);
1520 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1521
1522 for (i = 0; i < conf->copies; i++) {
1523 if (r10_bio->devs[i].bio)
1524 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1525 if (r10_bio->devs[i].repl_bio)
1526 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1527 }
1528 one_write_done(r10_bio);
1529}
1530
1531static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1532{
1533 struct r10conf *conf = mddev->private;
1534 struct r10bio *r10_bio;
1535
1536 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1537
1538 r10_bio->master_bio = bio;
1539 r10_bio->sectors = sectors;
1540
1541 r10_bio->mddev = mddev;
1542 r10_bio->sector = bio->bi_iter.bi_sector;
1543 r10_bio->state = 0;
1544 r10_bio->read_slot = -1;
1545 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1546 conf->geo.raid_disks);
1547
1548 if (bio_data_dir(bio) == READ)
1549 raid10_read_request(mddev, bio, r10_bio, true);
1550 else
1551 raid10_write_request(mddev, bio, r10_bio);
1552}
1553
1554static void raid_end_discard_bio(struct r10bio *r10bio)
1555{
1556 struct r10conf *conf = r10bio->mddev->private;
1557 struct r10bio *first_r10bio;
1558
1559 while (atomic_dec_and_test(&r10bio->remaining)) {
1560
1561 allow_barrier(conf);
1562
1563 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1564 first_r10bio = (struct r10bio *)r10bio->master_bio;
1565 free_r10bio(r10bio);
1566 r10bio = first_r10bio;
1567 } else {
1568 md_write_end(r10bio->mddev);
1569 bio_endio(r10bio->master_bio);
1570 free_r10bio(r10bio);
1571 break;
1572 }
1573 }
1574}
1575
1576static void raid10_end_discard_request(struct bio *bio)
1577{
1578 struct r10bio *r10_bio = bio->bi_private;
1579 struct r10conf *conf = r10_bio->mddev->private;
1580 struct md_rdev *rdev = NULL;
1581 int dev;
1582 int slot, repl;
1583
1584 /*
1585 * We don't care the return value of discard bio
1586 */
1587 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1588 set_bit(R10BIO_Uptodate, &r10_bio->state);
1589
1590 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1591 rdev = repl ? conf->mirrors[dev].replacement :
1592 conf->mirrors[dev].rdev;
1593
1594 raid_end_discard_bio(r10_bio);
1595 rdev_dec_pending(rdev, conf->mddev);
1596}
1597
1598/*
1599 * There are some limitations to handle discard bio
1600 * 1st, the discard size is bigger than stripe_size*2.
1601 * 2st, if the discard bio spans reshape progress, we use the old way to
1602 * handle discard bio
1603 */
1604static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1605{
1606 struct r10conf *conf = mddev->private;
1607 struct geom *geo = &conf->geo;
1608 int far_copies = geo->far_copies;
1609 bool first_copy = true;
1610 struct r10bio *r10_bio, *first_r10bio;
1611 struct bio *split;
1612 int disk;
1613 sector_t chunk;
1614 unsigned int stripe_size;
1615 unsigned int stripe_data_disks;
1616 sector_t split_size;
1617 sector_t bio_start, bio_end;
1618 sector_t first_stripe_index, last_stripe_index;
1619 sector_t start_disk_offset;
1620 unsigned int start_disk_index;
1621 sector_t end_disk_offset;
1622 unsigned int end_disk_index;
1623 unsigned int remainder;
1624
1625 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1626 return -EAGAIN;
1627
1628 if (WARN_ON_ONCE(bio->bi_opf & REQ_NOWAIT)) {
1629 bio_wouldblock_error(bio);
1630 return 0;
1631 }
1632 wait_barrier(conf, false);
1633
1634 /*
1635 * Check reshape again to avoid reshape happens after checking
1636 * MD_RECOVERY_RESHAPE and before wait_barrier
1637 */
1638 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1639 goto out;
1640
1641 if (geo->near_copies)
1642 stripe_data_disks = geo->raid_disks / geo->near_copies +
1643 geo->raid_disks % geo->near_copies;
1644 else
1645 stripe_data_disks = geo->raid_disks;
1646
1647 stripe_size = stripe_data_disks << geo->chunk_shift;
1648
1649 bio_start = bio->bi_iter.bi_sector;
1650 bio_end = bio_end_sector(bio);
1651
1652 /*
1653 * Maybe one discard bio is smaller than strip size or across one
1654 * stripe and discard region is larger than one stripe size. For far
1655 * offset layout, if the discard region is not aligned with stripe
1656 * size, there is hole when we submit discard bio to member disk.
1657 * For simplicity, we only handle discard bio which discard region
1658 * is bigger than stripe_size * 2
1659 */
1660 if (bio_sectors(bio) < stripe_size*2)
1661 goto out;
1662
1663 /*
1664 * Keep bio aligned with strip size.
1665 */
1666 div_u64_rem(bio_start, stripe_size, &remainder);
1667 if (remainder) {
1668 split_size = stripe_size - remainder;
1669 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1670 bio_chain(split, bio);
1671 allow_barrier(conf);
1672 /* Resend the fist split part */
1673 submit_bio_noacct(split);
1674 wait_barrier(conf, false);
1675 }
1676 div_u64_rem(bio_end, stripe_size, &remainder);
1677 if (remainder) {
1678 split_size = bio_sectors(bio) - remainder;
1679 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1680 bio_chain(split, bio);
1681 allow_barrier(conf);
1682 /* Resend the second split part */
1683 submit_bio_noacct(bio);
1684 bio = split;
1685 wait_barrier(conf, false);
1686 }
1687
1688 bio_start = bio->bi_iter.bi_sector;
1689 bio_end = bio_end_sector(bio);
1690
1691 /*
1692 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1693 * One stripe contains the chunks from all member disk (one chunk from
1694 * one disk at the same HBA address). For layout detail, see 'man md 4'
1695 */
1696 chunk = bio_start >> geo->chunk_shift;
1697 chunk *= geo->near_copies;
1698 first_stripe_index = chunk;
1699 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1700 if (geo->far_offset)
1701 first_stripe_index *= geo->far_copies;
1702 start_disk_offset = (bio_start & geo->chunk_mask) +
1703 (first_stripe_index << geo->chunk_shift);
1704
1705 chunk = bio_end >> geo->chunk_shift;
1706 chunk *= geo->near_copies;
1707 last_stripe_index = chunk;
1708 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1709 if (geo->far_offset)
1710 last_stripe_index *= geo->far_copies;
1711 end_disk_offset = (bio_end & geo->chunk_mask) +
1712 (last_stripe_index << geo->chunk_shift);
1713
1714retry_discard:
1715 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1716 r10_bio->mddev = mddev;
1717 r10_bio->state = 0;
1718 r10_bio->sectors = 0;
1719 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1720 wait_blocked_dev(mddev, r10_bio);
1721
1722 /*
1723 * For far layout it needs more than one r10bio to cover all regions.
1724 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1725 * to record the discard bio. Other r10bio->master_bio record the first
1726 * r10bio. The first r10bio only release after all other r10bios finish.
1727 * The discard bio returns only first r10bio finishes
1728 */
1729 if (first_copy) {
1730 r10_bio->master_bio = bio;
1731 set_bit(R10BIO_Discard, &r10_bio->state);
1732 first_copy = false;
1733 first_r10bio = r10_bio;
1734 } else
1735 r10_bio->master_bio = (struct bio *)first_r10bio;
1736
1737 /*
1738 * first select target devices under rcu_lock and
1739 * inc refcount on their rdev. Record them by setting
1740 * bios[x] to bio
1741 */
1742 for (disk = 0; disk < geo->raid_disks; disk++) {
1743 struct md_rdev *rdev, *rrdev;
1744
1745 rdev = conf->mirrors[disk].rdev;
1746 rrdev = conf->mirrors[disk].replacement;
1747 r10_bio->devs[disk].bio = NULL;
1748 r10_bio->devs[disk].repl_bio = NULL;
1749
1750 if (rdev && (test_bit(Faulty, &rdev->flags)))
1751 rdev = NULL;
1752 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1753 rrdev = NULL;
1754 if (!rdev && !rrdev)
1755 continue;
1756
1757 if (rdev) {
1758 r10_bio->devs[disk].bio = bio;
1759 atomic_inc(&rdev->nr_pending);
1760 }
1761 if (rrdev) {
1762 r10_bio->devs[disk].repl_bio = bio;
1763 atomic_inc(&rrdev->nr_pending);
1764 }
1765 }
1766
1767 atomic_set(&r10_bio->remaining, 1);
1768 for (disk = 0; disk < geo->raid_disks; disk++) {
1769 sector_t dev_start, dev_end;
1770 struct bio *mbio, *rbio = NULL;
1771
1772 /*
1773 * Now start to calculate the start and end address for each disk.
1774 * The space between dev_start and dev_end is the discard region.
1775 *
1776 * For dev_start, it needs to consider three conditions:
1777 * 1st, the disk is before start_disk, you can imagine the disk in
1778 * the next stripe. So the dev_start is the start address of next
1779 * stripe.
1780 * 2st, the disk is after start_disk, it means the disk is at the
1781 * same stripe of first disk
1782 * 3st, the first disk itself, we can use start_disk_offset directly
1783 */
1784 if (disk < start_disk_index)
1785 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1786 else if (disk > start_disk_index)
1787 dev_start = first_stripe_index * mddev->chunk_sectors;
1788 else
1789 dev_start = start_disk_offset;
1790
1791 if (disk < end_disk_index)
1792 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1793 else if (disk > end_disk_index)
1794 dev_end = last_stripe_index * mddev->chunk_sectors;
1795 else
1796 dev_end = end_disk_offset;
1797
1798 /*
1799 * It only handles discard bio which size is >= stripe size, so
1800 * dev_end > dev_start all the time.
1801 * It doesn't need to use rcu lock to get rdev here. We already
1802 * add rdev->nr_pending in the first loop.
1803 */
1804 if (r10_bio->devs[disk].bio) {
1805 struct md_rdev *rdev = conf->mirrors[disk].rdev;
1806 mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1807 &mddev->bio_set);
1808 mbio->bi_end_io = raid10_end_discard_request;
1809 mbio->bi_private = r10_bio;
1810 r10_bio->devs[disk].bio = mbio;
1811 r10_bio->devs[disk].devnum = disk;
1812 atomic_inc(&r10_bio->remaining);
1813 md_submit_discard_bio(mddev, rdev, mbio,
1814 dev_start + choose_data_offset(r10_bio, rdev),
1815 dev_end - dev_start);
1816 bio_endio(mbio);
1817 }
1818 if (r10_bio->devs[disk].repl_bio) {
1819 struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1820 rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1821 &mddev->bio_set);
1822 rbio->bi_end_io = raid10_end_discard_request;
1823 rbio->bi_private = r10_bio;
1824 r10_bio->devs[disk].repl_bio = rbio;
1825 r10_bio->devs[disk].devnum = disk;
1826 atomic_inc(&r10_bio->remaining);
1827 md_submit_discard_bio(mddev, rrdev, rbio,
1828 dev_start + choose_data_offset(r10_bio, rrdev),
1829 dev_end - dev_start);
1830 bio_endio(rbio);
1831 }
1832 }
1833
1834 if (!geo->far_offset && --far_copies) {
1835 first_stripe_index += geo->stride >> geo->chunk_shift;
1836 start_disk_offset += geo->stride;
1837 last_stripe_index += geo->stride >> geo->chunk_shift;
1838 end_disk_offset += geo->stride;
1839 atomic_inc(&first_r10bio->remaining);
1840 raid_end_discard_bio(r10_bio);
1841 wait_barrier(conf, false);
1842 goto retry_discard;
1843 }
1844
1845 raid_end_discard_bio(r10_bio);
1846
1847 return 0;
1848out:
1849 allow_barrier(conf);
1850 return -EAGAIN;
1851}
1852
1853static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1854{
1855 struct r10conf *conf = mddev->private;
1856 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1857 int chunk_sects = chunk_mask + 1;
1858 int sectors = bio_sectors(bio);
1859
1860 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1861 && md_flush_request(mddev, bio))
1862 return true;
1863
1864 if (!md_write_start(mddev, bio))
1865 return false;
1866
1867 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1868 if (!raid10_handle_discard(mddev, bio))
1869 return true;
1870
1871 /*
1872 * If this request crosses a chunk boundary, we need to split
1873 * it.
1874 */
1875 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1876 sectors > chunk_sects
1877 && (conf->geo.near_copies < conf->geo.raid_disks
1878 || conf->prev.near_copies <
1879 conf->prev.raid_disks)))
1880 sectors = chunk_sects -
1881 (bio->bi_iter.bi_sector &
1882 (chunk_sects - 1));
1883 __make_request(mddev, bio, sectors);
1884
1885 /* In case raid10d snuck in to freeze_array */
1886 wake_up_barrier(conf);
1887 return true;
1888}
1889
1890static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1891{
1892 struct r10conf *conf = mddev->private;
1893 int i;
1894
1895 lockdep_assert_held(&mddev->lock);
1896
1897 if (conf->geo.near_copies < conf->geo.raid_disks)
1898 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1899 if (conf->geo.near_copies > 1)
1900 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1901 if (conf->geo.far_copies > 1) {
1902 if (conf->geo.far_offset)
1903 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1904 else
1905 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1906 if (conf->geo.far_set_size != conf->geo.raid_disks)
1907 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1908 }
1909 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1910 conf->geo.raid_disks - mddev->degraded);
1911 for (i = 0; i < conf->geo.raid_disks; i++) {
1912 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1913
1914 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1915 }
1916 seq_printf(seq, "]");
1917}
1918
1919/* check if there are enough drives for
1920 * every block to appear on atleast one.
1921 * Don't consider the device numbered 'ignore'
1922 * as we might be about to remove it.
1923 */
1924static int _enough(struct r10conf *conf, int previous, int ignore)
1925{
1926 int first = 0;
1927 int has_enough = 0;
1928 int disks, ncopies;
1929 if (previous) {
1930 disks = conf->prev.raid_disks;
1931 ncopies = conf->prev.near_copies;
1932 } else {
1933 disks = conf->geo.raid_disks;
1934 ncopies = conf->geo.near_copies;
1935 }
1936
1937 do {
1938 int n = conf->copies;
1939 int cnt = 0;
1940 int this = first;
1941 while (n--) {
1942 struct md_rdev *rdev;
1943 if (this != ignore &&
1944 (rdev = conf->mirrors[this].rdev) &&
1945 test_bit(In_sync, &rdev->flags))
1946 cnt++;
1947 this = (this+1) % disks;
1948 }
1949 if (cnt == 0)
1950 goto out;
1951 first = (first + ncopies) % disks;
1952 } while (first != 0);
1953 has_enough = 1;
1954out:
1955 return has_enough;
1956}
1957
1958static int enough(struct r10conf *conf, int ignore)
1959{
1960 /* when calling 'enough', both 'prev' and 'geo' must
1961 * be stable.
1962 * This is ensured if ->reconfig_mutex or ->device_lock
1963 * is held.
1964 */
1965 return _enough(conf, 0, ignore) &&
1966 _enough(conf, 1, ignore);
1967}
1968
1969/**
1970 * raid10_error() - RAID10 error handler.
1971 * @mddev: affected md device.
1972 * @rdev: member device to fail.
1973 *
1974 * The routine acknowledges &rdev failure and determines new @mddev state.
1975 * If it failed, then:
1976 * - &MD_BROKEN flag is set in &mddev->flags.
1977 * Otherwise, it must be degraded:
1978 * - recovery is interrupted.
1979 * - &mddev->degraded is bumped.
1980 *
1981 * @rdev is marked as &Faulty excluding case when array is failed and
1982 * &mddev->fail_last_dev is off.
1983 */
1984static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1985{
1986 struct r10conf *conf = mddev->private;
1987 unsigned long flags;
1988
1989 spin_lock_irqsave(&conf->device_lock, flags);
1990
1991 if (test_bit(In_sync, &rdev->flags) && !enough(conf, rdev->raid_disk)) {
1992 set_bit(MD_BROKEN, &mddev->flags);
1993
1994 if (!mddev->fail_last_dev) {
1995 spin_unlock_irqrestore(&conf->device_lock, flags);
1996 return;
1997 }
1998 }
1999 if (test_and_clear_bit(In_sync, &rdev->flags))
2000 mddev->degraded++;
2001
2002 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2003 set_bit(Blocked, &rdev->flags);
2004 set_bit(Faulty, &rdev->flags);
2005 set_mask_bits(&mddev->sb_flags, 0,
2006 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2007 spin_unlock_irqrestore(&conf->device_lock, flags);
2008 pr_crit("md/raid10:%s: Disk failure on %pg, disabling device.\n"
2009 "md/raid10:%s: Operation continuing on %d devices.\n",
2010 mdname(mddev), rdev->bdev,
2011 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
2012}
2013
2014static void print_conf(struct r10conf *conf)
2015{
2016 int i;
2017 struct md_rdev *rdev;
2018
2019 pr_debug("RAID10 conf printout:\n");
2020 if (!conf) {
2021 pr_debug("(!conf)\n");
2022 return;
2023 }
2024 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2025 conf->geo.raid_disks);
2026
2027 lockdep_assert_held(&conf->mddev->reconfig_mutex);
2028 for (i = 0; i < conf->geo.raid_disks; i++) {
2029 rdev = conf->mirrors[i].rdev;
2030 if (rdev)
2031 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
2032 i, !test_bit(In_sync, &rdev->flags),
2033 !test_bit(Faulty, &rdev->flags),
2034 rdev->bdev);
2035 }
2036}
2037
2038static void close_sync(struct r10conf *conf)
2039{
2040 wait_barrier(conf, false);
2041 allow_barrier(conf);
2042
2043 mempool_exit(&conf->r10buf_pool);
2044}
2045
2046static int raid10_spare_active(struct mddev *mddev)
2047{
2048 int i;
2049 struct r10conf *conf = mddev->private;
2050 struct raid10_info *tmp;
2051 int count = 0;
2052 unsigned long flags;
2053
2054 /*
2055 * Find all non-in_sync disks within the RAID10 configuration
2056 * and mark them in_sync
2057 */
2058 for (i = 0; i < conf->geo.raid_disks; i++) {
2059 tmp = conf->mirrors + i;
2060 if (tmp->replacement
2061 && tmp->replacement->recovery_offset == MaxSector
2062 && !test_bit(Faulty, &tmp->replacement->flags)
2063 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2064 /* Replacement has just become active */
2065 if (!tmp->rdev
2066 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2067 count++;
2068 if (tmp->rdev) {
2069 /* Replaced device not technically faulty,
2070 * but we need to be sure it gets removed
2071 * and never re-added.
2072 */
2073 set_bit(Faulty, &tmp->rdev->flags);
2074 sysfs_notify_dirent_safe(
2075 tmp->rdev->sysfs_state);
2076 }
2077 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2078 } else if (tmp->rdev
2079 && tmp->rdev->recovery_offset == MaxSector
2080 && !test_bit(Faulty, &tmp->rdev->flags)
2081 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2082 count++;
2083 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2084 }
2085 }
2086 spin_lock_irqsave(&conf->device_lock, flags);
2087 mddev->degraded -= count;
2088 spin_unlock_irqrestore(&conf->device_lock, flags);
2089
2090 print_conf(conf);
2091 return count;
2092}
2093
2094static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2095{
2096 struct r10conf *conf = mddev->private;
2097 int err = -EEXIST;
2098 int mirror, repl_slot = -1;
2099 int first = 0;
2100 int last = conf->geo.raid_disks - 1;
2101 struct raid10_info *p;
2102
2103 if (mddev->recovery_cp < MaxSector)
2104 /* only hot-add to in-sync arrays, as recovery is
2105 * very different from resync
2106 */
2107 return -EBUSY;
2108 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2109 return -EINVAL;
2110
2111 if (md_integrity_add_rdev(rdev, mddev))
2112 return -ENXIO;
2113
2114 if (rdev->raid_disk >= 0)
2115 first = last = rdev->raid_disk;
2116
2117 if (rdev->saved_raid_disk >= first &&
2118 rdev->saved_raid_disk < conf->geo.raid_disks &&
2119 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2120 mirror = rdev->saved_raid_disk;
2121 else
2122 mirror = first;
2123 for ( ; mirror <= last ; mirror++) {
2124 p = &conf->mirrors[mirror];
2125 if (p->recovery_disabled == mddev->recovery_disabled)
2126 continue;
2127 if (p->rdev) {
2128 if (test_bit(WantReplacement, &p->rdev->flags) &&
2129 p->replacement == NULL && repl_slot < 0)
2130 repl_slot = mirror;
2131 continue;
2132 }
2133
2134 if (mddev->gendisk)
2135 disk_stack_limits(mddev->gendisk, rdev->bdev,
2136 rdev->data_offset << 9);
2137
2138 p->head_position = 0;
2139 p->recovery_disabled = mddev->recovery_disabled - 1;
2140 rdev->raid_disk = mirror;
2141 err = 0;
2142 if (rdev->saved_raid_disk != mirror)
2143 conf->fullsync = 1;
2144 WRITE_ONCE(p->rdev, rdev);
2145 break;
2146 }
2147
2148 if (err && repl_slot >= 0) {
2149 p = &conf->mirrors[repl_slot];
2150 clear_bit(In_sync, &rdev->flags);
2151 set_bit(Replacement, &rdev->flags);
2152 rdev->raid_disk = repl_slot;
2153 err = 0;
2154 if (mddev->gendisk)
2155 disk_stack_limits(mddev->gendisk, rdev->bdev,
2156 rdev->data_offset << 9);
2157 conf->fullsync = 1;
2158 WRITE_ONCE(p->replacement, rdev);
2159 }
2160
2161 print_conf(conf);
2162 return err;
2163}
2164
2165static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2166{
2167 struct r10conf *conf = mddev->private;
2168 int err = 0;
2169 int number = rdev->raid_disk;
2170 struct md_rdev **rdevp;
2171 struct raid10_info *p;
2172
2173 print_conf(conf);
2174 if (unlikely(number >= mddev->raid_disks))
2175 return 0;
2176 p = conf->mirrors + number;
2177 if (rdev == p->rdev)
2178 rdevp = &p->rdev;
2179 else if (rdev == p->replacement)
2180 rdevp = &p->replacement;
2181 else
2182 return 0;
2183
2184 if (test_bit(In_sync, &rdev->flags) ||
2185 atomic_read(&rdev->nr_pending)) {
2186 err = -EBUSY;
2187 goto abort;
2188 }
2189 /* Only remove non-faulty devices if recovery
2190 * is not possible.
2191 */
2192 if (!test_bit(Faulty, &rdev->flags) &&
2193 mddev->recovery_disabled != p->recovery_disabled &&
2194 (!p->replacement || p->replacement == rdev) &&
2195 number < conf->geo.raid_disks &&
2196 enough(conf, -1)) {
2197 err = -EBUSY;
2198 goto abort;
2199 }
2200 WRITE_ONCE(*rdevp, NULL);
2201 if (p->replacement) {
2202 /* We must have just cleared 'rdev' */
2203 WRITE_ONCE(p->rdev, p->replacement);
2204 clear_bit(Replacement, &p->replacement->flags);
2205 WRITE_ONCE(p->replacement, NULL);
2206 }
2207
2208 clear_bit(WantReplacement, &rdev->flags);
2209 err = md_integrity_register(mddev);
2210
2211abort:
2212
2213 print_conf(conf);
2214 return err;
2215}
2216
2217static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2218{
2219 struct r10conf *conf = r10_bio->mddev->private;
2220
2221 if (!bio->bi_status)
2222 set_bit(R10BIO_Uptodate, &r10_bio->state);
2223 else
2224 /* The write handler will notice the lack of
2225 * R10BIO_Uptodate and record any errors etc
2226 */
2227 atomic_add(r10_bio->sectors,
2228 &conf->mirrors[d].rdev->corrected_errors);
2229
2230 /* for reconstruct, we always reschedule after a read.
2231 * for resync, only after all reads
2232 */
2233 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2234 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2235 atomic_dec_and_test(&r10_bio->remaining)) {
2236 /* we have read all the blocks,
2237 * do the comparison in process context in raid10d
2238 */
2239 reschedule_retry(r10_bio);
2240 }
2241}
2242
2243static void end_sync_read(struct bio *bio)
2244{
2245 struct r10bio *r10_bio = get_resync_r10bio(bio);
2246 struct r10conf *conf = r10_bio->mddev->private;
2247 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2248
2249 __end_sync_read(r10_bio, bio, d);
2250}
2251
2252static void end_reshape_read(struct bio *bio)
2253{
2254 /* reshape read bio isn't allocated from r10buf_pool */
2255 struct r10bio *r10_bio = bio->bi_private;
2256
2257 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2258}
2259
2260static void end_sync_request(struct r10bio *r10_bio)
2261{
2262 struct mddev *mddev = r10_bio->mddev;
2263
2264 while (atomic_dec_and_test(&r10_bio->remaining)) {
2265 if (r10_bio->master_bio == NULL) {
2266 /* the primary of several recovery bios */
2267 sector_t s = r10_bio->sectors;
2268 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2269 test_bit(R10BIO_WriteError, &r10_bio->state))
2270 reschedule_retry(r10_bio);
2271 else
2272 put_buf(r10_bio);
2273 md_done_sync(mddev, s, 1);
2274 break;
2275 } else {
2276 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2277 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2278 test_bit(R10BIO_WriteError, &r10_bio->state))
2279 reschedule_retry(r10_bio);
2280 else
2281 put_buf(r10_bio);
2282 r10_bio = r10_bio2;
2283 }
2284 }
2285}
2286
2287static void end_sync_write(struct bio *bio)
2288{
2289 struct r10bio *r10_bio = get_resync_r10bio(bio);
2290 struct mddev *mddev = r10_bio->mddev;
2291 struct r10conf *conf = mddev->private;
2292 int d;
2293 sector_t first_bad;
2294 int bad_sectors;
2295 int slot;
2296 int repl;
2297 struct md_rdev *rdev = NULL;
2298
2299 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2300 if (repl)
2301 rdev = conf->mirrors[d].replacement;
2302 else
2303 rdev = conf->mirrors[d].rdev;
2304
2305 if (bio->bi_status) {
2306 if (repl)
2307 md_error(mddev, rdev);
2308 else {
2309 set_bit(WriteErrorSeen, &rdev->flags);
2310 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2311 set_bit(MD_RECOVERY_NEEDED,
2312 &rdev->mddev->recovery);
2313 set_bit(R10BIO_WriteError, &r10_bio->state);
2314 }
2315 } else if (is_badblock(rdev,
2316 r10_bio->devs[slot].addr,
2317 r10_bio->sectors,
2318 &first_bad, &bad_sectors))
2319 set_bit(R10BIO_MadeGood, &r10_bio->state);
2320
2321 rdev_dec_pending(rdev, mddev);
2322
2323 end_sync_request(r10_bio);
2324}
2325
2326/*
2327 * Note: sync and recover and handled very differently for raid10
2328 * This code is for resync.
2329 * For resync, we read through virtual addresses and read all blocks.
2330 * If there is any error, we schedule a write. The lowest numbered
2331 * drive is authoritative.
2332 * However requests come for physical address, so we need to map.
2333 * For every physical address there are raid_disks/copies virtual addresses,
2334 * which is always are least one, but is not necessarly an integer.
2335 * This means that a physical address can span multiple chunks, so we may
2336 * have to submit multiple io requests for a single sync request.
2337 */
2338/*
2339 * We check if all blocks are in-sync and only write to blocks that
2340 * aren't in sync
2341 */
2342static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2343{
2344 struct r10conf *conf = mddev->private;
2345 int i, first;
2346 struct bio *tbio, *fbio;
2347 int vcnt;
2348 struct page **tpages, **fpages;
2349
2350 atomic_set(&r10_bio->remaining, 1);
2351
2352 /* find the first device with a block */
2353 for (i=0; i<conf->copies; i++)
2354 if (!r10_bio->devs[i].bio->bi_status)
2355 break;
2356
2357 if (i == conf->copies)
2358 goto done;
2359
2360 first = i;
2361 fbio = r10_bio->devs[i].bio;
2362 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2363 fbio->bi_iter.bi_idx = 0;
2364 fpages = get_resync_pages(fbio)->pages;
2365
2366 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2367 /* now find blocks with errors */
2368 for (i=0 ; i < conf->copies ; i++) {
2369 int j, d;
2370 struct md_rdev *rdev;
2371 struct resync_pages *rp;
2372
2373 tbio = r10_bio->devs[i].bio;
2374
2375 if (tbio->bi_end_io != end_sync_read)
2376 continue;
2377 if (i == first)
2378 continue;
2379
2380 tpages = get_resync_pages(tbio)->pages;
2381 d = r10_bio->devs[i].devnum;
2382 rdev = conf->mirrors[d].rdev;
2383 if (!r10_bio->devs[i].bio->bi_status) {
2384 /* We know that the bi_io_vec layout is the same for
2385 * both 'first' and 'i', so we just compare them.
2386 * All vec entries are PAGE_SIZE;
2387 */
2388 int sectors = r10_bio->sectors;
2389 for (j = 0; j < vcnt; j++) {
2390 int len = PAGE_SIZE;
2391 if (sectors < (len / 512))
2392 len = sectors * 512;
2393 if (memcmp(page_address(fpages[j]),
2394 page_address(tpages[j]),
2395 len))
2396 break;
2397 sectors -= len/512;
2398 }
2399 if (j == vcnt)
2400 continue;
2401 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2402 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2403 /* Don't fix anything. */
2404 continue;
2405 } else if (test_bit(FailFast, &rdev->flags)) {
2406 /* Just give up on this device */
2407 md_error(rdev->mddev, rdev);
2408 continue;
2409 }
2410 /* Ok, we need to write this bio, either to correct an
2411 * inconsistency or to correct an unreadable block.
2412 * First we need to fixup bv_offset, bv_len and
2413 * bi_vecs, as the read request might have corrupted these
2414 */
2415 rp = get_resync_pages(tbio);
2416 bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2417
2418 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2419
2420 rp->raid_bio = r10_bio;
2421 tbio->bi_private = rp;
2422 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2423 tbio->bi_end_io = end_sync_write;
2424
2425 bio_copy_data(tbio, fbio);
2426
2427 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2428 atomic_inc(&r10_bio->remaining);
2429 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2430
2431 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2432 tbio->bi_opf |= MD_FAILFAST;
2433 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2434 submit_bio_noacct(tbio);
2435 }
2436
2437 /* Now write out to any replacement devices
2438 * that are active
2439 */
2440 for (i = 0; i < conf->copies; i++) {
2441 int d;
2442
2443 tbio = r10_bio->devs[i].repl_bio;
2444 if (!tbio || !tbio->bi_end_io)
2445 continue;
2446 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2447 && r10_bio->devs[i].bio != fbio)
2448 bio_copy_data(tbio, fbio);
2449 d = r10_bio->devs[i].devnum;
2450 atomic_inc(&r10_bio->remaining);
2451 md_sync_acct(conf->mirrors[d].replacement->bdev,
2452 bio_sectors(tbio));
2453 submit_bio_noacct(tbio);
2454 }
2455
2456done:
2457 if (atomic_dec_and_test(&r10_bio->remaining)) {
2458 md_done_sync(mddev, r10_bio->sectors, 1);
2459 put_buf(r10_bio);
2460 }
2461}
2462
2463/*
2464 * Now for the recovery code.
2465 * Recovery happens across physical sectors.
2466 * We recover all non-is_sync drives by finding the virtual address of
2467 * each, and then choose a working drive that also has that virt address.
2468 * There is a separate r10_bio for each non-in_sync drive.
2469 * Only the first two slots are in use. The first for reading,
2470 * The second for writing.
2471 *
2472 */
2473static void fix_recovery_read_error(struct r10bio *r10_bio)
2474{
2475 /* We got a read error during recovery.
2476 * We repeat the read in smaller page-sized sections.
2477 * If a read succeeds, write it to the new device or record
2478 * a bad block if we cannot.
2479 * If a read fails, record a bad block on both old and
2480 * new devices.
2481 */
2482 struct mddev *mddev = r10_bio->mddev;
2483 struct r10conf *conf = mddev->private;
2484 struct bio *bio = r10_bio->devs[0].bio;
2485 sector_t sect = 0;
2486 int sectors = r10_bio->sectors;
2487 int idx = 0;
2488 int dr = r10_bio->devs[0].devnum;
2489 int dw = r10_bio->devs[1].devnum;
2490 struct page **pages = get_resync_pages(bio)->pages;
2491
2492 while (sectors) {
2493 int s = sectors;
2494 struct md_rdev *rdev;
2495 sector_t addr;
2496 int ok;
2497
2498 if (s > (PAGE_SIZE>>9))
2499 s = PAGE_SIZE >> 9;
2500
2501 rdev = conf->mirrors[dr].rdev;
2502 addr = r10_bio->devs[0].addr + sect,
2503 ok = sync_page_io(rdev,
2504 addr,
2505 s << 9,
2506 pages[idx],
2507 REQ_OP_READ, false);
2508 if (ok) {
2509 rdev = conf->mirrors[dw].rdev;
2510 addr = r10_bio->devs[1].addr + sect;
2511 ok = sync_page_io(rdev,
2512 addr,
2513 s << 9,
2514 pages[idx],
2515 REQ_OP_WRITE, false);
2516 if (!ok) {
2517 set_bit(WriteErrorSeen, &rdev->flags);
2518 if (!test_and_set_bit(WantReplacement,
2519 &rdev->flags))
2520 set_bit(MD_RECOVERY_NEEDED,
2521 &rdev->mddev->recovery);
2522 }
2523 }
2524 if (!ok) {
2525 /* We don't worry if we cannot set a bad block -
2526 * it really is bad so there is no loss in not
2527 * recording it yet
2528 */
2529 rdev_set_badblocks(rdev, addr, s, 0);
2530
2531 if (rdev != conf->mirrors[dw].rdev) {
2532 /* need bad block on destination too */
2533 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2534 addr = r10_bio->devs[1].addr + sect;
2535 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2536 if (!ok) {
2537 /* just abort the recovery */
2538 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2539 mdname(mddev));
2540
2541 conf->mirrors[dw].recovery_disabled
2542 = mddev->recovery_disabled;
2543 set_bit(MD_RECOVERY_INTR,
2544 &mddev->recovery);
2545 break;
2546 }
2547 }
2548 }
2549
2550 sectors -= s;
2551 sect += s;
2552 idx++;
2553 }
2554}
2555
2556static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2557{
2558 struct r10conf *conf = mddev->private;
2559 int d;
2560 struct bio *wbio = r10_bio->devs[1].bio;
2561 struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2562
2563 /* Need to test wbio2->bi_end_io before we call
2564 * submit_bio_noacct as if the former is NULL,
2565 * the latter is free to free wbio2.
2566 */
2567 if (wbio2 && !wbio2->bi_end_io)
2568 wbio2 = NULL;
2569
2570 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2571 fix_recovery_read_error(r10_bio);
2572 if (wbio->bi_end_io)
2573 end_sync_request(r10_bio);
2574 if (wbio2)
2575 end_sync_request(r10_bio);
2576 return;
2577 }
2578
2579 /*
2580 * share the pages with the first bio
2581 * and submit the write request
2582 */
2583 d = r10_bio->devs[1].devnum;
2584 if (wbio->bi_end_io) {
2585 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2586 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2587 submit_bio_noacct(wbio);
2588 }
2589 if (wbio2) {
2590 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2591 md_sync_acct(conf->mirrors[d].replacement->bdev,
2592 bio_sectors(wbio2));
2593 submit_bio_noacct(wbio2);
2594 }
2595}
2596
2597static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2598 int sectors, struct page *page, enum req_op op)
2599{
2600 sector_t first_bad;
2601 int bad_sectors;
2602
2603 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2604 && (op == REQ_OP_READ || test_bit(WriteErrorSeen, &rdev->flags)))
2605 return -1;
2606 if (sync_page_io(rdev, sector, sectors << 9, page, op, false))
2607 /* success */
2608 return 1;
2609 if (op == REQ_OP_WRITE) {
2610 set_bit(WriteErrorSeen, &rdev->flags);
2611 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2612 set_bit(MD_RECOVERY_NEEDED,
2613 &rdev->mddev->recovery);
2614 }
2615 /* need to record an error - either for the block or the device */
2616 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2617 md_error(rdev->mddev, rdev);
2618 return 0;
2619}
2620
2621/*
2622 * This is a kernel thread which:
2623 *
2624 * 1. Retries failed read operations on working mirrors.
2625 * 2. Updates the raid superblock when problems encounter.
2626 * 3. Performs writes following reads for array synchronising.
2627 */
2628
2629static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2630{
2631 int sect = 0; /* Offset from r10_bio->sector */
2632 int sectors = r10_bio->sectors, slot = r10_bio->read_slot;
2633 struct md_rdev *rdev;
2634 int d = r10_bio->devs[slot].devnum;
2635
2636 /* still own a reference to this rdev, so it cannot
2637 * have been cleared recently.
2638 */
2639 rdev = conf->mirrors[d].rdev;
2640
2641 if (test_bit(Faulty, &rdev->flags))
2642 /* drive has already been failed, just ignore any
2643 more fix_read_error() attempts */
2644 return;
2645
2646 if (exceed_read_errors(mddev, rdev)) {
2647 r10_bio->devs[slot].bio = IO_BLOCKED;
2648 return;
2649 }
2650
2651 while(sectors) {
2652 int s = sectors;
2653 int sl = slot;
2654 int success = 0;
2655 int start;
2656
2657 if (s > (PAGE_SIZE>>9))
2658 s = PAGE_SIZE >> 9;
2659
2660 do {
2661 sector_t first_bad;
2662 int bad_sectors;
2663
2664 d = r10_bio->devs[sl].devnum;
2665 rdev = conf->mirrors[d].rdev;
2666 if (rdev &&
2667 test_bit(In_sync, &rdev->flags) &&
2668 !test_bit(Faulty, &rdev->flags) &&
2669 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2670 &first_bad, &bad_sectors) == 0) {
2671 atomic_inc(&rdev->nr_pending);
2672 success = sync_page_io(rdev,
2673 r10_bio->devs[sl].addr +
2674 sect,
2675 s<<9,
2676 conf->tmppage,
2677 REQ_OP_READ, false);
2678 rdev_dec_pending(rdev, mddev);
2679 if (success)
2680 break;
2681 }
2682 sl++;
2683 if (sl == conf->copies)
2684 sl = 0;
2685 } while (sl != slot);
2686
2687 if (!success) {
2688 /* Cannot read from anywhere, just mark the block
2689 * as bad on the first device to discourage future
2690 * reads.
2691 */
2692 int dn = r10_bio->devs[slot].devnum;
2693 rdev = conf->mirrors[dn].rdev;
2694
2695 if (!rdev_set_badblocks(
2696 rdev,
2697 r10_bio->devs[slot].addr
2698 + sect,
2699 s, 0)) {
2700 md_error(mddev, rdev);
2701 r10_bio->devs[slot].bio
2702 = IO_BLOCKED;
2703 }
2704 break;
2705 }
2706
2707 start = sl;
2708 /* write it back and re-read */
2709 while (sl != slot) {
2710 if (sl==0)
2711 sl = conf->copies;
2712 sl--;
2713 d = r10_bio->devs[sl].devnum;
2714 rdev = conf->mirrors[d].rdev;
2715 if (!rdev ||
2716 test_bit(Faulty, &rdev->flags) ||
2717 !test_bit(In_sync, &rdev->flags))
2718 continue;
2719
2720 atomic_inc(&rdev->nr_pending);
2721 if (r10_sync_page_io(rdev,
2722 r10_bio->devs[sl].addr +
2723 sect,
2724 s, conf->tmppage, REQ_OP_WRITE)
2725 == 0) {
2726 /* Well, this device is dead */
2727 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %pg)\n",
2728 mdname(mddev), s,
2729 (unsigned long long)(
2730 sect +
2731 choose_data_offset(r10_bio,
2732 rdev)),
2733 rdev->bdev);
2734 pr_notice("md/raid10:%s: %pg: failing drive\n",
2735 mdname(mddev),
2736 rdev->bdev);
2737 }
2738 rdev_dec_pending(rdev, mddev);
2739 }
2740 sl = start;
2741 while (sl != slot) {
2742 if (sl==0)
2743 sl = conf->copies;
2744 sl--;
2745 d = r10_bio->devs[sl].devnum;
2746 rdev = conf->mirrors[d].rdev;
2747 if (!rdev ||
2748 test_bit(Faulty, &rdev->flags) ||
2749 !test_bit(In_sync, &rdev->flags))
2750 continue;
2751
2752 atomic_inc(&rdev->nr_pending);
2753 switch (r10_sync_page_io(rdev,
2754 r10_bio->devs[sl].addr +
2755 sect,
2756 s, conf->tmppage, REQ_OP_READ)) {
2757 case 0:
2758 /* Well, this device is dead */
2759 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %pg)\n",
2760 mdname(mddev), s,
2761 (unsigned long long)(
2762 sect +
2763 choose_data_offset(r10_bio, rdev)),
2764 rdev->bdev);
2765 pr_notice("md/raid10:%s: %pg: failing drive\n",
2766 mdname(mddev),
2767 rdev->bdev);
2768 break;
2769 case 1:
2770 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %pg)\n",
2771 mdname(mddev), s,
2772 (unsigned long long)(
2773 sect +
2774 choose_data_offset(r10_bio, rdev)),
2775 rdev->bdev);
2776 atomic_add(s, &rdev->corrected_errors);
2777 }
2778
2779 rdev_dec_pending(rdev, mddev);
2780 }
2781
2782 sectors -= s;
2783 sect += s;
2784 }
2785}
2786
2787static int narrow_write_error(struct r10bio *r10_bio, int i)
2788{
2789 struct bio *bio = r10_bio->master_bio;
2790 struct mddev *mddev = r10_bio->mddev;
2791 struct r10conf *conf = mddev->private;
2792 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2793 /* bio has the data to be written to slot 'i' where
2794 * we just recently had a write error.
2795 * We repeatedly clone the bio and trim down to one block,
2796 * then try the write. Where the write fails we record
2797 * a bad block.
2798 * It is conceivable that the bio doesn't exactly align with
2799 * blocks. We must handle this.
2800 *
2801 * We currently own a reference to the rdev.
2802 */
2803
2804 int block_sectors;
2805 sector_t sector;
2806 int sectors;
2807 int sect_to_write = r10_bio->sectors;
2808 int ok = 1;
2809
2810 if (rdev->badblocks.shift < 0)
2811 return 0;
2812
2813 block_sectors = roundup(1 << rdev->badblocks.shift,
2814 bdev_logical_block_size(rdev->bdev) >> 9);
2815 sector = r10_bio->sector;
2816 sectors = ((r10_bio->sector + block_sectors)
2817 & ~(sector_t)(block_sectors - 1))
2818 - sector;
2819
2820 while (sect_to_write) {
2821 struct bio *wbio;
2822 sector_t wsector;
2823 if (sectors > sect_to_write)
2824 sectors = sect_to_write;
2825 /* Write at 'sector' for 'sectors' */
2826 wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2827 &mddev->bio_set);
2828 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2829 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2830 wbio->bi_iter.bi_sector = wsector +
2831 choose_data_offset(r10_bio, rdev);
2832 wbio->bi_opf = REQ_OP_WRITE;
2833
2834 if (submit_bio_wait(wbio) < 0)
2835 /* Failure! */
2836 ok = rdev_set_badblocks(rdev, wsector,
2837 sectors, 0)
2838 && ok;
2839
2840 bio_put(wbio);
2841 sect_to_write -= sectors;
2842 sector += sectors;
2843 sectors = block_sectors;
2844 }
2845 return ok;
2846}
2847
2848static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2849{
2850 int slot = r10_bio->read_slot;
2851 struct bio *bio;
2852 struct r10conf *conf = mddev->private;
2853 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2854
2855 /* we got a read error. Maybe the drive is bad. Maybe just
2856 * the block and we can fix it.
2857 * We freeze all other IO, and try reading the block from
2858 * other devices. When we find one, we re-write
2859 * and check it that fixes the read error.
2860 * This is all done synchronously while the array is
2861 * frozen.
2862 */
2863 bio = r10_bio->devs[slot].bio;
2864 bio_put(bio);
2865 r10_bio->devs[slot].bio = NULL;
2866
2867 if (mddev->ro)
2868 r10_bio->devs[slot].bio = IO_BLOCKED;
2869 else if (!test_bit(FailFast, &rdev->flags)) {
2870 freeze_array(conf, 1);
2871 fix_read_error(conf, mddev, r10_bio);
2872 unfreeze_array(conf);
2873 } else
2874 md_error(mddev, rdev);
2875
2876 rdev_dec_pending(rdev, mddev);
2877 r10_bio->state = 0;
2878 raid10_read_request(mddev, r10_bio->master_bio, r10_bio, false);
2879 /*
2880 * allow_barrier after re-submit to ensure no sync io
2881 * can be issued while regular io pending.
2882 */
2883 allow_barrier(conf);
2884}
2885
2886static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2887{
2888 /* Some sort of write request has finished and it
2889 * succeeded in writing where we thought there was a
2890 * bad block. So forget the bad block.
2891 * Or possibly if failed and we need to record
2892 * a bad block.
2893 */
2894 int m;
2895 struct md_rdev *rdev;
2896
2897 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2898 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2899 for (m = 0; m < conf->copies; m++) {
2900 int dev = r10_bio->devs[m].devnum;
2901 rdev = conf->mirrors[dev].rdev;
2902 if (r10_bio->devs[m].bio == NULL ||
2903 r10_bio->devs[m].bio->bi_end_io == NULL)
2904 continue;
2905 if (!r10_bio->devs[m].bio->bi_status) {
2906 rdev_clear_badblocks(
2907 rdev,
2908 r10_bio->devs[m].addr,
2909 r10_bio->sectors, 0);
2910 } else {
2911 if (!rdev_set_badblocks(
2912 rdev,
2913 r10_bio->devs[m].addr,
2914 r10_bio->sectors, 0))
2915 md_error(conf->mddev, rdev);
2916 }
2917 rdev = conf->mirrors[dev].replacement;
2918 if (r10_bio->devs[m].repl_bio == NULL ||
2919 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2920 continue;
2921
2922 if (!r10_bio->devs[m].repl_bio->bi_status) {
2923 rdev_clear_badblocks(
2924 rdev,
2925 r10_bio->devs[m].addr,
2926 r10_bio->sectors, 0);
2927 } else {
2928 if (!rdev_set_badblocks(
2929 rdev,
2930 r10_bio->devs[m].addr,
2931 r10_bio->sectors, 0))
2932 md_error(conf->mddev, rdev);
2933 }
2934 }
2935 put_buf(r10_bio);
2936 } else {
2937 bool fail = false;
2938 for (m = 0; m < conf->copies; m++) {
2939 int dev = r10_bio->devs[m].devnum;
2940 struct bio *bio = r10_bio->devs[m].bio;
2941 rdev = conf->mirrors[dev].rdev;
2942 if (bio == IO_MADE_GOOD) {
2943 rdev_clear_badblocks(
2944 rdev,
2945 r10_bio->devs[m].addr,
2946 r10_bio->sectors, 0);
2947 rdev_dec_pending(rdev, conf->mddev);
2948 } else if (bio != NULL && bio->bi_status) {
2949 fail = true;
2950 if (!narrow_write_error(r10_bio, m)) {
2951 md_error(conf->mddev, rdev);
2952 set_bit(R10BIO_Degraded,
2953 &r10_bio->state);
2954 }
2955 rdev_dec_pending(rdev, conf->mddev);
2956 }
2957 bio = r10_bio->devs[m].repl_bio;
2958 rdev = conf->mirrors[dev].replacement;
2959 if (rdev && bio == IO_MADE_GOOD) {
2960 rdev_clear_badblocks(
2961 rdev,
2962 r10_bio->devs[m].addr,
2963 r10_bio->sectors, 0);
2964 rdev_dec_pending(rdev, conf->mddev);
2965 }
2966 }
2967 if (fail) {
2968 spin_lock_irq(&conf->device_lock);
2969 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2970 conf->nr_queued++;
2971 spin_unlock_irq(&conf->device_lock);
2972 /*
2973 * In case freeze_array() is waiting for condition
2974 * nr_pending == nr_queued + extra to be true.
2975 */
2976 wake_up(&conf->wait_barrier);
2977 md_wakeup_thread(conf->mddev->thread);
2978 } else {
2979 if (test_bit(R10BIO_WriteError,
2980 &r10_bio->state))
2981 close_write(r10_bio);
2982 raid_end_bio_io(r10_bio);
2983 }
2984 }
2985}
2986
2987static void raid10d(struct md_thread *thread)
2988{
2989 struct mddev *mddev = thread->mddev;
2990 struct r10bio *r10_bio;
2991 unsigned long flags;
2992 struct r10conf *conf = mddev->private;
2993 struct list_head *head = &conf->retry_list;
2994 struct blk_plug plug;
2995
2996 md_check_recovery(mddev);
2997
2998 if (!list_empty_careful(&conf->bio_end_io_list) &&
2999 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3000 LIST_HEAD(tmp);
3001 spin_lock_irqsave(&conf->device_lock, flags);
3002 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3003 while (!list_empty(&conf->bio_end_io_list)) {
3004 list_move(conf->bio_end_io_list.prev, &tmp);
3005 conf->nr_queued--;
3006 }
3007 }
3008 spin_unlock_irqrestore(&conf->device_lock, flags);
3009 while (!list_empty(&tmp)) {
3010 r10_bio = list_first_entry(&tmp, struct r10bio,
3011 retry_list);
3012 list_del(&r10_bio->retry_list);
3013 if (mddev->degraded)
3014 set_bit(R10BIO_Degraded, &r10_bio->state);
3015
3016 if (test_bit(R10BIO_WriteError,
3017 &r10_bio->state))
3018 close_write(r10_bio);
3019 raid_end_bio_io(r10_bio);
3020 }
3021 }
3022
3023 blk_start_plug(&plug);
3024 for (;;) {
3025
3026 flush_pending_writes(conf);
3027
3028 spin_lock_irqsave(&conf->device_lock, flags);
3029 if (list_empty(head)) {
3030 spin_unlock_irqrestore(&conf->device_lock, flags);
3031 break;
3032 }
3033 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3034 list_del(head->prev);
3035 conf->nr_queued--;
3036 spin_unlock_irqrestore(&conf->device_lock, flags);
3037
3038 mddev = r10_bio->mddev;
3039 conf = mddev->private;
3040 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3041 test_bit(R10BIO_WriteError, &r10_bio->state))
3042 handle_write_completed(conf, r10_bio);
3043 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3044 reshape_request_write(mddev, r10_bio);
3045 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3046 sync_request_write(mddev, r10_bio);
3047 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3048 recovery_request_write(mddev, r10_bio);
3049 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3050 handle_read_error(mddev, r10_bio);
3051 else
3052 WARN_ON_ONCE(1);
3053
3054 cond_resched();
3055 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3056 md_check_recovery(mddev);
3057 }
3058 blk_finish_plug(&plug);
3059}
3060
3061static int init_resync(struct r10conf *conf)
3062{
3063 int ret, buffs, i;
3064
3065 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3066 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3067 conf->have_replacement = 0;
3068 for (i = 0; i < conf->geo.raid_disks; i++)
3069 if (conf->mirrors[i].replacement)
3070 conf->have_replacement = 1;
3071 ret = mempool_init(&conf->r10buf_pool, buffs,
3072 r10buf_pool_alloc, r10buf_pool_free, conf);
3073 if (ret)
3074 return ret;
3075 conf->next_resync = 0;
3076 return 0;
3077}
3078
3079static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3080{
3081 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3082 struct rsync_pages *rp;
3083 struct bio *bio;
3084 int nalloc;
3085 int i;
3086
3087 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3088 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3089 nalloc = conf->copies; /* resync */
3090 else
3091 nalloc = 2; /* recovery */
3092
3093 for (i = 0; i < nalloc; i++) {
3094 bio = r10bio->devs[i].bio;
3095 rp = bio->bi_private;
3096 bio_reset(bio, NULL, 0);
3097 bio->bi_private = rp;
3098 bio = r10bio->devs[i].repl_bio;
3099 if (bio) {
3100 rp = bio->bi_private;
3101 bio_reset(bio, NULL, 0);
3102 bio->bi_private = rp;
3103 }
3104 }
3105 return r10bio;
3106}
3107
3108/*
3109 * Set cluster_sync_high since we need other nodes to add the
3110 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3111 */
3112static void raid10_set_cluster_sync_high(struct r10conf *conf)
3113{
3114 sector_t window_size;
3115 int extra_chunk, chunks;
3116
3117 /*
3118 * First, here we define "stripe" as a unit which across
3119 * all member devices one time, so we get chunks by use
3120 * raid_disks / near_copies. Otherwise, if near_copies is
3121 * close to raid_disks, then resync window could increases
3122 * linearly with the increase of raid_disks, which means
3123 * we will suspend a really large IO window while it is not
3124 * necessary. If raid_disks is not divisible by near_copies,
3125 * an extra chunk is needed to ensure the whole "stripe" is
3126 * covered.
3127 */
3128
3129 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3130 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3131 extra_chunk = 0;
3132 else
3133 extra_chunk = 1;
3134 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3135
3136 /*
3137 * At least use a 32M window to align with raid1's resync window
3138 */
3139 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3140 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3141
3142 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3143}
3144
3145/*
3146 * perform a "sync" on one "block"
3147 *
3148 * We need to make sure that no normal I/O request - particularly write
3149 * requests - conflict with active sync requests.
3150 *
3151 * This is achieved by tracking pending requests and a 'barrier' concept
3152 * that can be installed to exclude normal IO requests.
3153 *
3154 * Resync and recovery are handled very differently.
3155 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3156 *
3157 * For resync, we iterate over virtual addresses, read all copies,
3158 * and update if there are differences. If only one copy is live,
3159 * skip it.
3160 * For recovery, we iterate over physical addresses, read a good
3161 * value for each non-in_sync drive, and over-write.
3162 *
3163 * So, for recovery we may have several outstanding complex requests for a
3164 * given address, one for each out-of-sync device. We model this by allocating
3165 * a number of r10_bio structures, one for each out-of-sync device.
3166 * As we setup these structures, we collect all bio's together into a list
3167 * which we then process collectively to add pages, and then process again
3168 * to pass to submit_bio_noacct.
3169 *
3170 * The r10_bio structures are linked using a borrowed master_bio pointer.
3171 * This link is counted in ->remaining. When the r10_bio that points to NULL
3172 * has its remaining count decremented to 0, the whole complex operation
3173 * is complete.
3174 *
3175 */
3176
3177static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3178 int *skipped)
3179{
3180 struct r10conf *conf = mddev->private;
3181 struct r10bio *r10_bio;
3182 struct bio *biolist = NULL, *bio;
3183 sector_t max_sector, nr_sectors;
3184 int i;
3185 int max_sync;
3186 sector_t sync_blocks;
3187 sector_t sectors_skipped = 0;
3188 int chunks_skipped = 0;
3189 sector_t chunk_mask = conf->geo.chunk_mask;
3190 int page_idx = 0;
3191 int error_disk = -1;
3192
3193 /*
3194 * Allow skipping a full rebuild for incremental assembly
3195 * of a clean array, like RAID1 does.
3196 */
3197 if (mddev->bitmap == NULL &&
3198 mddev->recovery_cp == MaxSector &&
3199 mddev->reshape_position == MaxSector &&
3200 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3201 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3202 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3203 conf->fullsync == 0) {
3204 *skipped = 1;
3205 return mddev->dev_sectors - sector_nr;
3206 }
3207
3208 if (!mempool_initialized(&conf->r10buf_pool))
3209 if (init_resync(conf))
3210 return 0;
3211
3212 skipped:
3213 max_sector = mddev->dev_sectors;
3214 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
3215 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3216 max_sector = mddev->resync_max_sectors;
3217 if (sector_nr >= max_sector) {
3218 conf->cluster_sync_low = 0;
3219 conf->cluster_sync_high = 0;
3220
3221 /* If we aborted, we need to abort the
3222 * sync on the 'current' bitmap chucks (there can
3223 * be several when recovering multiple devices).
3224 * as we may have started syncing it but not finished.
3225 * We can find the current address in
3226 * mddev->curr_resync, but for recovery,
3227 * we need to convert that to several
3228 * virtual addresses.
3229 */
3230 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3231 end_reshape(conf);
3232 close_sync(conf);
3233 return 0;
3234 }
3235
3236 if (mddev->curr_resync < max_sector) { /* aborted */
3237 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3238 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3239 &sync_blocks, 1);
3240 else for (i = 0; i < conf->geo.raid_disks; i++) {
3241 sector_t sect =
3242 raid10_find_virt(conf, mddev->curr_resync, i);
3243 md_bitmap_end_sync(mddev->bitmap, sect,
3244 &sync_blocks, 1);
3245 }
3246 } else {
3247 /* completed sync */
3248 if ((!mddev->bitmap || conf->fullsync)
3249 && conf->have_replacement
3250 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3251 /* Completed a full sync so the replacements
3252 * are now fully recovered.
3253 */
3254 for (i = 0; i < conf->geo.raid_disks; i++) {
3255 struct md_rdev *rdev =
3256 conf->mirrors[i].replacement;
3257
3258 if (rdev)
3259 rdev->recovery_offset = MaxSector;
3260 }
3261 }
3262 conf->fullsync = 0;
3263 }
3264 md_bitmap_close_sync(mddev->bitmap);
3265 close_sync(conf);
3266 *skipped = 1;
3267 return sectors_skipped;
3268 }
3269
3270 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3271 return reshape_request(mddev, sector_nr, skipped);
3272
3273 if (chunks_skipped >= conf->geo.raid_disks) {
3274 pr_err("md/raid10:%s: %s fails\n", mdname(mddev),
3275 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ? "resync" : "recovery");
3276 if (error_disk >= 0 &&
3277 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3278 /*
3279 * recovery fails, set mirrors.recovery_disabled,
3280 * device shouldn't be added to there.
3281 */
3282 conf->mirrors[error_disk].recovery_disabled =
3283 mddev->recovery_disabled;
3284 return 0;
3285 }
3286 /*
3287 * if there has been nothing to do on any drive,
3288 * then there is nothing to do at all.
3289 */
3290 *skipped = 1;
3291 return (max_sector - sector_nr) + sectors_skipped;
3292 }
3293
3294 if (max_sector > mddev->resync_max)
3295 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3296
3297 /* make sure whole request will fit in a chunk - if chunks
3298 * are meaningful
3299 */
3300 if (conf->geo.near_copies < conf->geo.raid_disks &&
3301 max_sector > (sector_nr | chunk_mask))
3302 max_sector = (sector_nr | chunk_mask) + 1;
3303
3304 /*
3305 * If there is non-resync activity waiting for a turn, then let it
3306 * though before starting on this new sync request.
3307 */
3308 if (conf->nr_waiting)
3309 schedule_timeout_uninterruptible(1);
3310
3311 /* Again, very different code for resync and recovery.
3312 * Both must result in an r10bio with a list of bios that
3313 * have bi_end_io, bi_sector, bi_bdev set,
3314 * and bi_private set to the r10bio.
3315 * For recovery, we may actually create several r10bios
3316 * with 2 bios in each, that correspond to the bios in the main one.
3317 * In this case, the subordinate r10bios link back through a
3318 * borrowed master_bio pointer, and the counter in the master
3319 * includes a ref from each subordinate.
3320 */
3321 /* First, we decide what to do and set ->bi_end_io
3322 * To end_sync_read if we want to read, and
3323 * end_sync_write if we will want to write.
3324 */
3325
3326 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3327 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3328 /* recovery... the complicated one */
3329 int j;
3330 r10_bio = NULL;
3331
3332 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3333 int still_degraded;
3334 struct r10bio *rb2;
3335 sector_t sect;
3336 int must_sync;
3337 int any_working;
3338 struct raid10_info *mirror = &conf->mirrors[i];
3339 struct md_rdev *mrdev, *mreplace;
3340
3341 mrdev = mirror->rdev;
3342 mreplace = mirror->replacement;
3343
3344 if (mrdev && (test_bit(Faulty, &mrdev->flags) ||
3345 test_bit(In_sync, &mrdev->flags)))
3346 mrdev = NULL;
3347 if (mreplace && test_bit(Faulty, &mreplace->flags))
3348 mreplace = NULL;
3349
3350 if (!mrdev && !mreplace)
3351 continue;
3352
3353 still_degraded = 0;
3354 /* want to reconstruct this device */
3355 rb2 = r10_bio;
3356 sect = raid10_find_virt(conf, sector_nr, i);
3357 if (sect >= mddev->resync_max_sectors)
3358 /* last stripe is not complete - don't
3359 * try to recover this sector.
3360 */
3361 continue;
3362 /* Unless we are doing a full sync, or a replacement
3363 * we only need to recover the block if it is set in
3364 * the bitmap
3365 */
3366 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3367 &sync_blocks, 1);
3368 if (sync_blocks < max_sync)
3369 max_sync = sync_blocks;
3370 if (!must_sync &&
3371 mreplace == NULL &&
3372 !conf->fullsync) {
3373 /* yep, skip the sync_blocks here, but don't assume
3374 * that there will never be anything to do here
3375 */
3376 chunks_skipped = -1;
3377 continue;
3378 }
3379 if (mrdev)
3380 atomic_inc(&mrdev->nr_pending);
3381 if (mreplace)
3382 atomic_inc(&mreplace->nr_pending);
3383
3384 r10_bio = raid10_alloc_init_r10buf(conf);
3385 r10_bio->state = 0;
3386 raise_barrier(conf, rb2 != NULL);
3387 atomic_set(&r10_bio->remaining, 0);
3388
3389 r10_bio->master_bio = (struct bio*)rb2;
3390 if (rb2)
3391 atomic_inc(&rb2->remaining);
3392 r10_bio->mddev = mddev;
3393 set_bit(R10BIO_IsRecover, &r10_bio->state);
3394 r10_bio->sector = sect;
3395
3396 raid10_find_phys(conf, r10_bio);
3397
3398 /* Need to check if the array will still be
3399 * degraded
3400 */
3401 for (j = 0; j < conf->geo.raid_disks; j++) {
3402 struct md_rdev *rdev = conf->mirrors[j].rdev;
3403
3404 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3405 still_degraded = 1;
3406 break;
3407 }
3408 }
3409
3410 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3411 &sync_blocks, still_degraded);
3412
3413 any_working = 0;
3414 for (j=0; j<conf->copies;j++) {
3415 int k;
3416 int d = r10_bio->devs[j].devnum;
3417 sector_t from_addr, to_addr;
3418 struct md_rdev *rdev = conf->mirrors[d].rdev;
3419 sector_t sector, first_bad;
3420 int bad_sectors;
3421 if (!rdev ||
3422 !test_bit(In_sync, &rdev->flags))
3423 continue;
3424 /* This is where we read from */
3425 any_working = 1;
3426 sector = r10_bio->devs[j].addr;
3427
3428 if (is_badblock(rdev, sector, max_sync,
3429 &first_bad, &bad_sectors)) {
3430 if (first_bad > sector)
3431 max_sync = first_bad - sector;
3432 else {
3433 bad_sectors -= (sector
3434 - first_bad);
3435 if (max_sync > bad_sectors)
3436 max_sync = bad_sectors;
3437 continue;
3438 }
3439 }
3440 bio = r10_bio->devs[0].bio;
3441 bio->bi_next = biolist;
3442 biolist = bio;
3443 bio->bi_end_io = end_sync_read;
3444 bio->bi_opf = REQ_OP_READ;
3445 if (test_bit(FailFast, &rdev->flags))
3446 bio->bi_opf |= MD_FAILFAST;
3447 from_addr = r10_bio->devs[j].addr;
3448 bio->bi_iter.bi_sector = from_addr +
3449 rdev->data_offset;
3450 bio_set_dev(bio, rdev->bdev);
3451 atomic_inc(&rdev->nr_pending);
3452 /* and we write to 'i' (if not in_sync) */
3453
3454 for (k=0; k<conf->copies; k++)
3455 if (r10_bio->devs[k].devnum == i)
3456 break;
3457 BUG_ON(k == conf->copies);
3458 to_addr = r10_bio->devs[k].addr;
3459 r10_bio->devs[0].devnum = d;
3460 r10_bio->devs[0].addr = from_addr;
3461 r10_bio->devs[1].devnum = i;
3462 r10_bio->devs[1].addr = to_addr;
3463
3464 if (mrdev) {
3465 bio = r10_bio->devs[1].bio;
3466 bio->bi_next = biolist;
3467 biolist = bio;
3468 bio->bi_end_io = end_sync_write;
3469 bio->bi_opf = REQ_OP_WRITE;
3470 bio->bi_iter.bi_sector = to_addr
3471 + mrdev->data_offset;
3472 bio_set_dev(bio, mrdev->bdev);
3473 atomic_inc(&r10_bio->remaining);
3474 } else
3475 r10_bio->devs[1].bio->bi_end_io = NULL;
3476
3477 /* and maybe write to replacement */
3478 bio = r10_bio->devs[1].repl_bio;
3479 if (bio)
3480 bio->bi_end_io = NULL;
3481 /* Note: if replace is not NULL, then bio
3482 * cannot be NULL as r10buf_pool_alloc will
3483 * have allocated it.
3484 */
3485 if (!mreplace)
3486 break;
3487 bio->bi_next = biolist;
3488 biolist = bio;
3489 bio->bi_end_io = end_sync_write;
3490 bio->bi_opf = REQ_OP_WRITE;
3491 bio->bi_iter.bi_sector = to_addr +
3492 mreplace->data_offset;
3493 bio_set_dev(bio, mreplace->bdev);
3494 atomic_inc(&r10_bio->remaining);
3495 break;
3496 }
3497 if (j == conf->copies) {
3498 /* Cannot recover, so abort the recovery or
3499 * record a bad block */
3500 if (any_working) {
3501 /* problem is that there are bad blocks
3502 * on other device(s)
3503 */
3504 int k;
3505 for (k = 0; k < conf->copies; k++)
3506 if (r10_bio->devs[k].devnum == i)
3507 break;
3508 if (mrdev && !test_bit(In_sync,
3509 &mrdev->flags)
3510 && !rdev_set_badblocks(
3511 mrdev,
3512 r10_bio->devs[k].addr,
3513 max_sync, 0))
3514 any_working = 0;
3515 if (mreplace &&
3516 !rdev_set_badblocks(
3517 mreplace,
3518 r10_bio->devs[k].addr,
3519 max_sync, 0))
3520 any_working = 0;
3521 }
3522 if (!any_working) {
3523 if (!test_and_set_bit(MD_RECOVERY_INTR,
3524 &mddev->recovery))
3525 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3526 mdname(mddev));
3527 mirror->recovery_disabled
3528 = mddev->recovery_disabled;
3529 } else {
3530 error_disk = i;
3531 }
3532 put_buf(r10_bio);
3533 if (rb2)
3534 atomic_dec(&rb2->remaining);
3535 r10_bio = rb2;
3536 if (mrdev)
3537 rdev_dec_pending(mrdev, mddev);
3538 if (mreplace)
3539 rdev_dec_pending(mreplace, mddev);
3540 break;
3541 }
3542 if (mrdev)
3543 rdev_dec_pending(mrdev, mddev);
3544 if (mreplace)
3545 rdev_dec_pending(mreplace, mddev);
3546 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3547 /* Only want this if there is elsewhere to
3548 * read from. 'j' is currently the first
3549 * readable copy.
3550 */
3551 int targets = 1;
3552 for (; j < conf->copies; j++) {
3553 int d = r10_bio->devs[j].devnum;
3554 if (conf->mirrors[d].rdev &&
3555 test_bit(In_sync,
3556 &conf->mirrors[d].rdev->flags))
3557 targets++;
3558 }
3559 if (targets == 1)
3560 r10_bio->devs[0].bio->bi_opf
3561 &= ~MD_FAILFAST;
3562 }
3563 }
3564 if (biolist == NULL) {
3565 while (r10_bio) {
3566 struct r10bio *rb2 = r10_bio;
3567 r10_bio = (struct r10bio*) rb2->master_bio;
3568 rb2->master_bio = NULL;
3569 put_buf(rb2);
3570 }
3571 goto giveup;
3572 }
3573 } else {
3574 /* resync. Schedule a read for every block at this virt offset */
3575 int count = 0;
3576
3577 /*
3578 * Since curr_resync_completed could probably not update in
3579 * time, and we will set cluster_sync_low based on it.
3580 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3581 * safety reason, which ensures curr_resync_completed is
3582 * updated in bitmap_cond_end_sync.
3583 */
3584 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3585 mddev_is_clustered(mddev) &&
3586 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3587
3588 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3589 &sync_blocks, mddev->degraded) &&
3590 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3591 &mddev->recovery)) {
3592 /* We can skip this block */
3593 *skipped = 1;
3594 return sync_blocks + sectors_skipped;
3595 }
3596 if (sync_blocks < max_sync)
3597 max_sync = sync_blocks;
3598 r10_bio = raid10_alloc_init_r10buf(conf);
3599 r10_bio->state = 0;
3600
3601 r10_bio->mddev = mddev;
3602 atomic_set(&r10_bio->remaining, 0);
3603 raise_barrier(conf, 0);
3604 conf->next_resync = sector_nr;
3605
3606 r10_bio->master_bio = NULL;
3607 r10_bio->sector = sector_nr;
3608 set_bit(R10BIO_IsSync, &r10_bio->state);
3609 raid10_find_phys(conf, r10_bio);
3610 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3611
3612 for (i = 0; i < conf->copies; i++) {
3613 int d = r10_bio->devs[i].devnum;
3614 sector_t first_bad, sector;
3615 int bad_sectors;
3616 struct md_rdev *rdev;
3617
3618 if (r10_bio->devs[i].repl_bio)
3619 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3620
3621 bio = r10_bio->devs[i].bio;
3622 bio->bi_status = BLK_STS_IOERR;
3623 rdev = conf->mirrors[d].rdev;
3624 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3625 continue;
3626
3627 sector = r10_bio->devs[i].addr;
3628 if (is_badblock(rdev, sector, max_sync,
3629 &first_bad, &bad_sectors)) {
3630 if (first_bad > sector)
3631 max_sync = first_bad - sector;
3632 else {
3633 bad_sectors -= (sector - first_bad);
3634 if (max_sync > bad_sectors)
3635 max_sync = bad_sectors;
3636 continue;
3637 }
3638 }
3639 atomic_inc(&rdev->nr_pending);
3640 atomic_inc(&r10_bio->remaining);
3641 bio->bi_next = biolist;
3642 biolist = bio;
3643 bio->bi_end_io = end_sync_read;
3644 bio->bi_opf = REQ_OP_READ;
3645 if (test_bit(FailFast, &rdev->flags))
3646 bio->bi_opf |= MD_FAILFAST;
3647 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3648 bio_set_dev(bio, rdev->bdev);
3649 count++;
3650
3651 rdev = conf->mirrors[d].replacement;
3652 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3653 continue;
3654
3655 atomic_inc(&rdev->nr_pending);
3656
3657 /* Need to set up for writing to the replacement */
3658 bio = r10_bio->devs[i].repl_bio;
3659 bio->bi_status = BLK_STS_IOERR;
3660
3661 sector = r10_bio->devs[i].addr;
3662 bio->bi_next = biolist;
3663 biolist = bio;
3664 bio->bi_end_io = end_sync_write;
3665 bio->bi_opf = REQ_OP_WRITE;
3666 if (test_bit(FailFast, &rdev->flags))
3667 bio->bi_opf |= MD_FAILFAST;
3668 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3669 bio_set_dev(bio, rdev->bdev);
3670 count++;
3671 }
3672
3673 if (count < 2) {
3674 for (i=0; i<conf->copies; i++) {
3675 int d = r10_bio->devs[i].devnum;
3676 if (r10_bio->devs[i].bio->bi_end_io)
3677 rdev_dec_pending(conf->mirrors[d].rdev,
3678 mddev);
3679 if (r10_bio->devs[i].repl_bio &&
3680 r10_bio->devs[i].repl_bio->bi_end_io)
3681 rdev_dec_pending(
3682 conf->mirrors[d].replacement,
3683 mddev);
3684 }
3685 put_buf(r10_bio);
3686 biolist = NULL;
3687 goto giveup;
3688 }
3689 }
3690
3691 nr_sectors = 0;
3692 if (sector_nr + max_sync < max_sector)
3693 max_sector = sector_nr + max_sync;
3694 do {
3695 struct page *page;
3696 int len = PAGE_SIZE;
3697 if (sector_nr + (len>>9) > max_sector)
3698 len = (max_sector - sector_nr) << 9;
3699 if (len == 0)
3700 break;
3701 for (bio= biolist ; bio ; bio=bio->bi_next) {
3702 struct resync_pages *rp = get_resync_pages(bio);
3703 page = resync_fetch_page(rp, page_idx);
3704 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
3705 bio->bi_status = BLK_STS_RESOURCE;
3706 bio_endio(bio);
3707 goto giveup;
3708 }
3709 }
3710 nr_sectors += len>>9;
3711 sector_nr += len>>9;
3712 } while (++page_idx < RESYNC_PAGES);
3713 r10_bio->sectors = nr_sectors;
3714
3715 if (mddev_is_clustered(mddev) &&
3716 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3717 /* It is resync not recovery */
3718 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3719 conf->cluster_sync_low = mddev->curr_resync_completed;
3720 raid10_set_cluster_sync_high(conf);
3721 /* Send resync message */
3722 md_cluster_ops->resync_info_update(mddev,
3723 conf->cluster_sync_low,
3724 conf->cluster_sync_high);
3725 }
3726 } else if (mddev_is_clustered(mddev)) {
3727 /* This is recovery not resync */
3728 sector_t sect_va1, sect_va2;
3729 bool broadcast_msg = false;
3730
3731 for (i = 0; i < conf->geo.raid_disks; i++) {
3732 /*
3733 * sector_nr is a device address for recovery, so we
3734 * need translate it to array address before compare
3735 * with cluster_sync_high.
3736 */
3737 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3738
3739 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3740 broadcast_msg = true;
3741 /*
3742 * curr_resync_completed is similar as
3743 * sector_nr, so make the translation too.
3744 */
3745 sect_va2 = raid10_find_virt(conf,
3746 mddev->curr_resync_completed, i);
3747
3748 if (conf->cluster_sync_low == 0 ||
3749 conf->cluster_sync_low > sect_va2)
3750 conf->cluster_sync_low = sect_va2;
3751 }
3752 }
3753 if (broadcast_msg) {
3754 raid10_set_cluster_sync_high(conf);
3755 md_cluster_ops->resync_info_update(mddev,
3756 conf->cluster_sync_low,
3757 conf->cluster_sync_high);
3758 }
3759 }
3760
3761 while (biolist) {
3762 bio = biolist;
3763 biolist = biolist->bi_next;
3764
3765 bio->bi_next = NULL;
3766 r10_bio = get_resync_r10bio(bio);
3767 r10_bio->sectors = nr_sectors;
3768
3769 if (bio->bi_end_io == end_sync_read) {
3770 md_sync_acct_bio(bio, nr_sectors);
3771 bio->bi_status = 0;
3772 submit_bio_noacct(bio);
3773 }
3774 }
3775
3776 if (sectors_skipped)
3777 /* pretend they weren't skipped, it makes
3778 * no important difference in this case
3779 */
3780 md_done_sync(mddev, sectors_skipped, 1);
3781
3782 return sectors_skipped + nr_sectors;
3783 giveup:
3784 /* There is nowhere to write, so all non-sync
3785 * drives must be failed or in resync, all drives
3786 * have a bad block, so try the next chunk...
3787 */
3788 if (sector_nr + max_sync < max_sector)
3789 max_sector = sector_nr + max_sync;
3790
3791 sectors_skipped += (max_sector - sector_nr);
3792 chunks_skipped ++;
3793 sector_nr = max_sector;
3794 goto skipped;
3795}
3796
3797static sector_t
3798raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3799{
3800 sector_t size;
3801 struct r10conf *conf = mddev->private;
3802
3803 if (!raid_disks)
3804 raid_disks = min(conf->geo.raid_disks,
3805 conf->prev.raid_disks);
3806 if (!sectors)
3807 sectors = conf->dev_sectors;
3808
3809 size = sectors >> conf->geo.chunk_shift;
3810 sector_div(size, conf->geo.far_copies);
3811 size = size * raid_disks;
3812 sector_div(size, conf->geo.near_copies);
3813
3814 return size << conf->geo.chunk_shift;
3815}
3816
3817static void calc_sectors(struct r10conf *conf, sector_t size)
3818{
3819 /* Calculate the number of sectors-per-device that will
3820 * actually be used, and set conf->dev_sectors and
3821 * conf->stride
3822 */
3823
3824 size = size >> conf->geo.chunk_shift;
3825 sector_div(size, conf->geo.far_copies);
3826 size = size * conf->geo.raid_disks;
3827 sector_div(size, conf->geo.near_copies);
3828 /* 'size' is now the number of chunks in the array */
3829 /* calculate "used chunks per device" */
3830 size = size * conf->copies;
3831
3832 /* We need to round up when dividing by raid_disks to
3833 * get the stride size.
3834 */
3835 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3836
3837 conf->dev_sectors = size << conf->geo.chunk_shift;
3838
3839 if (conf->geo.far_offset)
3840 conf->geo.stride = 1 << conf->geo.chunk_shift;
3841 else {
3842 sector_div(size, conf->geo.far_copies);
3843 conf->geo.stride = size << conf->geo.chunk_shift;
3844 }
3845}
3846
3847enum geo_type {geo_new, geo_old, geo_start};
3848static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3849{
3850 int nc, fc, fo;
3851 int layout, chunk, disks;
3852 switch (new) {
3853 case geo_old:
3854 layout = mddev->layout;
3855 chunk = mddev->chunk_sectors;
3856 disks = mddev->raid_disks - mddev->delta_disks;
3857 break;
3858 case geo_new:
3859 layout = mddev->new_layout;
3860 chunk = mddev->new_chunk_sectors;
3861 disks = mddev->raid_disks;
3862 break;
3863 default: /* avoid 'may be unused' warnings */
3864 case geo_start: /* new when starting reshape - raid_disks not
3865 * updated yet. */
3866 layout = mddev->new_layout;
3867 chunk = mddev->new_chunk_sectors;
3868 disks = mddev->raid_disks + mddev->delta_disks;
3869 break;
3870 }
3871 if (layout >> 19)
3872 return -1;
3873 if (chunk < (PAGE_SIZE >> 9) ||
3874 !is_power_of_2(chunk))
3875 return -2;
3876 nc = layout & 255;
3877 fc = (layout >> 8) & 255;
3878 fo = layout & (1<<16);
3879 geo->raid_disks = disks;
3880 geo->near_copies = nc;
3881 geo->far_copies = fc;
3882 geo->far_offset = fo;
3883 switch (layout >> 17) {
3884 case 0: /* original layout. simple but not always optimal */
3885 geo->far_set_size = disks;
3886 break;
3887 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3888 * actually using this, but leave code here just in case.*/
3889 geo->far_set_size = disks/fc;
3890 WARN(geo->far_set_size < fc,
3891 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3892 break;
3893 case 2: /* "improved" layout fixed to match documentation */
3894 geo->far_set_size = fc * nc;
3895 break;
3896 default: /* Not a valid layout */
3897 return -1;
3898 }
3899 geo->chunk_mask = chunk - 1;
3900 geo->chunk_shift = ffz(~chunk);
3901 return nc*fc;
3902}
3903
3904static void raid10_free_conf(struct r10conf *conf)
3905{
3906 if (!conf)
3907 return;
3908
3909 mempool_exit(&conf->r10bio_pool);
3910 kfree(conf->mirrors);
3911 kfree(conf->mirrors_old);
3912 kfree(conf->mirrors_new);
3913 safe_put_page(conf->tmppage);
3914 bioset_exit(&conf->bio_split);
3915 kfree(conf);
3916}
3917
3918static struct r10conf *setup_conf(struct mddev *mddev)
3919{
3920 struct r10conf *conf = NULL;
3921 int err = -EINVAL;
3922 struct geom geo;
3923 int copies;
3924
3925 copies = setup_geo(&geo, mddev, geo_new);
3926
3927 if (copies == -2) {
3928 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3929 mdname(mddev), PAGE_SIZE);
3930 goto out;
3931 }
3932
3933 if (copies < 2 || copies > mddev->raid_disks) {
3934 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3935 mdname(mddev), mddev->new_layout);
3936 goto out;
3937 }
3938
3939 err = -ENOMEM;
3940 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3941 if (!conf)
3942 goto out;
3943
3944 /* FIXME calc properly */
3945 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3946 sizeof(struct raid10_info),
3947 GFP_KERNEL);
3948 if (!conf->mirrors)
3949 goto out;
3950
3951 conf->tmppage = alloc_page(GFP_KERNEL);
3952 if (!conf->tmppage)
3953 goto out;
3954
3955 conf->geo = geo;
3956 conf->copies = copies;
3957 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3958 rbio_pool_free, conf);
3959 if (err)
3960 goto out;
3961
3962 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3963 if (err)
3964 goto out;
3965
3966 calc_sectors(conf, mddev->dev_sectors);
3967 if (mddev->reshape_position == MaxSector) {
3968 conf->prev = conf->geo;
3969 conf->reshape_progress = MaxSector;
3970 } else {
3971 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3972 err = -EINVAL;
3973 goto out;
3974 }
3975 conf->reshape_progress = mddev->reshape_position;
3976 if (conf->prev.far_offset)
3977 conf->prev.stride = 1 << conf->prev.chunk_shift;
3978 else
3979 /* far_copies must be 1 */
3980 conf->prev.stride = conf->dev_sectors;
3981 }
3982 conf->reshape_safe = conf->reshape_progress;
3983 spin_lock_init(&conf->device_lock);
3984 INIT_LIST_HEAD(&conf->retry_list);
3985 INIT_LIST_HEAD(&conf->bio_end_io_list);
3986
3987 seqlock_init(&conf->resync_lock);
3988 init_waitqueue_head(&conf->wait_barrier);
3989 atomic_set(&conf->nr_pending, 0);
3990
3991 err = -ENOMEM;
3992 rcu_assign_pointer(conf->thread,
3993 md_register_thread(raid10d, mddev, "raid10"));
3994 if (!conf->thread)
3995 goto out;
3996
3997 conf->mddev = mddev;
3998 return conf;
3999
4000 out:
4001 raid10_free_conf(conf);
4002 return ERR_PTR(err);
4003}
4004
4005static void raid10_set_io_opt(struct r10conf *conf)
4006{
4007 int raid_disks = conf->geo.raid_disks;
4008
4009 if (!(conf->geo.raid_disks % conf->geo.near_copies))
4010 raid_disks /= conf->geo.near_copies;
4011 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
4012 raid_disks);
4013}
4014
4015static int raid10_run(struct mddev *mddev)
4016{
4017 struct r10conf *conf;
4018 int i, disk_idx;
4019 struct raid10_info *disk;
4020 struct md_rdev *rdev;
4021 sector_t size;
4022 sector_t min_offset_diff = 0;
4023 int first = 1;
4024
4025 if (mddev->private == NULL) {
4026 conf = setup_conf(mddev);
4027 if (IS_ERR(conf))
4028 return PTR_ERR(conf);
4029 mddev->private = conf;
4030 }
4031 conf = mddev->private;
4032 if (!conf)
4033 goto out;
4034
4035 rcu_assign_pointer(mddev->thread, conf->thread);
4036 rcu_assign_pointer(conf->thread, NULL);
4037
4038 if (mddev_is_clustered(conf->mddev)) {
4039 int fc, fo;
4040
4041 fc = (mddev->layout >> 8) & 255;
4042 fo = mddev->layout & (1<<16);
4043 if (fc > 1 || fo > 0) {
4044 pr_err("only near layout is supported by clustered"
4045 " raid10\n");
4046 goto out_free_conf;
4047 }
4048 }
4049
4050 if (mddev->queue) {
4051 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
4052 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
4053 raid10_set_io_opt(conf);
4054 }
4055
4056 rdev_for_each(rdev, mddev) {
4057 long long diff;
4058
4059 disk_idx = rdev->raid_disk;
4060 if (disk_idx < 0)
4061 continue;
4062 if (disk_idx >= conf->geo.raid_disks &&
4063 disk_idx >= conf->prev.raid_disks)
4064 continue;
4065 disk = conf->mirrors + disk_idx;
4066
4067 if (test_bit(Replacement, &rdev->flags)) {
4068 if (disk->replacement)
4069 goto out_free_conf;
4070 disk->replacement = rdev;
4071 } else {
4072 if (disk->rdev)
4073 goto out_free_conf;
4074 disk->rdev = rdev;
4075 }
4076 diff = (rdev->new_data_offset - rdev->data_offset);
4077 if (!mddev->reshape_backwards)
4078 diff = -diff;
4079 if (diff < 0)
4080 diff = 0;
4081 if (first || diff < min_offset_diff)
4082 min_offset_diff = diff;
4083
4084 if (mddev->gendisk)
4085 disk_stack_limits(mddev->gendisk, rdev->bdev,
4086 rdev->data_offset << 9);
4087
4088 disk->head_position = 0;
4089 first = 0;
4090 }
4091
4092 /* need to check that every block has at least one working mirror */
4093 if (!enough(conf, -1)) {
4094 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4095 mdname(mddev));
4096 goto out_free_conf;
4097 }
4098
4099 if (conf->reshape_progress != MaxSector) {
4100 /* must ensure that shape change is supported */
4101 if (conf->geo.far_copies != 1 &&
4102 conf->geo.far_offset == 0)
4103 goto out_free_conf;
4104 if (conf->prev.far_copies != 1 &&
4105 conf->prev.far_offset == 0)
4106 goto out_free_conf;
4107 }
4108
4109 mddev->degraded = 0;
4110 for (i = 0;
4111 i < conf->geo.raid_disks
4112 || i < conf->prev.raid_disks;
4113 i++) {
4114
4115 disk = conf->mirrors + i;
4116
4117 if (!disk->rdev && disk->replacement) {
4118 /* The replacement is all we have - use it */
4119 disk->rdev = disk->replacement;
4120 disk->replacement = NULL;
4121 clear_bit(Replacement, &disk->rdev->flags);
4122 }
4123
4124 if (!disk->rdev ||
4125 !test_bit(In_sync, &disk->rdev->flags)) {
4126 disk->head_position = 0;
4127 mddev->degraded++;
4128 if (disk->rdev &&
4129 disk->rdev->saved_raid_disk < 0)
4130 conf->fullsync = 1;
4131 }
4132
4133 if (disk->replacement &&
4134 !test_bit(In_sync, &disk->replacement->flags) &&
4135 disk->replacement->saved_raid_disk < 0) {
4136 conf->fullsync = 1;
4137 }
4138
4139 disk->recovery_disabled = mddev->recovery_disabled - 1;
4140 }
4141
4142 if (mddev->recovery_cp != MaxSector)
4143 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4144 mdname(mddev));
4145 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4146 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4147 conf->geo.raid_disks);
4148 /*
4149 * Ok, everything is just fine now
4150 */
4151 mddev->dev_sectors = conf->dev_sectors;
4152 size = raid10_size(mddev, 0, 0);
4153 md_set_array_sectors(mddev, size);
4154 mddev->resync_max_sectors = size;
4155 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4156
4157 if (md_integrity_register(mddev))
4158 goto out_free_conf;
4159
4160 if (conf->reshape_progress != MaxSector) {
4161 unsigned long before_length, after_length;
4162
4163 before_length = ((1 << conf->prev.chunk_shift) *
4164 conf->prev.far_copies);
4165 after_length = ((1 << conf->geo.chunk_shift) *
4166 conf->geo.far_copies);
4167
4168 if (max(before_length, after_length) > min_offset_diff) {
4169 /* This cannot work */
4170 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4171 goto out_free_conf;
4172 }
4173 conf->offset_diff = min_offset_diff;
4174
4175 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4176 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4177 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4178 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4179 }
4180
4181 return 0;
4182
4183out_free_conf:
4184 md_unregister_thread(mddev, &mddev->thread);
4185 raid10_free_conf(conf);
4186 mddev->private = NULL;
4187out:
4188 return -EIO;
4189}
4190
4191static void raid10_free(struct mddev *mddev, void *priv)
4192{
4193 raid10_free_conf(priv);
4194}
4195
4196static void raid10_quiesce(struct mddev *mddev, int quiesce)
4197{
4198 struct r10conf *conf = mddev->private;
4199
4200 if (quiesce)
4201 raise_barrier(conf, 0);
4202 else
4203 lower_barrier(conf);
4204}
4205
4206static int raid10_resize(struct mddev *mddev, sector_t sectors)
4207{
4208 /* Resize of 'far' arrays is not supported.
4209 * For 'near' and 'offset' arrays we can set the
4210 * number of sectors used to be an appropriate multiple
4211 * of the chunk size.
4212 * For 'offset', this is far_copies*chunksize.
4213 * For 'near' the multiplier is the LCM of
4214 * near_copies and raid_disks.
4215 * So if far_copies > 1 && !far_offset, fail.
4216 * Else find LCM(raid_disks, near_copy)*far_copies and
4217 * multiply by chunk_size. Then round to this number.
4218 * This is mostly done by raid10_size()
4219 */
4220 struct r10conf *conf = mddev->private;
4221 sector_t oldsize, size;
4222
4223 if (mddev->reshape_position != MaxSector)
4224 return -EBUSY;
4225
4226 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4227 return -EINVAL;
4228
4229 oldsize = raid10_size(mddev, 0, 0);
4230 size = raid10_size(mddev, sectors, 0);
4231 if (mddev->external_size &&
4232 mddev->array_sectors > size)
4233 return -EINVAL;
4234 if (mddev->bitmap) {
4235 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4236 if (ret)
4237 return ret;
4238 }
4239 md_set_array_sectors(mddev, size);
4240 if (sectors > mddev->dev_sectors &&
4241 mddev->recovery_cp > oldsize) {
4242 mddev->recovery_cp = oldsize;
4243 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4244 }
4245 calc_sectors(conf, sectors);
4246 mddev->dev_sectors = conf->dev_sectors;
4247 mddev->resync_max_sectors = size;
4248 return 0;
4249}
4250
4251static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4252{
4253 struct md_rdev *rdev;
4254 struct r10conf *conf;
4255
4256 if (mddev->degraded > 0) {
4257 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4258 mdname(mddev));
4259 return ERR_PTR(-EINVAL);
4260 }
4261 sector_div(size, devs);
4262
4263 /* Set new parameters */
4264 mddev->new_level = 10;
4265 /* new layout: far_copies = 1, near_copies = 2 */
4266 mddev->new_layout = (1<<8) + 2;
4267 mddev->new_chunk_sectors = mddev->chunk_sectors;
4268 mddev->delta_disks = mddev->raid_disks;
4269 mddev->raid_disks *= 2;
4270 /* make sure it will be not marked as dirty */
4271 mddev->recovery_cp = MaxSector;
4272 mddev->dev_sectors = size;
4273
4274 conf = setup_conf(mddev);
4275 if (!IS_ERR(conf)) {
4276 rdev_for_each(rdev, mddev)
4277 if (rdev->raid_disk >= 0) {
4278 rdev->new_raid_disk = rdev->raid_disk * 2;
4279 rdev->sectors = size;
4280 }
4281 }
4282
4283 return conf;
4284}
4285
4286static void *raid10_takeover(struct mddev *mddev)
4287{
4288 struct r0conf *raid0_conf;
4289
4290 /* raid10 can take over:
4291 * raid0 - providing it has only two drives
4292 */
4293 if (mddev->level == 0) {
4294 /* for raid0 takeover only one zone is supported */
4295 raid0_conf = mddev->private;
4296 if (raid0_conf->nr_strip_zones > 1) {
4297 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4298 mdname(mddev));
4299 return ERR_PTR(-EINVAL);
4300 }
4301 return raid10_takeover_raid0(mddev,
4302 raid0_conf->strip_zone->zone_end,
4303 raid0_conf->strip_zone->nb_dev);
4304 }
4305 return ERR_PTR(-EINVAL);
4306}
4307
4308static int raid10_check_reshape(struct mddev *mddev)
4309{
4310 /* Called when there is a request to change
4311 * - layout (to ->new_layout)
4312 * - chunk size (to ->new_chunk_sectors)
4313 * - raid_disks (by delta_disks)
4314 * or when trying to restart a reshape that was ongoing.
4315 *
4316 * We need to validate the request and possibly allocate
4317 * space if that might be an issue later.
4318 *
4319 * Currently we reject any reshape of a 'far' mode array,
4320 * allow chunk size to change if new is generally acceptable,
4321 * allow raid_disks to increase, and allow
4322 * a switch between 'near' mode and 'offset' mode.
4323 */
4324 struct r10conf *conf = mddev->private;
4325 struct geom geo;
4326
4327 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4328 return -EINVAL;
4329
4330 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4331 /* mustn't change number of copies */
4332 return -EINVAL;
4333 if (geo.far_copies > 1 && !geo.far_offset)
4334 /* Cannot switch to 'far' mode */
4335 return -EINVAL;
4336
4337 if (mddev->array_sectors & geo.chunk_mask)
4338 /* not factor of array size */
4339 return -EINVAL;
4340
4341 if (!enough(conf, -1))
4342 return -EINVAL;
4343
4344 kfree(conf->mirrors_new);
4345 conf->mirrors_new = NULL;
4346 if (mddev->delta_disks > 0) {
4347 /* allocate new 'mirrors' list */
4348 conf->mirrors_new =
4349 kcalloc(mddev->raid_disks + mddev->delta_disks,
4350 sizeof(struct raid10_info),
4351 GFP_KERNEL);
4352 if (!conf->mirrors_new)
4353 return -ENOMEM;
4354 }
4355 return 0;
4356}
4357
4358/*
4359 * Need to check if array has failed when deciding whether to:
4360 * - start an array
4361 * - remove non-faulty devices
4362 * - add a spare
4363 * - allow a reshape
4364 * This determination is simple when no reshape is happening.
4365 * However if there is a reshape, we need to carefully check
4366 * both the before and after sections.
4367 * This is because some failed devices may only affect one
4368 * of the two sections, and some non-in_sync devices may
4369 * be insync in the section most affected by failed devices.
4370 */
4371static int calc_degraded(struct r10conf *conf)
4372{
4373 int degraded, degraded2;
4374 int i;
4375
4376 degraded = 0;
4377 /* 'prev' section first */
4378 for (i = 0; i < conf->prev.raid_disks; i++) {
4379 struct md_rdev *rdev = conf->mirrors[i].rdev;
4380
4381 if (!rdev || test_bit(Faulty, &rdev->flags))
4382 degraded++;
4383 else if (!test_bit(In_sync, &rdev->flags))
4384 /* When we can reduce the number of devices in
4385 * an array, this might not contribute to
4386 * 'degraded'. It does now.
4387 */
4388 degraded++;
4389 }
4390 if (conf->geo.raid_disks == conf->prev.raid_disks)
4391 return degraded;
4392 degraded2 = 0;
4393 for (i = 0; i < conf->geo.raid_disks; i++) {
4394 struct md_rdev *rdev = conf->mirrors[i].rdev;
4395
4396 if (!rdev || test_bit(Faulty, &rdev->flags))
4397 degraded2++;
4398 else if (!test_bit(In_sync, &rdev->flags)) {
4399 /* If reshape is increasing the number of devices,
4400 * this section has already been recovered, so
4401 * it doesn't contribute to degraded.
4402 * else it does.
4403 */
4404 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4405 degraded2++;
4406 }
4407 }
4408 if (degraded2 > degraded)
4409 return degraded2;
4410 return degraded;
4411}
4412
4413static int raid10_start_reshape(struct mddev *mddev)
4414{
4415 /* A 'reshape' has been requested. This commits
4416 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4417 * This also checks if there are enough spares and adds them
4418 * to the array.
4419 * We currently require enough spares to make the final
4420 * array non-degraded. We also require that the difference
4421 * between old and new data_offset - on each device - is
4422 * enough that we never risk over-writing.
4423 */
4424
4425 unsigned long before_length, after_length;
4426 sector_t min_offset_diff = 0;
4427 int first = 1;
4428 struct geom new;
4429 struct r10conf *conf = mddev->private;
4430 struct md_rdev *rdev;
4431 int spares = 0;
4432 int ret;
4433
4434 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4435 return -EBUSY;
4436
4437 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4438 return -EINVAL;
4439
4440 before_length = ((1 << conf->prev.chunk_shift) *
4441 conf->prev.far_copies);
4442 after_length = ((1 << conf->geo.chunk_shift) *
4443 conf->geo.far_copies);
4444
4445 rdev_for_each(rdev, mddev) {
4446 if (!test_bit(In_sync, &rdev->flags)
4447 && !test_bit(Faulty, &rdev->flags))
4448 spares++;
4449 if (rdev->raid_disk >= 0) {
4450 long long diff = (rdev->new_data_offset
4451 - rdev->data_offset);
4452 if (!mddev->reshape_backwards)
4453 diff = -diff;
4454 if (diff < 0)
4455 diff = 0;
4456 if (first || diff < min_offset_diff)
4457 min_offset_diff = diff;
4458 first = 0;
4459 }
4460 }
4461
4462 if (max(before_length, after_length) > min_offset_diff)
4463 return -EINVAL;
4464
4465 if (spares < mddev->delta_disks)
4466 return -EINVAL;
4467
4468 conf->offset_diff = min_offset_diff;
4469 spin_lock_irq(&conf->device_lock);
4470 if (conf->mirrors_new) {
4471 memcpy(conf->mirrors_new, conf->mirrors,
4472 sizeof(struct raid10_info)*conf->prev.raid_disks);
4473 smp_mb();
4474 kfree(conf->mirrors_old);
4475 conf->mirrors_old = conf->mirrors;
4476 conf->mirrors = conf->mirrors_new;
4477 conf->mirrors_new = NULL;
4478 }
4479 setup_geo(&conf->geo, mddev, geo_start);
4480 smp_mb();
4481 if (mddev->reshape_backwards) {
4482 sector_t size = raid10_size(mddev, 0, 0);
4483 if (size < mddev->array_sectors) {
4484 spin_unlock_irq(&conf->device_lock);
4485 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4486 mdname(mddev));
4487 return -EINVAL;
4488 }
4489 mddev->resync_max_sectors = size;
4490 conf->reshape_progress = size;
4491 } else
4492 conf->reshape_progress = 0;
4493 conf->reshape_safe = conf->reshape_progress;
4494 spin_unlock_irq(&conf->device_lock);
4495
4496 if (mddev->delta_disks && mddev->bitmap) {
4497 struct mdp_superblock_1 *sb = NULL;
4498 sector_t oldsize, newsize;
4499
4500 oldsize = raid10_size(mddev, 0, 0);
4501 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4502
4503 if (!mddev_is_clustered(mddev)) {
4504 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4505 if (ret)
4506 goto abort;
4507 else
4508 goto out;
4509 }
4510
4511 rdev_for_each(rdev, mddev) {
4512 if (rdev->raid_disk > -1 &&
4513 !test_bit(Faulty, &rdev->flags))
4514 sb = page_address(rdev->sb_page);
4515 }
4516
4517 /*
4518 * some node is already performing reshape, and no need to
4519 * call md_bitmap_resize again since it should be called when
4520 * receiving BITMAP_RESIZE msg
4521 */
4522 if ((sb && (le32_to_cpu(sb->feature_map) &
4523 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4524 goto out;
4525
4526 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4527 if (ret)
4528 goto abort;
4529
4530 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4531 if (ret) {
4532 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4533 goto abort;
4534 }
4535 }
4536out:
4537 if (mddev->delta_disks > 0) {
4538 rdev_for_each(rdev, mddev)
4539 if (rdev->raid_disk < 0 &&
4540 !test_bit(Faulty, &rdev->flags)) {
4541 if (raid10_add_disk(mddev, rdev) == 0) {
4542 if (rdev->raid_disk >=
4543 conf->prev.raid_disks)
4544 set_bit(In_sync, &rdev->flags);
4545 else
4546 rdev->recovery_offset = 0;
4547
4548 /* Failure here is OK */
4549 sysfs_link_rdev(mddev, rdev);
4550 }
4551 } else if (rdev->raid_disk >= conf->prev.raid_disks
4552 && !test_bit(Faulty, &rdev->flags)) {
4553 /* This is a spare that was manually added */
4554 set_bit(In_sync, &rdev->flags);
4555 }
4556 }
4557 /* When a reshape changes the number of devices,
4558 * ->degraded is measured against the larger of the
4559 * pre and post numbers.
4560 */
4561 spin_lock_irq(&conf->device_lock);
4562 mddev->degraded = calc_degraded(conf);
4563 spin_unlock_irq(&conf->device_lock);
4564 mddev->raid_disks = conf->geo.raid_disks;
4565 mddev->reshape_position = conf->reshape_progress;
4566 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4567
4568 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4569 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4570 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4571 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4572 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4573 conf->reshape_checkpoint = jiffies;
4574 md_new_event();
4575 return 0;
4576
4577abort:
4578 mddev->recovery = 0;
4579 spin_lock_irq(&conf->device_lock);
4580 conf->geo = conf->prev;
4581 mddev->raid_disks = conf->geo.raid_disks;
4582 rdev_for_each(rdev, mddev)
4583 rdev->new_data_offset = rdev->data_offset;
4584 smp_wmb();
4585 conf->reshape_progress = MaxSector;
4586 conf->reshape_safe = MaxSector;
4587 mddev->reshape_position = MaxSector;
4588 spin_unlock_irq(&conf->device_lock);
4589 return ret;
4590}
4591
4592/* Calculate the last device-address that could contain
4593 * any block from the chunk that includes the array-address 's'
4594 * and report the next address.
4595 * i.e. the address returned will be chunk-aligned and after
4596 * any data that is in the chunk containing 's'.
4597 */
4598static sector_t last_dev_address(sector_t s, struct geom *geo)
4599{
4600 s = (s | geo->chunk_mask) + 1;
4601 s >>= geo->chunk_shift;
4602 s *= geo->near_copies;
4603 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4604 s *= geo->far_copies;
4605 s <<= geo->chunk_shift;
4606 return s;
4607}
4608
4609/* Calculate the first device-address that could contain
4610 * any block from the chunk that includes the array-address 's'.
4611 * This too will be the start of a chunk
4612 */
4613static sector_t first_dev_address(sector_t s, struct geom *geo)
4614{
4615 s >>= geo->chunk_shift;
4616 s *= geo->near_copies;
4617 sector_div(s, geo->raid_disks);
4618 s *= geo->far_copies;
4619 s <<= geo->chunk_shift;
4620 return s;
4621}
4622
4623static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4624 int *skipped)
4625{
4626 /* We simply copy at most one chunk (smallest of old and new)
4627 * at a time, possibly less if that exceeds RESYNC_PAGES,
4628 * or we hit a bad block or something.
4629 * This might mean we pause for normal IO in the middle of
4630 * a chunk, but that is not a problem as mddev->reshape_position
4631 * can record any location.
4632 *
4633 * If we will want to write to a location that isn't
4634 * yet recorded as 'safe' (i.e. in metadata on disk) then
4635 * we need to flush all reshape requests and update the metadata.
4636 *
4637 * When reshaping forwards (e.g. to more devices), we interpret
4638 * 'safe' as the earliest block which might not have been copied
4639 * down yet. We divide this by previous stripe size and multiply
4640 * by previous stripe length to get lowest device offset that we
4641 * cannot write to yet.
4642 * We interpret 'sector_nr' as an address that we want to write to.
4643 * From this we use last_device_address() to find where we might
4644 * write to, and first_device_address on the 'safe' position.
4645 * If this 'next' write position is after the 'safe' position,
4646 * we must update the metadata to increase the 'safe' position.
4647 *
4648 * When reshaping backwards, we round in the opposite direction
4649 * and perform the reverse test: next write position must not be
4650 * less than current safe position.
4651 *
4652 * In all this the minimum difference in data offsets
4653 * (conf->offset_diff - always positive) allows a bit of slack,
4654 * so next can be after 'safe', but not by more than offset_diff
4655 *
4656 * We need to prepare all the bios here before we start any IO
4657 * to ensure the size we choose is acceptable to all devices.
4658 * The means one for each copy for write-out and an extra one for
4659 * read-in.
4660 * We store the read-in bio in ->master_bio and the others in
4661 * ->devs[x].bio and ->devs[x].repl_bio.
4662 */
4663 struct r10conf *conf = mddev->private;
4664 struct r10bio *r10_bio;
4665 sector_t next, safe, last;
4666 int max_sectors;
4667 int nr_sectors;
4668 int s;
4669 struct md_rdev *rdev;
4670 int need_flush = 0;
4671 struct bio *blist;
4672 struct bio *bio, *read_bio;
4673 int sectors_done = 0;
4674 struct page **pages;
4675
4676 if (sector_nr == 0) {
4677 /* If restarting in the middle, skip the initial sectors */
4678 if (mddev->reshape_backwards &&
4679 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4680 sector_nr = (raid10_size(mddev, 0, 0)
4681 - conf->reshape_progress);
4682 } else if (!mddev->reshape_backwards &&
4683 conf->reshape_progress > 0)
4684 sector_nr = conf->reshape_progress;
4685 if (sector_nr) {
4686 mddev->curr_resync_completed = sector_nr;
4687 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4688 *skipped = 1;
4689 return sector_nr;
4690 }
4691 }
4692
4693 /* We don't use sector_nr to track where we are up to
4694 * as that doesn't work well for ->reshape_backwards.
4695 * So just use ->reshape_progress.
4696 */
4697 if (mddev->reshape_backwards) {
4698 /* 'next' is the earliest device address that we might
4699 * write to for this chunk in the new layout
4700 */
4701 next = first_dev_address(conf->reshape_progress - 1,
4702 &conf->geo);
4703
4704 /* 'safe' is the last device address that we might read from
4705 * in the old layout after a restart
4706 */
4707 safe = last_dev_address(conf->reshape_safe - 1,
4708 &conf->prev);
4709
4710 if (next + conf->offset_diff < safe)
4711 need_flush = 1;
4712
4713 last = conf->reshape_progress - 1;
4714 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4715 & conf->prev.chunk_mask);
4716 if (sector_nr + RESYNC_SECTORS < last)
4717 sector_nr = last + 1 - RESYNC_SECTORS;
4718 } else {
4719 /* 'next' is after the last device address that we
4720 * might write to for this chunk in the new layout
4721 */
4722 next = last_dev_address(conf->reshape_progress, &conf->geo);
4723
4724 /* 'safe' is the earliest device address that we might
4725 * read from in the old layout after a restart
4726 */
4727 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4728
4729 /* Need to update metadata if 'next' might be beyond 'safe'
4730 * as that would possibly corrupt data
4731 */
4732 if (next > safe + conf->offset_diff)
4733 need_flush = 1;
4734
4735 sector_nr = conf->reshape_progress;
4736 last = sector_nr | (conf->geo.chunk_mask
4737 & conf->prev.chunk_mask);
4738
4739 if (sector_nr + RESYNC_SECTORS <= last)
4740 last = sector_nr + RESYNC_SECTORS - 1;
4741 }
4742
4743 if (need_flush ||
4744 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4745 /* Need to update reshape_position in metadata */
4746 wait_barrier(conf, false);
4747 mddev->reshape_position = conf->reshape_progress;
4748 if (mddev->reshape_backwards)
4749 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4750 - conf->reshape_progress;
4751 else
4752 mddev->curr_resync_completed = conf->reshape_progress;
4753 conf->reshape_checkpoint = jiffies;
4754 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4755 md_wakeup_thread(mddev->thread);
4756 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4757 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4758 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4759 allow_barrier(conf);
4760 return sectors_done;
4761 }
4762 conf->reshape_safe = mddev->reshape_position;
4763 allow_barrier(conf);
4764 }
4765
4766 raise_barrier(conf, 0);
4767read_more:
4768 /* Now schedule reads for blocks from sector_nr to last */
4769 r10_bio = raid10_alloc_init_r10buf(conf);
4770 r10_bio->state = 0;
4771 raise_barrier(conf, 1);
4772 atomic_set(&r10_bio->remaining, 0);
4773 r10_bio->mddev = mddev;
4774 r10_bio->sector = sector_nr;
4775 set_bit(R10BIO_IsReshape, &r10_bio->state);
4776 r10_bio->sectors = last - sector_nr + 1;
4777 rdev = read_balance(conf, r10_bio, &max_sectors);
4778 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4779
4780 if (!rdev) {
4781 /* Cannot read from here, so need to record bad blocks
4782 * on all the target devices.
4783 */
4784 // FIXME
4785 mempool_free(r10_bio, &conf->r10buf_pool);
4786 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4787 return sectors_done;
4788 }
4789
4790 read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4791 GFP_KERNEL, &mddev->bio_set);
4792 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4793 + rdev->data_offset);
4794 read_bio->bi_private = r10_bio;
4795 read_bio->bi_end_io = end_reshape_read;
4796 r10_bio->master_bio = read_bio;
4797 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4798
4799 /*
4800 * Broadcast RESYNC message to other nodes, so all nodes would not
4801 * write to the region to avoid conflict.
4802 */
4803 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4804 struct mdp_superblock_1 *sb = NULL;
4805 int sb_reshape_pos = 0;
4806
4807 conf->cluster_sync_low = sector_nr;
4808 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4809 sb = page_address(rdev->sb_page);
4810 if (sb) {
4811 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4812 /*
4813 * Set cluster_sync_low again if next address for array
4814 * reshape is less than cluster_sync_low. Since we can't
4815 * update cluster_sync_low until it has finished reshape.
4816 */
4817 if (sb_reshape_pos < conf->cluster_sync_low)
4818 conf->cluster_sync_low = sb_reshape_pos;
4819 }
4820
4821 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4822 conf->cluster_sync_high);
4823 }
4824
4825 /* Now find the locations in the new layout */
4826 __raid10_find_phys(&conf->geo, r10_bio);
4827
4828 blist = read_bio;
4829 read_bio->bi_next = NULL;
4830
4831 for (s = 0; s < conf->copies*2; s++) {
4832 struct bio *b;
4833 int d = r10_bio->devs[s/2].devnum;
4834 struct md_rdev *rdev2;
4835 if (s&1) {
4836 rdev2 = conf->mirrors[d].replacement;
4837 b = r10_bio->devs[s/2].repl_bio;
4838 } else {
4839 rdev2 = conf->mirrors[d].rdev;
4840 b = r10_bio->devs[s/2].bio;
4841 }
4842 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4843 continue;
4844
4845 bio_set_dev(b, rdev2->bdev);
4846 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4847 rdev2->new_data_offset;
4848 b->bi_end_io = end_reshape_write;
4849 b->bi_opf = REQ_OP_WRITE;
4850 b->bi_next = blist;
4851 blist = b;
4852 }
4853
4854 /* Now add as many pages as possible to all of these bios. */
4855
4856 nr_sectors = 0;
4857 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4858 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4859 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4860 int len = (max_sectors - s) << 9;
4861 if (len > PAGE_SIZE)
4862 len = PAGE_SIZE;
4863 for (bio = blist; bio ; bio = bio->bi_next) {
4864 if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
4865 bio->bi_status = BLK_STS_RESOURCE;
4866 bio_endio(bio);
4867 return sectors_done;
4868 }
4869 }
4870 sector_nr += len >> 9;
4871 nr_sectors += len >> 9;
4872 }
4873 r10_bio->sectors = nr_sectors;
4874
4875 /* Now submit the read */
4876 md_sync_acct_bio(read_bio, r10_bio->sectors);
4877 atomic_inc(&r10_bio->remaining);
4878 read_bio->bi_next = NULL;
4879 submit_bio_noacct(read_bio);
4880 sectors_done += nr_sectors;
4881 if (sector_nr <= last)
4882 goto read_more;
4883
4884 lower_barrier(conf);
4885
4886 /* Now that we have done the whole section we can
4887 * update reshape_progress
4888 */
4889 if (mddev->reshape_backwards)
4890 conf->reshape_progress -= sectors_done;
4891 else
4892 conf->reshape_progress += sectors_done;
4893
4894 return sectors_done;
4895}
4896
4897static void end_reshape_request(struct r10bio *r10_bio);
4898static int handle_reshape_read_error(struct mddev *mddev,
4899 struct r10bio *r10_bio);
4900static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4901{
4902 /* Reshape read completed. Hopefully we have a block
4903 * to write out.
4904 * If we got a read error then we do sync 1-page reads from
4905 * elsewhere until we find the data - or give up.
4906 */
4907 struct r10conf *conf = mddev->private;
4908 int s;
4909
4910 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4911 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4912 /* Reshape has been aborted */
4913 md_done_sync(mddev, r10_bio->sectors, 0);
4914 return;
4915 }
4916
4917 /* We definitely have the data in the pages, schedule the
4918 * writes.
4919 */
4920 atomic_set(&r10_bio->remaining, 1);
4921 for (s = 0; s < conf->copies*2; s++) {
4922 struct bio *b;
4923 int d = r10_bio->devs[s/2].devnum;
4924 struct md_rdev *rdev;
4925 if (s&1) {
4926 rdev = conf->mirrors[d].replacement;
4927 b = r10_bio->devs[s/2].repl_bio;
4928 } else {
4929 rdev = conf->mirrors[d].rdev;
4930 b = r10_bio->devs[s/2].bio;
4931 }
4932 if (!rdev || test_bit(Faulty, &rdev->flags))
4933 continue;
4934
4935 atomic_inc(&rdev->nr_pending);
4936 md_sync_acct_bio(b, r10_bio->sectors);
4937 atomic_inc(&r10_bio->remaining);
4938 b->bi_next = NULL;
4939 submit_bio_noacct(b);
4940 }
4941 end_reshape_request(r10_bio);
4942}
4943
4944static void end_reshape(struct r10conf *conf)
4945{
4946 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4947 return;
4948
4949 spin_lock_irq(&conf->device_lock);
4950 conf->prev = conf->geo;
4951 md_finish_reshape(conf->mddev);
4952 smp_wmb();
4953 conf->reshape_progress = MaxSector;
4954 conf->reshape_safe = MaxSector;
4955 spin_unlock_irq(&conf->device_lock);
4956
4957 if (conf->mddev->queue)
4958 raid10_set_io_opt(conf);
4959 conf->fullsync = 0;
4960}
4961
4962static void raid10_update_reshape_pos(struct mddev *mddev)
4963{
4964 struct r10conf *conf = mddev->private;
4965 sector_t lo, hi;
4966
4967 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4968 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4969 || mddev->reshape_position == MaxSector)
4970 conf->reshape_progress = mddev->reshape_position;
4971 else
4972 WARN_ON_ONCE(1);
4973}
4974
4975static int handle_reshape_read_error(struct mddev *mddev,
4976 struct r10bio *r10_bio)
4977{
4978 /* Use sync reads to get the blocks from somewhere else */
4979 int sectors = r10_bio->sectors;
4980 struct r10conf *conf = mddev->private;
4981 struct r10bio *r10b;
4982 int slot = 0;
4983 int idx = 0;
4984 struct page **pages;
4985
4986 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4987 if (!r10b) {
4988 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4989 return -ENOMEM;
4990 }
4991
4992 /* reshape IOs share pages from .devs[0].bio */
4993 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4994
4995 r10b->sector = r10_bio->sector;
4996 __raid10_find_phys(&conf->prev, r10b);
4997
4998 while (sectors) {
4999 int s = sectors;
5000 int success = 0;
5001 int first_slot = slot;
5002
5003 if (s > (PAGE_SIZE >> 9))
5004 s = PAGE_SIZE >> 9;
5005
5006 while (!success) {
5007 int d = r10b->devs[slot].devnum;
5008 struct md_rdev *rdev = conf->mirrors[d].rdev;
5009 sector_t addr;
5010 if (rdev == NULL ||
5011 test_bit(Faulty, &rdev->flags) ||
5012 !test_bit(In_sync, &rdev->flags))
5013 goto failed;
5014
5015 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5016 atomic_inc(&rdev->nr_pending);
5017 success = sync_page_io(rdev,
5018 addr,
5019 s << 9,
5020 pages[idx],
5021 REQ_OP_READ, false);
5022 rdev_dec_pending(rdev, mddev);
5023 if (success)
5024 break;
5025 failed:
5026 slot++;
5027 if (slot >= conf->copies)
5028 slot = 0;
5029 if (slot == first_slot)
5030 break;
5031 }
5032 if (!success) {
5033 /* couldn't read this block, must give up */
5034 set_bit(MD_RECOVERY_INTR,
5035 &mddev->recovery);
5036 kfree(r10b);
5037 return -EIO;
5038 }
5039 sectors -= s;
5040 idx++;
5041 }
5042 kfree(r10b);
5043 return 0;
5044}
5045
5046static void end_reshape_write(struct bio *bio)
5047{
5048 struct r10bio *r10_bio = get_resync_r10bio(bio);
5049 struct mddev *mddev = r10_bio->mddev;
5050 struct r10conf *conf = mddev->private;
5051 int d;
5052 int slot;
5053 int repl;
5054 struct md_rdev *rdev = NULL;
5055
5056 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5057 rdev = repl ? conf->mirrors[d].replacement :
5058 conf->mirrors[d].rdev;
5059
5060 if (bio->bi_status) {
5061 /* FIXME should record badblock */
5062 md_error(mddev, rdev);
5063 }
5064
5065 rdev_dec_pending(rdev, mddev);
5066 end_reshape_request(r10_bio);
5067}
5068
5069static void end_reshape_request(struct r10bio *r10_bio)
5070{
5071 if (!atomic_dec_and_test(&r10_bio->remaining))
5072 return;
5073 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5074 bio_put(r10_bio->master_bio);
5075 put_buf(r10_bio);
5076}
5077
5078static void raid10_finish_reshape(struct mddev *mddev)
5079{
5080 struct r10conf *conf = mddev->private;
5081
5082 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5083 return;
5084
5085 if (mddev->delta_disks > 0) {
5086 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5087 mddev->recovery_cp = mddev->resync_max_sectors;
5088 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5089 }
5090 mddev->resync_max_sectors = mddev->array_sectors;
5091 } else {
5092 int d;
5093 for (d = conf->geo.raid_disks ;
5094 d < conf->geo.raid_disks - mddev->delta_disks;
5095 d++) {
5096 struct md_rdev *rdev = conf->mirrors[d].rdev;
5097 if (rdev)
5098 clear_bit(In_sync, &rdev->flags);
5099 rdev = conf->mirrors[d].replacement;
5100 if (rdev)
5101 clear_bit(In_sync, &rdev->flags);
5102 }
5103 }
5104 mddev->layout = mddev->new_layout;
5105 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5106 mddev->reshape_position = MaxSector;
5107 mddev->delta_disks = 0;
5108 mddev->reshape_backwards = 0;
5109}
5110
5111static struct md_personality raid10_personality =
5112{
5113 .name = "raid10",
5114 .level = 10,
5115 .owner = THIS_MODULE,
5116 .make_request = raid10_make_request,
5117 .run = raid10_run,
5118 .free = raid10_free,
5119 .status = raid10_status,
5120 .error_handler = raid10_error,
5121 .hot_add_disk = raid10_add_disk,
5122 .hot_remove_disk= raid10_remove_disk,
5123 .spare_active = raid10_spare_active,
5124 .sync_request = raid10_sync_request,
5125 .quiesce = raid10_quiesce,
5126 .size = raid10_size,
5127 .resize = raid10_resize,
5128 .takeover = raid10_takeover,
5129 .check_reshape = raid10_check_reshape,
5130 .start_reshape = raid10_start_reshape,
5131 .finish_reshape = raid10_finish_reshape,
5132 .update_reshape_pos = raid10_update_reshape_pos,
5133};
5134
5135static int __init raid_init(void)
5136{
5137 return register_md_personality(&raid10_personality);
5138}
5139
5140static void raid_exit(void)
5141{
5142 unregister_md_personality(&raid10_personality);
5143}
5144
5145module_init(raid_init);
5146module_exit(raid_exit);
5147MODULE_LICENSE("GPL");
5148MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5149MODULE_ALIAS("md-personality-9"); /* RAID10 */
5150MODULE_ALIAS("md-raid10");
5151MODULE_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");