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#include "raid10.h"
23#include "raid0.h"
24#include "md-bitmap.h"
25
26/*
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
29 * chunk_size
30 * raid_disks
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
34 * use_far_sets (stored in bit 17 of layout )
35 * use_far_sets_bugfixed (stored in bit 18 of layout )
36 *
37 * The data to be stored is divided into chunks using chunksize. Each device
38 * is divided into far_copies sections. In each section, chunks are laid out
39 * in a style similar to raid0, but near_copies copies of each chunk is stored
40 * (each on a different drive). The starting device for each section is offset
41 * near_copies from the starting device of the previous section. Thus there
42 * are (near_copies * far_copies) of each chunk, and each is on a different
43 * drive. near_copies and far_copies must be at least one, and their product
44 * is at most raid_disks.
45 *
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of being very far
48 * apart on disk, there are adjacent stripes.
49 *
50 * The far and offset algorithms are handled slightly differently if
51 * 'use_far_sets' is true. In this case, the array's devices are grouped into
52 * sets that are (near_copies * far_copies) in size. The far copied stripes
53 * are still shifted by 'near_copies' devices, but this shifting stays confined
54 * to the set rather than the entire array. This is done to improve the number
55 * of device combinations that can fail without causing the array to fail.
56 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
57 * on a device):
58 * A B C D A B C D E
59 * ... ...
60 * D A B C E A B C D
61 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
62 * [A B] [C D] [A B] [C D E]
63 * |...| |...| |...| | ... |
64 * [B A] [D C] [B A] [E C D]
65 */
66
67static void allow_barrier(struct r10conf *conf);
68static void lower_barrier(struct r10conf *conf);
69static int _enough(struct r10conf *conf, int previous, int ignore);
70static int enough(struct r10conf *conf, int ignore);
71static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
72 int *skipped);
73static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
74static void end_reshape_write(struct bio *bio);
75static void end_reshape(struct r10conf *conf);
76
77#define raid10_log(md, fmt, args...) \
78 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
79
80#include "raid1-10.c"
81
82/*
83 * for resync bio, r10bio pointer can be retrieved from the per-bio
84 * 'struct resync_pages'.
85 */
86static inline struct r10bio *get_resync_r10bio(struct bio *bio)
87{
88 return get_resync_pages(bio)->raid_bio;
89}
90
91static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
92{
93 struct r10conf *conf = data;
94 int size = offsetof(struct r10bio, devs[conf->copies]);
95
96 /* allocate a r10bio with room for raid_disks entries in the
97 * bios array */
98 return kzalloc(size, gfp_flags);
99}
100
101#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
102/* amount of memory to reserve for resync requests */
103#define RESYNC_WINDOW (1024*1024)
104/* maximum number of concurrent requests, memory permitting */
105#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
106#define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
107#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
108
109/*
110 * When performing a resync, we need to read and compare, so
111 * we need as many pages are there are copies.
112 * When performing a recovery, we need 2 bios, one for read,
113 * one for write (we recover only one drive per r10buf)
114 *
115 */
116static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
117{
118 struct r10conf *conf = data;
119 struct r10bio *r10_bio;
120 struct bio *bio;
121 int j;
122 int nalloc, nalloc_rp;
123 struct resync_pages *rps;
124
125 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
126 if (!r10_bio)
127 return NULL;
128
129 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
130 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
131 nalloc = conf->copies; /* resync */
132 else
133 nalloc = 2; /* recovery */
134
135 /* allocate once for all bios */
136 if (!conf->have_replacement)
137 nalloc_rp = nalloc;
138 else
139 nalloc_rp = nalloc * 2;
140 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
141 if (!rps)
142 goto out_free_r10bio;
143
144 /*
145 * Allocate bios.
146 */
147 for (j = nalloc ; j-- ; ) {
148 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
149 if (!bio)
150 goto out_free_bio;
151 r10_bio->devs[j].bio = bio;
152 if (!conf->have_replacement)
153 continue;
154 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
155 if (!bio)
156 goto out_free_bio;
157 r10_bio->devs[j].repl_bio = bio;
158 }
159 /*
160 * Allocate RESYNC_PAGES data pages and attach them
161 * where needed.
162 */
163 for (j = 0; j < nalloc; j++) {
164 struct bio *rbio = r10_bio->devs[j].repl_bio;
165 struct resync_pages *rp, *rp_repl;
166
167 rp = &rps[j];
168 if (rbio)
169 rp_repl = &rps[nalloc + j];
170
171 bio = r10_bio->devs[j].bio;
172
173 if (!j || test_bit(MD_RECOVERY_SYNC,
174 &conf->mddev->recovery)) {
175 if (resync_alloc_pages(rp, gfp_flags))
176 goto out_free_pages;
177 } else {
178 memcpy(rp, &rps[0], sizeof(*rp));
179 resync_get_all_pages(rp);
180 }
181
182 rp->raid_bio = r10_bio;
183 bio->bi_private = rp;
184 if (rbio) {
185 memcpy(rp_repl, rp, sizeof(*rp));
186 rbio->bi_private = rp_repl;
187 }
188 }
189
190 return r10_bio;
191
192out_free_pages:
193 while (--j >= 0)
194 resync_free_pages(&rps[j * 2]);
195
196 j = 0;
197out_free_bio:
198 for ( ; j < nalloc; j++) {
199 if (r10_bio->devs[j].bio)
200 bio_put(r10_bio->devs[j].bio);
201 if (r10_bio->devs[j].repl_bio)
202 bio_put(r10_bio->devs[j].repl_bio);
203 }
204 kfree(rps);
205out_free_r10bio:
206 rbio_pool_free(r10_bio, conf);
207 return NULL;
208}
209
210static void r10buf_pool_free(void *__r10_bio, void *data)
211{
212 struct r10conf *conf = data;
213 struct r10bio *r10bio = __r10_bio;
214 int j;
215 struct resync_pages *rp = NULL;
216
217 for (j = conf->copies; j--; ) {
218 struct bio *bio = r10bio->devs[j].bio;
219
220 if (bio) {
221 rp = get_resync_pages(bio);
222 resync_free_pages(rp);
223 bio_put(bio);
224 }
225
226 bio = r10bio->devs[j].repl_bio;
227 if (bio)
228 bio_put(bio);
229 }
230
231 /* resync pages array stored in the 1st bio's .bi_private */
232 kfree(rp);
233
234 rbio_pool_free(r10bio, conf);
235}
236
237static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
238{
239 int i;
240
241 for (i = 0; i < conf->copies; i++) {
242 struct bio **bio = & r10_bio->devs[i].bio;
243 if (!BIO_SPECIAL(*bio))
244 bio_put(*bio);
245 *bio = NULL;
246 bio = &r10_bio->devs[i].repl_bio;
247 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
248 bio_put(*bio);
249 *bio = NULL;
250 }
251}
252
253static void free_r10bio(struct r10bio *r10_bio)
254{
255 struct r10conf *conf = r10_bio->mddev->private;
256
257 put_all_bios(conf, r10_bio);
258 mempool_free(r10_bio, &conf->r10bio_pool);
259}
260
261static void put_buf(struct r10bio *r10_bio)
262{
263 struct r10conf *conf = r10_bio->mddev->private;
264
265 mempool_free(r10_bio, &conf->r10buf_pool);
266
267 lower_barrier(conf);
268}
269
270static void reschedule_retry(struct r10bio *r10_bio)
271{
272 unsigned long flags;
273 struct mddev *mddev = r10_bio->mddev;
274 struct r10conf *conf = mddev->private;
275
276 spin_lock_irqsave(&conf->device_lock, flags);
277 list_add(&r10_bio->retry_list, &conf->retry_list);
278 conf->nr_queued ++;
279 spin_unlock_irqrestore(&conf->device_lock, flags);
280
281 /* wake up frozen array... */
282 wake_up(&conf->wait_barrier);
283
284 md_wakeup_thread(mddev->thread);
285}
286
287/*
288 * raid_end_bio_io() is called when we have finished servicing a mirrored
289 * operation and are ready to return a success/failure code to the buffer
290 * cache layer.
291 */
292static void raid_end_bio_io(struct r10bio *r10_bio)
293{
294 struct bio *bio = r10_bio->master_bio;
295 struct r10conf *conf = r10_bio->mddev->private;
296
297 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
298 bio->bi_status = BLK_STS_IOERR;
299
300 bio_endio(bio);
301 /*
302 * Wake up any possible resync thread that waits for the device
303 * to go idle.
304 */
305 allow_barrier(conf);
306
307 free_r10bio(r10_bio);
308}
309
310/*
311 * Update disk head position estimator based on IRQ completion info.
312 */
313static inline void update_head_pos(int slot, struct r10bio *r10_bio)
314{
315 struct r10conf *conf = r10_bio->mddev->private;
316
317 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
318 r10_bio->devs[slot].addr + (r10_bio->sectors);
319}
320
321/*
322 * Find the disk number which triggered given bio
323 */
324static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
325 struct bio *bio, int *slotp, int *replp)
326{
327 int slot;
328 int repl = 0;
329
330 for (slot = 0; slot < conf->copies; slot++) {
331 if (r10_bio->devs[slot].bio == bio)
332 break;
333 if (r10_bio->devs[slot].repl_bio == bio) {
334 repl = 1;
335 break;
336 }
337 }
338
339 BUG_ON(slot == conf->copies);
340 update_head_pos(slot, r10_bio);
341
342 if (slotp)
343 *slotp = slot;
344 if (replp)
345 *replp = repl;
346 return r10_bio->devs[slot].devnum;
347}
348
349static void raid10_end_read_request(struct bio *bio)
350{
351 int uptodate = !bio->bi_status;
352 struct r10bio *r10_bio = bio->bi_private;
353 int slot;
354 struct md_rdev *rdev;
355 struct r10conf *conf = r10_bio->mddev->private;
356
357 slot = r10_bio->read_slot;
358 rdev = r10_bio->devs[slot].rdev;
359 /*
360 * this branch is our 'one mirror IO has finished' event handler:
361 */
362 update_head_pos(slot, r10_bio);
363
364 if (uptodate) {
365 /*
366 * Set R10BIO_Uptodate in our master bio, so that
367 * we will return a good error code to the higher
368 * levels even if IO on some other mirrored buffer fails.
369 *
370 * The 'master' represents the composite IO operation to
371 * user-side. So if something waits for IO, then it will
372 * wait for the 'master' bio.
373 */
374 set_bit(R10BIO_Uptodate, &r10_bio->state);
375 } else {
376 /* If all other devices that store this block have
377 * failed, we want to return the error upwards rather
378 * than fail the last device. Here we redefine
379 * "uptodate" to mean "Don't want to retry"
380 */
381 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
382 rdev->raid_disk))
383 uptodate = 1;
384 }
385 if (uptodate) {
386 raid_end_bio_io(r10_bio);
387 rdev_dec_pending(rdev, conf->mddev);
388 } else {
389 /*
390 * oops, read error - keep the refcount on the rdev
391 */
392 char b[BDEVNAME_SIZE];
393 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
394 mdname(conf->mddev),
395 bdevname(rdev->bdev, b),
396 (unsigned long long)r10_bio->sector);
397 set_bit(R10BIO_ReadError, &r10_bio->state);
398 reschedule_retry(r10_bio);
399 }
400}
401
402static void close_write(struct r10bio *r10_bio)
403{
404 /* clear the bitmap if all writes complete successfully */
405 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
406 r10_bio->sectors,
407 !test_bit(R10BIO_Degraded, &r10_bio->state),
408 0);
409 md_write_end(r10_bio->mddev);
410}
411
412static void one_write_done(struct r10bio *r10_bio)
413{
414 if (atomic_dec_and_test(&r10_bio->remaining)) {
415 if (test_bit(R10BIO_WriteError, &r10_bio->state))
416 reschedule_retry(r10_bio);
417 else {
418 close_write(r10_bio);
419 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
420 reschedule_retry(r10_bio);
421 else
422 raid_end_bio_io(r10_bio);
423 }
424 }
425}
426
427static void raid10_end_write_request(struct bio *bio)
428{
429 struct r10bio *r10_bio = bio->bi_private;
430 int dev;
431 int dec_rdev = 1;
432 struct r10conf *conf = r10_bio->mddev->private;
433 int slot, repl;
434 struct md_rdev *rdev = NULL;
435 struct bio *to_put = NULL;
436 bool discard_error;
437
438 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
439
440 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
441
442 if (repl)
443 rdev = conf->mirrors[dev].replacement;
444 if (!rdev) {
445 smp_rmb();
446 repl = 0;
447 rdev = conf->mirrors[dev].rdev;
448 }
449 /*
450 * this branch is our 'one mirror IO has finished' event handler:
451 */
452 if (bio->bi_status && !discard_error) {
453 if (repl)
454 /* Never record new bad blocks to replacement,
455 * just fail it.
456 */
457 md_error(rdev->mddev, rdev);
458 else {
459 set_bit(WriteErrorSeen, &rdev->flags);
460 if (!test_and_set_bit(WantReplacement, &rdev->flags))
461 set_bit(MD_RECOVERY_NEEDED,
462 &rdev->mddev->recovery);
463
464 dec_rdev = 0;
465 if (test_bit(FailFast, &rdev->flags) &&
466 (bio->bi_opf & MD_FAILFAST)) {
467 md_error(rdev->mddev, rdev);
468 }
469
470 /*
471 * When the device is faulty, it is not necessary to
472 * handle write error.
473 * For failfast, this is the only remaining device,
474 * We need to retry the write without FailFast.
475 */
476 if (!test_bit(Faulty, &rdev->flags))
477 set_bit(R10BIO_WriteError, &r10_bio->state);
478 else {
479 r10_bio->devs[slot].bio = NULL;
480 to_put = bio;
481 dec_rdev = 1;
482 }
483 }
484 } else {
485 /*
486 * Set R10BIO_Uptodate in our master bio, so that
487 * we will return a good error code for to the higher
488 * levels even if IO on some other mirrored buffer fails.
489 *
490 * The 'master' represents the composite IO operation to
491 * user-side. So if something waits for IO, then it will
492 * wait for the 'master' bio.
493 */
494 sector_t first_bad;
495 int bad_sectors;
496
497 /*
498 * Do not set R10BIO_Uptodate if the current device is
499 * rebuilding or Faulty. This is because we cannot use
500 * such device for properly reading the data back (we could
501 * potentially use it, if the current write would have felt
502 * before rdev->recovery_offset, but for simplicity we don't
503 * check this here.
504 */
505 if (test_bit(In_sync, &rdev->flags) &&
506 !test_bit(Faulty, &rdev->flags))
507 set_bit(R10BIO_Uptodate, &r10_bio->state);
508
509 /* Maybe we can clear some bad blocks. */
510 if (is_badblock(rdev,
511 r10_bio->devs[slot].addr,
512 r10_bio->sectors,
513 &first_bad, &bad_sectors) && !discard_error) {
514 bio_put(bio);
515 if (repl)
516 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
517 else
518 r10_bio->devs[slot].bio = IO_MADE_GOOD;
519 dec_rdev = 0;
520 set_bit(R10BIO_MadeGood, &r10_bio->state);
521 }
522 }
523
524 /*
525 *
526 * Let's see if all mirrored write operations have finished
527 * already.
528 */
529 one_write_done(r10_bio);
530 if (dec_rdev)
531 rdev_dec_pending(rdev, conf->mddev);
532 if (to_put)
533 bio_put(to_put);
534}
535
536/*
537 * RAID10 layout manager
538 * As well as the chunksize and raid_disks count, there are two
539 * parameters: near_copies and far_copies.
540 * near_copies * far_copies must be <= raid_disks.
541 * Normally one of these will be 1.
542 * If both are 1, we get raid0.
543 * If near_copies == raid_disks, we get raid1.
544 *
545 * Chunks are laid out in raid0 style with near_copies copies of the
546 * first chunk, followed by near_copies copies of the next chunk and
547 * so on.
548 * If far_copies > 1, then after 1/far_copies of the array has been assigned
549 * as described above, we start again with a device offset of near_copies.
550 * So we effectively have another copy of the whole array further down all
551 * the drives, but with blocks on different drives.
552 * With this layout, and block is never stored twice on the one device.
553 *
554 * raid10_find_phys finds the sector offset of a given virtual sector
555 * on each device that it is on.
556 *
557 * raid10_find_virt does the reverse mapping, from a device and a
558 * sector offset to a virtual address
559 */
560
561static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
562{
563 int n,f;
564 sector_t sector;
565 sector_t chunk;
566 sector_t stripe;
567 int dev;
568 int slot = 0;
569 int last_far_set_start, last_far_set_size;
570
571 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
572 last_far_set_start *= geo->far_set_size;
573
574 last_far_set_size = geo->far_set_size;
575 last_far_set_size += (geo->raid_disks % geo->far_set_size);
576
577 /* now calculate first sector/dev */
578 chunk = r10bio->sector >> geo->chunk_shift;
579 sector = r10bio->sector & geo->chunk_mask;
580
581 chunk *= geo->near_copies;
582 stripe = chunk;
583 dev = sector_div(stripe, geo->raid_disks);
584 if (geo->far_offset)
585 stripe *= geo->far_copies;
586
587 sector += stripe << geo->chunk_shift;
588
589 /* and calculate all the others */
590 for (n = 0; n < geo->near_copies; n++) {
591 int d = dev;
592 int set;
593 sector_t s = sector;
594 r10bio->devs[slot].devnum = d;
595 r10bio->devs[slot].addr = s;
596 slot++;
597
598 for (f = 1; f < geo->far_copies; f++) {
599 set = d / geo->far_set_size;
600 d += geo->near_copies;
601
602 if ((geo->raid_disks % geo->far_set_size) &&
603 (d > last_far_set_start)) {
604 d -= last_far_set_start;
605 d %= last_far_set_size;
606 d += last_far_set_start;
607 } else {
608 d %= geo->far_set_size;
609 d += geo->far_set_size * set;
610 }
611 s += geo->stride;
612 r10bio->devs[slot].devnum = d;
613 r10bio->devs[slot].addr = s;
614 slot++;
615 }
616 dev++;
617 if (dev >= geo->raid_disks) {
618 dev = 0;
619 sector += (geo->chunk_mask + 1);
620 }
621 }
622}
623
624static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
625{
626 struct geom *geo = &conf->geo;
627
628 if (conf->reshape_progress != MaxSector &&
629 ((r10bio->sector >= conf->reshape_progress) !=
630 conf->mddev->reshape_backwards)) {
631 set_bit(R10BIO_Previous, &r10bio->state);
632 geo = &conf->prev;
633 } else
634 clear_bit(R10BIO_Previous, &r10bio->state);
635
636 __raid10_find_phys(geo, r10bio);
637}
638
639static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
640{
641 sector_t offset, chunk, vchunk;
642 /* Never use conf->prev as this is only called during resync
643 * or recovery, so reshape isn't happening
644 */
645 struct geom *geo = &conf->geo;
646 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
647 int far_set_size = geo->far_set_size;
648 int last_far_set_start;
649
650 if (geo->raid_disks % geo->far_set_size) {
651 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
652 last_far_set_start *= geo->far_set_size;
653
654 if (dev >= last_far_set_start) {
655 far_set_size = geo->far_set_size;
656 far_set_size += (geo->raid_disks % geo->far_set_size);
657 far_set_start = last_far_set_start;
658 }
659 }
660
661 offset = sector & geo->chunk_mask;
662 if (geo->far_offset) {
663 int fc;
664 chunk = sector >> geo->chunk_shift;
665 fc = sector_div(chunk, geo->far_copies);
666 dev -= fc * geo->near_copies;
667 if (dev < far_set_start)
668 dev += far_set_size;
669 } else {
670 while (sector >= geo->stride) {
671 sector -= geo->stride;
672 if (dev < (geo->near_copies + far_set_start))
673 dev += far_set_size - geo->near_copies;
674 else
675 dev -= geo->near_copies;
676 }
677 chunk = sector >> geo->chunk_shift;
678 }
679 vchunk = chunk * geo->raid_disks + dev;
680 sector_div(vchunk, geo->near_copies);
681 return (vchunk << geo->chunk_shift) + offset;
682}
683
684/*
685 * This routine returns the disk from which the requested read should
686 * be done. There is a per-array 'next expected sequential IO' sector
687 * number - if this matches on the next IO then we use the last disk.
688 * There is also a per-disk 'last know head position' sector that is
689 * maintained from IRQ contexts, both the normal and the resync IO
690 * completion handlers update this position correctly. If there is no
691 * perfect sequential match then we pick the disk whose head is closest.
692 *
693 * If there are 2 mirrors in the same 2 devices, performance degrades
694 * because position is mirror, not device based.
695 *
696 * The rdev for the device selected will have nr_pending incremented.
697 */
698
699/*
700 * FIXME: possibly should rethink readbalancing and do it differently
701 * depending on near_copies / far_copies geometry.
702 */
703static struct md_rdev *read_balance(struct r10conf *conf,
704 struct r10bio *r10_bio,
705 int *max_sectors)
706{
707 const sector_t this_sector = r10_bio->sector;
708 int disk, slot;
709 int sectors = r10_bio->sectors;
710 int best_good_sectors;
711 sector_t new_distance, best_dist;
712 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
713 int do_balance;
714 int best_dist_slot, best_pending_slot;
715 bool has_nonrot_disk = false;
716 unsigned int min_pending;
717 struct geom *geo = &conf->geo;
718
719 raid10_find_phys(conf, r10_bio);
720 rcu_read_lock();
721 best_dist_slot = -1;
722 min_pending = UINT_MAX;
723 best_dist_rdev = NULL;
724 best_pending_rdev = NULL;
725 best_dist = MaxSector;
726 best_good_sectors = 0;
727 do_balance = 1;
728 clear_bit(R10BIO_FailFast, &r10_bio->state);
729 /*
730 * Check if we can balance. We can balance on the whole
731 * device if no resync is going on (recovery is ok), or below
732 * the resync window. We take the first readable disk when
733 * above the resync window.
734 */
735 if ((conf->mddev->recovery_cp < MaxSector
736 && (this_sector + sectors >= conf->next_resync)) ||
737 (mddev_is_clustered(conf->mddev) &&
738 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
739 this_sector + sectors)))
740 do_balance = 0;
741
742 for (slot = 0; slot < conf->copies ; slot++) {
743 sector_t first_bad;
744 int bad_sectors;
745 sector_t dev_sector;
746 unsigned int pending;
747 bool nonrot;
748
749 if (r10_bio->devs[slot].bio == IO_BLOCKED)
750 continue;
751 disk = r10_bio->devs[slot].devnum;
752 rdev = rcu_dereference(conf->mirrors[disk].replacement);
753 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
754 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
755 rdev = rcu_dereference(conf->mirrors[disk].rdev);
756 if (rdev == NULL ||
757 test_bit(Faulty, &rdev->flags))
758 continue;
759 if (!test_bit(In_sync, &rdev->flags) &&
760 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
761 continue;
762
763 dev_sector = r10_bio->devs[slot].addr;
764 if (is_badblock(rdev, dev_sector, sectors,
765 &first_bad, &bad_sectors)) {
766 if (best_dist < MaxSector)
767 /* Already have a better slot */
768 continue;
769 if (first_bad <= dev_sector) {
770 /* Cannot read here. If this is the
771 * 'primary' device, then we must not read
772 * beyond 'bad_sectors' from another device.
773 */
774 bad_sectors -= (dev_sector - first_bad);
775 if (!do_balance && sectors > bad_sectors)
776 sectors = bad_sectors;
777 if (best_good_sectors > sectors)
778 best_good_sectors = sectors;
779 } else {
780 sector_t good_sectors =
781 first_bad - dev_sector;
782 if (good_sectors > best_good_sectors) {
783 best_good_sectors = good_sectors;
784 best_dist_slot = slot;
785 best_dist_rdev = rdev;
786 }
787 if (!do_balance)
788 /* Must read from here */
789 break;
790 }
791 continue;
792 } else
793 best_good_sectors = sectors;
794
795 if (!do_balance)
796 break;
797
798 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
799 has_nonrot_disk |= nonrot;
800 pending = atomic_read(&rdev->nr_pending);
801 if (min_pending > pending && nonrot) {
802 min_pending = pending;
803 best_pending_slot = slot;
804 best_pending_rdev = rdev;
805 }
806
807 if (best_dist_slot >= 0)
808 /* At least 2 disks to choose from so failfast is OK */
809 set_bit(R10BIO_FailFast, &r10_bio->state);
810 /* This optimisation is debatable, and completely destroys
811 * sequential read speed for 'far copies' arrays. So only
812 * keep it for 'near' arrays, and review those later.
813 */
814 if (geo->near_copies > 1 && !pending)
815 new_distance = 0;
816
817 /* for far > 1 always use the lowest address */
818 else if (geo->far_copies > 1)
819 new_distance = r10_bio->devs[slot].addr;
820 else
821 new_distance = abs(r10_bio->devs[slot].addr -
822 conf->mirrors[disk].head_position);
823
824 if (new_distance < best_dist) {
825 best_dist = new_distance;
826 best_dist_slot = slot;
827 best_dist_rdev = rdev;
828 }
829 }
830 if (slot >= conf->copies) {
831 if (has_nonrot_disk) {
832 slot = best_pending_slot;
833 rdev = best_pending_rdev;
834 } else {
835 slot = best_dist_slot;
836 rdev = best_dist_rdev;
837 }
838 }
839
840 if (slot >= 0) {
841 atomic_inc(&rdev->nr_pending);
842 r10_bio->read_slot = slot;
843 } else
844 rdev = NULL;
845 rcu_read_unlock();
846 *max_sectors = best_good_sectors;
847
848 return rdev;
849}
850
851static int raid10_congested(struct mddev *mddev, int bits)
852{
853 struct r10conf *conf = mddev->private;
854 int i, ret = 0;
855
856 if ((bits & (1 << WB_async_congested)) &&
857 conf->pending_count >= max_queued_requests)
858 return 1;
859
860 rcu_read_lock();
861 for (i = 0;
862 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
863 && ret == 0;
864 i++) {
865 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
866 if (rdev && !test_bit(Faulty, &rdev->flags)) {
867 struct request_queue *q = bdev_get_queue(rdev->bdev);
868
869 ret |= bdi_congested(q->backing_dev_info, bits);
870 }
871 }
872 rcu_read_unlock();
873 return ret;
874}
875
876static void flush_pending_writes(struct r10conf *conf)
877{
878 /* Any writes that have been queued but are awaiting
879 * bitmap updates get flushed here.
880 */
881 spin_lock_irq(&conf->device_lock);
882
883 if (conf->pending_bio_list.head) {
884 struct blk_plug plug;
885 struct bio *bio;
886
887 bio = bio_list_get(&conf->pending_bio_list);
888 conf->pending_count = 0;
889 spin_unlock_irq(&conf->device_lock);
890
891 /*
892 * As this is called in a wait_event() loop (see freeze_array),
893 * current->state might be TASK_UNINTERRUPTIBLE which will
894 * cause a warning when we prepare to wait again. As it is
895 * rare that this path is taken, it is perfectly safe to force
896 * us to go around the wait_event() loop again, so the warning
897 * is a false-positive. Silence the warning by resetting
898 * thread state
899 */
900 __set_current_state(TASK_RUNNING);
901
902 blk_start_plug(&plug);
903 /* flush any pending bitmap writes to disk
904 * before proceeding w/ I/O */
905 md_bitmap_unplug(conf->mddev->bitmap);
906 wake_up(&conf->wait_barrier);
907
908 while (bio) { /* submit pending writes */
909 struct bio *next = bio->bi_next;
910 struct md_rdev *rdev = (void*)bio->bi_disk;
911 bio->bi_next = NULL;
912 bio_set_dev(bio, rdev->bdev);
913 if (test_bit(Faulty, &rdev->flags)) {
914 bio_io_error(bio);
915 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
916 !blk_queue_discard(bio->bi_disk->queue)))
917 /* Just ignore it */
918 bio_endio(bio);
919 else
920 generic_make_request(bio);
921 bio = next;
922 }
923 blk_finish_plug(&plug);
924 } else
925 spin_unlock_irq(&conf->device_lock);
926}
927
928/* Barriers....
929 * Sometimes we need to suspend IO while we do something else,
930 * either some resync/recovery, or reconfigure the array.
931 * To do this we raise a 'barrier'.
932 * The 'barrier' is a counter that can be raised multiple times
933 * to count how many activities are happening which preclude
934 * normal IO.
935 * We can only raise the barrier if there is no pending IO.
936 * i.e. if nr_pending == 0.
937 * We choose only to raise the barrier if no-one is waiting for the
938 * barrier to go down. This means that as soon as an IO request
939 * is ready, no other operations which require a barrier will start
940 * until the IO request has had a chance.
941 *
942 * So: regular IO calls 'wait_barrier'. When that returns there
943 * is no backgroup IO happening, It must arrange to call
944 * allow_barrier when it has finished its IO.
945 * backgroup IO calls must call raise_barrier. Once that returns
946 * there is no normal IO happeing. It must arrange to call
947 * lower_barrier when the particular background IO completes.
948 */
949
950static void raise_barrier(struct r10conf *conf, int force)
951{
952 BUG_ON(force && !conf->barrier);
953 spin_lock_irq(&conf->resync_lock);
954
955 /* Wait until no block IO is waiting (unless 'force') */
956 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
957 conf->resync_lock);
958
959 /* block any new IO from starting */
960 conf->barrier++;
961
962 /* Now wait for all pending IO to complete */
963 wait_event_lock_irq(conf->wait_barrier,
964 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
965 conf->resync_lock);
966
967 spin_unlock_irq(&conf->resync_lock);
968}
969
970static void lower_barrier(struct r10conf *conf)
971{
972 unsigned long flags;
973 spin_lock_irqsave(&conf->resync_lock, flags);
974 conf->barrier--;
975 spin_unlock_irqrestore(&conf->resync_lock, flags);
976 wake_up(&conf->wait_barrier);
977}
978
979static void wait_barrier(struct r10conf *conf)
980{
981 spin_lock_irq(&conf->resync_lock);
982 if (conf->barrier) {
983 conf->nr_waiting++;
984 /* Wait for the barrier to drop.
985 * However if there are already pending
986 * requests (preventing the barrier from
987 * rising completely), and the
988 * pre-process bio queue isn't empty,
989 * then don't wait, as we need to empty
990 * that queue to get the nr_pending
991 * count down.
992 */
993 raid10_log(conf->mddev, "wait barrier");
994 wait_event_lock_irq(conf->wait_barrier,
995 !conf->barrier ||
996 (atomic_read(&conf->nr_pending) &&
997 current->bio_list &&
998 (!bio_list_empty(¤t->bio_list[0]) ||
999 !bio_list_empty(¤t->bio_list[1]))),
1000 conf->resync_lock);
1001 conf->nr_waiting--;
1002 if (!conf->nr_waiting)
1003 wake_up(&conf->wait_barrier);
1004 }
1005 atomic_inc(&conf->nr_pending);
1006 spin_unlock_irq(&conf->resync_lock);
1007}
1008
1009static void allow_barrier(struct r10conf *conf)
1010{
1011 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1012 (conf->array_freeze_pending))
1013 wake_up(&conf->wait_barrier);
1014}
1015
1016static void freeze_array(struct r10conf *conf, int extra)
1017{
1018 /* stop syncio and normal IO and wait for everything to
1019 * go quiet.
1020 * We increment barrier and nr_waiting, and then
1021 * wait until nr_pending match nr_queued+extra
1022 * This is called in the context of one normal IO request
1023 * that has failed. Thus any sync request that might be pending
1024 * will be blocked by nr_pending, and we need to wait for
1025 * pending IO requests to complete or be queued for re-try.
1026 * Thus the number queued (nr_queued) plus this request (extra)
1027 * must match the number of pending IOs (nr_pending) before
1028 * we continue.
1029 */
1030 spin_lock_irq(&conf->resync_lock);
1031 conf->array_freeze_pending++;
1032 conf->barrier++;
1033 conf->nr_waiting++;
1034 wait_event_lock_irq_cmd(conf->wait_barrier,
1035 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1036 conf->resync_lock,
1037 flush_pending_writes(conf));
1038
1039 conf->array_freeze_pending--;
1040 spin_unlock_irq(&conf->resync_lock);
1041}
1042
1043static void unfreeze_array(struct r10conf *conf)
1044{
1045 /* reverse the effect of the freeze */
1046 spin_lock_irq(&conf->resync_lock);
1047 conf->barrier--;
1048 conf->nr_waiting--;
1049 wake_up(&conf->wait_barrier);
1050 spin_unlock_irq(&conf->resync_lock);
1051}
1052
1053static sector_t choose_data_offset(struct r10bio *r10_bio,
1054 struct md_rdev *rdev)
1055{
1056 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1057 test_bit(R10BIO_Previous, &r10_bio->state))
1058 return rdev->data_offset;
1059 else
1060 return rdev->new_data_offset;
1061}
1062
1063struct raid10_plug_cb {
1064 struct blk_plug_cb cb;
1065 struct bio_list pending;
1066 int pending_cnt;
1067};
1068
1069static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1070{
1071 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1072 cb);
1073 struct mddev *mddev = plug->cb.data;
1074 struct r10conf *conf = mddev->private;
1075 struct bio *bio;
1076
1077 if (from_schedule || current->bio_list) {
1078 spin_lock_irq(&conf->device_lock);
1079 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1080 conf->pending_count += plug->pending_cnt;
1081 spin_unlock_irq(&conf->device_lock);
1082 wake_up(&conf->wait_barrier);
1083 md_wakeup_thread(mddev->thread);
1084 kfree(plug);
1085 return;
1086 }
1087
1088 /* we aren't scheduling, so we can do the write-out directly. */
1089 bio = bio_list_get(&plug->pending);
1090 md_bitmap_unplug(mddev->bitmap);
1091 wake_up(&conf->wait_barrier);
1092
1093 while (bio) { /* submit pending writes */
1094 struct bio *next = bio->bi_next;
1095 struct md_rdev *rdev = (void*)bio->bi_disk;
1096 bio->bi_next = NULL;
1097 bio_set_dev(bio, rdev->bdev);
1098 if (test_bit(Faulty, &rdev->flags)) {
1099 bio_io_error(bio);
1100 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1101 !blk_queue_discard(bio->bi_disk->queue)))
1102 /* Just ignore it */
1103 bio_endio(bio);
1104 else
1105 generic_make_request(bio);
1106 bio = next;
1107 }
1108 kfree(plug);
1109}
1110
1111/*
1112 * 1. Register the new request and wait if the reconstruction thread has put
1113 * up a bar for new requests. Continue immediately if no resync is active
1114 * currently.
1115 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1116 */
1117static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1118 struct bio *bio, sector_t sectors)
1119{
1120 wait_barrier(conf);
1121 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1122 bio->bi_iter.bi_sector < conf->reshape_progress &&
1123 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1124 raid10_log(conf->mddev, "wait reshape");
1125 allow_barrier(conf);
1126 wait_event(conf->wait_barrier,
1127 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1128 conf->reshape_progress >= bio->bi_iter.bi_sector +
1129 sectors);
1130 wait_barrier(conf);
1131 }
1132}
1133
1134static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1135 struct r10bio *r10_bio)
1136{
1137 struct r10conf *conf = mddev->private;
1138 struct bio *read_bio;
1139 const int op = bio_op(bio);
1140 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1141 int max_sectors;
1142 struct md_rdev *rdev;
1143 char b[BDEVNAME_SIZE];
1144 int slot = r10_bio->read_slot;
1145 struct md_rdev *err_rdev = NULL;
1146 gfp_t gfp = GFP_NOIO;
1147
1148 if (r10_bio->devs[slot].rdev) {
1149 /*
1150 * This is an error retry, but we cannot
1151 * safely dereference the rdev in the r10_bio,
1152 * we must use the one in conf.
1153 * If it has already been disconnected (unlikely)
1154 * we lose the device name in error messages.
1155 */
1156 int disk;
1157 /*
1158 * As we are blocking raid10, it is a little safer to
1159 * use __GFP_HIGH.
1160 */
1161 gfp = GFP_NOIO | __GFP_HIGH;
1162
1163 rcu_read_lock();
1164 disk = r10_bio->devs[slot].devnum;
1165 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1166 if (err_rdev)
1167 bdevname(err_rdev->bdev, b);
1168 else {
1169 strcpy(b, "???");
1170 /* This never gets dereferenced */
1171 err_rdev = r10_bio->devs[slot].rdev;
1172 }
1173 rcu_read_unlock();
1174 }
1175
1176 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1177 rdev = read_balance(conf, r10_bio, &max_sectors);
1178 if (!rdev) {
1179 if (err_rdev) {
1180 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1181 mdname(mddev), b,
1182 (unsigned long long)r10_bio->sector);
1183 }
1184 raid_end_bio_io(r10_bio);
1185 return;
1186 }
1187 if (err_rdev)
1188 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1189 mdname(mddev),
1190 bdevname(rdev->bdev, b),
1191 (unsigned long long)r10_bio->sector);
1192 if (max_sectors < bio_sectors(bio)) {
1193 struct bio *split = bio_split(bio, max_sectors,
1194 gfp, &conf->bio_split);
1195 bio_chain(split, bio);
1196 allow_barrier(conf);
1197 generic_make_request(bio);
1198 wait_barrier(conf);
1199 bio = split;
1200 r10_bio->master_bio = bio;
1201 r10_bio->sectors = max_sectors;
1202 }
1203 slot = r10_bio->read_slot;
1204
1205 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1206
1207 r10_bio->devs[slot].bio = read_bio;
1208 r10_bio->devs[slot].rdev = rdev;
1209
1210 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1211 choose_data_offset(r10_bio, rdev);
1212 bio_set_dev(read_bio, rdev->bdev);
1213 read_bio->bi_end_io = raid10_end_read_request;
1214 bio_set_op_attrs(read_bio, op, do_sync);
1215 if (test_bit(FailFast, &rdev->flags) &&
1216 test_bit(R10BIO_FailFast, &r10_bio->state))
1217 read_bio->bi_opf |= MD_FAILFAST;
1218 read_bio->bi_private = r10_bio;
1219
1220 if (mddev->gendisk)
1221 trace_block_bio_remap(read_bio->bi_disk->queue,
1222 read_bio, disk_devt(mddev->gendisk),
1223 r10_bio->sector);
1224 generic_make_request(read_bio);
1225 return;
1226}
1227
1228static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1229 struct bio *bio, bool replacement,
1230 int n_copy)
1231{
1232 const int op = bio_op(bio);
1233 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1234 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1235 unsigned long flags;
1236 struct blk_plug_cb *cb;
1237 struct raid10_plug_cb *plug = NULL;
1238 struct r10conf *conf = mddev->private;
1239 struct md_rdev *rdev;
1240 int devnum = r10_bio->devs[n_copy].devnum;
1241 struct bio *mbio;
1242
1243 if (replacement) {
1244 rdev = conf->mirrors[devnum].replacement;
1245 if (rdev == NULL) {
1246 /* Replacement just got moved to main 'rdev' */
1247 smp_mb();
1248 rdev = conf->mirrors[devnum].rdev;
1249 }
1250 } else
1251 rdev = conf->mirrors[devnum].rdev;
1252
1253 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1254 if (replacement)
1255 r10_bio->devs[n_copy].repl_bio = mbio;
1256 else
1257 r10_bio->devs[n_copy].bio = mbio;
1258
1259 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1260 choose_data_offset(r10_bio, rdev));
1261 bio_set_dev(mbio, rdev->bdev);
1262 mbio->bi_end_io = raid10_end_write_request;
1263 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1264 if (!replacement && test_bit(FailFast,
1265 &conf->mirrors[devnum].rdev->flags)
1266 && enough(conf, devnum))
1267 mbio->bi_opf |= MD_FAILFAST;
1268 mbio->bi_private = r10_bio;
1269
1270 if (conf->mddev->gendisk)
1271 trace_block_bio_remap(mbio->bi_disk->queue,
1272 mbio, disk_devt(conf->mddev->gendisk),
1273 r10_bio->sector);
1274 /* flush_pending_writes() needs access to the rdev so...*/
1275 mbio->bi_disk = (void *)rdev;
1276
1277 atomic_inc(&r10_bio->remaining);
1278
1279 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1280 if (cb)
1281 plug = container_of(cb, struct raid10_plug_cb, cb);
1282 else
1283 plug = NULL;
1284 if (plug) {
1285 bio_list_add(&plug->pending, mbio);
1286 plug->pending_cnt++;
1287 } else {
1288 spin_lock_irqsave(&conf->device_lock, flags);
1289 bio_list_add(&conf->pending_bio_list, mbio);
1290 conf->pending_count++;
1291 spin_unlock_irqrestore(&conf->device_lock, flags);
1292 md_wakeup_thread(mddev->thread);
1293 }
1294}
1295
1296static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1297 struct r10bio *r10_bio)
1298{
1299 struct r10conf *conf = mddev->private;
1300 int i;
1301 struct md_rdev *blocked_rdev;
1302 sector_t sectors;
1303 int max_sectors;
1304
1305 if ((mddev_is_clustered(mddev) &&
1306 md_cluster_ops->area_resyncing(mddev, WRITE,
1307 bio->bi_iter.bi_sector,
1308 bio_end_sector(bio)))) {
1309 DEFINE_WAIT(w);
1310 for (;;) {
1311 prepare_to_wait(&conf->wait_barrier,
1312 &w, TASK_IDLE);
1313 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1314 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1315 break;
1316 schedule();
1317 }
1318 finish_wait(&conf->wait_barrier, &w);
1319 }
1320
1321 sectors = r10_bio->sectors;
1322 regular_request_wait(mddev, conf, bio, sectors);
1323 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1324 (mddev->reshape_backwards
1325 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1326 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1327 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1328 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1329 /* Need to update reshape_position in metadata */
1330 mddev->reshape_position = conf->reshape_progress;
1331 set_mask_bits(&mddev->sb_flags, 0,
1332 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1333 md_wakeup_thread(mddev->thread);
1334 raid10_log(conf->mddev, "wait reshape metadata");
1335 wait_event(mddev->sb_wait,
1336 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1337
1338 conf->reshape_safe = mddev->reshape_position;
1339 }
1340
1341 if (conf->pending_count >= max_queued_requests) {
1342 md_wakeup_thread(mddev->thread);
1343 raid10_log(mddev, "wait queued");
1344 wait_event(conf->wait_barrier,
1345 conf->pending_count < max_queued_requests);
1346 }
1347 /* first select target devices under rcu_lock and
1348 * inc refcount on their rdev. Record them by setting
1349 * bios[x] to bio
1350 * If there are known/acknowledged bad blocks on any device
1351 * on which we have seen a write error, we want to avoid
1352 * writing to those blocks. This potentially requires several
1353 * writes to write around the bad blocks. Each set of writes
1354 * gets its own r10_bio with a set of bios attached.
1355 */
1356
1357 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1358 raid10_find_phys(conf, r10_bio);
1359retry_write:
1360 blocked_rdev = NULL;
1361 rcu_read_lock();
1362 max_sectors = r10_bio->sectors;
1363
1364 for (i = 0; i < conf->copies; i++) {
1365 int d = r10_bio->devs[i].devnum;
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1367 struct md_rdev *rrdev = rcu_dereference(
1368 conf->mirrors[d].replacement);
1369 if (rdev == rrdev)
1370 rrdev = NULL;
1371 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1372 atomic_inc(&rdev->nr_pending);
1373 blocked_rdev = rdev;
1374 break;
1375 }
1376 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1377 atomic_inc(&rrdev->nr_pending);
1378 blocked_rdev = rrdev;
1379 break;
1380 }
1381 if (rdev && (test_bit(Faulty, &rdev->flags)))
1382 rdev = NULL;
1383 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1384 rrdev = NULL;
1385
1386 r10_bio->devs[i].bio = NULL;
1387 r10_bio->devs[i].repl_bio = NULL;
1388
1389 if (!rdev && !rrdev) {
1390 set_bit(R10BIO_Degraded, &r10_bio->state);
1391 continue;
1392 }
1393 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1394 sector_t first_bad;
1395 sector_t dev_sector = r10_bio->devs[i].addr;
1396 int bad_sectors;
1397 int is_bad;
1398
1399 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1400 &first_bad, &bad_sectors);
1401 if (is_bad < 0) {
1402 /* Mustn't write here until the bad block
1403 * is acknowledged
1404 */
1405 atomic_inc(&rdev->nr_pending);
1406 set_bit(BlockedBadBlocks, &rdev->flags);
1407 blocked_rdev = rdev;
1408 break;
1409 }
1410 if (is_bad && first_bad <= dev_sector) {
1411 /* Cannot write here at all */
1412 bad_sectors -= (dev_sector - first_bad);
1413 if (bad_sectors < max_sectors)
1414 /* Mustn't write more than bad_sectors
1415 * to other devices yet
1416 */
1417 max_sectors = bad_sectors;
1418 /* We don't set R10BIO_Degraded as that
1419 * only applies if the disk is missing,
1420 * so it might be re-added, and we want to
1421 * know to recover this chunk.
1422 * In this case the device is here, and the
1423 * fact that this chunk is not in-sync is
1424 * recorded in the bad block log.
1425 */
1426 continue;
1427 }
1428 if (is_bad) {
1429 int good_sectors = first_bad - dev_sector;
1430 if (good_sectors < max_sectors)
1431 max_sectors = good_sectors;
1432 }
1433 }
1434 if (rdev) {
1435 r10_bio->devs[i].bio = bio;
1436 atomic_inc(&rdev->nr_pending);
1437 }
1438 if (rrdev) {
1439 r10_bio->devs[i].repl_bio = bio;
1440 atomic_inc(&rrdev->nr_pending);
1441 }
1442 }
1443 rcu_read_unlock();
1444
1445 if (unlikely(blocked_rdev)) {
1446 /* Have to wait for this device to get unblocked, then retry */
1447 int j;
1448 int d;
1449
1450 for (j = 0; j < i; j++) {
1451 if (r10_bio->devs[j].bio) {
1452 d = r10_bio->devs[j].devnum;
1453 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1454 }
1455 if (r10_bio->devs[j].repl_bio) {
1456 struct md_rdev *rdev;
1457 d = r10_bio->devs[j].devnum;
1458 rdev = conf->mirrors[d].replacement;
1459 if (!rdev) {
1460 /* Race with remove_disk */
1461 smp_mb();
1462 rdev = conf->mirrors[d].rdev;
1463 }
1464 rdev_dec_pending(rdev, mddev);
1465 }
1466 }
1467 allow_barrier(conf);
1468 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1469 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1470 wait_barrier(conf);
1471 goto retry_write;
1472 }
1473
1474 if (max_sectors < r10_bio->sectors)
1475 r10_bio->sectors = max_sectors;
1476
1477 if (r10_bio->sectors < bio_sectors(bio)) {
1478 struct bio *split = bio_split(bio, r10_bio->sectors,
1479 GFP_NOIO, &conf->bio_split);
1480 bio_chain(split, bio);
1481 allow_barrier(conf);
1482 generic_make_request(bio);
1483 wait_barrier(conf);
1484 bio = split;
1485 r10_bio->master_bio = bio;
1486 }
1487
1488 atomic_set(&r10_bio->remaining, 1);
1489 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1490
1491 for (i = 0; i < conf->copies; i++) {
1492 if (r10_bio->devs[i].bio)
1493 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1494 if (r10_bio->devs[i].repl_bio)
1495 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1496 }
1497 one_write_done(r10_bio);
1498}
1499
1500static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1501{
1502 struct r10conf *conf = mddev->private;
1503 struct r10bio *r10_bio;
1504
1505 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1506
1507 r10_bio->master_bio = bio;
1508 r10_bio->sectors = sectors;
1509
1510 r10_bio->mddev = mddev;
1511 r10_bio->sector = bio->bi_iter.bi_sector;
1512 r10_bio->state = 0;
1513 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1514
1515 if (bio_data_dir(bio) == READ)
1516 raid10_read_request(mddev, bio, r10_bio);
1517 else
1518 raid10_write_request(mddev, bio, r10_bio);
1519}
1520
1521static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1522{
1523 struct r10conf *conf = mddev->private;
1524 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1525 int chunk_sects = chunk_mask + 1;
1526 int sectors = bio_sectors(bio);
1527
1528 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1529 md_flush_request(mddev, bio);
1530 return true;
1531 }
1532
1533 if (!md_write_start(mddev, bio))
1534 return false;
1535
1536 /*
1537 * If this request crosses a chunk boundary, we need to split
1538 * it.
1539 */
1540 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1541 sectors > chunk_sects
1542 && (conf->geo.near_copies < conf->geo.raid_disks
1543 || conf->prev.near_copies <
1544 conf->prev.raid_disks)))
1545 sectors = chunk_sects -
1546 (bio->bi_iter.bi_sector &
1547 (chunk_sects - 1));
1548 __make_request(mddev, bio, sectors);
1549
1550 /* In case raid10d snuck in to freeze_array */
1551 wake_up(&conf->wait_barrier);
1552 return true;
1553}
1554
1555static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1556{
1557 struct r10conf *conf = mddev->private;
1558 int i;
1559
1560 if (conf->geo.near_copies < conf->geo.raid_disks)
1561 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1562 if (conf->geo.near_copies > 1)
1563 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1564 if (conf->geo.far_copies > 1) {
1565 if (conf->geo.far_offset)
1566 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1567 else
1568 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1569 if (conf->geo.far_set_size != conf->geo.raid_disks)
1570 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1571 }
1572 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1573 conf->geo.raid_disks - mddev->degraded);
1574 rcu_read_lock();
1575 for (i = 0; i < conf->geo.raid_disks; i++) {
1576 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1577 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1578 }
1579 rcu_read_unlock();
1580 seq_printf(seq, "]");
1581}
1582
1583/* check if there are enough drives for
1584 * every block to appear on atleast one.
1585 * Don't consider the device numbered 'ignore'
1586 * as we might be about to remove it.
1587 */
1588static int _enough(struct r10conf *conf, int previous, int ignore)
1589{
1590 int first = 0;
1591 int has_enough = 0;
1592 int disks, ncopies;
1593 if (previous) {
1594 disks = conf->prev.raid_disks;
1595 ncopies = conf->prev.near_copies;
1596 } else {
1597 disks = conf->geo.raid_disks;
1598 ncopies = conf->geo.near_copies;
1599 }
1600
1601 rcu_read_lock();
1602 do {
1603 int n = conf->copies;
1604 int cnt = 0;
1605 int this = first;
1606 while (n--) {
1607 struct md_rdev *rdev;
1608 if (this != ignore &&
1609 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1610 test_bit(In_sync, &rdev->flags))
1611 cnt++;
1612 this = (this+1) % disks;
1613 }
1614 if (cnt == 0)
1615 goto out;
1616 first = (first + ncopies) % disks;
1617 } while (first != 0);
1618 has_enough = 1;
1619out:
1620 rcu_read_unlock();
1621 return has_enough;
1622}
1623
1624static int enough(struct r10conf *conf, int ignore)
1625{
1626 /* when calling 'enough', both 'prev' and 'geo' must
1627 * be stable.
1628 * This is ensured if ->reconfig_mutex or ->device_lock
1629 * is held.
1630 */
1631 return _enough(conf, 0, ignore) &&
1632 _enough(conf, 1, ignore);
1633}
1634
1635static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1636{
1637 char b[BDEVNAME_SIZE];
1638 struct r10conf *conf = mddev->private;
1639 unsigned long flags;
1640
1641 /*
1642 * If it is not operational, then we have already marked it as dead
1643 * else if it is the last working disks with "fail_last_dev == false",
1644 * ignore the error, let the next level up know.
1645 * else mark the drive as failed
1646 */
1647 spin_lock_irqsave(&conf->device_lock, flags);
1648 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1649 && !enough(conf, rdev->raid_disk)) {
1650 /*
1651 * Don't fail the drive, just return an IO error.
1652 */
1653 spin_unlock_irqrestore(&conf->device_lock, flags);
1654 return;
1655 }
1656 if (test_and_clear_bit(In_sync, &rdev->flags))
1657 mddev->degraded++;
1658 /*
1659 * If recovery is running, make sure it aborts.
1660 */
1661 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1662 set_bit(Blocked, &rdev->flags);
1663 set_bit(Faulty, &rdev->flags);
1664 set_mask_bits(&mddev->sb_flags, 0,
1665 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1666 spin_unlock_irqrestore(&conf->device_lock, flags);
1667 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1668 "md/raid10:%s: Operation continuing on %d devices.\n",
1669 mdname(mddev), bdevname(rdev->bdev, b),
1670 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1671}
1672
1673static void print_conf(struct r10conf *conf)
1674{
1675 int i;
1676 struct md_rdev *rdev;
1677
1678 pr_debug("RAID10 conf printout:\n");
1679 if (!conf) {
1680 pr_debug("(!conf)\n");
1681 return;
1682 }
1683 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1684 conf->geo.raid_disks);
1685
1686 /* This is only called with ->reconfix_mutex held, so
1687 * rcu protection of rdev is not needed */
1688 for (i = 0; i < conf->geo.raid_disks; i++) {
1689 char b[BDEVNAME_SIZE];
1690 rdev = conf->mirrors[i].rdev;
1691 if (rdev)
1692 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1693 i, !test_bit(In_sync, &rdev->flags),
1694 !test_bit(Faulty, &rdev->flags),
1695 bdevname(rdev->bdev,b));
1696 }
1697}
1698
1699static void close_sync(struct r10conf *conf)
1700{
1701 wait_barrier(conf);
1702 allow_barrier(conf);
1703
1704 mempool_exit(&conf->r10buf_pool);
1705}
1706
1707static int raid10_spare_active(struct mddev *mddev)
1708{
1709 int i;
1710 struct r10conf *conf = mddev->private;
1711 struct raid10_info *tmp;
1712 int count = 0;
1713 unsigned long flags;
1714
1715 /*
1716 * Find all non-in_sync disks within the RAID10 configuration
1717 * and mark them in_sync
1718 */
1719 for (i = 0; i < conf->geo.raid_disks; i++) {
1720 tmp = conf->mirrors + i;
1721 if (tmp->replacement
1722 && tmp->replacement->recovery_offset == MaxSector
1723 && !test_bit(Faulty, &tmp->replacement->flags)
1724 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1725 /* Replacement has just become active */
1726 if (!tmp->rdev
1727 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1728 count++;
1729 if (tmp->rdev) {
1730 /* Replaced device not technically faulty,
1731 * but we need to be sure it gets removed
1732 * and never re-added.
1733 */
1734 set_bit(Faulty, &tmp->rdev->flags);
1735 sysfs_notify_dirent_safe(
1736 tmp->rdev->sysfs_state);
1737 }
1738 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1739 } else if (tmp->rdev
1740 && tmp->rdev->recovery_offset == MaxSector
1741 && !test_bit(Faulty, &tmp->rdev->flags)
1742 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1743 count++;
1744 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1745 }
1746 }
1747 spin_lock_irqsave(&conf->device_lock, flags);
1748 mddev->degraded -= count;
1749 spin_unlock_irqrestore(&conf->device_lock, flags);
1750
1751 print_conf(conf);
1752 return count;
1753}
1754
1755static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1756{
1757 struct r10conf *conf = mddev->private;
1758 int err = -EEXIST;
1759 int mirror;
1760 int first = 0;
1761 int last = conf->geo.raid_disks - 1;
1762
1763 if (mddev->recovery_cp < MaxSector)
1764 /* only hot-add to in-sync arrays, as recovery is
1765 * very different from resync
1766 */
1767 return -EBUSY;
1768 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1769 return -EINVAL;
1770
1771 if (md_integrity_add_rdev(rdev, mddev))
1772 return -ENXIO;
1773
1774 if (rdev->raid_disk >= 0)
1775 first = last = rdev->raid_disk;
1776
1777 if (rdev->saved_raid_disk >= first &&
1778 rdev->saved_raid_disk < conf->geo.raid_disks &&
1779 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1780 mirror = rdev->saved_raid_disk;
1781 else
1782 mirror = first;
1783 for ( ; mirror <= last ; mirror++) {
1784 struct raid10_info *p = &conf->mirrors[mirror];
1785 if (p->recovery_disabled == mddev->recovery_disabled)
1786 continue;
1787 if (p->rdev) {
1788 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1789 p->replacement != NULL)
1790 continue;
1791 clear_bit(In_sync, &rdev->flags);
1792 set_bit(Replacement, &rdev->flags);
1793 rdev->raid_disk = mirror;
1794 err = 0;
1795 if (mddev->gendisk)
1796 disk_stack_limits(mddev->gendisk, rdev->bdev,
1797 rdev->data_offset << 9);
1798 conf->fullsync = 1;
1799 rcu_assign_pointer(p->replacement, rdev);
1800 break;
1801 }
1802
1803 if (mddev->gendisk)
1804 disk_stack_limits(mddev->gendisk, rdev->bdev,
1805 rdev->data_offset << 9);
1806
1807 p->head_position = 0;
1808 p->recovery_disabled = mddev->recovery_disabled - 1;
1809 rdev->raid_disk = mirror;
1810 err = 0;
1811 if (rdev->saved_raid_disk != mirror)
1812 conf->fullsync = 1;
1813 rcu_assign_pointer(p->rdev, rdev);
1814 break;
1815 }
1816 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1817 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1818
1819 print_conf(conf);
1820 return err;
1821}
1822
1823static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1824{
1825 struct r10conf *conf = mddev->private;
1826 int err = 0;
1827 int number = rdev->raid_disk;
1828 struct md_rdev **rdevp;
1829 struct raid10_info *p = conf->mirrors + number;
1830
1831 print_conf(conf);
1832 if (rdev == p->rdev)
1833 rdevp = &p->rdev;
1834 else if (rdev == p->replacement)
1835 rdevp = &p->replacement;
1836 else
1837 return 0;
1838
1839 if (test_bit(In_sync, &rdev->flags) ||
1840 atomic_read(&rdev->nr_pending)) {
1841 err = -EBUSY;
1842 goto abort;
1843 }
1844 /* Only remove non-faulty devices if recovery
1845 * is not possible.
1846 */
1847 if (!test_bit(Faulty, &rdev->flags) &&
1848 mddev->recovery_disabled != p->recovery_disabled &&
1849 (!p->replacement || p->replacement == rdev) &&
1850 number < conf->geo.raid_disks &&
1851 enough(conf, -1)) {
1852 err = -EBUSY;
1853 goto abort;
1854 }
1855 *rdevp = NULL;
1856 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1857 synchronize_rcu();
1858 if (atomic_read(&rdev->nr_pending)) {
1859 /* lost the race, try later */
1860 err = -EBUSY;
1861 *rdevp = rdev;
1862 goto abort;
1863 }
1864 }
1865 if (p->replacement) {
1866 /* We must have just cleared 'rdev' */
1867 p->rdev = p->replacement;
1868 clear_bit(Replacement, &p->replacement->flags);
1869 smp_mb(); /* Make sure other CPUs may see both as identical
1870 * but will never see neither -- if they are careful.
1871 */
1872 p->replacement = NULL;
1873 }
1874
1875 clear_bit(WantReplacement, &rdev->flags);
1876 err = md_integrity_register(mddev);
1877
1878abort:
1879
1880 print_conf(conf);
1881 return err;
1882}
1883
1884static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1885{
1886 struct r10conf *conf = r10_bio->mddev->private;
1887
1888 if (!bio->bi_status)
1889 set_bit(R10BIO_Uptodate, &r10_bio->state);
1890 else
1891 /* The write handler will notice the lack of
1892 * R10BIO_Uptodate and record any errors etc
1893 */
1894 atomic_add(r10_bio->sectors,
1895 &conf->mirrors[d].rdev->corrected_errors);
1896
1897 /* for reconstruct, we always reschedule after a read.
1898 * for resync, only after all reads
1899 */
1900 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1901 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1902 atomic_dec_and_test(&r10_bio->remaining)) {
1903 /* we have read all the blocks,
1904 * do the comparison in process context in raid10d
1905 */
1906 reschedule_retry(r10_bio);
1907 }
1908}
1909
1910static void end_sync_read(struct bio *bio)
1911{
1912 struct r10bio *r10_bio = get_resync_r10bio(bio);
1913 struct r10conf *conf = r10_bio->mddev->private;
1914 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1915
1916 __end_sync_read(r10_bio, bio, d);
1917}
1918
1919static void end_reshape_read(struct bio *bio)
1920{
1921 /* reshape read bio isn't allocated from r10buf_pool */
1922 struct r10bio *r10_bio = bio->bi_private;
1923
1924 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1925}
1926
1927static void end_sync_request(struct r10bio *r10_bio)
1928{
1929 struct mddev *mddev = r10_bio->mddev;
1930
1931 while (atomic_dec_and_test(&r10_bio->remaining)) {
1932 if (r10_bio->master_bio == NULL) {
1933 /* the primary of several recovery bios */
1934 sector_t s = r10_bio->sectors;
1935 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1936 test_bit(R10BIO_WriteError, &r10_bio->state))
1937 reschedule_retry(r10_bio);
1938 else
1939 put_buf(r10_bio);
1940 md_done_sync(mddev, s, 1);
1941 break;
1942 } else {
1943 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1944 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1945 test_bit(R10BIO_WriteError, &r10_bio->state))
1946 reschedule_retry(r10_bio);
1947 else
1948 put_buf(r10_bio);
1949 r10_bio = r10_bio2;
1950 }
1951 }
1952}
1953
1954static void end_sync_write(struct bio *bio)
1955{
1956 struct r10bio *r10_bio = get_resync_r10bio(bio);
1957 struct mddev *mddev = r10_bio->mddev;
1958 struct r10conf *conf = mddev->private;
1959 int d;
1960 sector_t first_bad;
1961 int bad_sectors;
1962 int slot;
1963 int repl;
1964 struct md_rdev *rdev = NULL;
1965
1966 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1967 if (repl)
1968 rdev = conf->mirrors[d].replacement;
1969 else
1970 rdev = conf->mirrors[d].rdev;
1971
1972 if (bio->bi_status) {
1973 if (repl)
1974 md_error(mddev, rdev);
1975 else {
1976 set_bit(WriteErrorSeen, &rdev->flags);
1977 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1978 set_bit(MD_RECOVERY_NEEDED,
1979 &rdev->mddev->recovery);
1980 set_bit(R10BIO_WriteError, &r10_bio->state);
1981 }
1982 } else if (is_badblock(rdev,
1983 r10_bio->devs[slot].addr,
1984 r10_bio->sectors,
1985 &first_bad, &bad_sectors))
1986 set_bit(R10BIO_MadeGood, &r10_bio->state);
1987
1988 rdev_dec_pending(rdev, mddev);
1989
1990 end_sync_request(r10_bio);
1991}
1992
1993/*
1994 * Note: sync and recover and handled very differently for raid10
1995 * This code is for resync.
1996 * For resync, we read through virtual addresses and read all blocks.
1997 * If there is any error, we schedule a write. The lowest numbered
1998 * drive is authoritative.
1999 * However requests come for physical address, so we need to map.
2000 * For every physical address there are raid_disks/copies virtual addresses,
2001 * which is always are least one, but is not necessarly an integer.
2002 * This means that a physical address can span multiple chunks, so we may
2003 * have to submit multiple io requests for a single sync request.
2004 */
2005/*
2006 * We check if all blocks are in-sync and only write to blocks that
2007 * aren't in sync
2008 */
2009static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2010{
2011 struct r10conf *conf = mddev->private;
2012 int i, first;
2013 struct bio *tbio, *fbio;
2014 int vcnt;
2015 struct page **tpages, **fpages;
2016
2017 atomic_set(&r10_bio->remaining, 1);
2018
2019 /* find the first device with a block */
2020 for (i=0; i<conf->copies; i++)
2021 if (!r10_bio->devs[i].bio->bi_status)
2022 break;
2023
2024 if (i == conf->copies)
2025 goto done;
2026
2027 first = i;
2028 fbio = r10_bio->devs[i].bio;
2029 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2030 fbio->bi_iter.bi_idx = 0;
2031 fpages = get_resync_pages(fbio)->pages;
2032
2033 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2034 /* now find blocks with errors */
2035 for (i=0 ; i < conf->copies ; i++) {
2036 int j, d;
2037 struct md_rdev *rdev;
2038 struct resync_pages *rp;
2039
2040 tbio = r10_bio->devs[i].bio;
2041
2042 if (tbio->bi_end_io != end_sync_read)
2043 continue;
2044 if (i == first)
2045 continue;
2046
2047 tpages = get_resync_pages(tbio)->pages;
2048 d = r10_bio->devs[i].devnum;
2049 rdev = conf->mirrors[d].rdev;
2050 if (!r10_bio->devs[i].bio->bi_status) {
2051 /* We know that the bi_io_vec layout is the same for
2052 * both 'first' and 'i', so we just compare them.
2053 * All vec entries are PAGE_SIZE;
2054 */
2055 int sectors = r10_bio->sectors;
2056 for (j = 0; j < vcnt; j++) {
2057 int len = PAGE_SIZE;
2058 if (sectors < (len / 512))
2059 len = sectors * 512;
2060 if (memcmp(page_address(fpages[j]),
2061 page_address(tpages[j]),
2062 len))
2063 break;
2064 sectors -= len/512;
2065 }
2066 if (j == vcnt)
2067 continue;
2068 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2069 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2070 /* Don't fix anything. */
2071 continue;
2072 } else if (test_bit(FailFast, &rdev->flags)) {
2073 /* Just give up on this device */
2074 md_error(rdev->mddev, rdev);
2075 continue;
2076 }
2077 /* Ok, we need to write this bio, either to correct an
2078 * inconsistency or to correct an unreadable block.
2079 * First we need to fixup bv_offset, bv_len and
2080 * bi_vecs, as the read request might have corrupted these
2081 */
2082 rp = get_resync_pages(tbio);
2083 bio_reset(tbio);
2084
2085 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2086
2087 rp->raid_bio = r10_bio;
2088 tbio->bi_private = rp;
2089 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2090 tbio->bi_end_io = end_sync_write;
2091 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2092
2093 bio_copy_data(tbio, fbio);
2094
2095 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2096 atomic_inc(&r10_bio->remaining);
2097 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2098
2099 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2100 tbio->bi_opf |= MD_FAILFAST;
2101 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2102 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2103 generic_make_request(tbio);
2104 }
2105
2106 /* Now write out to any replacement devices
2107 * that are active
2108 */
2109 for (i = 0; i < conf->copies; i++) {
2110 int d;
2111
2112 tbio = r10_bio->devs[i].repl_bio;
2113 if (!tbio || !tbio->bi_end_io)
2114 continue;
2115 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2116 && r10_bio->devs[i].bio != fbio)
2117 bio_copy_data(tbio, fbio);
2118 d = r10_bio->devs[i].devnum;
2119 atomic_inc(&r10_bio->remaining);
2120 md_sync_acct(conf->mirrors[d].replacement->bdev,
2121 bio_sectors(tbio));
2122 generic_make_request(tbio);
2123 }
2124
2125done:
2126 if (atomic_dec_and_test(&r10_bio->remaining)) {
2127 md_done_sync(mddev, r10_bio->sectors, 1);
2128 put_buf(r10_bio);
2129 }
2130}
2131
2132/*
2133 * Now for the recovery code.
2134 * Recovery happens across physical sectors.
2135 * We recover all non-is_sync drives by finding the virtual address of
2136 * each, and then choose a working drive that also has that virt address.
2137 * There is a separate r10_bio for each non-in_sync drive.
2138 * Only the first two slots are in use. The first for reading,
2139 * The second for writing.
2140 *
2141 */
2142static void fix_recovery_read_error(struct r10bio *r10_bio)
2143{
2144 /* We got a read error during recovery.
2145 * We repeat the read in smaller page-sized sections.
2146 * If a read succeeds, write it to the new device or record
2147 * a bad block if we cannot.
2148 * If a read fails, record a bad block on both old and
2149 * new devices.
2150 */
2151 struct mddev *mddev = r10_bio->mddev;
2152 struct r10conf *conf = mddev->private;
2153 struct bio *bio = r10_bio->devs[0].bio;
2154 sector_t sect = 0;
2155 int sectors = r10_bio->sectors;
2156 int idx = 0;
2157 int dr = r10_bio->devs[0].devnum;
2158 int dw = r10_bio->devs[1].devnum;
2159 struct page **pages = get_resync_pages(bio)->pages;
2160
2161 while (sectors) {
2162 int s = sectors;
2163 struct md_rdev *rdev;
2164 sector_t addr;
2165 int ok;
2166
2167 if (s > (PAGE_SIZE>>9))
2168 s = PAGE_SIZE >> 9;
2169
2170 rdev = conf->mirrors[dr].rdev;
2171 addr = r10_bio->devs[0].addr + sect,
2172 ok = sync_page_io(rdev,
2173 addr,
2174 s << 9,
2175 pages[idx],
2176 REQ_OP_READ, 0, false);
2177 if (ok) {
2178 rdev = conf->mirrors[dw].rdev;
2179 addr = r10_bio->devs[1].addr + sect;
2180 ok = sync_page_io(rdev,
2181 addr,
2182 s << 9,
2183 pages[idx],
2184 REQ_OP_WRITE, 0, false);
2185 if (!ok) {
2186 set_bit(WriteErrorSeen, &rdev->flags);
2187 if (!test_and_set_bit(WantReplacement,
2188 &rdev->flags))
2189 set_bit(MD_RECOVERY_NEEDED,
2190 &rdev->mddev->recovery);
2191 }
2192 }
2193 if (!ok) {
2194 /* We don't worry if we cannot set a bad block -
2195 * it really is bad so there is no loss in not
2196 * recording it yet
2197 */
2198 rdev_set_badblocks(rdev, addr, s, 0);
2199
2200 if (rdev != conf->mirrors[dw].rdev) {
2201 /* need bad block on destination too */
2202 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2203 addr = r10_bio->devs[1].addr + sect;
2204 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2205 if (!ok) {
2206 /* just abort the recovery */
2207 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2208 mdname(mddev));
2209
2210 conf->mirrors[dw].recovery_disabled
2211 = mddev->recovery_disabled;
2212 set_bit(MD_RECOVERY_INTR,
2213 &mddev->recovery);
2214 break;
2215 }
2216 }
2217 }
2218
2219 sectors -= s;
2220 sect += s;
2221 idx++;
2222 }
2223}
2224
2225static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2226{
2227 struct r10conf *conf = mddev->private;
2228 int d;
2229 struct bio *wbio, *wbio2;
2230
2231 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2232 fix_recovery_read_error(r10_bio);
2233 end_sync_request(r10_bio);
2234 return;
2235 }
2236
2237 /*
2238 * share the pages with the first bio
2239 * and submit the write request
2240 */
2241 d = r10_bio->devs[1].devnum;
2242 wbio = r10_bio->devs[1].bio;
2243 wbio2 = r10_bio->devs[1].repl_bio;
2244 /* Need to test wbio2->bi_end_io before we call
2245 * generic_make_request as if the former is NULL,
2246 * the latter is free to free wbio2.
2247 */
2248 if (wbio2 && !wbio2->bi_end_io)
2249 wbio2 = NULL;
2250 if (wbio->bi_end_io) {
2251 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2252 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2253 generic_make_request(wbio);
2254 }
2255 if (wbio2) {
2256 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2257 md_sync_acct(conf->mirrors[d].replacement->bdev,
2258 bio_sectors(wbio2));
2259 generic_make_request(wbio2);
2260 }
2261}
2262
2263/*
2264 * Used by fix_read_error() to decay the per rdev read_errors.
2265 * We halve the read error count for every hour that has elapsed
2266 * since the last recorded read error.
2267 *
2268 */
2269static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2270{
2271 long cur_time_mon;
2272 unsigned long hours_since_last;
2273 unsigned int read_errors = atomic_read(&rdev->read_errors);
2274
2275 cur_time_mon = ktime_get_seconds();
2276
2277 if (rdev->last_read_error == 0) {
2278 /* first time we've seen a read error */
2279 rdev->last_read_error = cur_time_mon;
2280 return;
2281 }
2282
2283 hours_since_last = (long)(cur_time_mon -
2284 rdev->last_read_error) / 3600;
2285
2286 rdev->last_read_error = cur_time_mon;
2287
2288 /*
2289 * if hours_since_last is > the number of bits in read_errors
2290 * just set read errors to 0. We do this to avoid
2291 * overflowing the shift of read_errors by hours_since_last.
2292 */
2293 if (hours_since_last >= 8 * sizeof(read_errors))
2294 atomic_set(&rdev->read_errors, 0);
2295 else
2296 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2297}
2298
2299static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2300 int sectors, struct page *page, int rw)
2301{
2302 sector_t first_bad;
2303 int bad_sectors;
2304
2305 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2306 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2307 return -1;
2308 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2309 /* success */
2310 return 1;
2311 if (rw == WRITE) {
2312 set_bit(WriteErrorSeen, &rdev->flags);
2313 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2314 set_bit(MD_RECOVERY_NEEDED,
2315 &rdev->mddev->recovery);
2316 }
2317 /* need to record an error - either for the block or the device */
2318 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2319 md_error(rdev->mddev, rdev);
2320 return 0;
2321}
2322
2323/*
2324 * This is a kernel thread which:
2325 *
2326 * 1. Retries failed read operations on working mirrors.
2327 * 2. Updates the raid superblock when problems encounter.
2328 * 3. Performs writes following reads for array synchronising.
2329 */
2330
2331static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2332{
2333 int sect = 0; /* Offset from r10_bio->sector */
2334 int sectors = r10_bio->sectors;
2335 struct md_rdev *rdev;
2336 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2337 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2338
2339 /* still own a reference to this rdev, so it cannot
2340 * have been cleared recently.
2341 */
2342 rdev = conf->mirrors[d].rdev;
2343
2344 if (test_bit(Faulty, &rdev->flags))
2345 /* drive has already been failed, just ignore any
2346 more fix_read_error() attempts */
2347 return;
2348
2349 check_decay_read_errors(mddev, rdev);
2350 atomic_inc(&rdev->read_errors);
2351 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2352 char b[BDEVNAME_SIZE];
2353 bdevname(rdev->bdev, b);
2354
2355 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2356 mdname(mddev), b,
2357 atomic_read(&rdev->read_errors), max_read_errors);
2358 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2359 mdname(mddev), b);
2360 md_error(mddev, rdev);
2361 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2362 return;
2363 }
2364
2365 while(sectors) {
2366 int s = sectors;
2367 int sl = r10_bio->read_slot;
2368 int success = 0;
2369 int start;
2370
2371 if (s > (PAGE_SIZE>>9))
2372 s = PAGE_SIZE >> 9;
2373
2374 rcu_read_lock();
2375 do {
2376 sector_t first_bad;
2377 int bad_sectors;
2378
2379 d = r10_bio->devs[sl].devnum;
2380 rdev = rcu_dereference(conf->mirrors[d].rdev);
2381 if (rdev &&
2382 test_bit(In_sync, &rdev->flags) &&
2383 !test_bit(Faulty, &rdev->flags) &&
2384 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2385 &first_bad, &bad_sectors) == 0) {
2386 atomic_inc(&rdev->nr_pending);
2387 rcu_read_unlock();
2388 success = sync_page_io(rdev,
2389 r10_bio->devs[sl].addr +
2390 sect,
2391 s<<9,
2392 conf->tmppage,
2393 REQ_OP_READ, 0, false);
2394 rdev_dec_pending(rdev, mddev);
2395 rcu_read_lock();
2396 if (success)
2397 break;
2398 }
2399 sl++;
2400 if (sl == conf->copies)
2401 sl = 0;
2402 } while (!success && sl != r10_bio->read_slot);
2403 rcu_read_unlock();
2404
2405 if (!success) {
2406 /* Cannot read from anywhere, just mark the block
2407 * as bad on the first device to discourage future
2408 * reads.
2409 */
2410 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2411 rdev = conf->mirrors[dn].rdev;
2412
2413 if (!rdev_set_badblocks(
2414 rdev,
2415 r10_bio->devs[r10_bio->read_slot].addr
2416 + sect,
2417 s, 0)) {
2418 md_error(mddev, rdev);
2419 r10_bio->devs[r10_bio->read_slot].bio
2420 = IO_BLOCKED;
2421 }
2422 break;
2423 }
2424
2425 start = sl;
2426 /* write it back and re-read */
2427 rcu_read_lock();
2428 while (sl != r10_bio->read_slot) {
2429 char b[BDEVNAME_SIZE];
2430
2431 if (sl==0)
2432 sl = conf->copies;
2433 sl--;
2434 d = r10_bio->devs[sl].devnum;
2435 rdev = rcu_dereference(conf->mirrors[d].rdev);
2436 if (!rdev ||
2437 test_bit(Faulty, &rdev->flags) ||
2438 !test_bit(In_sync, &rdev->flags))
2439 continue;
2440
2441 atomic_inc(&rdev->nr_pending);
2442 rcu_read_unlock();
2443 if (r10_sync_page_io(rdev,
2444 r10_bio->devs[sl].addr +
2445 sect,
2446 s, conf->tmppage, WRITE)
2447 == 0) {
2448 /* Well, this device is dead */
2449 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2450 mdname(mddev), s,
2451 (unsigned long long)(
2452 sect +
2453 choose_data_offset(r10_bio,
2454 rdev)),
2455 bdevname(rdev->bdev, b));
2456 pr_notice("md/raid10:%s: %s: failing drive\n",
2457 mdname(mddev),
2458 bdevname(rdev->bdev, b));
2459 }
2460 rdev_dec_pending(rdev, mddev);
2461 rcu_read_lock();
2462 }
2463 sl = start;
2464 while (sl != r10_bio->read_slot) {
2465 char b[BDEVNAME_SIZE];
2466
2467 if (sl==0)
2468 sl = conf->copies;
2469 sl--;
2470 d = r10_bio->devs[sl].devnum;
2471 rdev = rcu_dereference(conf->mirrors[d].rdev);
2472 if (!rdev ||
2473 test_bit(Faulty, &rdev->flags) ||
2474 !test_bit(In_sync, &rdev->flags))
2475 continue;
2476
2477 atomic_inc(&rdev->nr_pending);
2478 rcu_read_unlock();
2479 switch (r10_sync_page_io(rdev,
2480 r10_bio->devs[sl].addr +
2481 sect,
2482 s, conf->tmppage,
2483 READ)) {
2484 case 0:
2485 /* Well, this device is dead */
2486 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2487 mdname(mddev), s,
2488 (unsigned long long)(
2489 sect +
2490 choose_data_offset(r10_bio, rdev)),
2491 bdevname(rdev->bdev, b));
2492 pr_notice("md/raid10:%s: %s: failing drive\n",
2493 mdname(mddev),
2494 bdevname(rdev->bdev, b));
2495 break;
2496 case 1:
2497 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2498 mdname(mddev), s,
2499 (unsigned long long)(
2500 sect +
2501 choose_data_offset(r10_bio, rdev)),
2502 bdevname(rdev->bdev, b));
2503 atomic_add(s, &rdev->corrected_errors);
2504 }
2505
2506 rdev_dec_pending(rdev, mddev);
2507 rcu_read_lock();
2508 }
2509 rcu_read_unlock();
2510
2511 sectors -= s;
2512 sect += s;
2513 }
2514}
2515
2516static int narrow_write_error(struct r10bio *r10_bio, int i)
2517{
2518 struct bio *bio = r10_bio->master_bio;
2519 struct mddev *mddev = r10_bio->mddev;
2520 struct r10conf *conf = mddev->private;
2521 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2522 /* bio has the data to be written to slot 'i' where
2523 * we just recently had a write error.
2524 * We repeatedly clone the bio and trim down to one block,
2525 * then try the write. Where the write fails we record
2526 * a bad block.
2527 * It is conceivable that the bio doesn't exactly align with
2528 * blocks. We must handle this.
2529 *
2530 * We currently own a reference to the rdev.
2531 */
2532
2533 int block_sectors;
2534 sector_t sector;
2535 int sectors;
2536 int sect_to_write = r10_bio->sectors;
2537 int ok = 1;
2538
2539 if (rdev->badblocks.shift < 0)
2540 return 0;
2541
2542 block_sectors = roundup(1 << rdev->badblocks.shift,
2543 bdev_logical_block_size(rdev->bdev) >> 9);
2544 sector = r10_bio->sector;
2545 sectors = ((r10_bio->sector + block_sectors)
2546 & ~(sector_t)(block_sectors - 1))
2547 - sector;
2548
2549 while (sect_to_write) {
2550 struct bio *wbio;
2551 sector_t wsector;
2552 if (sectors > sect_to_write)
2553 sectors = sect_to_write;
2554 /* Write at 'sector' for 'sectors' */
2555 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2556 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2557 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2558 wbio->bi_iter.bi_sector = wsector +
2559 choose_data_offset(r10_bio, rdev);
2560 bio_set_dev(wbio, rdev->bdev);
2561 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2562
2563 if (submit_bio_wait(wbio) < 0)
2564 /* Failure! */
2565 ok = rdev_set_badblocks(rdev, wsector,
2566 sectors, 0)
2567 && ok;
2568
2569 bio_put(wbio);
2570 sect_to_write -= sectors;
2571 sector += sectors;
2572 sectors = block_sectors;
2573 }
2574 return ok;
2575}
2576
2577static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2578{
2579 int slot = r10_bio->read_slot;
2580 struct bio *bio;
2581 struct r10conf *conf = mddev->private;
2582 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2583
2584 /* we got a read error. Maybe the drive is bad. Maybe just
2585 * the block and we can fix it.
2586 * We freeze all other IO, and try reading the block from
2587 * other devices. When we find one, we re-write
2588 * and check it that fixes the read error.
2589 * This is all done synchronously while the array is
2590 * frozen.
2591 */
2592 bio = r10_bio->devs[slot].bio;
2593 bio_put(bio);
2594 r10_bio->devs[slot].bio = NULL;
2595
2596 if (mddev->ro)
2597 r10_bio->devs[slot].bio = IO_BLOCKED;
2598 else if (!test_bit(FailFast, &rdev->flags)) {
2599 freeze_array(conf, 1);
2600 fix_read_error(conf, mddev, r10_bio);
2601 unfreeze_array(conf);
2602 } else
2603 md_error(mddev, rdev);
2604
2605 rdev_dec_pending(rdev, mddev);
2606 allow_barrier(conf);
2607 r10_bio->state = 0;
2608 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2609}
2610
2611static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2612{
2613 /* Some sort of write request has finished and it
2614 * succeeded in writing where we thought there was a
2615 * bad block. So forget the bad block.
2616 * Or possibly if failed and we need to record
2617 * a bad block.
2618 */
2619 int m;
2620 struct md_rdev *rdev;
2621
2622 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2623 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2624 for (m = 0; m < conf->copies; m++) {
2625 int dev = r10_bio->devs[m].devnum;
2626 rdev = conf->mirrors[dev].rdev;
2627 if (r10_bio->devs[m].bio == NULL ||
2628 r10_bio->devs[m].bio->bi_end_io == NULL)
2629 continue;
2630 if (!r10_bio->devs[m].bio->bi_status) {
2631 rdev_clear_badblocks(
2632 rdev,
2633 r10_bio->devs[m].addr,
2634 r10_bio->sectors, 0);
2635 } else {
2636 if (!rdev_set_badblocks(
2637 rdev,
2638 r10_bio->devs[m].addr,
2639 r10_bio->sectors, 0))
2640 md_error(conf->mddev, rdev);
2641 }
2642 rdev = conf->mirrors[dev].replacement;
2643 if (r10_bio->devs[m].repl_bio == NULL ||
2644 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2645 continue;
2646
2647 if (!r10_bio->devs[m].repl_bio->bi_status) {
2648 rdev_clear_badblocks(
2649 rdev,
2650 r10_bio->devs[m].addr,
2651 r10_bio->sectors, 0);
2652 } else {
2653 if (!rdev_set_badblocks(
2654 rdev,
2655 r10_bio->devs[m].addr,
2656 r10_bio->sectors, 0))
2657 md_error(conf->mddev, rdev);
2658 }
2659 }
2660 put_buf(r10_bio);
2661 } else {
2662 bool fail = false;
2663 for (m = 0; m < conf->copies; m++) {
2664 int dev = r10_bio->devs[m].devnum;
2665 struct bio *bio = r10_bio->devs[m].bio;
2666 rdev = conf->mirrors[dev].rdev;
2667 if (bio == IO_MADE_GOOD) {
2668 rdev_clear_badblocks(
2669 rdev,
2670 r10_bio->devs[m].addr,
2671 r10_bio->sectors, 0);
2672 rdev_dec_pending(rdev, conf->mddev);
2673 } else if (bio != NULL && bio->bi_status) {
2674 fail = true;
2675 if (!narrow_write_error(r10_bio, m)) {
2676 md_error(conf->mddev, rdev);
2677 set_bit(R10BIO_Degraded,
2678 &r10_bio->state);
2679 }
2680 rdev_dec_pending(rdev, conf->mddev);
2681 }
2682 bio = r10_bio->devs[m].repl_bio;
2683 rdev = conf->mirrors[dev].replacement;
2684 if (rdev && bio == IO_MADE_GOOD) {
2685 rdev_clear_badblocks(
2686 rdev,
2687 r10_bio->devs[m].addr,
2688 r10_bio->sectors, 0);
2689 rdev_dec_pending(rdev, conf->mddev);
2690 }
2691 }
2692 if (fail) {
2693 spin_lock_irq(&conf->device_lock);
2694 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2695 conf->nr_queued++;
2696 spin_unlock_irq(&conf->device_lock);
2697 /*
2698 * In case freeze_array() is waiting for condition
2699 * nr_pending == nr_queued + extra to be true.
2700 */
2701 wake_up(&conf->wait_barrier);
2702 md_wakeup_thread(conf->mddev->thread);
2703 } else {
2704 if (test_bit(R10BIO_WriteError,
2705 &r10_bio->state))
2706 close_write(r10_bio);
2707 raid_end_bio_io(r10_bio);
2708 }
2709 }
2710}
2711
2712static void raid10d(struct md_thread *thread)
2713{
2714 struct mddev *mddev = thread->mddev;
2715 struct r10bio *r10_bio;
2716 unsigned long flags;
2717 struct r10conf *conf = mddev->private;
2718 struct list_head *head = &conf->retry_list;
2719 struct blk_plug plug;
2720
2721 md_check_recovery(mddev);
2722
2723 if (!list_empty_careful(&conf->bio_end_io_list) &&
2724 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2725 LIST_HEAD(tmp);
2726 spin_lock_irqsave(&conf->device_lock, flags);
2727 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2728 while (!list_empty(&conf->bio_end_io_list)) {
2729 list_move(conf->bio_end_io_list.prev, &tmp);
2730 conf->nr_queued--;
2731 }
2732 }
2733 spin_unlock_irqrestore(&conf->device_lock, flags);
2734 while (!list_empty(&tmp)) {
2735 r10_bio = list_first_entry(&tmp, struct r10bio,
2736 retry_list);
2737 list_del(&r10_bio->retry_list);
2738 if (mddev->degraded)
2739 set_bit(R10BIO_Degraded, &r10_bio->state);
2740
2741 if (test_bit(R10BIO_WriteError,
2742 &r10_bio->state))
2743 close_write(r10_bio);
2744 raid_end_bio_io(r10_bio);
2745 }
2746 }
2747
2748 blk_start_plug(&plug);
2749 for (;;) {
2750
2751 flush_pending_writes(conf);
2752
2753 spin_lock_irqsave(&conf->device_lock, flags);
2754 if (list_empty(head)) {
2755 spin_unlock_irqrestore(&conf->device_lock, flags);
2756 break;
2757 }
2758 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2759 list_del(head->prev);
2760 conf->nr_queued--;
2761 spin_unlock_irqrestore(&conf->device_lock, flags);
2762
2763 mddev = r10_bio->mddev;
2764 conf = mddev->private;
2765 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2766 test_bit(R10BIO_WriteError, &r10_bio->state))
2767 handle_write_completed(conf, r10_bio);
2768 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2769 reshape_request_write(mddev, r10_bio);
2770 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2771 sync_request_write(mddev, r10_bio);
2772 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2773 recovery_request_write(mddev, r10_bio);
2774 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2775 handle_read_error(mddev, r10_bio);
2776 else
2777 WARN_ON_ONCE(1);
2778
2779 cond_resched();
2780 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2781 md_check_recovery(mddev);
2782 }
2783 blk_finish_plug(&plug);
2784}
2785
2786static int init_resync(struct r10conf *conf)
2787{
2788 int ret, buffs, i;
2789
2790 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2791 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2792 conf->have_replacement = 0;
2793 for (i = 0; i < conf->geo.raid_disks; i++)
2794 if (conf->mirrors[i].replacement)
2795 conf->have_replacement = 1;
2796 ret = mempool_init(&conf->r10buf_pool, buffs,
2797 r10buf_pool_alloc, r10buf_pool_free, conf);
2798 if (ret)
2799 return ret;
2800 conf->next_resync = 0;
2801 return 0;
2802}
2803
2804static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2805{
2806 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2807 struct rsync_pages *rp;
2808 struct bio *bio;
2809 int nalloc;
2810 int i;
2811
2812 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2813 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2814 nalloc = conf->copies; /* resync */
2815 else
2816 nalloc = 2; /* recovery */
2817
2818 for (i = 0; i < nalloc; i++) {
2819 bio = r10bio->devs[i].bio;
2820 rp = bio->bi_private;
2821 bio_reset(bio);
2822 bio->bi_private = rp;
2823 bio = r10bio->devs[i].repl_bio;
2824 if (bio) {
2825 rp = bio->bi_private;
2826 bio_reset(bio);
2827 bio->bi_private = rp;
2828 }
2829 }
2830 return r10bio;
2831}
2832
2833/*
2834 * Set cluster_sync_high since we need other nodes to add the
2835 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2836 */
2837static void raid10_set_cluster_sync_high(struct r10conf *conf)
2838{
2839 sector_t window_size;
2840 int extra_chunk, chunks;
2841
2842 /*
2843 * First, here we define "stripe" as a unit which across
2844 * all member devices one time, so we get chunks by use
2845 * raid_disks / near_copies. Otherwise, if near_copies is
2846 * close to raid_disks, then resync window could increases
2847 * linearly with the increase of raid_disks, which means
2848 * we will suspend a really large IO window while it is not
2849 * necessary. If raid_disks is not divisible by near_copies,
2850 * an extra chunk is needed to ensure the whole "stripe" is
2851 * covered.
2852 */
2853
2854 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2855 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2856 extra_chunk = 0;
2857 else
2858 extra_chunk = 1;
2859 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2860
2861 /*
2862 * At least use a 32M window to align with raid1's resync window
2863 */
2864 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2865 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2866
2867 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2868}
2869
2870/*
2871 * perform a "sync" on one "block"
2872 *
2873 * We need to make sure that no normal I/O request - particularly write
2874 * requests - conflict with active sync requests.
2875 *
2876 * This is achieved by tracking pending requests and a 'barrier' concept
2877 * that can be installed to exclude normal IO requests.
2878 *
2879 * Resync and recovery are handled very differently.
2880 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2881 *
2882 * For resync, we iterate over virtual addresses, read all copies,
2883 * and update if there are differences. If only one copy is live,
2884 * skip it.
2885 * For recovery, we iterate over physical addresses, read a good
2886 * value for each non-in_sync drive, and over-write.
2887 *
2888 * So, for recovery we may have several outstanding complex requests for a
2889 * given address, one for each out-of-sync device. We model this by allocating
2890 * a number of r10_bio structures, one for each out-of-sync device.
2891 * As we setup these structures, we collect all bio's together into a list
2892 * which we then process collectively to add pages, and then process again
2893 * to pass to generic_make_request.
2894 *
2895 * The r10_bio structures are linked using a borrowed master_bio pointer.
2896 * This link is counted in ->remaining. When the r10_bio that points to NULL
2897 * has its remaining count decremented to 0, the whole complex operation
2898 * is complete.
2899 *
2900 */
2901
2902static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2903 int *skipped)
2904{
2905 struct r10conf *conf = mddev->private;
2906 struct r10bio *r10_bio;
2907 struct bio *biolist = NULL, *bio;
2908 sector_t max_sector, nr_sectors;
2909 int i;
2910 int max_sync;
2911 sector_t sync_blocks;
2912 sector_t sectors_skipped = 0;
2913 int chunks_skipped = 0;
2914 sector_t chunk_mask = conf->geo.chunk_mask;
2915 int page_idx = 0;
2916
2917 if (!mempool_initialized(&conf->r10buf_pool))
2918 if (init_resync(conf))
2919 return 0;
2920
2921 /*
2922 * Allow skipping a full rebuild for incremental assembly
2923 * of a clean array, like RAID1 does.
2924 */
2925 if (mddev->bitmap == NULL &&
2926 mddev->recovery_cp == MaxSector &&
2927 mddev->reshape_position == MaxSector &&
2928 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2929 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2930 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2931 conf->fullsync == 0) {
2932 *skipped = 1;
2933 return mddev->dev_sectors - sector_nr;
2934 }
2935
2936 skipped:
2937 max_sector = mddev->dev_sectors;
2938 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2939 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2940 max_sector = mddev->resync_max_sectors;
2941 if (sector_nr >= max_sector) {
2942 conf->cluster_sync_low = 0;
2943 conf->cluster_sync_high = 0;
2944
2945 /* If we aborted, we need to abort the
2946 * sync on the 'current' bitmap chucks (there can
2947 * be several when recovering multiple devices).
2948 * as we may have started syncing it but not finished.
2949 * We can find the current address in
2950 * mddev->curr_resync, but for recovery,
2951 * we need to convert that to several
2952 * virtual addresses.
2953 */
2954 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2955 end_reshape(conf);
2956 close_sync(conf);
2957 return 0;
2958 }
2959
2960 if (mddev->curr_resync < max_sector) { /* aborted */
2961 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2962 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2963 &sync_blocks, 1);
2964 else for (i = 0; i < conf->geo.raid_disks; i++) {
2965 sector_t sect =
2966 raid10_find_virt(conf, mddev->curr_resync, i);
2967 md_bitmap_end_sync(mddev->bitmap, sect,
2968 &sync_blocks, 1);
2969 }
2970 } else {
2971 /* completed sync */
2972 if ((!mddev->bitmap || conf->fullsync)
2973 && conf->have_replacement
2974 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2975 /* Completed a full sync so the replacements
2976 * are now fully recovered.
2977 */
2978 rcu_read_lock();
2979 for (i = 0; i < conf->geo.raid_disks; i++) {
2980 struct md_rdev *rdev =
2981 rcu_dereference(conf->mirrors[i].replacement);
2982 if (rdev)
2983 rdev->recovery_offset = MaxSector;
2984 }
2985 rcu_read_unlock();
2986 }
2987 conf->fullsync = 0;
2988 }
2989 md_bitmap_close_sync(mddev->bitmap);
2990 close_sync(conf);
2991 *skipped = 1;
2992 return sectors_skipped;
2993 }
2994
2995 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2996 return reshape_request(mddev, sector_nr, skipped);
2997
2998 if (chunks_skipped >= conf->geo.raid_disks) {
2999 /* if there has been nothing to do on any drive,
3000 * then there is nothing to do at all..
3001 */
3002 *skipped = 1;
3003 return (max_sector - sector_nr) + sectors_skipped;
3004 }
3005
3006 if (max_sector > mddev->resync_max)
3007 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3008
3009 /* make sure whole request will fit in a chunk - if chunks
3010 * are meaningful
3011 */
3012 if (conf->geo.near_copies < conf->geo.raid_disks &&
3013 max_sector > (sector_nr | chunk_mask))
3014 max_sector = (sector_nr | chunk_mask) + 1;
3015
3016 /*
3017 * If there is non-resync activity waiting for a turn, then let it
3018 * though before starting on this new sync request.
3019 */
3020 if (conf->nr_waiting)
3021 schedule_timeout_uninterruptible(1);
3022
3023 /* Again, very different code for resync and recovery.
3024 * Both must result in an r10bio with a list of bios that
3025 * have bi_end_io, bi_sector, bi_disk set,
3026 * and bi_private set to the r10bio.
3027 * For recovery, we may actually create several r10bios
3028 * with 2 bios in each, that correspond to the bios in the main one.
3029 * In this case, the subordinate r10bios link back through a
3030 * borrowed master_bio pointer, and the counter in the master
3031 * includes a ref from each subordinate.
3032 */
3033 /* First, we decide what to do and set ->bi_end_io
3034 * To end_sync_read if we want to read, and
3035 * end_sync_write if we will want to write.
3036 */
3037
3038 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3039 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3040 /* recovery... the complicated one */
3041 int j;
3042 r10_bio = NULL;
3043
3044 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3045 int still_degraded;
3046 struct r10bio *rb2;
3047 sector_t sect;
3048 int must_sync;
3049 int any_working;
3050 int need_recover = 0;
3051 int need_replace = 0;
3052 struct raid10_info *mirror = &conf->mirrors[i];
3053 struct md_rdev *mrdev, *mreplace;
3054
3055 rcu_read_lock();
3056 mrdev = rcu_dereference(mirror->rdev);
3057 mreplace = rcu_dereference(mirror->replacement);
3058
3059 if (mrdev != NULL &&
3060 !test_bit(Faulty, &mrdev->flags) &&
3061 !test_bit(In_sync, &mrdev->flags))
3062 need_recover = 1;
3063 if (mreplace != NULL &&
3064 !test_bit(Faulty, &mreplace->flags))
3065 need_replace = 1;
3066
3067 if (!need_recover && !need_replace) {
3068 rcu_read_unlock();
3069 continue;
3070 }
3071
3072 still_degraded = 0;
3073 /* want to reconstruct this device */
3074 rb2 = r10_bio;
3075 sect = raid10_find_virt(conf, sector_nr, i);
3076 if (sect >= mddev->resync_max_sectors) {
3077 /* last stripe is not complete - don't
3078 * try to recover this sector.
3079 */
3080 rcu_read_unlock();
3081 continue;
3082 }
3083 if (mreplace && test_bit(Faulty, &mreplace->flags))
3084 mreplace = NULL;
3085 /* Unless we are doing a full sync, or a replacement
3086 * we only need to recover the block if it is set in
3087 * the bitmap
3088 */
3089 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3090 &sync_blocks, 1);
3091 if (sync_blocks < max_sync)
3092 max_sync = sync_blocks;
3093 if (!must_sync &&
3094 mreplace == NULL &&
3095 !conf->fullsync) {
3096 /* yep, skip the sync_blocks here, but don't assume
3097 * that there will never be anything to do here
3098 */
3099 chunks_skipped = -1;
3100 rcu_read_unlock();
3101 continue;
3102 }
3103 atomic_inc(&mrdev->nr_pending);
3104 if (mreplace)
3105 atomic_inc(&mreplace->nr_pending);
3106 rcu_read_unlock();
3107
3108 r10_bio = raid10_alloc_init_r10buf(conf);
3109 r10_bio->state = 0;
3110 raise_barrier(conf, rb2 != NULL);
3111 atomic_set(&r10_bio->remaining, 0);
3112
3113 r10_bio->master_bio = (struct bio*)rb2;
3114 if (rb2)
3115 atomic_inc(&rb2->remaining);
3116 r10_bio->mddev = mddev;
3117 set_bit(R10BIO_IsRecover, &r10_bio->state);
3118 r10_bio->sector = sect;
3119
3120 raid10_find_phys(conf, r10_bio);
3121
3122 /* Need to check if the array will still be
3123 * degraded
3124 */
3125 rcu_read_lock();
3126 for (j = 0; j < conf->geo.raid_disks; j++) {
3127 struct md_rdev *rdev = rcu_dereference(
3128 conf->mirrors[j].rdev);
3129 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3130 still_degraded = 1;
3131 break;
3132 }
3133 }
3134
3135 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3136 &sync_blocks, still_degraded);
3137
3138 any_working = 0;
3139 for (j=0; j<conf->copies;j++) {
3140 int k;
3141 int d = r10_bio->devs[j].devnum;
3142 sector_t from_addr, to_addr;
3143 struct md_rdev *rdev =
3144 rcu_dereference(conf->mirrors[d].rdev);
3145 sector_t sector, first_bad;
3146 int bad_sectors;
3147 if (!rdev ||
3148 !test_bit(In_sync, &rdev->flags))
3149 continue;
3150 /* This is where we read from */
3151 any_working = 1;
3152 sector = r10_bio->devs[j].addr;
3153
3154 if (is_badblock(rdev, sector, max_sync,
3155 &first_bad, &bad_sectors)) {
3156 if (first_bad > sector)
3157 max_sync = first_bad - sector;
3158 else {
3159 bad_sectors -= (sector
3160 - first_bad);
3161 if (max_sync > bad_sectors)
3162 max_sync = bad_sectors;
3163 continue;
3164 }
3165 }
3166 bio = r10_bio->devs[0].bio;
3167 bio->bi_next = biolist;
3168 biolist = bio;
3169 bio->bi_end_io = end_sync_read;
3170 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3171 if (test_bit(FailFast, &rdev->flags))
3172 bio->bi_opf |= MD_FAILFAST;
3173 from_addr = r10_bio->devs[j].addr;
3174 bio->bi_iter.bi_sector = from_addr +
3175 rdev->data_offset;
3176 bio_set_dev(bio, rdev->bdev);
3177 atomic_inc(&rdev->nr_pending);
3178 /* and we write to 'i' (if not in_sync) */
3179
3180 for (k=0; k<conf->copies; k++)
3181 if (r10_bio->devs[k].devnum == i)
3182 break;
3183 BUG_ON(k == conf->copies);
3184 to_addr = r10_bio->devs[k].addr;
3185 r10_bio->devs[0].devnum = d;
3186 r10_bio->devs[0].addr = from_addr;
3187 r10_bio->devs[1].devnum = i;
3188 r10_bio->devs[1].addr = to_addr;
3189
3190 if (need_recover) {
3191 bio = r10_bio->devs[1].bio;
3192 bio->bi_next = biolist;
3193 biolist = bio;
3194 bio->bi_end_io = end_sync_write;
3195 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3196 bio->bi_iter.bi_sector = to_addr
3197 + mrdev->data_offset;
3198 bio_set_dev(bio, mrdev->bdev);
3199 atomic_inc(&r10_bio->remaining);
3200 } else
3201 r10_bio->devs[1].bio->bi_end_io = NULL;
3202
3203 /* and maybe write to replacement */
3204 bio = r10_bio->devs[1].repl_bio;
3205 if (bio)
3206 bio->bi_end_io = NULL;
3207 /* Note: if need_replace, then bio
3208 * cannot be NULL as r10buf_pool_alloc will
3209 * have allocated it.
3210 */
3211 if (!need_replace)
3212 break;
3213 bio->bi_next = biolist;
3214 biolist = bio;
3215 bio->bi_end_io = end_sync_write;
3216 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3217 bio->bi_iter.bi_sector = to_addr +
3218 mreplace->data_offset;
3219 bio_set_dev(bio, mreplace->bdev);
3220 atomic_inc(&r10_bio->remaining);
3221 break;
3222 }
3223 rcu_read_unlock();
3224 if (j == conf->copies) {
3225 /* Cannot recover, so abort the recovery or
3226 * record a bad block */
3227 if (any_working) {
3228 /* problem is that there are bad blocks
3229 * on other device(s)
3230 */
3231 int k;
3232 for (k = 0; k < conf->copies; k++)
3233 if (r10_bio->devs[k].devnum == i)
3234 break;
3235 if (!test_bit(In_sync,
3236 &mrdev->flags)
3237 && !rdev_set_badblocks(
3238 mrdev,
3239 r10_bio->devs[k].addr,
3240 max_sync, 0))
3241 any_working = 0;
3242 if (mreplace &&
3243 !rdev_set_badblocks(
3244 mreplace,
3245 r10_bio->devs[k].addr,
3246 max_sync, 0))
3247 any_working = 0;
3248 }
3249 if (!any_working) {
3250 if (!test_and_set_bit(MD_RECOVERY_INTR,
3251 &mddev->recovery))
3252 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3253 mdname(mddev));
3254 mirror->recovery_disabled
3255 = mddev->recovery_disabled;
3256 }
3257 put_buf(r10_bio);
3258 if (rb2)
3259 atomic_dec(&rb2->remaining);
3260 r10_bio = rb2;
3261 rdev_dec_pending(mrdev, mddev);
3262 if (mreplace)
3263 rdev_dec_pending(mreplace, mddev);
3264 break;
3265 }
3266 rdev_dec_pending(mrdev, mddev);
3267 if (mreplace)
3268 rdev_dec_pending(mreplace, mddev);
3269 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3270 /* Only want this if there is elsewhere to
3271 * read from. 'j' is currently the first
3272 * readable copy.
3273 */
3274 int targets = 1;
3275 for (; j < conf->copies; j++) {
3276 int d = r10_bio->devs[j].devnum;
3277 if (conf->mirrors[d].rdev &&
3278 test_bit(In_sync,
3279 &conf->mirrors[d].rdev->flags))
3280 targets++;
3281 }
3282 if (targets == 1)
3283 r10_bio->devs[0].bio->bi_opf
3284 &= ~MD_FAILFAST;
3285 }
3286 }
3287 if (biolist == NULL) {
3288 while (r10_bio) {
3289 struct r10bio *rb2 = r10_bio;
3290 r10_bio = (struct r10bio*) rb2->master_bio;
3291 rb2->master_bio = NULL;
3292 put_buf(rb2);
3293 }
3294 goto giveup;
3295 }
3296 } else {
3297 /* resync. Schedule a read for every block at this virt offset */
3298 int count = 0;
3299
3300 /*
3301 * Since curr_resync_completed could probably not update in
3302 * time, and we will set cluster_sync_low based on it.
3303 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3304 * safety reason, which ensures curr_resync_completed is
3305 * updated in bitmap_cond_end_sync.
3306 */
3307 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3308 mddev_is_clustered(mddev) &&
3309 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3310
3311 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3312 &sync_blocks, mddev->degraded) &&
3313 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3314 &mddev->recovery)) {
3315 /* We can skip this block */
3316 *skipped = 1;
3317 return sync_blocks + sectors_skipped;
3318 }
3319 if (sync_blocks < max_sync)
3320 max_sync = sync_blocks;
3321 r10_bio = raid10_alloc_init_r10buf(conf);
3322 r10_bio->state = 0;
3323
3324 r10_bio->mddev = mddev;
3325 atomic_set(&r10_bio->remaining, 0);
3326 raise_barrier(conf, 0);
3327 conf->next_resync = sector_nr;
3328
3329 r10_bio->master_bio = NULL;
3330 r10_bio->sector = sector_nr;
3331 set_bit(R10BIO_IsSync, &r10_bio->state);
3332 raid10_find_phys(conf, r10_bio);
3333 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3334
3335 for (i = 0; i < conf->copies; i++) {
3336 int d = r10_bio->devs[i].devnum;
3337 sector_t first_bad, sector;
3338 int bad_sectors;
3339 struct md_rdev *rdev;
3340
3341 if (r10_bio->devs[i].repl_bio)
3342 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3343
3344 bio = r10_bio->devs[i].bio;
3345 bio->bi_status = BLK_STS_IOERR;
3346 rcu_read_lock();
3347 rdev = rcu_dereference(conf->mirrors[d].rdev);
3348 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3349 rcu_read_unlock();
3350 continue;
3351 }
3352 sector = r10_bio->devs[i].addr;
3353 if (is_badblock(rdev, sector, max_sync,
3354 &first_bad, &bad_sectors)) {
3355 if (first_bad > sector)
3356 max_sync = first_bad - sector;
3357 else {
3358 bad_sectors -= (sector - first_bad);
3359 if (max_sync > bad_sectors)
3360 max_sync = bad_sectors;
3361 rcu_read_unlock();
3362 continue;
3363 }
3364 }
3365 atomic_inc(&rdev->nr_pending);
3366 atomic_inc(&r10_bio->remaining);
3367 bio->bi_next = biolist;
3368 biolist = bio;
3369 bio->bi_end_io = end_sync_read;
3370 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3371 if (test_bit(FailFast, &rdev->flags))
3372 bio->bi_opf |= MD_FAILFAST;
3373 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3374 bio_set_dev(bio, rdev->bdev);
3375 count++;
3376
3377 rdev = rcu_dereference(conf->mirrors[d].replacement);
3378 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3379 rcu_read_unlock();
3380 continue;
3381 }
3382 atomic_inc(&rdev->nr_pending);
3383
3384 /* Need to set up for writing to the replacement */
3385 bio = r10_bio->devs[i].repl_bio;
3386 bio->bi_status = BLK_STS_IOERR;
3387
3388 sector = r10_bio->devs[i].addr;
3389 bio->bi_next = biolist;
3390 biolist = bio;
3391 bio->bi_end_io = end_sync_write;
3392 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3393 if (test_bit(FailFast, &rdev->flags))
3394 bio->bi_opf |= MD_FAILFAST;
3395 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3396 bio_set_dev(bio, rdev->bdev);
3397 count++;
3398 rcu_read_unlock();
3399 }
3400
3401 if (count < 2) {
3402 for (i=0; i<conf->copies; i++) {
3403 int d = r10_bio->devs[i].devnum;
3404 if (r10_bio->devs[i].bio->bi_end_io)
3405 rdev_dec_pending(conf->mirrors[d].rdev,
3406 mddev);
3407 if (r10_bio->devs[i].repl_bio &&
3408 r10_bio->devs[i].repl_bio->bi_end_io)
3409 rdev_dec_pending(
3410 conf->mirrors[d].replacement,
3411 mddev);
3412 }
3413 put_buf(r10_bio);
3414 biolist = NULL;
3415 goto giveup;
3416 }
3417 }
3418
3419 nr_sectors = 0;
3420 if (sector_nr + max_sync < max_sector)
3421 max_sector = sector_nr + max_sync;
3422 do {
3423 struct page *page;
3424 int len = PAGE_SIZE;
3425 if (sector_nr + (len>>9) > max_sector)
3426 len = (max_sector - sector_nr) << 9;
3427 if (len == 0)
3428 break;
3429 for (bio= biolist ; bio ; bio=bio->bi_next) {
3430 struct resync_pages *rp = get_resync_pages(bio);
3431 page = resync_fetch_page(rp, page_idx);
3432 /*
3433 * won't fail because the vec table is big enough
3434 * to hold all these pages
3435 */
3436 bio_add_page(bio, page, len, 0);
3437 }
3438 nr_sectors += len>>9;
3439 sector_nr += len>>9;
3440 } while (++page_idx < RESYNC_PAGES);
3441 r10_bio->sectors = nr_sectors;
3442
3443 if (mddev_is_clustered(mddev) &&
3444 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3445 /* It is resync not recovery */
3446 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3447 conf->cluster_sync_low = mddev->curr_resync_completed;
3448 raid10_set_cluster_sync_high(conf);
3449 /* Send resync message */
3450 md_cluster_ops->resync_info_update(mddev,
3451 conf->cluster_sync_low,
3452 conf->cluster_sync_high);
3453 }
3454 } else if (mddev_is_clustered(mddev)) {
3455 /* This is recovery not resync */
3456 sector_t sect_va1, sect_va2;
3457 bool broadcast_msg = false;
3458
3459 for (i = 0; i < conf->geo.raid_disks; i++) {
3460 /*
3461 * sector_nr is a device address for recovery, so we
3462 * need translate it to array address before compare
3463 * with cluster_sync_high.
3464 */
3465 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3466
3467 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3468 broadcast_msg = true;
3469 /*
3470 * curr_resync_completed is similar as
3471 * sector_nr, so make the translation too.
3472 */
3473 sect_va2 = raid10_find_virt(conf,
3474 mddev->curr_resync_completed, i);
3475
3476 if (conf->cluster_sync_low == 0 ||
3477 conf->cluster_sync_low > sect_va2)
3478 conf->cluster_sync_low = sect_va2;
3479 }
3480 }
3481 if (broadcast_msg) {
3482 raid10_set_cluster_sync_high(conf);
3483 md_cluster_ops->resync_info_update(mddev,
3484 conf->cluster_sync_low,
3485 conf->cluster_sync_high);
3486 }
3487 }
3488
3489 while (biolist) {
3490 bio = biolist;
3491 biolist = biolist->bi_next;
3492
3493 bio->bi_next = NULL;
3494 r10_bio = get_resync_r10bio(bio);
3495 r10_bio->sectors = nr_sectors;
3496
3497 if (bio->bi_end_io == end_sync_read) {
3498 md_sync_acct_bio(bio, nr_sectors);
3499 bio->bi_status = 0;
3500 generic_make_request(bio);
3501 }
3502 }
3503
3504 if (sectors_skipped)
3505 /* pretend they weren't skipped, it makes
3506 * no important difference in this case
3507 */
3508 md_done_sync(mddev, sectors_skipped, 1);
3509
3510 return sectors_skipped + nr_sectors;
3511 giveup:
3512 /* There is nowhere to write, so all non-sync
3513 * drives must be failed or in resync, all drives
3514 * have a bad block, so try the next chunk...
3515 */
3516 if (sector_nr + max_sync < max_sector)
3517 max_sector = sector_nr + max_sync;
3518
3519 sectors_skipped += (max_sector - sector_nr);
3520 chunks_skipped ++;
3521 sector_nr = max_sector;
3522 goto skipped;
3523}
3524
3525static sector_t
3526raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3527{
3528 sector_t size;
3529 struct r10conf *conf = mddev->private;
3530
3531 if (!raid_disks)
3532 raid_disks = min(conf->geo.raid_disks,
3533 conf->prev.raid_disks);
3534 if (!sectors)
3535 sectors = conf->dev_sectors;
3536
3537 size = sectors >> conf->geo.chunk_shift;
3538 sector_div(size, conf->geo.far_copies);
3539 size = size * raid_disks;
3540 sector_div(size, conf->geo.near_copies);
3541
3542 return size << conf->geo.chunk_shift;
3543}
3544
3545static void calc_sectors(struct r10conf *conf, sector_t size)
3546{
3547 /* Calculate the number of sectors-per-device that will
3548 * actually be used, and set conf->dev_sectors and
3549 * conf->stride
3550 */
3551
3552 size = size >> conf->geo.chunk_shift;
3553 sector_div(size, conf->geo.far_copies);
3554 size = size * conf->geo.raid_disks;
3555 sector_div(size, conf->geo.near_copies);
3556 /* 'size' is now the number of chunks in the array */
3557 /* calculate "used chunks per device" */
3558 size = size * conf->copies;
3559
3560 /* We need to round up when dividing by raid_disks to
3561 * get the stride size.
3562 */
3563 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3564
3565 conf->dev_sectors = size << conf->geo.chunk_shift;
3566
3567 if (conf->geo.far_offset)
3568 conf->geo.stride = 1 << conf->geo.chunk_shift;
3569 else {
3570 sector_div(size, conf->geo.far_copies);
3571 conf->geo.stride = size << conf->geo.chunk_shift;
3572 }
3573}
3574
3575enum geo_type {geo_new, geo_old, geo_start};
3576static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3577{
3578 int nc, fc, fo;
3579 int layout, chunk, disks;
3580 switch (new) {
3581 case geo_old:
3582 layout = mddev->layout;
3583 chunk = mddev->chunk_sectors;
3584 disks = mddev->raid_disks - mddev->delta_disks;
3585 break;
3586 case geo_new:
3587 layout = mddev->new_layout;
3588 chunk = mddev->new_chunk_sectors;
3589 disks = mddev->raid_disks;
3590 break;
3591 default: /* avoid 'may be unused' warnings */
3592 case geo_start: /* new when starting reshape - raid_disks not
3593 * updated yet. */
3594 layout = mddev->new_layout;
3595 chunk = mddev->new_chunk_sectors;
3596 disks = mddev->raid_disks + mddev->delta_disks;
3597 break;
3598 }
3599 if (layout >> 19)
3600 return -1;
3601 if (chunk < (PAGE_SIZE >> 9) ||
3602 !is_power_of_2(chunk))
3603 return -2;
3604 nc = layout & 255;
3605 fc = (layout >> 8) & 255;
3606 fo = layout & (1<<16);
3607 geo->raid_disks = disks;
3608 geo->near_copies = nc;
3609 geo->far_copies = fc;
3610 geo->far_offset = fo;
3611 switch (layout >> 17) {
3612 case 0: /* original layout. simple but not always optimal */
3613 geo->far_set_size = disks;
3614 break;
3615 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3616 * actually using this, but leave code here just in case.*/
3617 geo->far_set_size = disks/fc;
3618 WARN(geo->far_set_size < fc,
3619 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3620 break;
3621 case 2: /* "improved" layout fixed to match documentation */
3622 geo->far_set_size = fc * nc;
3623 break;
3624 default: /* Not a valid layout */
3625 return -1;
3626 }
3627 geo->chunk_mask = chunk - 1;
3628 geo->chunk_shift = ffz(~chunk);
3629 return nc*fc;
3630}
3631
3632static struct r10conf *setup_conf(struct mddev *mddev)
3633{
3634 struct r10conf *conf = NULL;
3635 int err = -EINVAL;
3636 struct geom geo;
3637 int copies;
3638
3639 copies = setup_geo(&geo, mddev, geo_new);
3640
3641 if (copies == -2) {
3642 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3643 mdname(mddev), PAGE_SIZE);
3644 goto out;
3645 }
3646
3647 if (copies < 2 || copies > mddev->raid_disks) {
3648 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3649 mdname(mddev), mddev->new_layout);
3650 goto out;
3651 }
3652
3653 err = -ENOMEM;
3654 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3655 if (!conf)
3656 goto out;
3657
3658 /* FIXME calc properly */
3659 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3660 sizeof(struct raid10_info),
3661 GFP_KERNEL);
3662 if (!conf->mirrors)
3663 goto out;
3664
3665 conf->tmppage = alloc_page(GFP_KERNEL);
3666 if (!conf->tmppage)
3667 goto out;
3668
3669 conf->geo = geo;
3670 conf->copies = copies;
3671 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3672 rbio_pool_free, conf);
3673 if (err)
3674 goto out;
3675
3676 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3677 if (err)
3678 goto out;
3679
3680 calc_sectors(conf, mddev->dev_sectors);
3681 if (mddev->reshape_position == MaxSector) {
3682 conf->prev = conf->geo;
3683 conf->reshape_progress = MaxSector;
3684 } else {
3685 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3686 err = -EINVAL;
3687 goto out;
3688 }
3689 conf->reshape_progress = mddev->reshape_position;
3690 if (conf->prev.far_offset)
3691 conf->prev.stride = 1 << conf->prev.chunk_shift;
3692 else
3693 /* far_copies must be 1 */
3694 conf->prev.stride = conf->dev_sectors;
3695 }
3696 conf->reshape_safe = conf->reshape_progress;
3697 spin_lock_init(&conf->device_lock);
3698 INIT_LIST_HEAD(&conf->retry_list);
3699 INIT_LIST_HEAD(&conf->bio_end_io_list);
3700
3701 spin_lock_init(&conf->resync_lock);
3702 init_waitqueue_head(&conf->wait_barrier);
3703 atomic_set(&conf->nr_pending, 0);
3704
3705 err = -ENOMEM;
3706 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3707 if (!conf->thread)
3708 goto out;
3709
3710 conf->mddev = mddev;
3711 return conf;
3712
3713 out:
3714 if (conf) {
3715 mempool_exit(&conf->r10bio_pool);
3716 kfree(conf->mirrors);
3717 safe_put_page(conf->tmppage);
3718 bioset_exit(&conf->bio_split);
3719 kfree(conf);
3720 }
3721 return ERR_PTR(err);
3722}
3723
3724static int raid10_run(struct mddev *mddev)
3725{
3726 struct r10conf *conf;
3727 int i, disk_idx, chunk_size;
3728 struct raid10_info *disk;
3729 struct md_rdev *rdev;
3730 sector_t size;
3731 sector_t min_offset_diff = 0;
3732 int first = 1;
3733 bool discard_supported = false;
3734
3735 if (mddev_init_writes_pending(mddev) < 0)
3736 return -ENOMEM;
3737
3738 if (mddev->private == NULL) {
3739 conf = setup_conf(mddev);
3740 if (IS_ERR(conf))
3741 return PTR_ERR(conf);
3742 mddev->private = conf;
3743 }
3744 conf = mddev->private;
3745 if (!conf)
3746 goto out;
3747
3748 if (mddev_is_clustered(conf->mddev)) {
3749 int fc, fo;
3750
3751 fc = (mddev->layout >> 8) & 255;
3752 fo = mddev->layout & (1<<16);
3753 if (fc > 1 || fo > 0) {
3754 pr_err("only near layout is supported by clustered"
3755 " raid10\n");
3756 goto out_free_conf;
3757 }
3758 }
3759
3760 mddev->thread = conf->thread;
3761 conf->thread = NULL;
3762
3763 chunk_size = mddev->chunk_sectors << 9;
3764 if (mddev->queue) {
3765 blk_queue_max_discard_sectors(mddev->queue,
3766 mddev->chunk_sectors);
3767 blk_queue_max_write_same_sectors(mddev->queue, 0);
3768 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3769 blk_queue_io_min(mddev->queue, chunk_size);
3770 if (conf->geo.raid_disks % conf->geo.near_copies)
3771 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3772 else
3773 blk_queue_io_opt(mddev->queue, chunk_size *
3774 (conf->geo.raid_disks / conf->geo.near_copies));
3775 }
3776
3777 rdev_for_each(rdev, mddev) {
3778 long long diff;
3779
3780 disk_idx = rdev->raid_disk;
3781 if (disk_idx < 0)
3782 continue;
3783 if (disk_idx >= conf->geo.raid_disks &&
3784 disk_idx >= conf->prev.raid_disks)
3785 continue;
3786 disk = conf->mirrors + disk_idx;
3787
3788 if (test_bit(Replacement, &rdev->flags)) {
3789 if (disk->replacement)
3790 goto out_free_conf;
3791 disk->replacement = rdev;
3792 } else {
3793 if (disk->rdev)
3794 goto out_free_conf;
3795 disk->rdev = rdev;
3796 }
3797 diff = (rdev->new_data_offset - rdev->data_offset);
3798 if (!mddev->reshape_backwards)
3799 diff = -diff;
3800 if (diff < 0)
3801 diff = 0;
3802 if (first || diff < min_offset_diff)
3803 min_offset_diff = diff;
3804
3805 if (mddev->gendisk)
3806 disk_stack_limits(mddev->gendisk, rdev->bdev,
3807 rdev->data_offset << 9);
3808
3809 disk->head_position = 0;
3810
3811 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3812 discard_supported = true;
3813 first = 0;
3814 }
3815
3816 if (mddev->queue) {
3817 if (discard_supported)
3818 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3819 mddev->queue);
3820 else
3821 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3822 mddev->queue);
3823 }
3824 /* need to check that every block has at least one working mirror */
3825 if (!enough(conf, -1)) {
3826 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3827 mdname(mddev));
3828 goto out_free_conf;
3829 }
3830
3831 if (conf->reshape_progress != MaxSector) {
3832 /* must ensure that shape change is supported */
3833 if (conf->geo.far_copies != 1 &&
3834 conf->geo.far_offset == 0)
3835 goto out_free_conf;
3836 if (conf->prev.far_copies != 1 &&
3837 conf->prev.far_offset == 0)
3838 goto out_free_conf;
3839 }
3840
3841 mddev->degraded = 0;
3842 for (i = 0;
3843 i < conf->geo.raid_disks
3844 || i < conf->prev.raid_disks;
3845 i++) {
3846
3847 disk = conf->mirrors + i;
3848
3849 if (!disk->rdev && disk->replacement) {
3850 /* The replacement is all we have - use it */
3851 disk->rdev = disk->replacement;
3852 disk->replacement = NULL;
3853 clear_bit(Replacement, &disk->rdev->flags);
3854 }
3855
3856 if (!disk->rdev ||
3857 !test_bit(In_sync, &disk->rdev->flags)) {
3858 disk->head_position = 0;
3859 mddev->degraded++;
3860 if (disk->rdev &&
3861 disk->rdev->saved_raid_disk < 0)
3862 conf->fullsync = 1;
3863 }
3864
3865 if (disk->replacement &&
3866 !test_bit(In_sync, &disk->replacement->flags) &&
3867 disk->replacement->saved_raid_disk < 0) {
3868 conf->fullsync = 1;
3869 }
3870
3871 disk->recovery_disabled = mddev->recovery_disabled - 1;
3872 }
3873
3874 if (mddev->recovery_cp != MaxSector)
3875 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3876 mdname(mddev));
3877 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3878 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3879 conf->geo.raid_disks);
3880 /*
3881 * Ok, everything is just fine now
3882 */
3883 mddev->dev_sectors = conf->dev_sectors;
3884 size = raid10_size(mddev, 0, 0);
3885 md_set_array_sectors(mddev, size);
3886 mddev->resync_max_sectors = size;
3887 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3888
3889 if (mddev->queue) {
3890 int stripe = conf->geo.raid_disks *
3891 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3892
3893 /* Calculate max read-ahead size.
3894 * We need to readahead at least twice a whole stripe....
3895 * maybe...
3896 */
3897 stripe /= conf->geo.near_copies;
3898 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3899 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3900 }
3901
3902 if (md_integrity_register(mddev))
3903 goto out_free_conf;
3904
3905 if (conf->reshape_progress != MaxSector) {
3906 unsigned long before_length, after_length;
3907
3908 before_length = ((1 << conf->prev.chunk_shift) *
3909 conf->prev.far_copies);
3910 after_length = ((1 << conf->geo.chunk_shift) *
3911 conf->geo.far_copies);
3912
3913 if (max(before_length, after_length) > min_offset_diff) {
3914 /* This cannot work */
3915 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3916 goto out_free_conf;
3917 }
3918 conf->offset_diff = min_offset_diff;
3919
3920 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3921 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3922 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3923 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3924 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3925 "reshape");
3926 if (!mddev->sync_thread)
3927 goto out_free_conf;
3928 }
3929
3930 return 0;
3931
3932out_free_conf:
3933 md_unregister_thread(&mddev->thread);
3934 mempool_exit(&conf->r10bio_pool);
3935 safe_put_page(conf->tmppage);
3936 kfree(conf->mirrors);
3937 kfree(conf);
3938 mddev->private = NULL;
3939out:
3940 return -EIO;
3941}
3942
3943static void raid10_free(struct mddev *mddev, void *priv)
3944{
3945 struct r10conf *conf = priv;
3946
3947 mempool_exit(&conf->r10bio_pool);
3948 safe_put_page(conf->tmppage);
3949 kfree(conf->mirrors);
3950 kfree(conf->mirrors_old);
3951 kfree(conf->mirrors_new);
3952 bioset_exit(&conf->bio_split);
3953 kfree(conf);
3954}
3955
3956static void raid10_quiesce(struct mddev *mddev, int quiesce)
3957{
3958 struct r10conf *conf = mddev->private;
3959
3960 if (quiesce)
3961 raise_barrier(conf, 0);
3962 else
3963 lower_barrier(conf);
3964}
3965
3966static int raid10_resize(struct mddev *mddev, sector_t sectors)
3967{
3968 /* Resize of 'far' arrays is not supported.
3969 * For 'near' and 'offset' arrays we can set the
3970 * number of sectors used to be an appropriate multiple
3971 * of the chunk size.
3972 * For 'offset', this is far_copies*chunksize.
3973 * For 'near' the multiplier is the LCM of
3974 * near_copies and raid_disks.
3975 * So if far_copies > 1 && !far_offset, fail.
3976 * Else find LCM(raid_disks, near_copy)*far_copies and
3977 * multiply by chunk_size. Then round to this number.
3978 * This is mostly done by raid10_size()
3979 */
3980 struct r10conf *conf = mddev->private;
3981 sector_t oldsize, size;
3982
3983 if (mddev->reshape_position != MaxSector)
3984 return -EBUSY;
3985
3986 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3987 return -EINVAL;
3988
3989 oldsize = raid10_size(mddev, 0, 0);
3990 size = raid10_size(mddev, sectors, 0);
3991 if (mddev->external_size &&
3992 mddev->array_sectors > size)
3993 return -EINVAL;
3994 if (mddev->bitmap) {
3995 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3996 if (ret)
3997 return ret;
3998 }
3999 md_set_array_sectors(mddev, size);
4000 if (sectors > mddev->dev_sectors &&
4001 mddev->recovery_cp > oldsize) {
4002 mddev->recovery_cp = oldsize;
4003 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4004 }
4005 calc_sectors(conf, sectors);
4006 mddev->dev_sectors = conf->dev_sectors;
4007 mddev->resync_max_sectors = size;
4008 return 0;
4009}
4010
4011static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4012{
4013 struct md_rdev *rdev;
4014 struct r10conf *conf;
4015
4016 if (mddev->degraded > 0) {
4017 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4018 mdname(mddev));
4019 return ERR_PTR(-EINVAL);
4020 }
4021 sector_div(size, devs);
4022
4023 /* Set new parameters */
4024 mddev->new_level = 10;
4025 /* new layout: far_copies = 1, near_copies = 2 */
4026 mddev->new_layout = (1<<8) + 2;
4027 mddev->new_chunk_sectors = mddev->chunk_sectors;
4028 mddev->delta_disks = mddev->raid_disks;
4029 mddev->raid_disks *= 2;
4030 /* make sure it will be not marked as dirty */
4031 mddev->recovery_cp = MaxSector;
4032 mddev->dev_sectors = size;
4033
4034 conf = setup_conf(mddev);
4035 if (!IS_ERR(conf)) {
4036 rdev_for_each(rdev, mddev)
4037 if (rdev->raid_disk >= 0) {
4038 rdev->new_raid_disk = rdev->raid_disk * 2;
4039 rdev->sectors = size;
4040 }
4041 conf->barrier = 1;
4042 }
4043
4044 return conf;
4045}
4046
4047static void *raid10_takeover(struct mddev *mddev)
4048{
4049 struct r0conf *raid0_conf;
4050
4051 /* raid10 can take over:
4052 * raid0 - providing it has only two drives
4053 */
4054 if (mddev->level == 0) {
4055 /* for raid0 takeover only one zone is supported */
4056 raid0_conf = mddev->private;
4057 if (raid0_conf->nr_strip_zones > 1) {
4058 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4059 mdname(mddev));
4060 return ERR_PTR(-EINVAL);
4061 }
4062 return raid10_takeover_raid0(mddev,
4063 raid0_conf->strip_zone->zone_end,
4064 raid0_conf->strip_zone->nb_dev);
4065 }
4066 return ERR_PTR(-EINVAL);
4067}
4068
4069static int raid10_check_reshape(struct mddev *mddev)
4070{
4071 /* Called when there is a request to change
4072 * - layout (to ->new_layout)
4073 * - chunk size (to ->new_chunk_sectors)
4074 * - raid_disks (by delta_disks)
4075 * or when trying to restart a reshape that was ongoing.
4076 *
4077 * We need to validate the request and possibly allocate
4078 * space if that might be an issue later.
4079 *
4080 * Currently we reject any reshape of a 'far' mode array,
4081 * allow chunk size to change if new is generally acceptable,
4082 * allow raid_disks to increase, and allow
4083 * a switch between 'near' mode and 'offset' mode.
4084 */
4085 struct r10conf *conf = mddev->private;
4086 struct geom geo;
4087
4088 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4089 return -EINVAL;
4090
4091 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4092 /* mustn't change number of copies */
4093 return -EINVAL;
4094 if (geo.far_copies > 1 && !geo.far_offset)
4095 /* Cannot switch to 'far' mode */
4096 return -EINVAL;
4097
4098 if (mddev->array_sectors & geo.chunk_mask)
4099 /* not factor of array size */
4100 return -EINVAL;
4101
4102 if (!enough(conf, -1))
4103 return -EINVAL;
4104
4105 kfree(conf->mirrors_new);
4106 conf->mirrors_new = NULL;
4107 if (mddev->delta_disks > 0) {
4108 /* allocate new 'mirrors' list */
4109 conf->mirrors_new =
4110 kcalloc(mddev->raid_disks + mddev->delta_disks,
4111 sizeof(struct raid10_info),
4112 GFP_KERNEL);
4113 if (!conf->mirrors_new)
4114 return -ENOMEM;
4115 }
4116 return 0;
4117}
4118
4119/*
4120 * Need to check if array has failed when deciding whether to:
4121 * - start an array
4122 * - remove non-faulty devices
4123 * - add a spare
4124 * - allow a reshape
4125 * This determination is simple when no reshape is happening.
4126 * However if there is a reshape, we need to carefully check
4127 * both the before and after sections.
4128 * This is because some failed devices may only affect one
4129 * of the two sections, and some non-in_sync devices may
4130 * be insync in the section most affected by failed devices.
4131 */
4132static int calc_degraded(struct r10conf *conf)
4133{
4134 int degraded, degraded2;
4135 int i;
4136
4137 rcu_read_lock();
4138 degraded = 0;
4139 /* 'prev' section first */
4140 for (i = 0; i < conf->prev.raid_disks; i++) {
4141 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4142 if (!rdev || test_bit(Faulty, &rdev->flags))
4143 degraded++;
4144 else if (!test_bit(In_sync, &rdev->flags))
4145 /* When we can reduce the number of devices in
4146 * an array, this might not contribute to
4147 * 'degraded'. It does now.
4148 */
4149 degraded++;
4150 }
4151 rcu_read_unlock();
4152 if (conf->geo.raid_disks == conf->prev.raid_disks)
4153 return degraded;
4154 rcu_read_lock();
4155 degraded2 = 0;
4156 for (i = 0; i < conf->geo.raid_disks; i++) {
4157 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4158 if (!rdev || test_bit(Faulty, &rdev->flags))
4159 degraded2++;
4160 else if (!test_bit(In_sync, &rdev->flags)) {
4161 /* If reshape is increasing the number of devices,
4162 * this section has already been recovered, so
4163 * it doesn't contribute to degraded.
4164 * else it does.
4165 */
4166 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4167 degraded2++;
4168 }
4169 }
4170 rcu_read_unlock();
4171 if (degraded2 > degraded)
4172 return degraded2;
4173 return degraded;
4174}
4175
4176static int raid10_start_reshape(struct mddev *mddev)
4177{
4178 /* A 'reshape' has been requested. This commits
4179 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4180 * This also checks if there are enough spares and adds them
4181 * to the array.
4182 * We currently require enough spares to make the final
4183 * array non-degraded. We also require that the difference
4184 * between old and new data_offset - on each device - is
4185 * enough that we never risk over-writing.
4186 */
4187
4188 unsigned long before_length, after_length;
4189 sector_t min_offset_diff = 0;
4190 int first = 1;
4191 struct geom new;
4192 struct r10conf *conf = mddev->private;
4193 struct md_rdev *rdev;
4194 int spares = 0;
4195 int ret;
4196
4197 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4198 return -EBUSY;
4199
4200 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4201 return -EINVAL;
4202
4203 before_length = ((1 << conf->prev.chunk_shift) *
4204 conf->prev.far_copies);
4205 after_length = ((1 << conf->geo.chunk_shift) *
4206 conf->geo.far_copies);
4207
4208 rdev_for_each(rdev, mddev) {
4209 if (!test_bit(In_sync, &rdev->flags)
4210 && !test_bit(Faulty, &rdev->flags))
4211 spares++;
4212 if (rdev->raid_disk >= 0) {
4213 long long diff = (rdev->new_data_offset
4214 - rdev->data_offset);
4215 if (!mddev->reshape_backwards)
4216 diff = -diff;
4217 if (diff < 0)
4218 diff = 0;
4219 if (first || diff < min_offset_diff)
4220 min_offset_diff = diff;
4221 first = 0;
4222 }
4223 }
4224
4225 if (max(before_length, after_length) > min_offset_diff)
4226 return -EINVAL;
4227
4228 if (spares < mddev->delta_disks)
4229 return -EINVAL;
4230
4231 conf->offset_diff = min_offset_diff;
4232 spin_lock_irq(&conf->device_lock);
4233 if (conf->mirrors_new) {
4234 memcpy(conf->mirrors_new, conf->mirrors,
4235 sizeof(struct raid10_info)*conf->prev.raid_disks);
4236 smp_mb();
4237 kfree(conf->mirrors_old);
4238 conf->mirrors_old = conf->mirrors;
4239 conf->mirrors = conf->mirrors_new;
4240 conf->mirrors_new = NULL;
4241 }
4242 setup_geo(&conf->geo, mddev, geo_start);
4243 smp_mb();
4244 if (mddev->reshape_backwards) {
4245 sector_t size = raid10_size(mddev, 0, 0);
4246 if (size < mddev->array_sectors) {
4247 spin_unlock_irq(&conf->device_lock);
4248 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4249 mdname(mddev));
4250 return -EINVAL;
4251 }
4252 mddev->resync_max_sectors = size;
4253 conf->reshape_progress = size;
4254 } else
4255 conf->reshape_progress = 0;
4256 conf->reshape_safe = conf->reshape_progress;
4257 spin_unlock_irq(&conf->device_lock);
4258
4259 if (mddev->delta_disks && mddev->bitmap) {
4260 struct mdp_superblock_1 *sb = NULL;
4261 sector_t oldsize, newsize;
4262
4263 oldsize = raid10_size(mddev, 0, 0);
4264 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4265
4266 if (!mddev_is_clustered(mddev)) {
4267 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4268 if (ret)
4269 goto abort;
4270 else
4271 goto out;
4272 }
4273
4274 rdev_for_each(rdev, mddev) {
4275 if (rdev->raid_disk > -1 &&
4276 !test_bit(Faulty, &rdev->flags))
4277 sb = page_address(rdev->sb_page);
4278 }
4279
4280 /*
4281 * some node is already performing reshape, and no need to
4282 * call md_bitmap_resize again since it should be called when
4283 * receiving BITMAP_RESIZE msg
4284 */
4285 if ((sb && (le32_to_cpu(sb->feature_map) &
4286 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4287 goto out;
4288
4289 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4290 if (ret)
4291 goto abort;
4292
4293 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4294 if (ret) {
4295 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4296 goto abort;
4297 }
4298 }
4299out:
4300 if (mddev->delta_disks > 0) {
4301 rdev_for_each(rdev, mddev)
4302 if (rdev->raid_disk < 0 &&
4303 !test_bit(Faulty, &rdev->flags)) {
4304 if (raid10_add_disk(mddev, rdev) == 0) {
4305 if (rdev->raid_disk >=
4306 conf->prev.raid_disks)
4307 set_bit(In_sync, &rdev->flags);
4308 else
4309 rdev->recovery_offset = 0;
4310
4311 if (sysfs_link_rdev(mddev, rdev))
4312 /* Failure here is OK */;
4313 }
4314 } else if (rdev->raid_disk >= conf->prev.raid_disks
4315 && !test_bit(Faulty, &rdev->flags)) {
4316 /* This is a spare that was manually added */
4317 set_bit(In_sync, &rdev->flags);
4318 }
4319 }
4320 /* When a reshape changes the number of devices,
4321 * ->degraded is measured against the larger of the
4322 * pre and post numbers.
4323 */
4324 spin_lock_irq(&conf->device_lock);
4325 mddev->degraded = calc_degraded(conf);
4326 spin_unlock_irq(&conf->device_lock);
4327 mddev->raid_disks = conf->geo.raid_disks;
4328 mddev->reshape_position = conf->reshape_progress;
4329 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4330
4331 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4332 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4333 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4334 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4335 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4336
4337 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4338 "reshape");
4339 if (!mddev->sync_thread) {
4340 ret = -EAGAIN;
4341 goto abort;
4342 }
4343 conf->reshape_checkpoint = jiffies;
4344 md_wakeup_thread(mddev->sync_thread);
4345 md_new_event(mddev);
4346 return 0;
4347
4348abort:
4349 mddev->recovery = 0;
4350 spin_lock_irq(&conf->device_lock);
4351 conf->geo = conf->prev;
4352 mddev->raid_disks = conf->geo.raid_disks;
4353 rdev_for_each(rdev, mddev)
4354 rdev->new_data_offset = rdev->data_offset;
4355 smp_wmb();
4356 conf->reshape_progress = MaxSector;
4357 conf->reshape_safe = MaxSector;
4358 mddev->reshape_position = MaxSector;
4359 spin_unlock_irq(&conf->device_lock);
4360 return ret;
4361}
4362
4363/* Calculate the last device-address that could contain
4364 * any block from the chunk that includes the array-address 's'
4365 * and report the next address.
4366 * i.e. the address returned will be chunk-aligned and after
4367 * any data that is in the chunk containing 's'.
4368 */
4369static sector_t last_dev_address(sector_t s, struct geom *geo)
4370{
4371 s = (s | geo->chunk_mask) + 1;
4372 s >>= geo->chunk_shift;
4373 s *= geo->near_copies;
4374 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4375 s *= geo->far_copies;
4376 s <<= geo->chunk_shift;
4377 return s;
4378}
4379
4380/* Calculate the first device-address that could contain
4381 * any block from the chunk that includes the array-address 's'.
4382 * This too will be the start of a chunk
4383 */
4384static sector_t first_dev_address(sector_t s, struct geom *geo)
4385{
4386 s >>= geo->chunk_shift;
4387 s *= geo->near_copies;
4388 sector_div(s, geo->raid_disks);
4389 s *= geo->far_copies;
4390 s <<= geo->chunk_shift;
4391 return s;
4392}
4393
4394static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4395 int *skipped)
4396{
4397 /* We simply copy at most one chunk (smallest of old and new)
4398 * at a time, possibly less if that exceeds RESYNC_PAGES,
4399 * or we hit a bad block or something.
4400 * This might mean we pause for normal IO in the middle of
4401 * a chunk, but that is not a problem as mddev->reshape_position
4402 * can record any location.
4403 *
4404 * If we will want to write to a location that isn't
4405 * yet recorded as 'safe' (i.e. in metadata on disk) then
4406 * we need to flush all reshape requests and update the metadata.
4407 *
4408 * When reshaping forwards (e.g. to more devices), we interpret
4409 * 'safe' as the earliest block which might not have been copied
4410 * down yet. We divide this by previous stripe size and multiply
4411 * by previous stripe length to get lowest device offset that we
4412 * cannot write to yet.
4413 * We interpret 'sector_nr' as an address that we want to write to.
4414 * From this we use last_device_address() to find where we might
4415 * write to, and first_device_address on the 'safe' position.
4416 * If this 'next' write position is after the 'safe' position,
4417 * we must update the metadata to increase the 'safe' position.
4418 *
4419 * When reshaping backwards, we round in the opposite direction
4420 * and perform the reverse test: next write position must not be
4421 * less than current safe position.
4422 *
4423 * In all this the minimum difference in data offsets
4424 * (conf->offset_diff - always positive) allows a bit of slack,
4425 * so next can be after 'safe', but not by more than offset_diff
4426 *
4427 * We need to prepare all the bios here before we start any IO
4428 * to ensure the size we choose is acceptable to all devices.
4429 * The means one for each copy for write-out and an extra one for
4430 * read-in.
4431 * We store the read-in bio in ->master_bio and the others in
4432 * ->devs[x].bio and ->devs[x].repl_bio.
4433 */
4434 struct r10conf *conf = mddev->private;
4435 struct r10bio *r10_bio;
4436 sector_t next, safe, last;
4437 int max_sectors;
4438 int nr_sectors;
4439 int s;
4440 struct md_rdev *rdev;
4441 int need_flush = 0;
4442 struct bio *blist;
4443 struct bio *bio, *read_bio;
4444 int sectors_done = 0;
4445 struct page **pages;
4446
4447 if (sector_nr == 0) {
4448 /* If restarting in the middle, skip the initial sectors */
4449 if (mddev->reshape_backwards &&
4450 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4451 sector_nr = (raid10_size(mddev, 0, 0)
4452 - conf->reshape_progress);
4453 } else if (!mddev->reshape_backwards &&
4454 conf->reshape_progress > 0)
4455 sector_nr = conf->reshape_progress;
4456 if (sector_nr) {
4457 mddev->curr_resync_completed = sector_nr;
4458 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4459 *skipped = 1;
4460 return sector_nr;
4461 }
4462 }
4463
4464 /* We don't use sector_nr to track where we are up to
4465 * as that doesn't work well for ->reshape_backwards.
4466 * So just use ->reshape_progress.
4467 */
4468 if (mddev->reshape_backwards) {
4469 /* 'next' is the earliest device address that we might
4470 * write to for this chunk in the new layout
4471 */
4472 next = first_dev_address(conf->reshape_progress - 1,
4473 &conf->geo);
4474
4475 /* 'safe' is the last device address that we might read from
4476 * in the old layout after a restart
4477 */
4478 safe = last_dev_address(conf->reshape_safe - 1,
4479 &conf->prev);
4480
4481 if (next + conf->offset_diff < safe)
4482 need_flush = 1;
4483
4484 last = conf->reshape_progress - 1;
4485 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4486 & conf->prev.chunk_mask);
4487 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4488 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4489 } else {
4490 /* 'next' is after the last device address that we
4491 * might write to for this chunk in the new layout
4492 */
4493 next = last_dev_address(conf->reshape_progress, &conf->geo);
4494
4495 /* 'safe' is the earliest device address that we might
4496 * read from in the old layout after a restart
4497 */
4498 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4499
4500 /* Need to update metadata if 'next' might be beyond 'safe'
4501 * as that would possibly corrupt data
4502 */
4503 if (next > safe + conf->offset_diff)
4504 need_flush = 1;
4505
4506 sector_nr = conf->reshape_progress;
4507 last = sector_nr | (conf->geo.chunk_mask
4508 & conf->prev.chunk_mask);
4509
4510 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4511 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4512 }
4513
4514 if (need_flush ||
4515 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4516 /* Need to update reshape_position in metadata */
4517 wait_barrier(conf);
4518 mddev->reshape_position = conf->reshape_progress;
4519 if (mddev->reshape_backwards)
4520 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4521 - conf->reshape_progress;
4522 else
4523 mddev->curr_resync_completed = conf->reshape_progress;
4524 conf->reshape_checkpoint = jiffies;
4525 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4526 md_wakeup_thread(mddev->thread);
4527 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4528 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4529 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4530 allow_barrier(conf);
4531 return sectors_done;
4532 }
4533 conf->reshape_safe = mddev->reshape_position;
4534 allow_barrier(conf);
4535 }
4536
4537 raise_barrier(conf, 0);
4538read_more:
4539 /* Now schedule reads for blocks from sector_nr to last */
4540 r10_bio = raid10_alloc_init_r10buf(conf);
4541 r10_bio->state = 0;
4542 raise_barrier(conf, 1);
4543 atomic_set(&r10_bio->remaining, 0);
4544 r10_bio->mddev = mddev;
4545 r10_bio->sector = sector_nr;
4546 set_bit(R10BIO_IsReshape, &r10_bio->state);
4547 r10_bio->sectors = last - sector_nr + 1;
4548 rdev = read_balance(conf, r10_bio, &max_sectors);
4549 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4550
4551 if (!rdev) {
4552 /* Cannot read from here, so need to record bad blocks
4553 * on all the target devices.
4554 */
4555 // FIXME
4556 mempool_free(r10_bio, &conf->r10buf_pool);
4557 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4558 return sectors_done;
4559 }
4560
4561 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4562
4563 bio_set_dev(read_bio, rdev->bdev);
4564 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4565 + rdev->data_offset);
4566 read_bio->bi_private = r10_bio;
4567 read_bio->bi_end_io = end_reshape_read;
4568 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4569 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4570 read_bio->bi_status = 0;
4571 read_bio->bi_vcnt = 0;
4572 read_bio->bi_iter.bi_size = 0;
4573 r10_bio->master_bio = read_bio;
4574 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4575
4576 /*
4577 * Broadcast RESYNC message to other nodes, so all nodes would not
4578 * write to the region to avoid conflict.
4579 */
4580 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4581 struct mdp_superblock_1 *sb = NULL;
4582 int sb_reshape_pos = 0;
4583
4584 conf->cluster_sync_low = sector_nr;
4585 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4586 sb = page_address(rdev->sb_page);
4587 if (sb) {
4588 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4589 /*
4590 * Set cluster_sync_low again if next address for array
4591 * reshape is less than cluster_sync_low. Since we can't
4592 * update cluster_sync_low until it has finished reshape.
4593 */
4594 if (sb_reshape_pos < conf->cluster_sync_low)
4595 conf->cluster_sync_low = sb_reshape_pos;
4596 }
4597
4598 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4599 conf->cluster_sync_high);
4600 }
4601
4602 /* Now find the locations in the new layout */
4603 __raid10_find_phys(&conf->geo, r10_bio);
4604
4605 blist = read_bio;
4606 read_bio->bi_next = NULL;
4607
4608 rcu_read_lock();
4609 for (s = 0; s < conf->copies*2; s++) {
4610 struct bio *b;
4611 int d = r10_bio->devs[s/2].devnum;
4612 struct md_rdev *rdev2;
4613 if (s&1) {
4614 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4615 b = r10_bio->devs[s/2].repl_bio;
4616 } else {
4617 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4618 b = r10_bio->devs[s/2].bio;
4619 }
4620 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4621 continue;
4622
4623 bio_set_dev(b, rdev2->bdev);
4624 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4625 rdev2->new_data_offset;
4626 b->bi_end_io = end_reshape_write;
4627 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4628 b->bi_next = blist;
4629 blist = b;
4630 }
4631
4632 /* Now add as many pages as possible to all of these bios. */
4633
4634 nr_sectors = 0;
4635 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4636 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4637 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4638 int len = (max_sectors - s) << 9;
4639 if (len > PAGE_SIZE)
4640 len = PAGE_SIZE;
4641 for (bio = blist; bio ; bio = bio->bi_next) {
4642 /*
4643 * won't fail because the vec table is big enough
4644 * to hold all these pages
4645 */
4646 bio_add_page(bio, page, len, 0);
4647 }
4648 sector_nr += len >> 9;
4649 nr_sectors += len >> 9;
4650 }
4651 rcu_read_unlock();
4652 r10_bio->sectors = nr_sectors;
4653
4654 /* Now submit the read */
4655 md_sync_acct_bio(read_bio, r10_bio->sectors);
4656 atomic_inc(&r10_bio->remaining);
4657 read_bio->bi_next = NULL;
4658 generic_make_request(read_bio);
4659 sectors_done += nr_sectors;
4660 if (sector_nr <= last)
4661 goto read_more;
4662
4663 lower_barrier(conf);
4664
4665 /* Now that we have done the whole section we can
4666 * update reshape_progress
4667 */
4668 if (mddev->reshape_backwards)
4669 conf->reshape_progress -= sectors_done;
4670 else
4671 conf->reshape_progress += sectors_done;
4672
4673 return sectors_done;
4674}
4675
4676static void end_reshape_request(struct r10bio *r10_bio);
4677static int handle_reshape_read_error(struct mddev *mddev,
4678 struct r10bio *r10_bio);
4679static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4680{
4681 /* Reshape read completed. Hopefully we have a block
4682 * to write out.
4683 * If we got a read error then we do sync 1-page reads from
4684 * elsewhere until we find the data - or give up.
4685 */
4686 struct r10conf *conf = mddev->private;
4687 int s;
4688
4689 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4690 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4691 /* Reshape has been aborted */
4692 md_done_sync(mddev, r10_bio->sectors, 0);
4693 return;
4694 }
4695
4696 /* We definitely have the data in the pages, schedule the
4697 * writes.
4698 */
4699 atomic_set(&r10_bio->remaining, 1);
4700 for (s = 0; s < conf->copies*2; s++) {
4701 struct bio *b;
4702 int d = r10_bio->devs[s/2].devnum;
4703 struct md_rdev *rdev;
4704 rcu_read_lock();
4705 if (s&1) {
4706 rdev = rcu_dereference(conf->mirrors[d].replacement);
4707 b = r10_bio->devs[s/2].repl_bio;
4708 } else {
4709 rdev = rcu_dereference(conf->mirrors[d].rdev);
4710 b = r10_bio->devs[s/2].bio;
4711 }
4712 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4713 rcu_read_unlock();
4714 continue;
4715 }
4716 atomic_inc(&rdev->nr_pending);
4717 rcu_read_unlock();
4718 md_sync_acct_bio(b, r10_bio->sectors);
4719 atomic_inc(&r10_bio->remaining);
4720 b->bi_next = NULL;
4721 generic_make_request(b);
4722 }
4723 end_reshape_request(r10_bio);
4724}
4725
4726static void end_reshape(struct r10conf *conf)
4727{
4728 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4729 return;
4730
4731 spin_lock_irq(&conf->device_lock);
4732 conf->prev = conf->geo;
4733 md_finish_reshape(conf->mddev);
4734 smp_wmb();
4735 conf->reshape_progress = MaxSector;
4736 conf->reshape_safe = MaxSector;
4737 spin_unlock_irq(&conf->device_lock);
4738
4739 /* read-ahead size must cover two whole stripes, which is
4740 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4741 */
4742 if (conf->mddev->queue) {
4743 int stripe = conf->geo.raid_disks *
4744 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4745 stripe /= conf->geo.near_copies;
4746 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4747 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4748 }
4749 conf->fullsync = 0;
4750}
4751
4752static void raid10_update_reshape_pos(struct mddev *mddev)
4753{
4754 struct r10conf *conf = mddev->private;
4755 sector_t lo, hi;
4756
4757 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4758 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4759 || mddev->reshape_position == MaxSector)
4760 conf->reshape_progress = mddev->reshape_position;
4761 else
4762 WARN_ON_ONCE(1);
4763}
4764
4765static int handle_reshape_read_error(struct mddev *mddev,
4766 struct r10bio *r10_bio)
4767{
4768 /* Use sync reads to get the blocks from somewhere else */
4769 int sectors = r10_bio->sectors;
4770 struct r10conf *conf = mddev->private;
4771 struct r10bio *r10b;
4772 int slot = 0;
4773 int idx = 0;
4774 struct page **pages;
4775
4776 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4777 if (!r10b) {
4778 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4779 return -ENOMEM;
4780 }
4781
4782 /* reshape IOs share pages from .devs[0].bio */
4783 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4784
4785 r10b->sector = r10_bio->sector;
4786 __raid10_find_phys(&conf->prev, r10b);
4787
4788 while (sectors) {
4789 int s = sectors;
4790 int success = 0;
4791 int first_slot = slot;
4792
4793 if (s > (PAGE_SIZE >> 9))
4794 s = PAGE_SIZE >> 9;
4795
4796 rcu_read_lock();
4797 while (!success) {
4798 int d = r10b->devs[slot].devnum;
4799 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4800 sector_t addr;
4801 if (rdev == NULL ||
4802 test_bit(Faulty, &rdev->flags) ||
4803 !test_bit(In_sync, &rdev->flags))
4804 goto failed;
4805
4806 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4807 atomic_inc(&rdev->nr_pending);
4808 rcu_read_unlock();
4809 success = sync_page_io(rdev,
4810 addr,
4811 s << 9,
4812 pages[idx],
4813 REQ_OP_READ, 0, false);
4814 rdev_dec_pending(rdev, mddev);
4815 rcu_read_lock();
4816 if (success)
4817 break;
4818 failed:
4819 slot++;
4820 if (slot >= conf->copies)
4821 slot = 0;
4822 if (slot == first_slot)
4823 break;
4824 }
4825 rcu_read_unlock();
4826 if (!success) {
4827 /* couldn't read this block, must give up */
4828 set_bit(MD_RECOVERY_INTR,
4829 &mddev->recovery);
4830 kfree(r10b);
4831 return -EIO;
4832 }
4833 sectors -= s;
4834 idx++;
4835 }
4836 kfree(r10b);
4837 return 0;
4838}
4839
4840static void end_reshape_write(struct bio *bio)
4841{
4842 struct r10bio *r10_bio = get_resync_r10bio(bio);
4843 struct mddev *mddev = r10_bio->mddev;
4844 struct r10conf *conf = mddev->private;
4845 int d;
4846 int slot;
4847 int repl;
4848 struct md_rdev *rdev = NULL;
4849
4850 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4851 if (repl)
4852 rdev = conf->mirrors[d].replacement;
4853 if (!rdev) {
4854 smp_mb();
4855 rdev = conf->mirrors[d].rdev;
4856 }
4857
4858 if (bio->bi_status) {
4859 /* FIXME should record badblock */
4860 md_error(mddev, rdev);
4861 }
4862
4863 rdev_dec_pending(rdev, mddev);
4864 end_reshape_request(r10_bio);
4865}
4866
4867static void end_reshape_request(struct r10bio *r10_bio)
4868{
4869 if (!atomic_dec_and_test(&r10_bio->remaining))
4870 return;
4871 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4872 bio_put(r10_bio->master_bio);
4873 put_buf(r10_bio);
4874}
4875
4876static void raid10_finish_reshape(struct mddev *mddev)
4877{
4878 struct r10conf *conf = mddev->private;
4879
4880 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4881 return;
4882
4883 if (mddev->delta_disks > 0) {
4884 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4885 mddev->recovery_cp = mddev->resync_max_sectors;
4886 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4887 }
4888 mddev->resync_max_sectors = mddev->array_sectors;
4889 } else {
4890 int d;
4891 rcu_read_lock();
4892 for (d = conf->geo.raid_disks ;
4893 d < conf->geo.raid_disks - mddev->delta_disks;
4894 d++) {
4895 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4896 if (rdev)
4897 clear_bit(In_sync, &rdev->flags);
4898 rdev = rcu_dereference(conf->mirrors[d].replacement);
4899 if (rdev)
4900 clear_bit(In_sync, &rdev->flags);
4901 }
4902 rcu_read_unlock();
4903 }
4904 mddev->layout = mddev->new_layout;
4905 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4906 mddev->reshape_position = MaxSector;
4907 mddev->delta_disks = 0;
4908 mddev->reshape_backwards = 0;
4909}
4910
4911static struct md_personality raid10_personality =
4912{
4913 .name = "raid10",
4914 .level = 10,
4915 .owner = THIS_MODULE,
4916 .make_request = raid10_make_request,
4917 .run = raid10_run,
4918 .free = raid10_free,
4919 .status = raid10_status,
4920 .error_handler = raid10_error,
4921 .hot_add_disk = raid10_add_disk,
4922 .hot_remove_disk= raid10_remove_disk,
4923 .spare_active = raid10_spare_active,
4924 .sync_request = raid10_sync_request,
4925 .quiesce = raid10_quiesce,
4926 .size = raid10_size,
4927 .resize = raid10_resize,
4928 .takeover = raid10_takeover,
4929 .check_reshape = raid10_check_reshape,
4930 .start_reshape = raid10_start_reshape,
4931 .finish_reshape = raid10_finish_reshape,
4932 .update_reshape_pos = raid10_update_reshape_pos,
4933 .congested = raid10_congested,
4934};
4935
4936static int __init raid_init(void)
4937{
4938 return register_md_personality(&raid10_personality);
4939}
4940
4941static void raid_exit(void)
4942{
4943 unregister_md_personality(&raid10_personality);
4944}
4945
4946module_init(raid_init);
4947module_exit(raid_exit);
4948MODULE_LICENSE("GPL");
4949MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4950MODULE_ALIAS("md-personality-9"); /* RAID10 */
4951MODULE_ALIAS("md-raid10");
4952MODULE_ALIAS("md-level-10");
4953
4954module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
1/*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
21#include <linux/slab.h>
22#include <linux/delay.h>
23#include <linux/blkdev.h>
24#include <linux/module.h>
25#include <linux/seq_file.h>
26#include <linux/ratelimit.h>
27#include <linux/kthread.h>
28#include "md.h"
29#include "raid10.h"
30#include "raid0.h"
31#include "bitmap.h"
32
33/*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
42 * use_far_sets_bugfixed (stored in bit 18 of layout )
43 *
44 * The data to be stored is divided into chunks using chunksize. Each device
45 * is divided into far_copies sections. In each section, chunks are laid out
46 * in a style similar to raid0, but near_copies copies of each chunk is stored
47 * (each on a different drive). The starting device for each section is offset
48 * near_copies from the starting device of the previous section. Thus there
49 * are (near_copies * far_copies) of each chunk, and each is on a different
50 * drive. near_copies and far_copies must be at least one, and their product
51 * is at most raid_disks.
52 *
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of being very far
55 * apart on disk, there are adjacent stripes.
56 *
57 * The far and offset algorithms are handled slightly differently if
58 * 'use_far_sets' is true. In this case, the array's devices are grouped into
59 * sets that are (near_copies * far_copies) in size. The far copied stripes
60 * are still shifted by 'near_copies' devices, but this shifting stays confined
61 * to the set rather than the entire array. This is done to improve the number
62 * of device combinations that can fail without causing the array to fail.
63 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
64 * on a device):
65 * A B C D A B C D E
66 * ... ...
67 * D A B C E A B C D
68 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69 * [A B] [C D] [A B] [C D E]
70 * |...| |...| |...| | ... |
71 * [B A] [D C] [B A] [E C D]
72 */
73
74/*
75 * Number of guaranteed r10bios in case of extreme VM load:
76 */
77#define NR_RAID10_BIOS 256
78
79/* when we get a read error on a read-only array, we redirect to another
80 * device without failing the first device, or trying to over-write to
81 * correct the read error. To keep track of bad blocks on a per-bio
82 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 */
84#define IO_BLOCKED ((struct bio *)1)
85/* When we successfully write to a known bad-block, we need to remove the
86 * bad-block marking which must be done from process context. So we record
87 * the success by setting devs[n].bio to IO_MADE_GOOD
88 */
89#define IO_MADE_GOOD ((struct bio *)2)
90
91#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92
93/* When there are this many requests queued to be written by
94 * the raid10 thread, we become 'congested' to provide back-pressure
95 * for writeback.
96 */
97static int max_queued_requests = 1024;
98
99static void allow_barrier(struct r10conf *conf);
100static void lower_barrier(struct r10conf *conf);
101static int _enough(struct r10conf *conf, int previous, int ignore);
102static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103 int *skipped);
104static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
105static void end_reshape_write(struct bio *bio);
106static void end_reshape(struct r10conf *conf);
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->copies]);
112
113 /* allocate a r10bio with room for raid_disks entries in the
114 * bios array */
115 return kzalloc(size, gfp_flags);
116}
117
118static void r10bio_pool_free(void *r10_bio, void *data)
119{
120 kfree(r10_bio);
121}
122
123/* Maximum size of each resync request */
124#define RESYNC_BLOCK_SIZE (64*1024)
125#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
126/* amount of memory to reserve for resync requests */
127#define RESYNC_WINDOW (1024*1024)
128/* maximum number of concurrent requests, memory permitting */
129#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
130
131/*
132 * When performing a resync, we need to read and compare, so
133 * we need as many pages are there are copies.
134 * When performing a recovery, we need 2 bios, one for read,
135 * one for write (we recover only one drive per r10buf)
136 *
137 */
138static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
139{
140 struct r10conf *conf = data;
141 struct page *page;
142 struct r10bio *r10_bio;
143 struct bio *bio;
144 int i, j;
145 int nalloc;
146
147 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
148 if (!r10_bio)
149 return NULL;
150
151 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
152 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
153 nalloc = conf->copies; /* resync */
154 else
155 nalloc = 2; /* recovery */
156
157 /*
158 * Allocate bios.
159 */
160 for (j = nalloc ; j-- ; ) {
161 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
162 if (!bio)
163 goto out_free_bio;
164 r10_bio->devs[j].bio = bio;
165 if (!conf->have_replacement)
166 continue;
167 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
168 if (!bio)
169 goto out_free_bio;
170 r10_bio->devs[j].repl_bio = bio;
171 }
172 /*
173 * Allocate RESYNC_PAGES data pages and attach them
174 * where needed.
175 */
176 for (j = 0 ; j < nalloc; j++) {
177 struct bio *rbio = r10_bio->devs[j].repl_bio;
178 bio = r10_bio->devs[j].bio;
179 for (i = 0; i < RESYNC_PAGES; i++) {
180 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
181 &conf->mddev->recovery)) {
182 /* we can share bv_page's during recovery
183 * and reshape */
184 struct bio *rbio = r10_bio->devs[0].bio;
185 page = rbio->bi_io_vec[i].bv_page;
186 get_page(page);
187 } else
188 page = alloc_page(gfp_flags);
189 if (unlikely(!page))
190 goto out_free_pages;
191
192 bio->bi_io_vec[i].bv_page = page;
193 if (rbio)
194 rbio->bi_io_vec[i].bv_page = page;
195 }
196 }
197
198 return r10_bio;
199
200out_free_pages:
201 for ( ; i > 0 ; i--)
202 safe_put_page(bio->bi_io_vec[i-1].bv_page);
203 while (j--)
204 for (i = 0; i < RESYNC_PAGES ; i++)
205 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
206 j = 0;
207out_free_bio:
208 for ( ; j < nalloc; j++) {
209 if (r10_bio->devs[j].bio)
210 bio_put(r10_bio->devs[j].bio);
211 if (r10_bio->devs[j].repl_bio)
212 bio_put(r10_bio->devs[j].repl_bio);
213 }
214 r10bio_pool_free(r10_bio, conf);
215 return NULL;
216}
217
218static void r10buf_pool_free(void *__r10_bio, void *data)
219{
220 int i;
221 struct r10conf *conf = data;
222 struct r10bio *r10bio = __r10_bio;
223 int j;
224
225 for (j=0; j < conf->copies; j++) {
226 struct bio *bio = r10bio->devs[j].bio;
227 if (bio) {
228 for (i = 0; i < RESYNC_PAGES; i++) {
229 safe_put_page(bio->bi_io_vec[i].bv_page);
230 bio->bi_io_vec[i].bv_page = NULL;
231 }
232 bio_put(bio);
233 }
234 bio = r10bio->devs[j].repl_bio;
235 if (bio)
236 bio_put(bio);
237 }
238 r10bio_pool_free(r10bio, conf);
239}
240
241static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
242{
243 int i;
244
245 for (i = 0; i < conf->copies; i++) {
246 struct bio **bio = & r10_bio->devs[i].bio;
247 if (!BIO_SPECIAL(*bio))
248 bio_put(*bio);
249 *bio = NULL;
250 bio = &r10_bio->devs[i].repl_bio;
251 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
252 bio_put(*bio);
253 *bio = NULL;
254 }
255}
256
257static void free_r10bio(struct r10bio *r10_bio)
258{
259 struct r10conf *conf = r10_bio->mddev->private;
260
261 put_all_bios(conf, r10_bio);
262 mempool_free(r10_bio, conf->r10bio_pool);
263}
264
265static void put_buf(struct r10bio *r10_bio)
266{
267 struct r10conf *conf = r10_bio->mddev->private;
268
269 mempool_free(r10_bio, conf->r10buf_pool);
270
271 lower_barrier(conf);
272}
273
274static void reschedule_retry(struct r10bio *r10_bio)
275{
276 unsigned long flags;
277 struct mddev *mddev = r10_bio->mddev;
278 struct r10conf *conf = mddev->private;
279
280 spin_lock_irqsave(&conf->device_lock, flags);
281 list_add(&r10_bio->retry_list, &conf->retry_list);
282 conf->nr_queued ++;
283 spin_unlock_irqrestore(&conf->device_lock, flags);
284
285 /* wake up frozen array... */
286 wake_up(&conf->wait_barrier);
287
288 md_wakeup_thread(mddev->thread);
289}
290
291/*
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
294 * cache layer.
295 */
296static void raid_end_bio_io(struct r10bio *r10_bio)
297{
298 struct bio *bio = r10_bio->master_bio;
299 int done;
300 struct r10conf *conf = r10_bio->mddev->private;
301
302 if (bio->bi_phys_segments) {
303 unsigned long flags;
304 spin_lock_irqsave(&conf->device_lock, flags);
305 bio->bi_phys_segments--;
306 done = (bio->bi_phys_segments == 0);
307 spin_unlock_irqrestore(&conf->device_lock, flags);
308 } else
309 done = 1;
310 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
311 bio->bi_error = -EIO;
312 if (done) {
313 bio_endio(bio);
314 /*
315 * Wake up any possible resync thread that waits for the device
316 * to go idle.
317 */
318 allow_barrier(conf);
319 }
320 free_r10bio(r10_bio);
321}
322
323/*
324 * Update disk head position estimator based on IRQ completion info.
325 */
326static inline void update_head_pos(int slot, struct r10bio *r10_bio)
327{
328 struct r10conf *conf = r10_bio->mddev->private;
329
330 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
331 r10_bio->devs[slot].addr + (r10_bio->sectors);
332}
333
334/*
335 * Find the disk number which triggered given bio
336 */
337static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
338 struct bio *bio, int *slotp, int *replp)
339{
340 int slot;
341 int repl = 0;
342
343 for (slot = 0; slot < conf->copies; slot++) {
344 if (r10_bio->devs[slot].bio == bio)
345 break;
346 if (r10_bio->devs[slot].repl_bio == bio) {
347 repl = 1;
348 break;
349 }
350 }
351
352 BUG_ON(slot == conf->copies);
353 update_head_pos(slot, r10_bio);
354
355 if (slotp)
356 *slotp = slot;
357 if (replp)
358 *replp = repl;
359 return r10_bio->devs[slot].devnum;
360}
361
362static void raid10_end_read_request(struct bio *bio)
363{
364 int uptodate = !bio->bi_error;
365 struct r10bio *r10_bio = bio->bi_private;
366 int slot, dev;
367 struct md_rdev *rdev;
368 struct r10conf *conf = r10_bio->mddev->private;
369
370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
372 rdev = r10_bio->devs[slot].rdev;
373 /*
374 * this branch is our 'one mirror IO has finished' event handler:
375 */
376 update_head_pos(slot, r10_bio);
377
378 if (uptodate) {
379 /*
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
383 *
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
387 */
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
389 } else {
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
394 */
395 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
396 rdev->raid_disk))
397 uptodate = 1;
398 }
399 if (uptodate) {
400 raid_end_bio_io(r10_bio);
401 rdev_dec_pending(rdev, conf->mddev);
402 } else {
403 /*
404 * oops, read error - keep the refcount on the rdev
405 */
406 char b[BDEVNAME_SIZE];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
409 mdname(conf->mddev),
410 bdevname(rdev->bdev, b),
411 (unsigned long long)r10_bio->sector);
412 set_bit(R10BIO_ReadError, &r10_bio->state);
413 reschedule_retry(r10_bio);
414 }
415}
416
417static void close_write(struct r10bio *r10_bio)
418{
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421 r10_bio->sectors,
422 !test_bit(R10BIO_Degraded, &r10_bio->state),
423 0);
424 md_write_end(r10_bio->mddev);
425}
426
427static void one_write_done(struct r10bio *r10_bio)
428{
429 if (atomic_dec_and_test(&r10_bio->remaining)) {
430 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431 reschedule_retry(r10_bio);
432 else {
433 close_write(r10_bio);
434 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435 reschedule_retry(r10_bio);
436 else
437 raid_end_bio_io(r10_bio);
438 }
439 }
440}
441
442static void raid10_end_write_request(struct bio *bio)
443{
444 struct r10bio *r10_bio = bio->bi_private;
445 int dev;
446 int dec_rdev = 1;
447 struct r10conf *conf = r10_bio->mddev->private;
448 int slot, repl;
449 struct md_rdev *rdev = NULL;
450
451 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
452
453 if (repl)
454 rdev = conf->mirrors[dev].replacement;
455 if (!rdev) {
456 smp_rmb();
457 repl = 0;
458 rdev = conf->mirrors[dev].rdev;
459 }
460 /*
461 * this branch is our 'one mirror IO has finished' event handler:
462 */
463 if (bio->bi_error) {
464 if (repl)
465 /* Never record new bad blocks to replacement,
466 * just fail it.
467 */
468 md_error(rdev->mddev, rdev);
469 else {
470 set_bit(WriteErrorSeen, &rdev->flags);
471 if (!test_and_set_bit(WantReplacement, &rdev->flags))
472 set_bit(MD_RECOVERY_NEEDED,
473 &rdev->mddev->recovery);
474 set_bit(R10BIO_WriteError, &r10_bio->state);
475 dec_rdev = 0;
476 }
477 } else {
478 /*
479 * Set R10BIO_Uptodate in our master bio, so that
480 * we will return a good error code for to the higher
481 * levels even if IO on some other mirrored buffer fails.
482 *
483 * The 'master' represents the composite IO operation to
484 * user-side. So if something waits for IO, then it will
485 * wait for the 'master' bio.
486 */
487 sector_t first_bad;
488 int bad_sectors;
489
490 /*
491 * Do not set R10BIO_Uptodate if the current device is
492 * rebuilding or Faulty. This is because we cannot use
493 * such device for properly reading the data back (we could
494 * potentially use it, if the current write would have felt
495 * before rdev->recovery_offset, but for simplicity we don't
496 * check this here.
497 */
498 if (test_bit(In_sync, &rdev->flags) &&
499 !test_bit(Faulty, &rdev->flags))
500 set_bit(R10BIO_Uptodate, &r10_bio->state);
501
502 /* Maybe we can clear some bad blocks. */
503 if (is_badblock(rdev,
504 r10_bio->devs[slot].addr,
505 r10_bio->sectors,
506 &first_bad, &bad_sectors)) {
507 bio_put(bio);
508 if (repl)
509 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
510 else
511 r10_bio->devs[slot].bio = IO_MADE_GOOD;
512 dec_rdev = 0;
513 set_bit(R10BIO_MadeGood, &r10_bio->state);
514 }
515 }
516
517 /*
518 *
519 * Let's see if all mirrored write operations have finished
520 * already.
521 */
522 one_write_done(r10_bio);
523 if (dec_rdev)
524 rdev_dec_pending(rdev, conf->mddev);
525}
526
527/*
528 * RAID10 layout manager
529 * As well as the chunksize and raid_disks count, there are two
530 * parameters: near_copies and far_copies.
531 * near_copies * far_copies must be <= raid_disks.
532 * Normally one of these will be 1.
533 * If both are 1, we get raid0.
534 * If near_copies == raid_disks, we get raid1.
535 *
536 * Chunks are laid out in raid0 style with near_copies copies of the
537 * first chunk, followed by near_copies copies of the next chunk and
538 * so on.
539 * If far_copies > 1, then after 1/far_copies of the array has been assigned
540 * as described above, we start again with a device offset of near_copies.
541 * So we effectively have another copy of the whole array further down all
542 * the drives, but with blocks on different drives.
543 * With this layout, and block is never stored twice on the one device.
544 *
545 * raid10_find_phys finds the sector offset of a given virtual sector
546 * on each device that it is on.
547 *
548 * raid10_find_virt does the reverse mapping, from a device and a
549 * sector offset to a virtual address
550 */
551
552static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
553{
554 int n,f;
555 sector_t sector;
556 sector_t chunk;
557 sector_t stripe;
558 int dev;
559 int slot = 0;
560 int last_far_set_start, last_far_set_size;
561
562 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
563 last_far_set_start *= geo->far_set_size;
564
565 last_far_set_size = geo->far_set_size;
566 last_far_set_size += (geo->raid_disks % geo->far_set_size);
567
568 /* now calculate first sector/dev */
569 chunk = r10bio->sector >> geo->chunk_shift;
570 sector = r10bio->sector & geo->chunk_mask;
571
572 chunk *= geo->near_copies;
573 stripe = chunk;
574 dev = sector_div(stripe, geo->raid_disks);
575 if (geo->far_offset)
576 stripe *= geo->far_copies;
577
578 sector += stripe << geo->chunk_shift;
579
580 /* and calculate all the others */
581 for (n = 0; n < geo->near_copies; n++) {
582 int d = dev;
583 int set;
584 sector_t s = sector;
585 r10bio->devs[slot].devnum = d;
586 r10bio->devs[slot].addr = s;
587 slot++;
588
589 for (f = 1; f < geo->far_copies; f++) {
590 set = d / geo->far_set_size;
591 d += geo->near_copies;
592
593 if ((geo->raid_disks % geo->far_set_size) &&
594 (d > last_far_set_start)) {
595 d -= last_far_set_start;
596 d %= last_far_set_size;
597 d += last_far_set_start;
598 } else {
599 d %= geo->far_set_size;
600 d += geo->far_set_size * set;
601 }
602 s += geo->stride;
603 r10bio->devs[slot].devnum = d;
604 r10bio->devs[slot].addr = s;
605 slot++;
606 }
607 dev++;
608 if (dev >= geo->raid_disks) {
609 dev = 0;
610 sector += (geo->chunk_mask + 1);
611 }
612 }
613}
614
615static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
616{
617 struct geom *geo = &conf->geo;
618
619 if (conf->reshape_progress != MaxSector &&
620 ((r10bio->sector >= conf->reshape_progress) !=
621 conf->mddev->reshape_backwards)) {
622 set_bit(R10BIO_Previous, &r10bio->state);
623 geo = &conf->prev;
624 } else
625 clear_bit(R10BIO_Previous, &r10bio->state);
626
627 __raid10_find_phys(geo, r10bio);
628}
629
630static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
631{
632 sector_t offset, chunk, vchunk;
633 /* Never use conf->prev as this is only called during resync
634 * or recovery, so reshape isn't happening
635 */
636 struct geom *geo = &conf->geo;
637 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
638 int far_set_size = geo->far_set_size;
639 int last_far_set_start;
640
641 if (geo->raid_disks % geo->far_set_size) {
642 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
643 last_far_set_start *= geo->far_set_size;
644
645 if (dev >= last_far_set_start) {
646 far_set_size = geo->far_set_size;
647 far_set_size += (geo->raid_disks % geo->far_set_size);
648 far_set_start = last_far_set_start;
649 }
650 }
651
652 offset = sector & geo->chunk_mask;
653 if (geo->far_offset) {
654 int fc;
655 chunk = sector >> geo->chunk_shift;
656 fc = sector_div(chunk, geo->far_copies);
657 dev -= fc * geo->near_copies;
658 if (dev < far_set_start)
659 dev += far_set_size;
660 } else {
661 while (sector >= geo->stride) {
662 sector -= geo->stride;
663 if (dev < (geo->near_copies + far_set_start))
664 dev += far_set_size - geo->near_copies;
665 else
666 dev -= geo->near_copies;
667 }
668 chunk = sector >> geo->chunk_shift;
669 }
670 vchunk = chunk * geo->raid_disks + dev;
671 sector_div(vchunk, geo->near_copies);
672 return (vchunk << geo->chunk_shift) + offset;
673}
674
675/*
676 * This routine returns the disk from which the requested read should
677 * be done. There is a per-array 'next expected sequential IO' sector
678 * number - if this matches on the next IO then we use the last disk.
679 * There is also a per-disk 'last know head position' sector that is
680 * maintained from IRQ contexts, both the normal and the resync IO
681 * completion handlers update this position correctly. If there is no
682 * perfect sequential match then we pick the disk whose head is closest.
683 *
684 * If there are 2 mirrors in the same 2 devices, performance degrades
685 * because position is mirror, not device based.
686 *
687 * The rdev for the device selected will have nr_pending incremented.
688 */
689
690/*
691 * FIXME: possibly should rethink readbalancing and do it differently
692 * depending on near_copies / far_copies geometry.
693 */
694static struct md_rdev *read_balance(struct r10conf *conf,
695 struct r10bio *r10_bio,
696 int *max_sectors)
697{
698 const sector_t this_sector = r10_bio->sector;
699 int disk, slot;
700 int sectors = r10_bio->sectors;
701 int best_good_sectors;
702 sector_t new_distance, best_dist;
703 struct md_rdev *best_rdev, *rdev = NULL;
704 int do_balance;
705 int best_slot;
706 struct geom *geo = &conf->geo;
707
708 raid10_find_phys(conf, r10_bio);
709 rcu_read_lock();
710retry:
711 sectors = r10_bio->sectors;
712 best_slot = -1;
713 best_rdev = NULL;
714 best_dist = MaxSector;
715 best_good_sectors = 0;
716 do_balance = 1;
717 /*
718 * Check if we can balance. We can balance on the whole
719 * device if no resync is going on (recovery is ok), or below
720 * the resync window. We take the first readable disk when
721 * above the resync window.
722 */
723 if (conf->mddev->recovery_cp < MaxSector
724 && (this_sector + sectors >= conf->next_resync))
725 do_balance = 0;
726
727 for (slot = 0; slot < conf->copies ; slot++) {
728 sector_t first_bad;
729 int bad_sectors;
730 sector_t dev_sector;
731
732 if (r10_bio->devs[slot].bio == IO_BLOCKED)
733 continue;
734 disk = r10_bio->devs[slot].devnum;
735 rdev = rcu_dereference(conf->mirrors[disk].replacement);
736 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
737 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
738 rdev = rcu_dereference(conf->mirrors[disk].rdev);
739 if (rdev == NULL ||
740 test_bit(Faulty, &rdev->flags))
741 continue;
742 if (!test_bit(In_sync, &rdev->flags) &&
743 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
744 continue;
745
746 dev_sector = r10_bio->devs[slot].addr;
747 if (is_badblock(rdev, dev_sector, sectors,
748 &first_bad, &bad_sectors)) {
749 if (best_dist < MaxSector)
750 /* Already have a better slot */
751 continue;
752 if (first_bad <= dev_sector) {
753 /* Cannot read here. If this is the
754 * 'primary' device, then we must not read
755 * beyond 'bad_sectors' from another device.
756 */
757 bad_sectors -= (dev_sector - first_bad);
758 if (!do_balance && sectors > bad_sectors)
759 sectors = bad_sectors;
760 if (best_good_sectors > sectors)
761 best_good_sectors = sectors;
762 } else {
763 sector_t good_sectors =
764 first_bad - dev_sector;
765 if (good_sectors > best_good_sectors) {
766 best_good_sectors = good_sectors;
767 best_slot = slot;
768 best_rdev = rdev;
769 }
770 if (!do_balance)
771 /* Must read from here */
772 break;
773 }
774 continue;
775 } else
776 best_good_sectors = sectors;
777
778 if (!do_balance)
779 break;
780
781 /* This optimisation is debatable, and completely destroys
782 * sequential read speed for 'far copies' arrays. So only
783 * keep it for 'near' arrays, and review those later.
784 */
785 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
786 break;
787
788 /* for far > 1 always use the lowest address */
789 if (geo->far_copies > 1)
790 new_distance = r10_bio->devs[slot].addr;
791 else
792 new_distance = abs(r10_bio->devs[slot].addr -
793 conf->mirrors[disk].head_position);
794 if (new_distance < best_dist) {
795 best_dist = new_distance;
796 best_slot = slot;
797 best_rdev = rdev;
798 }
799 }
800 if (slot >= conf->copies) {
801 slot = best_slot;
802 rdev = best_rdev;
803 }
804
805 if (slot >= 0) {
806 atomic_inc(&rdev->nr_pending);
807 if (test_bit(Faulty, &rdev->flags)) {
808 /* Cannot risk returning a device that failed
809 * before we inc'ed nr_pending
810 */
811 rdev_dec_pending(rdev, conf->mddev);
812 goto retry;
813 }
814 r10_bio->read_slot = slot;
815 } else
816 rdev = NULL;
817 rcu_read_unlock();
818 *max_sectors = best_good_sectors;
819
820 return rdev;
821}
822
823static int raid10_congested(struct mddev *mddev, int bits)
824{
825 struct r10conf *conf = mddev->private;
826 int i, ret = 0;
827
828 if ((bits & (1 << WB_async_congested)) &&
829 conf->pending_count >= max_queued_requests)
830 return 1;
831
832 rcu_read_lock();
833 for (i = 0;
834 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
835 && ret == 0;
836 i++) {
837 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
838 if (rdev && !test_bit(Faulty, &rdev->flags)) {
839 struct request_queue *q = bdev_get_queue(rdev->bdev);
840
841 ret |= bdi_congested(&q->backing_dev_info, bits);
842 }
843 }
844 rcu_read_unlock();
845 return ret;
846}
847
848static void flush_pending_writes(struct r10conf *conf)
849{
850 /* Any writes that have been queued but are awaiting
851 * bitmap updates get flushed here.
852 */
853 spin_lock_irq(&conf->device_lock);
854
855 if (conf->pending_bio_list.head) {
856 struct bio *bio;
857 bio = bio_list_get(&conf->pending_bio_list);
858 conf->pending_count = 0;
859 spin_unlock_irq(&conf->device_lock);
860 /* flush any pending bitmap writes to disk
861 * before proceeding w/ I/O */
862 bitmap_unplug(conf->mddev->bitmap);
863 wake_up(&conf->wait_barrier);
864
865 while (bio) { /* submit pending writes */
866 struct bio *next = bio->bi_next;
867 bio->bi_next = NULL;
868 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
869 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
870 /* Just ignore it */
871 bio_endio(bio);
872 else
873 generic_make_request(bio);
874 bio = next;
875 }
876 } else
877 spin_unlock_irq(&conf->device_lock);
878}
879
880/* Barriers....
881 * Sometimes we need to suspend IO while we do something else,
882 * either some resync/recovery, or reconfigure the array.
883 * To do this we raise a 'barrier'.
884 * The 'barrier' is a counter that can be raised multiple times
885 * to count how many activities are happening which preclude
886 * normal IO.
887 * We can only raise the barrier if there is no pending IO.
888 * i.e. if nr_pending == 0.
889 * We choose only to raise the barrier if no-one is waiting for the
890 * barrier to go down. This means that as soon as an IO request
891 * is ready, no other operations which require a barrier will start
892 * until the IO request has had a chance.
893 *
894 * So: regular IO calls 'wait_barrier'. When that returns there
895 * is no backgroup IO happening, It must arrange to call
896 * allow_barrier when it has finished its IO.
897 * backgroup IO calls must call raise_barrier. Once that returns
898 * there is no normal IO happeing. It must arrange to call
899 * lower_barrier when the particular background IO completes.
900 */
901
902static void raise_barrier(struct r10conf *conf, int force)
903{
904 BUG_ON(force && !conf->barrier);
905 spin_lock_irq(&conf->resync_lock);
906
907 /* Wait until no block IO is waiting (unless 'force') */
908 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
909 conf->resync_lock);
910
911 /* block any new IO from starting */
912 conf->barrier++;
913
914 /* Now wait for all pending IO to complete */
915 wait_event_lock_irq(conf->wait_barrier,
916 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
917 conf->resync_lock);
918
919 spin_unlock_irq(&conf->resync_lock);
920}
921
922static void lower_barrier(struct r10conf *conf)
923{
924 unsigned long flags;
925 spin_lock_irqsave(&conf->resync_lock, flags);
926 conf->barrier--;
927 spin_unlock_irqrestore(&conf->resync_lock, flags);
928 wake_up(&conf->wait_barrier);
929}
930
931static void wait_barrier(struct r10conf *conf)
932{
933 spin_lock_irq(&conf->resync_lock);
934 if (conf->barrier) {
935 conf->nr_waiting++;
936 /* Wait for the barrier to drop.
937 * However if there are already pending
938 * requests (preventing the barrier from
939 * rising completely), and the
940 * pre-process bio queue isn't empty,
941 * then don't wait, as we need to empty
942 * that queue to get the nr_pending
943 * count down.
944 */
945 wait_event_lock_irq(conf->wait_barrier,
946 !conf->barrier ||
947 (conf->nr_pending &&
948 current->bio_list &&
949 !bio_list_empty(current->bio_list)),
950 conf->resync_lock);
951 conf->nr_waiting--;
952 }
953 conf->nr_pending++;
954 spin_unlock_irq(&conf->resync_lock);
955}
956
957static void allow_barrier(struct r10conf *conf)
958{
959 unsigned long flags;
960 spin_lock_irqsave(&conf->resync_lock, flags);
961 conf->nr_pending--;
962 spin_unlock_irqrestore(&conf->resync_lock, flags);
963 wake_up(&conf->wait_barrier);
964}
965
966static void freeze_array(struct r10conf *conf, int extra)
967{
968 /* stop syncio and normal IO and wait for everything to
969 * go quiet.
970 * We increment barrier and nr_waiting, and then
971 * wait until nr_pending match nr_queued+extra
972 * This is called in the context of one normal IO request
973 * that has failed. Thus any sync request that might be pending
974 * will be blocked by nr_pending, and we need to wait for
975 * pending IO requests to complete or be queued for re-try.
976 * Thus the number queued (nr_queued) plus this request (extra)
977 * must match the number of pending IOs (nr_pending) before
978 * we continue.
979 */
980 spin_lock_irq(&conf->resync_lock);
981 conf->barrier++;
982 conf->nr_waiting++;
983 wait_event_lock_irq_cmd(conf->wait_barrier,
984 conf->nr_pending == conf->nr_queued+extra,
985 conf->resync_lock,
986 flush_pending_writes(conf));
987
988 spin_unlock_irq(&conf->resync_lock);
989}
990
991static void unfreeze_array(struct r10conf *conf)
992{
993 /* reverse the effect of the freeze */
994 spin_lock_irq(&conf->resync_lock);
995 conf->barrier--;
996 conf->nr_waiting--;
997 wake_up(&conf->wait_barrier);
998 spin_unlock_irq(&conf->resync_lock);
999}
1000
1001static sector_t choose_data_offset(struct r10bio *r10_bio,
1002 struct md_rdev *rdev)
1003{
1004 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1005 test_bit(R10BIO_Previous, &r10_bio->state))
1006 return rdev->data_offset;
1007 else
1008 return rdev->new_data_offset;
1009}
1010
1011struct raid10_plug_cb {
1012 struct blk_plug_cb cb;
1013 struct bio_list pending;
1014 int pending_cnt;
1015};
1016
1017static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1018{
1019 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1020 cb);
1021 struct mddev *mddev = plug->cb.data;
1022 struct r10conf *conf = mddev->private;
1023 struct bio *bio;
1024
1025 if (from_schedule || current->bio_list) {
1026 spin_lock_irq(&conf->device_lock);
1027 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1028 conf->pending_count += plug->pending_cnt;
1029 spin_unlock_irq(&conf->device_lock);
1030 wake_up(&conf->wait_barrier);
1031 md_wakeup_thread(mddev->thread);
1032 kfree(plug);
1033 return;
1034 }
1035
1036 /* we aren't scheduling, so we can do the write-out directly. */
1037 bio = bio_list_get(&plug->pending);
1038 bitmap_unplug(mddev->bitmap);
1039 wake_up(&conf->wait_barrier);
1040
1041 while (bio) { /* submit pending writes */
1042 struct bio *next = bio->bi_next;
1043 bio->bi_next = NULL;
1044 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1045 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1046 /* Just ignore it */
1047 bio_endio(bio);
1048 else
1049 generic_make_request(bio);
1050 bio = next;
1051 }
1052 kfree(plug);
1053}
1054
1055static void __make_request(struct mddev *mddev, struct bio *bio)
1056{
1057 struct r10conf *conf = mddev->private;
1058 struct r10bio *r10_bio;
1059 struct bio *read_bio;
1060 int i;
1061 const int rw = bio_data_dir(bio);
1062 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1063 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1064 const unsigned long do_discard = (bio->bi_rw
1065 & (REQ_DISCARD | REQ_SECURE));
1066 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1067 unsigned long flags;
1068 struct md_rdev *blocked_rdev;
1069 struct blk_plug_cb *cb;
1070 struct raid10_plug_cb *plug = NULL;
1071 int sectors_handled;
1072 int max_sectors;
1073 int sectors;
1074
1075 /*
1076 * Register the new request and wait if the reconstruction
1077 * thread has put up a bar for new requests.
1078 * Continue immediately if no resync is active currently.
1079 */
1080 wait_barrier(conf);
1081
1082 sectors = bio_sectors(bio);
1083 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1084 bio->bi_iter.bi_sector < conf->reshape_progress &&
1085 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1086 /* IO spans the reshape position. Need to wait for
1087 * reshape to pass
1088 */
1089 allow_barrier(conf);
1090 wait_event(conf->wait_barrier,
1091 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1092 conf->reshape_progress >= bio->bi_iter.bi_sector +
1093 sectors);
1094 wait_barrier(conf);
1095 }
1096 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1097 bio_data_dir(bio) == WRITE &&
1098 (mddev->reshape_backwards
1099 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1100 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1101 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1102 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1103 /* Need to update reshape_position in metadata */
1104 mddev->reshape_position = conf->reshape_progress;
1105 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1106 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1107 md_wakeup_thread(mddev->thread);
1108 wait_event(mddev->sb_wait,
1109 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1110
1111 conf->reshape_safe = mddev->reshape_position;
1112 }
1113
1114 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1115
1116 r10_bio->master_bio = bio;
1117 r10_bio->sectors = sectors;
1118
1119 r10_bio->mddev = mddev;
1120 r10_bio->sector = bio->bi_iter.bi_sector;
1121 r10_bio->state = 0;
1122
1123 /* We might need to issue multiple reads to different
1124 * devices if there are bad blocks around, so we keep
1125 * track of the number of reads in bio->bi_phys_segments.
1126 * If this is 0, there is only one r10_bio and no locking
1127 * will be needed when the request completes. If it is
1128 * non-zero, then it is the number of not-completed requests.
1129 */
1130 bio->bi_phys_segments = 0;
1131 bio_clear_flag(bio, BIO_SEG_VALID);
1132
1133 if (rw == READ) {
1134 /*
1135 * read balancing logic:
1136 */
1137 struct md_rdev *rdev;
1138 int slot;
1139
1140read_again:
1141 rdev = read_balance(conf, r10_bio, &max_sectors);
1142 if (!rdev) {
1143 raid_end_bio_io(r10_bio);
1144 return;
1145 }
1146 slot = r10_bio->read_slot;
1147
1148 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1149 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1150 max_sectors);
1151
1152 r10_bio->devs[slot].bio = read_bio;
1153 r10_bio->devs[slot].rdev = rdev;
1154
1155 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1156 choose_data_offset(r10_bio, rdev);
1157 read_bio->bi_bdev = rdev->bdev;
1158 read_bio->bi_end_io = raid10_end_read_request;
1159 read_bio->bi_rw = READ | do_sync;
1160 read_bio->bi_private = r10_bio;
1161
1162 if (max_sectors < r10_bio->sectors) {
1163 /* Could not read all from this device, so we will
1164 * need another r10_bio.
1165 */
1166 sectors_handled = (r10_bio->sector + max_sectors
1167 - bio->bi_iter.bi_sector);
1168 r10_bio->sectors = max_sectors;
1169 spin_lock_irq(&conf->device_lock);
1170 if (bio->bi_phys_segments == 0)
1171 bio->bi_phys_segments = 2;
1172 else
1173 bio->bi_phys_segments++;
1174 spin_unlock_irq(&conf->device_lock);
1175 /* Cannot call generic_make_request directly
1176 * as that will be queued in __generic_make_request
1177 * and subsequent mempool_alloc might block
1178 * waiting for it. so hand bio over to raid10d.
1179 */
1180 reschedule_retry(r10_bio);
1181
1182 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1183
1184 r10_bio->master_bio = bio;
1185 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1186 r10_bio->state = 0;
1187 r10_bio->mddev = mddev;
1188 r10_bio->sector = bio->bi_iter.bi_sector +
1189 sectors_handled;
1190 goto read_again;
1191 } else
1192 generic_make_request(read_bio);
1193 return;
1194 }
1195
1196 /*
1197 * WRITE:
1198 */
1199 if (conf->pending_count >= max_queued_requests) {
1200 md_wakeup_thread(mddev->thread);
1201 wait_event(conf->wait_barrier,
1202 conf->pending_count < max_queued_requests);
1203 }
1204 /* first select target devices under rcu_lock and
1205 * inc refcount on their rdev. Record them by setting
1206 * bios[x] to bio
1207 * If there are known/acknowledged bad blocks on any device
1208 * on which we have seen a write error, we want to avoid
1209 * writing to those blocks. This potentially requires several
1210 * writes to write around the bad blocks. Each set of writes
1211 * gets its own r10_bio with a set of bios attached. The number
1212 * of r10_bios is recored in bio->bi_phys_segments just as with
1213 * the read case.
1214 */
1215
1216 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1217 raid10_find_phys(conf, r10_bio);
1218retry_write:
1219 blocked_rdev = NULL;
1220 rcu_read_lock();
1221 max_sectors = r10_bio->sectors;
1222
1223 for (i = 0; i < conf->copies; i++) {
1224 int d = r10_bio->devs[i].devnum;
1225 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1226 struct md_rdev *rrdev = rcu_dereference(
1227 conf->mirrors[d].replacement);
1228 if (rdev == rrdev)
1229 rrdev = NULL;
1230 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1231 atomic_inc(&rdev->nr_pending);
1232 blocked_rdev = rdev;
1233 break;
1234 }
1235 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1236 atomic_inc(&rrdev->nr_pending);
1237 blocked_rdev = rrdev;
1238 break;
1239 }
1240 if (rdev && (test_bit(Faulty, &rdev->flags)))
1241 rdev = NULL;
1242 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1243 rrdev = NULL;
1244
1245 r10_bio->devs[i].bio = NULL;
1246 r10_bio->devs[i].repl_bio = NULL;
1247
1248 if (!rdev && !rrdev) {
1249 set_bit(R10BIO_Degraded, &r10_bio->state);
1250 continue;
1251 }
1252 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1253 sector_t first_bad;
1254 sector_t dev_sector = r10_bio->devs[i].addr;
1255 int bad_sectors;
1256 int is_bad;
1257
1258 is_bad = is_badblock(rdev, dev_sector,
1259 max_sectors,
1260 &first_bad, &bad_sectors);
1261 if (is_bad < 0) {
1262 /* Mustn't write here until the bad block
1263 * is acknowledged
1264 */
1265 atomic_inc(&rdev->nr_pending);
1266 set_bit(BlockedBadBlocks, &rdev->flags);
1267 blocked_rdev = rdev;
1268 break;
1269 }
1270 if (is_bad && first_bad <= dev_sector) {
1271 /* Cannot write here at all */
1272 bad_sectors -= (dev_sector - first_bad);
1273 if (bad_sectors < max_sectors)
1274 /* Mustn't write more than bad_sectors
1275 * to other devices yet
1276 */
1277 max_sectors = bad_sectors;
1278 /* We don't set R10BIO_Degraded as that
1279 * only applies if the disk is missing,
1280 * so it might be re-added, and we want to
1281 * know to recover this chunk.
1282 * In this case the device is here, and the
1283 * fact that this chunk is not in-sync is
1284 * recorded in the bad block log.
1285 */
1286 continue;
1287 }
1288 if (is_bad) {
1289 int good_sectors = first_bad - dev_sector;
1290 if (good_sectors < max_sectors)
1291 max_sectors = good_sectors;
1292 }
1293 }
1294 if (rdev) {
1295 r10_bio->devs[i].bio = bio;
1296 atomic_inc(&rdev->nr_pending);
1297 }
1298 if (rrdev) {
1299 r10_bio->devs[i].repl_bio = bio;
1300 atomic_inc(&rrdev->nr_pending);
1301 }
1302 }
1303 rcu_read_unlock();
1304
1305 if (unlikely(blocked_rdev)) {
1306 /* Have to wait for this device to get unblocked, then retry */
1307 int j;
1308 int d;
1309
1310 for (j = 0; j < i; j++) {
1311 if (r10_bio->devs[j].bio) {
1312 d = r10_bio->devs[j].devnum;
1313 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1314 }
1315 if (r10_bio->devs[j].repl_bio) {
1316 struct md_rdev *rdev;
1317 d = r10_bio->devs[j].devnum;
1318 rdev = conf->mirrors[d].replacement;
1319 if (!rdev) {
1320 /* Race with remove_disk */
1321 smp_mb();
1322 rdev = conf->mirrors[d].rdev;
1323 }
1324 rdev_dec_pending(rdev, mddev);
1325 }
1326 }
1327 allow_barrier(conf);
1328 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1329 wait_barrier(conf);
1330 goto retry_write;
1331 }
1332
1333 if (max_sectors < r10_bio->sectors) {
1334 /* We are splitting this into multiple parts, so
1335 * we need to prepare for allocating another r10_bio.
1336 */
1337 r10_bio->sectors = max_sectors;
1338 spin_lock_irq(&conf->device_lock);
1339 if (bio->bi_phys_segments == 0)
1340 bio->bi_phys_segments = 2;
1341 else
1342 bio->bi_phys_segments++;
1343 spin_unlock_irq(&conf->device_lock);
1344 }
1345 sectors_handled = r10_bio->sector + max_sectors -
1346 bio->bi_iter.bi_sector;
1347
1348 atomic_set(&r10_bio->remaining, 1);
1349 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1350
1351 for (i = 0; i < conf->copies; i++) {
1352 struct bio *mbio;
1353 int d = r10_bio->devs[i].devnum;
1354 if (r10_bio->devs[i].bio) {
1355 struct md_rdev *rdev = conf->mirrors[d].rdev;
1356 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1357 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1358 max_sectors);
1359 r10_bio->devs[i].bio = mbio;
1360
1361 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1362 choose_data_offset(r10_bio,
1363 rdev));
1364 mbio->bi_bdev = rdev->bdev;
1365 mbio->bi_end_io = raid10_end_write_request;
1366 mbio->bi_rw =
1367 WRITE | do_sync | do_fua | do_discard | do_same;
1368 mbio->bi_private = r10_bio;
1369
1370 atomic_inc(&r10_bio->remaining);
1371
1372 cb = blk_check_plugged(raid10_unplug, mddev,
1373 sizeof(*plug));
1374 if (cb)
1375 plug = container_of(cb, struct raid10_plug_cb,
1376 cb);
1377 else
1378 plug = NULL;
1379 spin_lock_irqsave(&conf->device_lock, flags);
1380 if (plug) {
1381 bio_list_add(&plug->pending, mbio);
1382 plug->pending_cnt++;
1383 } else {
1384 bio_list_add(&conf->pending_bio_list, mbio);
1385 conf->pending_count++;
1386 }
1387 spin_unlock_irqrestore(&conf->device_lock, flags);
1388 if (!plug)
1389 md_wakeup_thread(mddev->thread);
1390 }
1391
1392 if (r10_bio->devs[i].repl_bio) {
1393 struct md_rdev *rdev = conf->mirrors[d].replacement;
1394 if (rdev == NULL) {
1395 /* Replacement just got moved to main 'rdev' */
1396 smp_mb();
1397 rdev = conf->mirrors[d].rdev;
1398 }
1399 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1400 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1401 max_sectors);
1402 r10_bio->devs[i].repl_bio = mbio;
1403
1404 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1405 choose_data_offset(
1406 r10_bio, rdev));
1407 mbio->bi_bdev = rdev->bdev;
1408 mbio->bi_end_io = raid10_end_write_request;
1409 mbio->bi_rw =
1410 WRITE | do_sync | do_fua | do_discard | do_same;
1411 mbio->bi_private = r10_bio;
1412
1413 atomic_inc(&r10_bio->remaining);
1414 spin_lock_irqsave(&conf->device_lock, flags);
1415 bio_list_add(&conf->pending_bio_list, mbio);
1416 conf->pending_count++;
1417 spin_unlock_irqrestore(&conf->device_lock, flags);
1418 if (!mddev_check_plugged(mddev))
1419 md_wakeup_thread(mddev->thread);
1420 }
1421 }
1422
1423 /* Don't remove the bias on 'remaining' (one_write_done) until
1424 * after checking if we need to go around again.
1425 */
1426
1427 if (sectors_handled < bio_sectors(bio)) {
1428 one_write_done(r10_bio);
1429 /* We need another r10_bio. It has already been counted
1430 * in bio->bi_phys_segments.
1431 */
1432 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1433
1434 r10_bio->master_bio = bio;
1435 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1436
1437 r10_bio->mddev = mddev;
1438 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1439 r10_bio->state = 0;
1440 goto retry_write;
1441 }
1442 one_write_done(r10_bio);
1443}
1444
1445static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1446{
1447 struct r10conf *conf = mddev->private;
1448 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1449 int chunk_sects = chunk_mask + 1;
1450
1451 struct bio *split;
1452
1453 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1454 md_flush_request(mddev, bio);
1455 return;
1456 }
1457
1458 md_write_start(mddev, bio);
1459
1460 do {
1461
1462 /*
1463 * If this request crosses a chunk boundary, we need to split
1464 * it.
1465 */
1466 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1467 bio_sectors(bio) > chunk_sects
1468 && (conf->geo.near_copies < conf->geo.raid_disks
1469 || conf->prev.near_copies <
1470 conf->prev.raid_disks))) {
1471 split = bio_split(bio, chunk_sects -
1472 (bio->bi_iter.bi_sector &
1473 (chunk_sects - 1)),
1474 GFP_NOIO, fs_bio_set);
1475 bio_chain(split, bio);
1476 } else {
1477 split = bio;
1478 }
1479
1480 __make_request(mddev, split);
1481 } while (split != bio);
1482
1483 /* In case raid10d snuck in to freeze_array */
1484 wake_up(&conf->wait_barrier);
1485}
1486
1487static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1488{
1489 struct r10conf *conf = mddev->private;
1490 int i;
1491
1492 if (conf->geo.near_copies < conf->geo.raid_disks)
1493 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1494 if (conf->geo.near_copies > 1)
1495 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1496 if (conf->geo.far_copies > 1) {
1497 if (conf->geo.far_offset)
1498 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1499 else
1500 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1501 if (conf->geo.far_set_size != conf->geo.raid_disks)
1502 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1503 }
1504 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1505 conf->geo.raid_disks - mddev->degraded);
1506 for (i = 0; i < conf->geo.raid_disks; i++)
1507 seq_printf(seq, "%s",
1508 conf->mirrors[i].rdev &&
1509 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1510 seq_printf(seq, "]");
1511}
1512
1513/* check if there are enough drives for
1514 * every block to appear on atleast one.
1515 * Don't consider the device numbered 'ignore'
1516 * as we might be about to remove it.
1517 */
1518static int _enough(struct r10conf *conf, int previous, int ignore)
1519{
1520 int first = 0;
1521 int has_enough = 0;
1522 int disks, ncopies;
1523 if (previous) {
1524 disks = conf->prev.raid_disks;
1525 ncopies = conf->prev.near_copies;
1526 } else {
1527 disks = conf->geo.raid_disks;
1528 ncopies = conf->geo.near_copies;
1529 }
1530
1531 rcu_read_lock();
1532 do {
1533 int n = conf->copies;
1534 int cnt = 0;
1535 int this = first;
1536 while (n--) {
1537 struct md_rdev *rdev;
1538 if (this != ignore &&
1539 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1540 test_bit(In_sync, &rdev->flags))
1541 cnt++;
1542 this = (this+1) % disks;
1543 }
1544 if (cnt == 0)
1545 goto out;
1546 first = (first + ncopies) % disks;
1547 } while (first != 0);
1548 has_enough = 1;
1549out:
1550 rcu_read_unlock();
1551 return has_enough;
1552}
1553
1554static int enough(struct r10conf *conf, int ignore)
1555{
1556 /* when calling 'enough', both 'prev' and 'geo' must
1557 * be stable.
1558 * This is ensured if ->reconfig_mutex or ->device_lock
1559 * is held.
1560 */
1561 return _enough(conf, 0, ignore) &&
1562 _enough(conf, 1, ignore);
1563}
1564
1565static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1566{
1567 char b[BDEVNAME_SIZE];
1568 struct r10conf *conf = mddev->private;
1569 unsigned long flags;
1570
1571 /*
1572 * If it is not operational, then we have already marked it as dead
1573 * else if it is the last working disks, ignore the error, let the
1574 * next level up know.
1575 * else mark the drive as failed
1576 */
1577 spin_lock_irqsave(&conf->device_lock, flags);
1578 if (test_bit(In_sync, &rdev->flags)
1579 && !enough(conf, rdev->raid_disk)) {
1580 /*
1581 * Don't fail the drive, just return an IO error.
1582 */
1583 spin_unlock_irqrestore(&conf->device_lock, flags);
1584 return;
1585 }
1586 if (test_and_clear_bit(In_sync, &rdev->flags))
1587 mddev->degraded++;
1588 /*
1589 * If recovery is running, make sure it aborts.
1590 */
1591 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1592 set_bit(Blocked, &rdev->flags);
1593 set_bit(Faulty, &rdev->flags);
1594 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1595 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1596 spin_unlock_irqrestore(&conf->device_lock, flags);
1597 printk(KERN_ALERT
1598 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1599 "md/raid10:%s: Operation continuing on %d devices.\n",
1600 mdname(mddev), bdevname(rdev->bdev, b),
1601 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1602}
1603
1604static void print_conf(struct r10conf *conf)
1605{
1606 int i;
1607 struct raid10_info *tmp;
1608
1609 printk(KERN_DEBUG "RAID10 conf printout:\n");
1610 if (!conf) {
1611 printk(KERN_DEBUG "(!conf)\n");
1612 return;
1613 }
1614 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1615 conf->geo.raid_disks);
1616
1617 for (i = 0; i < conf->geo.raid_disks; i++) {
1618 char b[BDEVNAME_SIZE];
1619 tmp = conf->mirrors + i;
1620 if (tmp->rdev)
1621 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1622 i, !test_bit(In_sync, &tmp->rdev->flags),
1623 !test_bit(Faulty, &tmp->rdev->flags),
1624 bdevname(tmp->rdev->bdev,b));
1625 }
1626}
1627
1628static void close_sync(struct r10conf *conf)
1629{
1630 wait_barrier(conf);
1631 allow_barrier(conf);
1632
1633 mempool_destroy(conf->r10buf_pool);
1634 conf->r10buf_pool = NULL;
1635}
1636
1637static int raid10_spare_active(struct mddev *mddev)
1638{
1639 int i;
1640 struct r10conf *conf = mddev->private;
1641 struct raid10_info *tmp;
1642 int count = 0;
1643 unsigned long flags;
1644
1645 /*
1646 * Find all non-in_sync disks within the RAID10 configuration
1647 * and mark them in_sync
1648 */
1649 for (i = 0; i < conf->geo.raid_disks; i++) {
1650 tmp = conf->mirrors + i;
1651 if (tmp->replacement
1652 && tmp->replacement->recovery_offset == MaxSector
1653 && !test_bit(Faulty, &tmp->replacement->flags)
1654 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1655 /* Replacement has just become active */
1656 if (!tmp->rdev
1657 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1658 count++;
1659 if (tmp->rdev) {
1660 /* Replaced device not technically faulty,
1661 * but we need to be sure it gets removed
1662 * and never re-added.
1663 */
1664 set_bit(Faulty, &tmp->rdev->flags);
1665 sysfs_notify_dirent_safe(
1666 tmp->rdev->sysfs_state);
1667 }
1668 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1669 } else if (tmp->rdev
1670 && tmp->rdev->recovery_offset == MaxSector
1671 && !test_bit(Faulty, &tmp->rdev->flags)
1672 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1673 count++;
1674 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1675 }
1676 }
1677 spin_lock_irqsave(&conf->device_lock, flags);
1678 mddev->degraded -= count;
1679 spin_unlock_irqrestore(&conf->device_lock, flags);
1680
1681 print_conf(conf);
1682 return count;
1683}
1684
1685static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1686{
1687 struct r10conf *conf = mddev->private;
1688 int err = -EEXIST;
1689 int mirror;
1690 int first = 0;
1691 int last = conf->geo.raid_disks - 1;
1692
1693 if (mddev->recovery_cp < MaxSector)
1694 /* only hot-add to in-sync arrays, as recovery is
1695 * very different from resync
1696 */
1697 return -EBUSY;
1698 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1699 return -EINVAL;
1700
1701 if (md_integrity_add_rdev(rdev, mddev))
1702 return -ENXIO;
1703
1704 if (rdev->raid_disk >= 0)
1705 first = last = rdev->raid_disk;
1706
1707 if (rdev->saved_raid_disk >= first &&
1708 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1709 mirror = rdev->saved_raid_disk;
1710 else
1711 mirror = first;
1712 for ( ; mirror <= last ; mirror++) {
1713 struct raid10_info *p = &conf->mirrors[mirror];
1714 if (p->recovery_disabled == mddev->recovery_disabled)
1715 continue;
1716 if (p->rdev) {
1717 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1718 p->replacement != NULL)
1719 continue;
1720 clear_bit(In_sync, &rdev->flags);
1721 set_bit(Replacement, &rdev->flags);
1722 rdev->raid_disk = mirror;
1723 err = 0;
1724 if (mddev->gendisk)
1725 disk_stack_limits(mddev->gendisk, rdev->bdev,
1726 rdev->data_offset << 9);
1727 conf->fullsync = 1;
1728 rcu_assign_pointer(p->replacement, rdev);
1729 break;
1730 }
1731
1732 if (mddev->gendisk)
1733 disk_stack_limits(mddev->gendisk, rdev->bdev,
1734 rdev->data_offset << 9);
1735
1736 p->head_position = 0;
1737 p->recovery_disabled = mddev->recovery_disabled - 1;
1738 rdev->raid_disk = mirror;
1739 err = 0;
1740 if (rdev->saved_raid_disk != mirror)
1741 conf->fullsync = 1;
1742 rcu_assign_pointer(p->rdev, rdev);
1743 break;
1744 }
1745 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1746 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1747
1748 print_conf(conf);
1749 return err;
1750}
1751
1752static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1753{
1754 struct r10conf *conf = mddev->private;
1755 int err = 0;
1756 int number = rdev->raid_disk;
1757 struct md_rdev **rdevp;
1758 struct raid10_info *p = conf->mirrors + number;
1759
1760 print_conf(conf);
1761 if (rdev == p->rdev)
1762 rdevp = &p->rdev;
1763 else if (rdev == p->replacement)
1764 rdevp = &p->replacement;
1765 else
1766 return 0;
1767
1768 if (test_bit(In_sync, &rdev->flags) ||
1769 atomic_read(&rdev->nr_pending)) {
1770 err = -EBUSY;
1771 goto abort;
1772 }
1773 /* Only remove faulty devices if recovery
1774 * is not possible.
1775 */
1776 if (!test_bit(Faulty, &rdev->flags) &&
1777 mddev->recovery_disabled != p->recovery_disabled &&
1778 (!p->replacement || p->replacement == rdev) &&
1779 number < conf->geo.raid_disks &&
1780 enough(conf, -1)) {
1781 err = -EBUSY;
1782 goto abort;
1783 }
1784 *rdevp = NULL;
1785 synchronize_rcu();
1786 if (atomic_read(&rdev->nr_pending)) {
1787 /* lost the race, try later */
1788 err = -EBUSY;
1789 *rdevp = rdev;
1790 goto abort;
1791 } else if (p->replacement) {
1792 /* We must have just cleared 'rdev' */
1793 p->rdev = p->replacement;
1794 clear_bit(Replacement, &p->replacement->flags);
1795 smp_mb(); /* Make sure other CPUs may see both as identical
1796 * but will never see neither -- if they are careful.
1797 */
1798 p->replacement = NULL;
1799 clear_bit(WantReplacement, &rdev->flags);
1800 } else
1801 /* We might have just remove the Replacement as faulty
1802 * Clear the flag just in case
1803 */
1804 clear_bit(WantReplacement, &rdev->flags);
1805
1806 err = md_integrity_register(mddev);
1807
1808abort:
1809
1810 print_conf(conf);
1811 return err;
1812}
1813
1814static void end_sync_read(struct bio *bio)
1815{
1816 struct r10bio *r10_bio = bio->bi_private;
1817 struct r10conf *conf = r10_bio->mddev->private;
1818 int d;
1819
1820 if (bio == r10_bio->master_bio) {
1821 /* this is a reshape read */
1822 d = r10_bio->read_slot; /* really the read dev */
1823 } else
1824 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1825
1826 if (!bio->bi_error)
1827 set_bit(R10BIO_Uptodate, &r10_bio->state);
1828 else
1829 /* The write handler will notice the lack of
1830 * R10BIO_Uptodate and record any errors etc
1831 */
1832 atomic_add(r10_bio->sectors,
1833 &conf->mirrors[d].rdev->corrected_errors);
1834
1835 /* for reconstruct, we always reschedule after a read.
1836 * for resync, only after all reads
1837 */
1838 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1839 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1840 atomic_dec_and_test(&r10_bio->remaining)) {
1841 /* we have read all the blocks,
1842 * do the comparison in process context in raid10d
1843 */
1844 reschedule_retry(r10_bio);
1845 }
1846}
1847
1848static void end_sync_request(struct r10bio *r10_bio)
1849{
1850 struct mddev *mddev = r10_bio->mddev;
1851
1852 while (atomic_dec_and_test(&r10_bio->remaining)) {
1853 if (r10_bio->master_bio == NULL) {
1854 /* the primary of several recovery bios */
1855 sector_t s = r10_bio->sectors;
1856 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1857 test_bit(R10BIO_WriteError, &r10_bio->state))
1858 reschedule_retry(r10_bio);
1859 else
1860 put_buf(r10_bio);
1861 md_done_sync(mddev, s, 1);
1862 break;
1863 } else {
1864 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1865 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1866 test_bit(R10BIO_WriteError, &r10_bio->state))
1867 reschedule_retry(r10_bio);
1868 else
1869 put_buf(r10_bio);
1870 r10_bio = r10_bio2;
1871 }
1872 }
1873}
1874
1875static void end_sync_write(struct bio *bio)
1876{
1877 struct r10bio *r10_bio = bio->bi_private;
1878 struct mddev *mddev = r10_bio->mddev;
1879 struct r10conf *conf = mddev->private;
1880 int d;
1881 sector_t first_bad;
1882 int bad_sectors;
1883 int slot;
1884 int repl;
1885 struct md_rdev *rdev = NULL;
1886
1887 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1888 if (repl)
1889 rdev = conf->mirrors[d].replacement;
1890 else
1891 rdev = conf->mirrors[d].rdev;
1892
1893 if (bio->bi_error) {
1894 if (repl)
1895 md_error(mddev, rdev);
1896 else {
1897 set_bit(WriteErrorSeen, &rdev->flags);
1898 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1899 set_bit(MD_RECOVERY_NEEDED,
1900 &rdev->mddev->recovery);
1901 set_bit(R10BIO_WriteError, &r10_bio->state);
1902 }
1903 } else if (is_badblock(rdev,
1904 r10_bio->devs[slot].addr,
1905 r10_bio->sectors,
1906 &first_bad, &bad_sectors))
1907 set_bit(R10BIO_MadeGood, &r10_bio->state);
1908
1909 rdev_dec_pending(rdev, mddev);
1910
1911 end_sync_request(r10_bio);
1912}
1913
1914/*
1915 * Note: sync and recover and handled very differently for raid10
1916 * This code is for resync.
1917 * For resync, we read through virtual addresses and read all blocks.
1918 * If there is any error, we schedule a write. The lowest numbered
1919 * drive is authoritative.
1920 * However requests come for physical address, so we need to map.
1921 * For every physical address there are raid_disks/copies virtual addresses,
1922 * which is always are least one, but is not necessarly an integer.
1923 * This means that a physical address can span multiple chunks, so we may
1924 * have to submit multiple io requests for a single sync request.
1925 */
1926/*
1927 * We check if all blocks are in-sync and only write to blocks that
1928 * aren't in sync
1929 */
1930static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1931{
1932 struct r10conf *conf = mddev->private;
1933 int i, first;
1934 struct bio *tbio, *fbio;
1935 int vcnt;
1936
1937 atomic_set(&r10_bio->remaining, 1);
1938
1939 /* find the first device with a block */
1940 for (i=0; i<conf->copies; i++)
1941 if (!r10_bio->devs[i].bio->bi_error)
1942 break;
1943
1944 if (i == conf->copies)
1945 goto done;
1946
1947 first = i;
1948 fbio = r10_bio->devs[i].bio;
1949 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1950 fbio->bi_iter.bi_idx = 0;
1951
1952 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1953 /* now find blocks with errors */
1954 for (i=0 ; i < conf->copies ; i++) {
1955 int j, d;
1956
1957 tbio = r10_bio->devs[i].bio;
1958
1959 if (tbio->bi_end_io != end_sync_read)
1960 continue;
1961 if (i == first)
1962 continue;
1963 if (!r10_bio->devs[i].bio->bi_error) {
1964 /* We know that the bi_io_vec layout is the same for
1965 * both 'first' and 'i', so we just compare them.
1966 * All vec entries are PAGE_SIZE;
1967 */
1968 int sectors = r10_bio->sectors;
1969 for (j = 0; j < vcnt; j++) {
1970 int len = PAGE_SIZE;
1971 if (sectors < (len / 512))
1972 len = sectors * 512;
1973 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1974 page_address(tbio->bi_io_vec[j].bv_page),
1975 len))
1976 break;
1977 sectors -= len/512;
1978 }
1979 if (j == vcnt)
1980 continue;
1981 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1982 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1983 /* Don't fix anything. */
1984 continue;
1985 }
1986 /* Ok, we need to write this bio, either to correct an
1987 * inconsistency or to correct an unreadable block.
1988 * First we need to fixup bv_offset, bv_len and
1989 * bi_vecs, as the read request might have corrupted these
1990 */
1991 bio_reset(tbio);
1992
1993 tbio->bi_vcnt = vcnt;
1994 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
1995 tbio->bi_rw = WRITE;
1996 tbio->bi_private = r10_bio;
1997 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1998 tbio->bi_end_io = end_sync_write;
1999
2000 bio_copy_data(tbio, fbio);
2001
2002 d = r10_bio->devs[i].devnum;
2003 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2004 atomic_inc(&r10_bio->remaining);
2005 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2006
2007 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2008 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2009 generic_make_request(tbio);
2010 }
2011
2012 /* Now write out to any replacement devices
2013 * that are active
2014 */
2015 for (i = 0; i < conf->copies; i++) {
2016 int d;
2017
2018 tbio = r10_bio->devs[i].repl_bio;
2019 if (!tbio || !tbio->bi_end_io)
2020 continue;
2021 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2022 && r10_bio->devs[i].bio != fbio)
2023 bio_copy_data(tbio, fbio);
2024 d = r10_bio->devs[i].devnum;
2025 atomic_inc(&r10_bio->remaining);
2026 md_sync_acct(conf->mirrors[d].replacement->bdev,
2027 bio_sectors(tbio));
2028 generic_make_request(tbio);
2029 }
2030
2031done:
2032 if (atomic_dec_and_test(&r10_bio->remaining)) {
2033 md_done_sync(mddev, r10_bio->sectors, 1);
2034 put_buf(r10_bio);
2035 }
2036}
2037
2038/*
2039 * Now for the recovery code.
2040 * Recovery happens across physical sectors.
2041 * We recover all non-is_sync drives by finding the virtual address of
2042 * each, and then choose a working drive that also has that virt address.
2043 * There is a separate r10_bio for each non-in_sync drive.
2044 * Only the first two slots are in use. The first for reading,
2045 * The second for writing.
2046 *
2047 */
2048static void fix_recovery_read_error(struct r10bio *r10_bio)
2049{
2050 /* We got a read error during recovery.
2051 * We repeat the read in smaller page-sized sections.
2052 * If a read succeeds, write it to the new device or record
2053 * a bad block if we cannot.
2054 * If a read fails, record a bad block on both old and
2055 * new devices.
2056 */
2057 struct mddev *mddev = r10_bio->mddev;
2058 struct r10conf *conf = mddev->private;
2059 struct bio *bio = r10_bio->devs[0].bio;
2060 sector_t sect = 0;
2061 int sectors = r10_bio->sectors;
2062 int idx = 0;
2063 int dr = r10_bio->devs[0].devnum;
2064 int dw = r10_bio->devs[1].devnum;
2065
2066 while (sectors) {
2067 int s = sectors;
2068 struct md_rdev *rdev;
2069 sector_t addr;
2070 int ok;
2071
2072 if (s > (PAGE_SIZE>>9))
2073 s = PAGE_SIZE >> 9;
2074
2075 rdev = conf->mirrors[dr].rdev;
2076 addr = r10_bio->devs[0].addr + sect,
2077 ok = sync_page_io(rdev,
2078 addr,
2079 s << 9,
2080 bio->bi_io_vec[idx].bv_page,
2081 READ, false);
2082 if (ok) {
2083 rdev = conf->mirrors[dw].rdev;
2084 addr = r10_bio->devs[1].addr + sect;
2085 ok = sync_page_io(rdev,
2086 addr,
2087 s << 9,
2088 bio->bi_io_vec[idx].bv_page,
2089 WRITE, false);
2090 if (!ok) {
2091 set_bit(WriteErrorSeen, &rdev->flags);
2092 if (!test_and_set_bit(WantReplacement,
2093 &rdev->flags))
2094 set_bit(MD_RECOVERY_NEEDED,
2095 &rdev->mddev->recovery);
2096 }
2097 }
2098 if (!ok) {
2099 /* We don't worry if we cannot set a bad block -
2100 * it really is bad so there is no loss in not
2101 * recording it yet
2102 */
2103 rdev_set_badblocks(rdev, addr, s, 0);
2104
2105 if (rdev != conf->mirrors[dw].rdev) {
2106 /* need bad block on destination too */
2107 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2108 addr = r10_bio->devs[1].addr + sect;
2109 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2110 if (!ok) {
2111 /* just abort the recovery */
2112 printk(KERN_NOTICE
2113 "md/raid10:%s: recovery aborted"
2114 " due to read error\n",
2115 mdname(mddev));
2116
2117 conf->mirrors[dw].recovery_disabled
2118 = mddev->recovery_disabled;
2119 set_bit(MD_RECOVERY_INTR,
2120 &mddev->recovery);
2121 break;
2122 }
2123 }
2124 }
2125
2126 sectors -= s;
2127 sect += s;
2128 idx++;
2129 }
2130}
2131
2132static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2133{
2134 struct r10conf *conf = mddev->private;
2135 int d;
2136 struct bio *wbio, *wbio2;
2137
2138 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2139 fix_recovery_read_error(r10_bio);
2140 end_sync_request(r10_bio);
2141 return;
2142 }
2143
2144 /*
2145 * share the pages with the first bio
2146 * and submit the write request
2147 */
2148 d = r10_bio->devs[1].devnum;
2149 wbio = r10_bio->devs[1].bio;
2150 wbio2 = r10_bio->devs[1].repl_bio;
2151 /* Need to test wbio2->bi_end_io before we call
2152 * generic_make_request as if the former is NULL,
2153 * the latter is free to free wbio2.
2154 */
2155 if (wbio2 && !wbio2->bi_end_io)
2156 wbio2 = NULL;
2157 if (wbio->bi_end_io) {
2158 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2159 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2160 generic_make_request(wbio);
2161 }
2162 if (wbio2) {
2163 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2164 md_sync_acct(conf->mirrors[d].replacement->bdev,
2165 bio_sectors(wbio2));
2166 generic_make_request(wbio2);
2167 }
2168}
2169
2170/*
2171 * Used by fix_read_error() to decay the per rdev read_errors.
2172 * We halve the read error count for every hour that has elapsed
2173 * since the last recorded read error.
2174 *
2175 */
2176static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2177{
2178 struct timespec cur_time_mon;
2179 unsigned long hours_since_last;
2180 unsigned int read_errors = atomic_read(&rdev->read_errors);
2181
2182 ktime_get_ts(&cur_time_mon);
2183
2184 if (rdev->last_read_error.tv_sec == 0 &&
2185 rdev->last_read_error.tv_nsec == 0) {
2186 /* first time we've seen a read error */
2187 rdev->last_read_error = cur_time_mon;
2188 return;
2189 }
2190
2191 hours_since_last = (cur_time_mon.tv_sec -
2192 rdev->last_read_error.tv_sec) / 3600;
2193
2194 rdev->last_read_error = cur_time_mon;
2195
2196 /*
2197 * if hours_since_last is > the number of bits in read_errors
2198 * just set read errors to 0. We do this to avoid
2199 * overflowing the shift of read_errors by hours_since_last.
2200 */
2201 if (hours_since_last >= 8 * sizeof(read_errors))
2202 atomic_set(&rdev->read_errors, 0);
2203 else
2204 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2205}
2206
2207static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2208 int sectors, struct page *page, int rw)
2209{
2210 sector_t first_bad;
2211 int bad_sectors;
2212
2213 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2214 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2215 return -1;
2216 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2217 /* success */
2218 return 1;
2219 if (rw == WRITE) {
2220 set_bit(WriteErrorSeen, &rdev->flags);
2221 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2222 set_bit(MD_RECOVERY_NEEDED,
2223 &rdev->mddev->recovery);
2224 }
2225 /* need to record an error - either for the block or the device */
2226 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2227 md_error(rdev->mddev, rdev);
2228 return 0;
2229}
2230
2231/*
2232 * This is a kernel thread which:
2233 *
2234 * 1. Retries failed read operations on working mirrors.
2235 * 2. Updates the raid superblock when problems encounter.
2236 * 3. Performs writes following reads for array synchronising.
2237 */
2238
2239static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2240{
2241 int sect = 0; /* Offset from r10_bio->sector */
2242 int sectors = r10_bio->sectors;
2243 struct md_rdev*rdev;
2244 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2245 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2246
2247 /* still own a reference to this rdev, so it cannot
2248 * have been cleared recently.
2249 */
2250 rdev = conf->mirrors[d].rdev;
2251
2252 if (test_bit(Faulty, &rdev->flags))
2253 /* drive has already been failed, just ignore any
2254 more fix_read_error() attempts */
2255 return;
2256
2257 check_decay_read_errors(mddev, rdev);
2258 atomic_inc(&rdev->read_errors);
2259 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2260 char b[BDEVNAME_SIZE];
2261 bdevname(rdev->bdev, b);
2262
2263 printk(KERN_NOTICE
2264 "md/raid10:%s: %s: Raid device exceeded "
2265 "read_error threshold [cur %d:max %d]\n",
2266 mdname(mddev), b,
2267 atomic_read(&rdev->read_errors), max_read_errors);
2268 printk(KERN_NOTICE
2269 "md/raid10:%s: %s: Failing raid device\n",
2270 mdname(mddev), b);
2271 md_error(mddev, conf->mirrors[d].rdev);
2272 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2273 return;
2274 }
2275
2276 while(sectors) {
2277 int s = sectors;
2278 int sl = r10_bio->read_slot;
2279 int success = 0;
2280 int start;
2281
2282 if (s > (PAGE_SIZE>>9))
2283 s = PAGE_SIZE >> 9;
2284
2285 rcu_read_lock();
2286 do {
2287 sector_t first_bad;
2288 int bad_sectors;
2289
2290 d = r10_bio->devs[sl].devnum;
2291 rdev = rcu_dereference(conf->mirrors[d].rdev);
2292 if (rdev &&
2293 test_bit(In_sync, &rdev->flags) &&
2294 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2295 &first_bad, &bad_sectors) == 0) {
2296 atomic_inc(&rdev->nr_pending);
2297 rcu_read_unlock();
2298 success = sync_page_io(rdev,
2299 r10_bio->devs[sl].addr +
2300 sect,
2301 s<<9,
2302 conf->tmppage, READ, false);
2303 rdev_dec_pending(rdev, mddev);
2304 rcu_read_lock();
2305 if (success)
2306 break;
2307 }
2308 sl++;
2309 if (sl == conf->copies)
2310 sl = 0;
2311 } while (!success && sl != r10_bio->read_slot);
2312 rcu_read_unlock();
2313
2314 if (!success) {
2315 /* Cannot read from anywhere, just mark the block
2316 * as bad on the first device to discourage future
2317 * reads.
2318 */
2319 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2320 rdev = conf->mirrors[dn].rdev;
2321
2322 if (!rdev_set_badblocks(
2323 rdev,
2324 r10_bio->devs[r10_bio->read_slot].addr
2325 + sect,
2326 s, 0)) {
2327 md_error(mddev, rdev);
2328 r10_bio->devs[r10_bio->read_slot].bio
2329 = IO_BLOCKED;
2330 }
2331 break;
2332 }
2333
2334 start = sl;
2335 /* write it back and re-read */
2336 rcu_read_lock();
2337 while (sl != r10_bio->read_slot) {
2338 char b[BDEVNAME_SIZE];
2339
2340 if (sl==0)
2341 sl = conf->copies;
2342 sl--;
2343 d = r10_bio->devs[sl].devnum;
2344 rdev = rcu_dereference(conf->mirrors[d].rdev);
2345 if (!rdev ||
2346 !test_bit(In_sync, &rdev->flags))
2347 continue;
2348
2349 atomic_inc(&rdev->nr_pending);
2350 rcu_read_unlock();
2351 if (r10_sync_page_io(rdev,
2352 r10_bio->devs[sl].addr +
2353 sect,
2354 s, conf->tmppage, WRITE)
2355 == 0) {
2356 /* Well, this device is dead */
2357 printk(KERN_NOTICE
2358 "md/raid10:%s: read correction "
2359 "write failed"
2360 " (%d sectors at %llu on %s)\n",
2361 mdname(mddev), s,
2362 (unsigned long long)(
2363 sect +
2364 choose_data_offset(r10_bio,
2365 rdev)),
2366 bdevname(rdev->bdev, b));
2367 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2368 "drive\n",
2369 mdname(mddev),
2370 bdevname(rdev->bdev, b));
2371 }
2372 rdev_dec_pending(rdev, mddev);
2373 rcu_read_lock();
2374 }
2375 sl = start;
2376 while (sl != r10_bio->read_slot) {
2377 char b[BDEVNAME_SIZE];
2378
2379 if (sl==0)
2380 sl = conf->copies;
2381 sl--;
2382 d = r10_bio->devs[sl].devnum;
2383 rdev = rcu_dereference(conf->mirrors[d].rdev);
2384 if (!rdev ||
2385 !test_bit(In_sync, &rdev->flags))
2386 continue;
2387
2388 atomic_inc(&rdev->nr_pending);
2389 rcu_read_unlock();
2390 switch (r10_sync_page_io(rdev,
2391 r10_bio->devs[sl].addr +
2392 sect,
2393 s, conf->tmppage,
2394 READ)) {
2395 case 0:
2396 /* Well, this device is dead */
2397 printk(KERN_NOTICE
2398 "md/raid10:%s: unable to read back "
2399 "corrected sectors"
2400 " (%d sectors at %llu on %s)\n",
2401 mdname(mddev), s,
2402 (unsigned long long)(
2403 sect +
2404 choose_data_offset(r10_bio, rdev)),
2405 bdevname(rdev->bdev, b));
2406 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2407 "drive\n",
2408 mdname(mddev),
2409 bdevname(rdev->bdev, b));
2410 break;
2411 case 1:
2412 printk(KERN_INFO
2413 "md/raid10:%s: read error corrected"
2414 " (%d sectors at %llu on %s)\n",
2415 mdname(mddev), s,
2416 (unsigned long long)(
2417 sect +
2418 choose_data_offset(r10_bio, rdev)),
2419 bdevname(rdev->bdev, b));
2420 atomic_add(s, &rdev->corrected_errors);
2421 }
2422
2423 rdev_dec_pending(rdev, mddev);
2424 rcu_read_lock();
2425 }
2426 rcu_read_unlock();
2427
2428 sectors -= s;
2429 sect += s;
2430 }
2431}
2432
2433static int narrow_write_error(struct r10bio *r10_bio, int i)
2434{
2435 struct bio *bio = r10_bio->master_bio;
2436 struct mddev *mddev = r10_bio->mddev;
2437 struct r10conf *conf = mddev->private;
2438 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2439 /* bio has the data to be written to slot 'i' where
2440 * we just recently had a write error.
2441 * We repeatedly clone the bio and trim down to one block,
2442 * then try the write. Where the write fails we record
2443 * a bad block.
2444 * It is conceivable that the bio doesn't exactly align with
2445 * blocks. We must handle this.
2446 *
2447 * We currently own a reference to the rdev.
2448 */
2449
2450 int block_sectors;
2451 sector_t sector;
2452 int sectors;
2453 int sect_to_write = r10_bio->sectors;
2454 int ok = 1;
2455
2456 if (rdev->badblocks.shift < 0)
2457 return 0;
2458
2459 block_sectors = roundup(1 << rdev->badblocks.shift,
2460 bdev_logical_block_size(rdev->bdev) >> 9);
2461 sector = r10_bio->sector;
2462 sectors = ((r10_bio->sector + block_sectors)
2463 & ~(sector_t)(block_sectors - 1))
2464 - sector;
2465
2466 while (sect_to_write) {
2467 struct bio *wbio;
2468 if (sectors > sect_to_write)
2469 sectors = sect_to_write;
2470 /* Write at 'sector' for 'sectors' */
2471 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2472 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2473 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2474 choose_data_offset(r10_bio, rdev) +
2475 (sector - r10_bio->sector));
2476 wbio->bi_bdev = rdev->bdev;
2477 if (submit_bio_wait(WRITE, wbio) < 0)
2478 /* Failure! */
2479 ok = rdev_set_badblocks(rdev, sector,
2480 sectors, 0)
2481 && ok;
2482
2483 bio_put(wbio);
2484 sect_to_write -= sectors;
2485 sector += sectors;
2486 sectors = block_sectors;
2487 }
2488 return ok;
2489}
2490
2491static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2492{
2493 int slot = r10_bio->read_slot;
2494 struct bio *bio;
2495 struct r10conf *conf = mddev->private;
2496 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2497 char b[BDEVNAME_SIZE];
2498 unsigned long do_sync;
2499 int max_sectors;
2500
2501 /* we got a read error. Maybe the drive is bad. Maybe just
2502 * the block and we can fix it.
2503 * We freeze all other IO, and try reading the block from
2504 * other devices. When we find one, we re-write
2505 * and check it that fixes the read error.
2506 * This is all done synchronously while the array is
2507 * frozen.
2508 */
2509 bio = r10_bio->devs[slot].bio;
2510 bdevname(bio->bi_bdev, b);
2511 bio_put(bio);
2512 r10_bio->devs[slot].bio = NULL;
2513
2514 if (mddev->ro == 0) {
2515 freeze_array(conf, 1);
2516 fix_read_error(conf, mddev, r10_bio);
2517 unfreeze_array(conf);
2518 } else
2519 r10_bio->devs[slot].bio = IO_BLOCKED;
2520
2521 rdev_dec_pending(rdev, mddev);
2522
2523read_more:
2524 rdev = read_balance(conf, r10_bio, &max_sectors);
2525 if (rdev == NULL) {
2526 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2527 " read error for block %llu\n",
2528 mdname(mddev), b,
2529 (unsigned long long)r10_bio->sector);
2530 raid_end_bio_io(r10_bio);
2531 return;
2532 }
2533
2534 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2535 slot = r10_bio->read_slot;
2536 printk_ratelimited(
2537 KERN_ERR
2538 "md/raid10:%s: %s: redirecting "
2539 "sector %llu to another mirror\n",
2540 mdname(mddev),
2541 bdevname(rdev->bdev, b),
2542 (unsigned long long)r10_bio->sector);
2543 bio = bio_clone_mddev(r10_bio->master_bio,
2544 GFP_NOIO, mddev);
2545 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2546 r10_bio->devs[slot].bio = bio;
2547 r10_bio->devs[slot].rdev = rdev;
2548 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2549 + choose_data_offset(r10_bio, rdev);
2550 bio->bi_bdev = rdev->bdev;
2551 bio->bi_rw = READ | do_sync;
2552 bio->bi_private = r10_bio;
2553 bio->bi_end_io = raid10_end_read_request;
2554 if (max_sectors < r10_bio->sectors) {
2555 /* Drat - have to split this up more */
2556 struct bio *mbio = r10_bio->master_bio;
2557 int sectors_handled =
2558 r10_bio->sector + max_sectors
2559 - mbio->bi_iter.bi_sector;
2560 r10_bio->sectors = max_sectors;
2561 spin_lock_irq(&conf->device_lock);
2562 if (mbio->bi_phys_segments == 0)
2563 mbio->bi_phys_segments = 2;
2564 else
2565 mbio->bi_phys_segments++;
2566 spin_unlock_irq(&conf->device_lock);
2567 generic_make_request(bio);
2568
2569 r10_bio = mempool_alloc(conf->r10bio_pool,
2570 GFP_NOIO);
2571 r10_bio->master_bio = mbio;
2572 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2573 r10_bio->state = 0;
2574 set_bit(R10BIO_ReadError,
2575 &r10_bio->state);
2576 r10_bio->mddev = mddev;
2577 r10_bio->sector = mbio->bi_iter.bi_sector
2578 + sectors_handled;
2579
2580 goto read_more;
2581 } else
2582 generic_make_request(bio);
2583}
2584
2585static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2586{
2587 /* Some sort of write request has finished and it
2588 * succeeded in writing where we thought there was a
2589 * bad block. So forget the bad block.
2590 * Or possibly if failed and we need to record
2591 * a bad block.
2592 */
2593 int m;
2594 struct md_rdev *rdev;
2595
2596 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2597 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2598 for (m = 0; m < conf->copies; m++) {
2599 int dev = r10_bio->devs[m].devnum;
2600 rdev = conf->mirrors[dev].rdev;
2601 if (r10_bio->devs[m].bio == NULL)
2602 continue;
2603 if (!r10_bio->devs[m].bio->bi_error) {
2604 rdev_clear_badblocks(
2605 rdev,
2606 r10_bio->devs[m].addr,
2607 r10_bio->sectors, 0);
2608 } else {
2609 if (!rdev_set_badblocks(
2610 rdev,
2611 r10_bio->devs[m].addr,
2612 r10_bio->sectors, 0))
2613 md_error(conf->mddev, rdev);
2614 }
2615 rdev = conf->mirrors[dev].replacement;
2616 if (r10_bio->devs[m].repl_bio == NULL)
2617 continue;
2618
2619 if (!r10_bio->devs[m].repl_bio->bi_error) {
2620 rdev_clear_badblocks(
2621 rdev,
2622 r10_bio->devs[m].addr,
2623 r10_bio->sectors, 0);
2624 } else {
2625 if (!rdev_set_badblocks(
2626 rdev,
2627 r10_bio->devs[m].addr,
2628 r10_bio->sectors, 0))
2629 md_error(conf->mddev, rdev);
2630 }
2631 }
2632 put_buf(r10_bio);
2633 } else {
2634 bool fail = false;
2635 for (m = 0; m < conf->copies; m++) {
2636 int dev = r10_bio->devs[m].devnum;
2637 struct bio *bio = r10_bio->devs[m].bio;
2638 rdev = conf->mirrors[dev].rdev;
2639 if (bio == IO_MADE_GOOD) {
2640 rdev_clear_badblocks(
2641 rdev,
2642 r10_bio->devs[m].addr,
2643 r10_bio->sectors, 0);
2644 rdev_dec_pending(rdev, conf->mddev);
2645 } else if (bio != NULL && bio->bi_error) {
2646 fail = true;
2647 if (!narrow_write_error(r10_bio, m)) {
2648 md_error(conf->mddev, rdev);
2649 set_bit(R10BIO_Degraded,
2650 &r10_bio->state);
2651 }
2652 rdev_dec_pending(rdev, conf->mddev);
2653 }
2654 bio = r10_bio->devs[m].repl_bio;
2655 rdev = conf->mirrors[dev].replacement;
2656 if (rdev && bio == IO_MADE_GOOD) {
2657 rdev_clear_badblocks(
2658 rdev,
2659 r10_bio->devs[m].addr,
2660 r10_bio->sectors, 0);
2661 rdev_dec_pending(rdev, conf->mddev);
2662 }
2663 }
2664 if (fail) {
2665 spin_lock_irq(&conf->device_lock);
2666 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2667 conf->nr_queued++;
2668 spin_unlock_irq(&conf->device_lock);
2669 md_wakeup_thread(conf->mddev->thread);
2670 } else {
2671 if (test_bit(R10BIO_WriteError,
2672 &r10_bio->state))
2673 close_write(r10_bio);
2674 raid_end_bio_io(r10_bio);
2675 }
2676 }
2677}
2678
2679static void raid10d(struct md_thread *thread)
2680{
2681 struct mddev *mddev = thread->mddev;
2682 struct r10bio *r10_bio;
2683 unsigned long flags;
2684 struct r10conf *conf = mddev->private;
2685 struct list_head *head = &conf->retry_list;
2686 struct blk_plug plug;
2687
2688 md_check_recovery(mddev);
2689
2690 if (!list_empty_careful(&conf->bio_end_io_list) &&
2691 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2692 LIST_HEAD(tmp);
2693 spin_lock_irqsave(&conf->device_lock, flags);
2694 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2695 while (!list_empty(&conf->bio_end_io_list)) {
2696 list_move(conf->bio_end_io_list.prev, &tmp);
2697 conf->nr_queued--;
2698 }
2699 }
2700 spin_unlock_irqrestore(&conf->device_lock, flags);
2701 while (!list_empty(&tmp)) {
2702 r10_bio = list_first_entry(&tmp, struct r10bio,
2703 retry_list);
2704 list_del(&r10_bio->retry_list);
2705 if (mddev->degraded)
2706 set_bit(R10BIO_Degraded, &r10_bio->state);
2707
2708 if (test_bit(R10BIO_WriteError,
2709 &r10_bio->state))
2710 close_write(r10_bio);
2711 raid_end_bio_io(r10_bio);
2712 }
2713 }
2714
2715 blk_start_plug(&plug);
2716 for (;;) {
2717
2718 flush_pending_writes(conf);
2719
2720 spin_lock_irqsave(&conf->device_lock, flags);
2721 if (list_empty(head)) {
2722 spin_unlock_irqrestore(&conf->device_lock, flags);
2723 break;
2724 }
2725 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2726 list_del(head->prev);
2727 conf->nr_queued--;
2728 spin_unlock_irqrestore(&conf->device_lock, flags);
2729
2730 mddev = r10_bio->mddev;
2731 conf = mddev->private;
2732 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2733 test_bit(R10BIO_WriteError, &r10_bio->state))
2734 handle_write_completed(conf, r10_bio);
2735 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2736 reshape_request_write(mddev, r10_bio);
2737 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2738 sync_request_write(mddev, r10_bio);
2739 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2740 recovery_request_write(mddev, r10_bio);
2741 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2742 handle_read_error(mddev, r10_bio);
2743 else {
2744 /* just a partial read to be scheduled from a
2745 * separate context
2746 */
2747 int slot = r10_bio->read_slot;
2748 generic_make_request(r10_bio->devs[slot].bio);
2749 }
2750
2751 cond_resched();
2752 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2753 md_check_recovery(mddev);
2754 }
2755 blk_finish_plug(&plug);
2756}
2757
2758static int init_resync(struct r10conf *conf)
2759{
2760 int buffs;
2761 int i;
2762
2763 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2764 BUG_ON(conf->r10buf_pool);
2765 conf->have_replacement = 0;
2766 for (i = 0; i < conf->geo.raid_disks; i++)
2767 if (conf->mirrors[i].replacement)
2768 conf->have_replacement = 1;
2769 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2770 if (!conf->r10buf_pool)
2771 return -ENOMEM;
2772 conf->next_resync = 0;
2773 return 0;
2774}
2775
2776/*
2777 * perform a "sync" on one "block"
2778 *
2779 * We need to make sure that no normal I/O request - particularly write
2780 * requests - conflict with active sync requests.
2781 *
2782 * This is achieved by tracking pending requests and a 'barrier' concept
2783 * that can be installed to exclude normal IO requests.
2784 *
2785 * Resync and recovery are handled very differently.
2786 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2787 *
2788 * For resync, we iterate over virtual addresses, read all copies,
2789 * and update if there are differences. If only one copy is live,
2790 * skip it.
2791 * For recovery, we iterate over physical addresses, read a good
2792 * value for each non-in_sync drive, and over-write.
2793 *
2794 * So, for recovery we may have several outstanding complex requests for a
2795 * given address, one for each out-of-sync device. We model this by allocating
2796 * a number of r10_bio structures, one for each out-of-sync device.
2797 * As we setup these structures, we collect all bio's together into a list
2798 * which we then process collectively to add pages, and then process again
2799 * to pass to generic_make_request.
2800 *
2801 * The r10_bio structures are linked using a borrowed master_bio pointer.
2802 * This link is counted in ->remaining. When the r10_bio that points to NULL
2803 * has its remaining count decremented to 0, the whole complex operation
2804 * is complete.
2805 *
2806 */
2807
2808static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2809 int *skipped)
2810{
2811 struct r10conf *conf = mddev->private;
2812 struct r10bio *r10_bio;
2813 struct bio *biolist = NULL, *bio;
2814 sector_t max_sector, nr_sectors;
2815 int i;
2816 int max_sync;
2817 sector_t sync_blocks;
2818 sector_t sectors_skipped = 0;
2819 int chunks_skipped = 0;
2820 sector_t chunk_mask = conf->geo.chunk_mask;
2821
2822 if (!conf->r10buf_pool)
2823 if (init_resync(conf))
2824 return 0;
2825
2826 /*
2827 * Allow skipping a full rebuild for incremental assembly
2828 * of a clean array, like RAID1 does.
2829 */
2830 if (mddev->bitmap == NULL &&
2831 mddev->recovery_cp == MaxSector &&
2832 mddev->reshape_position == MaxSector &&
2833 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2834 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2835 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2836 conf->fullsync == 0) {
2837 *skipped = 1;
2838 return mddev->dev_sectors - sector_nr;
2839 }
2840
2841 skipped:
2842 max_sector = mddev->dev_sectors;
2843 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2844 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2845 max_sector = mddev->resync_max_sectors;
2846 if (sector_nr >= max_sector) {
2847 /* If we aborted, we need to abort the
2848 * sync on the 'current' bitmap chucks (there can
2849 * be several when recovering multiple devices).
2850 * as we may have started syncing it but not finished.
2851 * We can find the current address in
2852 * mddev->curr_resync, but for recovery,
2853 * we need to convert that to several
2854 * virtual addresses.
2855 */
2856 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2857 end_reshape(conf);
2858 close_sync(conf);
2859 return 0;
2860 }
2861
2862 if (mddev->curr_resync < max_sector) { /* aborted */
2863 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2864 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2865 &sync_blocks, 1);
2866 else for (i = 0; i < conf->geo.raid_disks; i++) {
2867 sector_t sect =
2868 raid10_find_virt(conf, mddev->curr_resync, i);
2869 bitmap_end_sync(mddev->bitmap, sect,
2870 &sync_blocks, 1);
2871 }
2872 } else {
2873 /* completed sync */
2874 if ((!mddev->bitmap || conf->fullsync)
2875 && conf->have_replacement
2876 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2877 /* Completed a full sync so the replacements
2878 * are now fully recovered.
2879 */
2880 for (i = 0; i < conf->geo.raid_disks; i++)
2881 if (conf->mirrors[i].replacement)
2882 conf->mirrors[i].replacement
2883 ->recovery_offset
2884 = MaxSector;
2885 }
2886 conf->fullsync = 0;
2887 }
2888 bitmap_close_sync(mddev->bitmap);
2889 close_sync(conf);
2890 *skipped = 1;
2891 return sectors_skipped;
2892 }
2893
2894 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2895 return reshape_request(mddev, sector_nr, skipped);
2896
2897 if (chunks_skipped >= conf->geo.raid_disks) {
2898 /* if there has been nothing to do on any drive,
2899 * then there is nothing to do at all..
2900 */
2901 *skipped = 1;
2902 return (max_sector - sector_nr) + sectors_skipped;
2903 }
2904
2905 if (max_sector > mddev->resync_max)
2906 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2907
2908 /* make sure whole request will fit in a chunk - if chunks
2909 * are meaningful
2910 */
2911 if (conf->geo.near_copies < conf->geo.raid_disks &&
2912 max_sector > (sector_nr | chunk_mask))
2913 max_sector = (sector_nr | chunk_mask) + 1;
2914
2915 /* Again, very different code for resync and recovery.
2916 * Both must result in an r10bio with a list of bios that
2917 * have bi_end_io, bi_sector, bi_bdev set,
2918 * and bi_private set to the r10bio.
2919 * For recovery, we may actually create several r10bios
2920 * with 2 bios in each, that correspond to the bios in the main one.
2921 * In this case, the subordinate r10bios link back through a
2922 * borrowed master_bio pointer, and the counter in the master
2923 * includes a ref from each subordinate.
2924 */
2925 /* First, we decide what to do and set ->bi_end_io
2926 * To end_sync_read if we want to read, and
2927 * end_sync_write if we will want to write.
2928 */
2929
2930 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2931 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2932 /* recovery... the complicated one */
2933 int j;
2934 r10_bio = NULL;
2935
2936 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2937 int still_degraded;
2938 struct r10bio *rb2;
2939 sector_t sect;
2940 int must_sync;
2941 int any_working;
2942 struct raid10_info *mirror = &conf->mirrors[i];
2943
2944 if ((mirror->rdev == NULL ||
2945 test_bit(In_sync, &mirror->rdev->flags))
2946 &&
2947 (mirror->replacement == NULL ||
2948 test_bit(Faulty,
2949 &mirror->replacement->flags)))
2950 continue;
2951
2952 still_degraded = 0;
2953 /* want to reconstruct this device */
2954 rb2 = r10_bio;
2955 sect = raid10_find_virt(conf, sector_nr, i);
2956 if (sect >= mddev->resync_max_sectors) {
2957 /* last stripe is not complete - don't
2958 * try to recover this sector.
2959 */
2960 continue;
2961 }
2962 /* Unless we are doing a full sync, or a replacement
2963 * we only need to recover the block if it is set in
2964 * the bitmap
2965 */
2966 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2967 &sync_blocks, 1);
2968 if (sync_blocks < max_sync)
2969 max_sync = sync_blocks;
2970 if (!must_sync &&
2971 mirror->replacement == NULL &&
2972 !conf->fullsync) {
2973 /* yep, skip the sync_blocks here, but don't assume
2974 * that there will never be anything to do here
2975 */
2976 chunks_skipped = -1;
2977 continue;
2978 }
2979
2980 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2981 r10_bio->state = 0;
2982 raise_barrier(conf, rb2 != NULL);
2983 atomic_set(&r10_bio->remaining, 0);
2984
2985 r10_bio->master_bio = (struct bio*)rb2;
2986 if (rb2)
2987 atomic_inc(&rb2->remaining);
2988 r10_bio->mddev = mddev;
2989 set_bit(R10BIO_IsRecover, &r10_bio->state);
2990 r10_bio->sector = sect;
2991
2992 raid10_find_phys(conf, r10_bio);
2993
2994 /* Need to check if the array will still be
2995 * degraded
2996 */
2997 for (j = 0; j < conf->geo.raid_disks; j++)
2998 if (conf->mirrors[j].rdev == NULL ||
2999 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3000 still_degraded = 1;
3001 break;
3002 }
3003
3004 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3005 &sync_blocks, still_degraded);
3006
3007 any_working = 0;
3008 for (j=0; j<conf->copies;j++) {
3009 int k;
3010 int d = r10_bio->devs[j].devnum;
3011 sector_t from_addr, to_addr;
3012 struct md_rdev *rdev;
3013 sector_t sector, first_bad;
3014 int bad_sectors;
3015 if (!conf->mirrors[d].rdev ||
3016 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3017 continue;
3018 /* This is where we read from */
3019 any_working = 1;
3020 rdev = conf->mirrors[d].rdev;
3021 sector = r10_bio->devs[j].addr;
3022
3023 if (is_badblock(rdev, sector, max_sync,
3024 &first_bad, &bad_sectors)) {
3025 if (first_bad > sector)
3026 max_sync = first_bad - sector;
3027 else {
3028 bad_sectors -= (sector
3029 - first_bad);
3030 if (max_sync > bad_sectors)
3031 max_sync = bad_sectors;
3032 continue;
3033 }
3034 }
3035 bio = r10_bio->devs[0].bio;
3036 bio_reset(bio);
3037 bio->bi_next = biolist;
3038 biolist = bio;
3039 bio->bi_private = r10_bio;
3040 bio->bi_end_io = end_sync_read;
3041 bio->bi_rw = READ;
3042 from_addr = r10_bio->devs[j].addr;
3043 bio->bi_iter.bi_sector = from_addr +
3044 rdev->data_offset;
3045 bio->bi_bdev = rdev->bdev;
3046 atomic_inc(&rdev->nr_pending);
3047 /* and we write to 'i' (if not in_sync) */
3048
3049 for (k=0; k<conf->copies; k++)
3050 if (r10_bio->devs[k].devnum == i)
3051 break;
3052 BUG_ON(k == conf->copies);
3053 to_addr = r10_bio->devs[k].addr;
3054 r10_bio->devs[0].devnum = d;
3055 r10_bio->devs[0].addr = from_addr;
3056 r10_bio->devs[1].devnum = i;
3057 r10_bio->devs[1].addr = to_addr;
3058
3059 rdev = mirror->rdev;
3060 if (!test_bit(In_sync, &rdev->flags)) {
3061 bio = r10_bio->devs[1].bio;
3062 bio_reset(bio);
3063 bio->bi_next = biolist;
3064 biolist = bio;
3065 bio->bi_private = r10_bio;
3066 bio->bi_end_io = end_sync_write;
3067 bio->bi_rw = WRITE;
3068 bio->bi_iter.bi_sector = to_addr
3069 + rdev->data_offset;
3070 bio->bi_bdev = rdev->bdev;
3071 atomic_inc(&r10_bio->remaining);
3072 } else
3073 r10_bio->devs[1].bio->bi_end_io = NULL;
3074
3075 /* and maybe write to replacement */
3076 bio = r10_bio->devs[1].repl_bio;
3077 if (bio)
3078 bio->bi_end_io = NULL;
3079 rdev = mirror->replacement;
3080 /* Note: if rdev != NULL, then bio
3081 * cannot be NULL as r10buf_pool_alloc will
3082 * have allocated it.
3083 * So the second test here is pointless.
3084 * But it keeps semantic-checkers happy, and
3085 * this comment keeps human reviewers
3086 * happy.
3087 */
3088 if (rdev == NULL || bio == NULL ||
3089 test_bit(Faulty, &rdev->flags))
3090 break;
3091 bio_reset(bio);
3092 bio->bi_next = biolist;
3093 biolist = bio;
3094 bio->bi_private = r10_bio;
3095 bio->bi_end_io = end_sync_write;
3096 bio->bi_rw = WRITE;
3097 bio->bi_iter.bi_sector = to_addr +
3098 rdev->data_offset;
3099 bio->bi_bdev = rdev->bdev;
3100 atomic_inc(&r10_bio->remaining);
3101 break;
3102 }
3103 if (j == conf->copies) {
3104 /* Cannot recover, so abort the recovery or
3105 * record a bad block */
3106 if (any_working) {
3107 /* problem is that there are bad blocks
3108 * on other device(s)
3109 */
3110 int k;
3111 for (k = 0; k < conf->copies; k++)
3112 if (r10_bio->devs[k].devnum == i)
3113 break;
3114 if (!test_bit(In_sync,
3115 &mirror->rdev->flags)
3116 && !rdev_set_badblocks(
3117 mirror->rdev,
3118 r10_bio->devs[k].addr,
3119 max_sync, 0))
3120 any_working = 0;
3121 if (mirror->replacement &&
3122 !rdev_set_badblocks(
3123 mirror->replacement,
3124 r10_bio->devs[k].addr,
3125 max_sync, 0))
3126 any_working = 0;
3127 }
3128 if (!any_working) {
3129 if (!test_and_set_bit(MD_RECOVERY_INTR,
3130 &mddev->recovery))
3131 printk(KERN_INFO "md/raid10:%s: insufficient "
3132 "working devices for recovery.\n",
3133 mdname(mddev));
3134 mirror->recovery_disabled
3135 = mddev->recovery_disabled;
3136 }
3137 put_buf(r10_bio);
3138 if (rb2)
3139 atomic_dec(&rb2->remaining);
3140 r10_bio = rb2;
3141 break;
3142 }
3143 }
3144 if (biolist == NULL) {
3145 while (r10_bio) {
3146 struct r10bio *rb2 = r10_bio;
3147 r10_bio = (struct r10bio*) rb2->master_bio;
3148 rb2->master_bio = NULL;
3149 put_buf(rb2);
3150 }
3151 goto giveup;
3152 }
3153 } else {
3154 /* resync. Schedule a read for every block at this virt offset */
3155 int count = 0;
3156
3157 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3158
3159 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3160 &sync_blocks, mddev->degraded) &&
3161 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3162 &mddev->recovery)) {
3163 /* We can skip this block */
3164 *skipped = 1;
3165 return sync_blocks + sectors_skipped;
3166 }
3167 if (sync_blocks < max_sync)
3168 max_sync = sync_blocks;
3169 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3170 r10_bio->state = 0;
3171
3172 r10_bio->mddev = mddev;
3173 atomic_set(&r10_bio->remaining, 0);
3174 raise_barrier(conf, 0);
3175 conf->next_resync = sector_nr;
3176
3177 r10_bio->master_bio = NULL;
3178 r10_bio->sector = sector_nr;
3179 set_bit(R10BIO_IsSync, &r10_bio->state);
3180 raid10_find_phys(conf, r10_bio);
3181 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3182
3183 for (i = 0; i < conf->copies; i++) {
3184 int d = r10_bio->devs[i].devnum;
3185 sector_t first_bad, sector;
3186 int bad_sectors;
3187
3188 if (r10_bio->devs[i].repl_bio)
3189 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3190
3191 bio = r10_bio->devs[i].bio;
3192 bio_reset(bio);
3193 bio->bi_error = -EIO;
3194 if (conf->mirrors[d].rdev == NULL ||
3195 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3196 continue;
3197 sector = r10_bio->devs[i].addr;
3198 if (is_badblock(conf->mirrors[d].rdev,
3199 sector, max_sync,
3200 &first_bad, &bad_sectors)) {
3201 if (first_bad > sector)
3202 max_sync = first_bad - sector;
3203 else {
3204 bad_sectors -= (sector - first_bad);
3205 if (max_sync > bad_sectors)
3206 max_sync = bad_sectors;
3207 continue;
3208 }
3209 }
3210 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3211 atomic_inc(&r10_bio->remaining);
3212 bio->bi_next = biolist;
3213 biolist = bio;
3214 bio->bi_private = r10_bio;
3215 bio->bi_end_io = end_sync_read;
3216 bio->bi_rw = READ;
3217 bio->bi_iter.bi_sector = sector +
3218 conf->mirrors[d].rdev->data_offset;
3219 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3220 count++;
3221
3222 if (conf->mirrors[d].replacement == NULL ||
3223 test_bit(Faulty,
3224 &conf->mirrors[d].replacement->flags))
3225 continue;
3226
3227 /* Need to set up for writing to the replacement */
3228 bio = r10_bio->devs[i].repl_bio;
3229 bio_reset(bio);
3230 bio->bi_error = -EIO;
3231
3232 sector = r10_bio->devs[i].addr;
3233 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3234 bio->bi_next = biolist;
3235 biolist = bio;
3236 bio->bi_private = r10_bio;
3237 bio->bi_end_io = end_sync_write;
3238 bio->bi_rw = WRITE;
3239 bio->bi_iter.bi_sector = sector +
3240 conf->mirrors[d].replacement->data_offset;
3241 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3242 count++;
3243 }
3244
3245 if (count < 2) {
3246 for (i=0; i<conf->copies; i++) {
3247 int d = r10_bio->devs[i].devnum;
3248 if (r10_bio->devs[i].bio->bi_end_io)
3249 rdev_dec_pending(conf->mirrors[d].rdev,
3250 mddev);
3251 if (r10_bio->devs[i].repl_bio &&
3252 r10_bio->devs[i].repl_bio->bi_end_io)
3253 rdev_dec_pending(
3254 conf->mirrors[d].replacement,
3255 mddev);
3256 }
3257 put_buf(r10_bio);
3258 biolist = NULL;
3259 goto giveup;
3260 }
3261 }
3262
3263 nr_sectors = 0;
3264 if (sector_nr + max_sync < max_sector)
3265 max_sector = sector_nr + max_sync;
3266 do {
3267 struct page *page;
3268 int len = PAGE_SIZE;
3269 if (sector_nr + (len>>9) > max_sector)
3270 len = (max_sector - sector_nr) << 9;
3271 if (len == 0)
3272 break;
3273 for (bio= biolist ; bio ; bio=bio->bi_next) {
3274 struct bio *bio2;
3275 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3276 if (bio_add_page(bio, page, len, 0))
3277 continue;
3278
3279 /* stop here */
3280 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3281 for (bio2 = biolist;
3282 bio2 && bio2 != bio;
3283 bio2 = bio2->bi_next) {
3284 /* remove last page from this bio */
3285 bio2->bi_vcnt--;
3286 bio2->bi_iter.bi_size -= len;
3287 bio_clear_flag(bio2, BIO_SEG_VALID);
3288 }
3289 goto bio_full;
3290 }
3291 nr_sectors += len>>9;
3292 sector_nr += len>>9;
3293 } while (biolist->bi_vcnt < RESYNC_PAGES);
3294 bio_full:
3295 r10_bio->sectors = nr_sectors;
3296
3297 while (biolist) {
3298 bio = biolist;
3299 biolist = biolist->bi_next;
3300
3301 bio->bi_next = NULL;
3302 r10_bio = bio->bi_private;
3303 r10_bio->sectors = nr_sectors;
3304
3305 if (bio->bi_end_io == end_sync_read) {
3306 md_sync_acct(bio->bi_bdev, nr_sectors);
3307 bio->bi_error = 0;
3308 generic_make_request(bio);
3309 }
3310 }
3311
3312 if (sectors_skipped)
3313 /* pretend they weren't skipped, it makes
3314 * no important difference in this case
3315 */
3316 md_done_sync(mddev, sectors_skipped, 1);
3317
3318 return sectors_skipped + nr_sectors;
3319 giveup:
3320 /* There is nowhere to write, so all non-sync
3321 * drives must be failed or in resync, all drives
3322 * have a bad block, so try the next chunk...
3323 */
3324 if (sector_nr + max_sync < max_sector)
3325 max_sector = sector_nr + max_sync;
3326
3327 sectors_skipped += (max_sector - sector_nr);
3328 chunks_skipped ++;
3329 sector_nr = max_sector;
3330 goto skipped;
3331}
3332
3333static sector_t
3334raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3335{
3336 sector_t size;
3337 struct r10conf *conf = mddev->private;
3338
3339 if (!raid_disks)
3340 raid_disks = min(conf->geo.raid_disks,
3341 conf->prev.raid_disks);
3342 if (!sectors)
3343 sectors = conf->dev_sectors;
3344
3345 size = sectors >> conf->geo.chunk_shift;
3346 sector_div(size, conf->geo.far_copies);
3347 size = size * raid_disks;
3348 sector_div(size, conf->geo.near_copies);
3349
3350 return size << conf->geo.chunk_shift;
3351}
3352
3353static void calc_sectors(struct r10conf *conf, sector_t size)
3354{
3355 /* Calculate the number of sectors-per-device that will
3356 * actually be used, and set conf->dev_sectors and
3357 * conf->stride
3358 */
3359
3360 size = size >> conf->geo.chunk_shift;
3361 sector_div(size, conf->geo.far_copies);
3362 size = size * conf->geo.raid_disks;
3363 sector_div(size, conf->geo.near_copies);
3364 /* 'size' is now the number of chunks in the array */
3365 /* calculate "used chunks per device" */
3366 size = size * conf->copies;
3367
3368 /* We need to round up when dividing by raid_disks to
3369 * get the stride size.
3370 */
3371 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3372
3373 conf->dev_sectors = size << conf->geo.chunk_shift;
3374
3375 if (conf->geo.far_offset)
3376 conf->geo.stride = 1 << conf->geo.chunk_shift;
3377 else {
3378 sector_div(size, conf->geo.far_copies);
3379 conf->geo.stride = size << conf->geo.chunk_shift;
3380 }
3381}
3382
3383enum geo_type {geo_new, geo_old, geo_start};
3384static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3385{
3386 int nc, fc, fo;
3387 int layout, chunk, disks;
3388 switch (new) {
3389 case geo_old:
3390 layout = mddev->layout;
3391 chunk = mddev->chunk_sectors;
3392 disks = mddev->raid_disks - mddev->delta_disks;
3393 break;
3394 case geo_new:
3395 layout = mddev->new_layout;
3396 chunk = mddev->new_chunk_sectors;
3397 disks = mddev->raid_disks;
3398 break;
3399 default: /* avoid 'may be unused' warnings */
3400 case geo_start: /* new when starting reshape - raid_disks not
3401 * updated yet. */
3402 layout = mddev->new_layout;
3403 chunk = mddev->new_chunk_sectors;
3404 disks = mddev->raid_disks + mddev->delta_disks;
3405 break;
3406 }
3407 if (layout >> 19)
3408 return -1;
3409 if (chunk < (PAGE_SIZE >> 9) ||
3410 !is_power_of_2(chunk))
3411 return -2;
3412 nc = layout & 255;
3413 fc = (layout >> 8) & 255;
3414 fo = layout & (1<<16);
3415 geo->raid_disks = disks;
3416 geo->near_copies = nc;
3417 geo->far_copies = fc;
3418 geo->far_offset = fo;
3419 switch (layout >> 17) {
3420 case 0: /* original layout. simple but not always optimal */
3421 geo->far_set_size = disks;
3422 break;
3423 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3424 * actually using this, but leave code here just in case.*/
3425 geo->far_set_size = disks/fc;
3426 WARN(geo->far_set_size < fc,
3427 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3428 break;
3429 case 2: /* "improved" layout fixed to match documentation */
3430 geo->far_set_size = fc * nc;
3431 break;
3432 default: /* Not a valid layout */
3433 return -1;
3434 }
3435 geo->chunk_mask = chunk - 1;
3436 geo->chunk_shift = ffz(~chunk);
3437 return nc*fc;
3438}
3439
3440static struct r10conf *setup_conf(struct mddev *mddev)
3441{
3442 struct r10conf *conf = NULL;
3443 int err = -EINVAL;
3444 struct geom geo;
3445 int copies;
3446
3447 copies = setup_geo(&geo, mddev, geo_new);
3448
3449 if (copies == -2) {
3450 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3451 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3452 mdname(mddev), PAGE_SIZE);
3453 goto out;
3454 }
3455
3456 if (copies < 2 || copies > mddev->raid_disks) {
3457 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3458 mdname(mddev), mddev->new_layout);
3459 goto out;
3460 }
3461
3462 err = -ENOMEM;
3463 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3464 if (!conf)
3465 goto out;
3466
3467 /* FIXME calc properly */
3468 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3469 max(0,-mddev->delta_disks)),
3470 GFP_KERNEL);
3471 if (!conf->mirrors)
3472 goto out;
3473
3474 conf->tmppage = alloc_page(GFP_KERNEL);
3475 if (!conf->tmppage)
3476 goto out;
3477
3478 conf->geo = geo;
3479 conf->copies = copies;
3480 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3481 r10bio_pool_free, conf);
3482 if (!conf->r10bio_pool)
3483 goto out;
3484
3485 calc_sectors(conf, mddev->dev_sectors);
3486 if (mddev->reshape_position == MaxSector) {
3487 conf->prev = conf->geo;
3488 conf->reshape_progress = MaxSector;
3489 } else {
3490 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3491 err = -EINVAL;
3492 goto out;
3493 }
3494 conf->reshape_progress = mddev->reshape_position;
3495 if (conf->prev.far_offset)
3496 conf->prev.stride = 1 << conf->prev.chunk_shift;
3497 else
3498 /* far_copies must be 1 */
3499 conf->prev.stride = conf->dev_sectors;
3500 }
3501 conf->reshape_safe = conf->reshape_progress;
3502 spin_lock_init(&conf->device_lock);
3503 INIT_LIST_HEAD(&conf->retry_list);
3504 INIT_LIST_HEAD(&conf->bio_end_io_list);
3505
3506 spin_lock_init(&conf->resync_lock);
3507 init_waitqueue_head(&conf->wait_barrier);
3508
3509 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3510 if (!conf->thread)
3511 goto out;
3512
3513 conf->mddev = mddev;
3514 return conf;
3515
3516 out:
3517 if (err == -ENOMEM)
3518 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3519 mdname(mddev));
3520 if (conf) {
3521 mempool_destroy(conf->r10bio_pool);
3522 kfree(conf->mirrors);
3523 safe_put_page(conf->tmppage);
3524 kfree(conf);
3525 }
3526 return ERR_PTR(err);
3527}
3528
3529static int raid10_run(struct mddev *mddev)
3530{
3531 struct r10conf *conf;
3532 int i, disk_idx, chunk_size;
3533 struct raid10_info *disk;
3534 struct md_rdev *rdev;
3535 sector_t size;
3536 sector_t min_offset_diff = 0;
3537 int first = 1;
3538 bool discard_supported = false;
3539
3540 if (mddev->private == NULL) {
3541 conf = setup_conf(mddev);
3542 if (IS_ERR(conf))
3543 return PTR_ERR(conf);
3544 mddev->private = conf;
3545 }
3546 conf = mddev->private;
3547 if (!conf)
3548 goto out;
3549
3550 mddev->thread = conf->thread;
3551 conf->thread = NULL;
3552
3553 chunk_size = mddev->chunk_sectors << 9;
3554 if (mddev->queue) {
3555 blk_queue_max_discard_sectors(mddev->queue,
3556 mddev->chunk_sectors);
3557 blk_queue_max_write_same_sectors(mddev->queue, 0);
3558 blk_queue_io_min(mddev->queue, chunk_size);
3559 if (conf->geo.raid_disks % conf->geo.near_copies)
3560 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3561 else
3562 blk_queue_io_opt(mddev->queue, chunk_size *
3563 (conf->geo.raid_disks / conf->geo.near_copies));
3564 }
3565
3566 rdev_for_each(rdev, mddev) {
3567 long long diff;
3568 struct request_queue *q;
3569
3570 disk_idx = rdev->raid_disk;
3571 if (disk_idx < 0)
3572 continue;
3573 if (disk_idx >= conf->geo.raid_disks &&
3574 disk_idx >= conf->prev.raid_disks)
3575 continue;
3576 disk = conf->mirrors + disk_idx;
3577
3578 if (test_bit(Replacement, &rdev->flags)) {
3579 if (disk->replacement)
3580 goto out_free_conf;
3581 disk->replacement = rdev;
3582 } else {
3583 if (disk->rdev)
3584 goto out_free_conf;
3585 disk->rdev = rdev;
3586 }
3587 q = bdev_get_queue(rdev->bdev);
3588 diff = (rdev->new_data_offset - rdev->data_offset);
3589 if (!mddev->reshape_backwards)
3590 diff = -diff;
3591 if (diff < 0)
3592 diff = 0;
3593 if (first || diff < min_offset_diff)
3594 min_offset_diff = diff;
3595
3596 if (mddev->gendisk)
3597 disk_stack_limits(mddev->gendisk, rdev->bdev,
3598 rdev->data_offset << 9);
3599
3600 disk->head_position = 0;
3601
3602 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3603 discard_supported = true;
3604 }
3605
3606 if (mddev->queue) {
3607 if (discard_supported)
3608 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3609 mddev->queue);
3610 else
3611 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3612 mddev->queue);
3613 }
3614 /* need to check that every block has at least one working mirror */
3615 if (!enough(conf, -1)) {
3616 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3617 mdname(mddev));
3618 goto out_free_conf;
3619 }
3620
3621 if (conf->reshape_progress != MaxSector) {
3622 /* must ensure that shape change is supported */
3623 if (conf->geo.far_copies != 1 &&
3624 conf->geo.far_offset == 0)
3625 goto out_free_conf;
3626 if (conf->prev.far_copies != 1 &&
3627 conf->prev.far_offset == 0)
3628 goto out_free_conf;
3629 }
3630
3631 mddev->degraded = 0;
3632 for (i = 0;
3633 i < conf->geo.raid_disks
3634 || i < conf->prev.raid_disks;
3635 i++) {
3636
3637 disk = conf->mirrors + i;
3638
3639 if (!disk->rdev && disk->replacement) {
3640 /* The replacement is all we have - use it */
3641 disk->rdev = disk->replacement;
3642 disk->replacement = NULL;
3643 clear_bit(Replacement, &disk->rdev->flags);
3644 }
3645
3646 if (!disk->rdev ||
3647 !test_bit(In_sync, &disk->rdev->flags)) {
3648 disk->head_position = 0;
3649 mddev->degraded++;
3650 if (disk->rdev &&
3651 disk->rdev->saved_raid_disk < 0)
3652 conf->fullsync = 1;
3653 }
3654 disk->recovery_disabled = mddev->recovery_disabled - 1;
3655 }
3656
3657 if (mddev->recovery_cp != MaxSector)
3658 printk(KERN_NOTICE "md/raid10:%s: not clean"
3659 " -- starting background reconstruction\n",
3660 mdname(mddev));
3661 printk(KERN_INFO
3662 "md/raid10:%s: active with %d out of %d devices\n",
3663 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3664 conf->geo.raid_disks);
3665 /*
3666 * Ok, everything is just fine now
3667 */
3668 mddev->dev_sectors = conf->dev_sectors;
3669 size = raid10_size(mddev, 0, 0);
3670 md_set_array_sectors(mddev, size);
3671 mddev->resync_max_sectors = size;
3672
3673 if (mddev->queue) {
3674 int stripe = conf->geo.raid_disks *
3675 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3676
3677 /* Calculate max read-ahead size.
3678 * We need to readahead at least twice a whole stripe....
3679 * maybe...
3680 */
3681 stripe /= conf->geo.near_copies;
3682 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3683 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3684 }
3685
3686 if (md_integrity_register(mddev))
3687 goto out_free_conf;
3688
3689 if (conf->reshape_progress != MaxSector) {
3690 unsigned long before_length, after_length;
3691
3692 before_length = ((1 << conf->prev.chunk_shift) *
3693 conf->prev.far_copies);
3694 after_length = ((1 << conf->geo.chunk_shift) *
3695 conf->geo.far_copies);
3696
3697 if (max(before_length, after_length) > min_offset_diff) {
3698 /* This cannot work */
3699 printk("md/raid10: offset difference not enough to continue reshape\n");
3700 goto out_free_conf;
3701 }
3702 conf->offset_diff = min_offset_diff;
3703
3704 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3705 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3706 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3707 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3708 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3709 "reshape");
3710 }
3711
3712 return 0;
3713
3714out_free_conf:
3715 md_unregister_thread(&mddev->thread);
3716 mempool_destroy(conf->r10bio_pool);
3717 safe_put_page(conf->tmppage);
3718 kfree(conf->mirrors);
3719 kfree(conf);
3720 mddev->private = NULL;
3721out:
3722 return -EIO;
3723}
3724
3725static void raid10_free(struct mddev *mddev, void *priv)
3726{
3727 struct r10conf *conf = priv;
3728
3729 mempool_destroy(conf->r10bio_pool);
3730 safe_put_page(conf->tmppage);
3731 kfree(conf->mirrors);
3732 kfree(conf->mirrors_old);
3733 kfree(conf->mirrors_new);
3734 kfree(conf);
3735}
3736
3737static void raid10_quiesce(struct mddev *mddev, int state)
3738{
3739 struct r10conf *conf = mddev->private;
3740
3741 switch(state) {
3742 case 1:
3743 raise_barrier(conf, 0);
3744 break;
3745 case 0:
3746 lower_barrier(conf);
3747 break;
3748 }
3749}
3750
3751static int raid10_resize(struct mddev *mddev, sector_t sectors)
3752{
3753 /* Resize of 'far' arrays is not supported.
3754 * For 'near' and 'offset' arrays we can set the
3755 * number of sectors used to be an appropriate multiple
3756 * of the chunk size.
3757 * For 'offset', this is far_copies*chunksize.
3758 * For 'near' the multiplier is the LCM of
3759 * near_copies and raid_disks.
3760 * So if far_copies > 1 && !far_offset, fail.
3761 * Else find LCM(raid_disks, near_copy)*far_copies and
3762 * multiply by chunk_size. Then round to this number.
3763 * This is mostly done by raid10_size()
3764 */
3765 struct r10conf *conf = mddev->private;
3766 sector_t oldsize, size;
3767
3768 if (mddev->reshape_position != MaxSector)
3769 return -EBUSY;
3770
3771 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3772 return -EINVAL;
3773
3774 oldsize = raid10_size(mddev, 0, 0);
3775 size = raid10_size(mddev, sectors, 0);
3776 if (mddev->external_size &&
3777 mddev->array_sectors > size)
3778 return -EINVAL;
3779 if (mddev->bitmap) {
3780 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3781 if (ret)
3782 return ret;
3783 }
3784 md_set_array_sectors(mddev, size);
3785 set_capacity(mddev->gendisk, mddev->array_sectors);
3786 revalidate_disk(mddev->gendisk);
3787 if (sectors > mddev->dev_sectors &&
3788 mddev->recovery_cp > oldsize) {
3789 mddev->recovery_cp = oldsize;
3790 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3791 }
3792 calc_sectors(conf, sectors);
3793 mddev->dev_sectors = conf->dev_sectors;
3794 mddev->resync_max_sectors = size;
3795 return 0;
3796}
3797
3798static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3799{
3800 struct md_rdev *rdev;
3801 struct r10conf *conf;
3802
3803 if (mddev->degraded > 0) {
3804 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3805 mdname(mddev));
3806 return ERR_PTR(-EINVAL);
3807 }
3808 sector_div(size, devs);
3809
3810 /* Set new parameters */
3811 mddev->new_level = 10;
3812 /* new layout: far_copies = 1, near_copies = 2 */
3813 mddev->new_layout = (1<<8) + 2;
3814 mddev->new_chunk_sectors = mddev->chunk_sectors;
3815 mddev->delta_disks = mddev->raid_disks;
3816 mddev->raid_disks *= 2;
3817 /* make sure it will be not marked as dirty */
3818 mddev->recovery_cp = MaxSector;
3819 mddev->dev_sectors = size;
3820
3821 conf = setup_conf(mddev);
3822 if (!IS_ERR(conf)) {
3823 rdev_for_each(rdev, mddev)
3824 if (rdev->raid_disk >= 0) {
3825 rdev->new_raid_disk = rdev->raid_disk * 2;
3826 rdev->sectors = size;
3827 }
3828 conf->barrier = 1;
3829 }
3830
3831 return conf;
3832}
3833
3834static void *raid10_takeover(struct mddev *mddev)
3835{
3836 struct r0conf *raid0_conf;
3837
3838 /* raid10 can take over:
3839 * raid0 - providing it has only two drives
3840 */
3841 if (mddev->level == 0) {
3842 /* for raid0 takeover only one zone is supported */
3843 raid0_conf = mddev->private;
3844 if (raid0_conf->nr_strip_zones > 1) {
3845 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3846 " with more than one zone.\n",
3847 mdname(mddev));
3848 return ERR_PTR(-EINVAL);
3849 }
3850 return raid10_takeover_raid0(mddev,
3851 raid0_conf->strip_zone->zone_end,
3852 raid0_conf->strip_zone->nb_dev);
3853 }
3854 return ERR_PTR(-EINVAL);
3855}
3856
3857static int raid10_check_reshape(struct mddev *mddev)
3858{
3859 /* Called when there is a request to change
3860 * - layout (to ->new_layout)
3861 * - chunk size (to ->new_chunk_sectors)
3862 * - raid_disks (by delta_disks)
3863 * or when trying to restart a reshape that was ongoing.
3864 *
3865 * We need to validate the request and possibly allocate
3866 * space if that might be an issue later.
3867 *
3868 * Currently we reject any reshape of a 'far' mode array,
3869 * allow chunk size to change if new is generally acceptable,
3870 * allow raid_disks to increase, and allow
3871 * a switch between 'near' mode and 'offset' mode.
3872 */
3873 struct r10conf *conf = mddev->private;
3874 struct geom geo;
3875
3876 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3877 return -EINVAL;
3878
3879 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3880 /* mustn't change number of copies */
3881 return -EINVAL;
3882 if (geo.far_copies > 1 && !geo.far_offset)
3883 /* Cannot switch to 'far' mode */
3884 return -EINVAL;
3885
3886 if (mddev->array_sectors & geo.chunk_mask)
3887 /* not factor of array size */
3888 return -EINVAL;
3889
3890 if (!enough(conf, -1))
3891 return -EINVAL;
3892
3893 kfree(conf->mirrors_new);
3894 conf->mirrors_new = NULL;
3895 if (mddev->delta_disks > 0) {
3896 /* allocate new 'mirrors' list */
3897 conf->mirrors_new = kzalloc(
3898 sizeof(struct raid10_info)
3899 *(mddev->raid_disks +
3900 mddev->delta_disks),
3901 GFP_KERNEL);
3902 if (!conf->mirrors_new)
3903 return -ENOMEM;
3904 }
3905 return 0;
3906}
3907
3908/*
3909 * Need to check if array has failed when deciding whether to:
3910 * - start an array
3911 * - remove non-faulty devices
3912 * - add a spare
3913 * - allow a reshape
3914 * This determination is simple when no reshape is happening.
3915 * However if there is a reshape, we need to carefully check
3916 * both the before and after sections.
3917 * This is because some failed devices may only affect one
3918 * of the two sections, and some non-in_sync devices may
3919 * be insync in the section most affected by failed devices.
3920 */
3921static int calc_degraded(struct r10conf *conf)
3922{
3923 int degraded, degraded2;
3924 int i;
3925
3926 rcu_read_lock();
3927 degraded = 0;
3928 /* 'prev' section first */
3929 for (i = 0; i < conf->prev.raid_disks; i++) {
3930 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3931 if (!rdev || test_bit(Faulty, &rdev->flags))
3932 degraded++;
3933 else if (!test_bit(In_sync, &rdev->flags))
3934 /* When we can reduce the number of devices in
3935 * an array, this might not contribute to
3936 * 'degraded'. It does now.
3937 */
3938 degraded++;
3939 }
3940 rcu_read_unlock();
3941 if (conf->geo.raid_disks == conf->prev.raid_disks)
3942 return degraded;
3943 rcu_read_lock();
3944 degraded2 = 0;
3945 for (i = 0; i < conf->geo.raid_disks; i++) {
3946 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3947 if (!rdev || test_bit(Faulty, &rdev->flags))
3948 degraded2++;
3949 else if (!test_bit(In_sync, &rdev->flags)) {
3950 /* If reshape is increasing the number of devices,
3951 * this section has already been recovered, so
3952 * it doesn't contribute to degraded.
3953 * else it does.
3954 */
3955 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3956 degraded2++;
3957 }
3958 }
3959 rcu_read_unlock();
3960 if (degraded2 > degraded)
3961 return degraded2;
3962 return degraded;
3963}
3964
3965static int raid10_start_reshape(struct mddev *mddev)
3966{
3967 /* A 'reshape' has been requested. This commits
3968 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3969 * This also checks if there are enough spares and adds them
3970 * to the array.
3971 * We currently require enough spares to make the final
3972 * array non-degraded. We also require that the difference
3973 * between old and new data_offset - on each device - is
3974 * enough that we never risk over-writing.
3975 */
3976
3977 unsigned long before_length, after_length;
3978 sector_t min_offset_diff = 0;
3979 int first = 1;
3980 struct geom new;
3981 struct r10conf *conf = mddev->private;
3982 struct md_rdev *rdev;
3983 int spares = 0;
3984 int ret;
3985
3986 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3987 return -EBUSY;
3988
3989 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3990 return -EINVAL;
3991
3992 before_length = ((1 << conf->prev.chunk_shift) *
3993 conf->prev.far_copies);
3994 after_length = ((1 << conf->geo.chunk_shift) *
3995 conf->geo.far_copies);
3996
3997 rdev_for_each(rdev, mddev) {
3998 if (!test_bit(In_sync, &rdev->flags)
3999 && !test_bit(Faulty, &rdev->flags))
4000 spares++;
4001 if (rdev->raid_disk >= 0) {
4002 long long diff = (rdev->new_data_offset
4003 - rdev->data_offset);
4004 if (!mddev->reshape_backwards)
4005 diff = -diff;
4006 if (diff < 0)
4007 diff = 0;
4008 if (first || diff < min_offset_diff)
4009 min_offset_diff = diff;
4010 }
4011 }
4012
4013 if (max(before_length, after_length) > min_offset_diff)
4014 return -EINVAL;
4015
4016 if (spares < mddev->delta_disks)
4017 return -EINVAL;
4018
4019 conf->offset_diff = min_offset_diff;
4020 spin_lock_irq(&conf->device_lock);
4021 if (conf->mirrors_new) {
4022 memcpy(conf->mirrors_new, conf->mirrors,
4023 sizeof(struct raid10_info)*conf->prev.raid_disks);
4024 smp_mb();
4025 kfree(conf->mirrors_old);
4026 conf->mirrors_old = conf->mirrors;
4027 conf->mirrors = conf->mirrors_new;
4028 conf->mirrors_new = NULL;
4029 }
4030 setup_geo(&conf->geo, mddev, geo_start);
4031 smp_mb();
4032 if (mddev->reshape_backwards) {
4033 sector_t size = raid10_size(mddev, 0, 0);
4034 if (size < mddev->array_sectors) {
4035 spin_unlock_irq(&conf->device_lock);
4036 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4037 mdname(mddev));
4038 return -EINVAL;
4039 }
4040 mddev->resync_max_sectors = size;
4041 conf->reshape_progress = size;
4042 } else
4043 conf->reshape_progress = 0;
4044 conf->reshape_safe = conf->reshape_progress;
4045 spin_unlock_irq(&conf->device_lock);
4046
4047 if (mddev->delta_disks && mddev->bitmap) {
4048 ret = bitmap_resize(mddev->bitmap,
4049 raid10_size(mddev, 0,
4050 conf->geo.raid_disks),
4051 0, 0);
4052 if (ret)
4053 goto abort;
4054 }
4055 if (mddev->delta_disks > 0) {
4056 rdev_for_each(rdev, mddev)
4057 if (rdev->raid_disk < 0 &&
4058 !test_bit(Faulty, &rdev->flags)) {
4059 if (raid10_add_disk(mddev, rdev) == 0) {
4060 if (rdev->raid_disk >=
4061 conf->prev.raid_disks)
4062 set_bit(In_sync, &rdev->flags);
4063 else
4064 rdev->recovery_offset = 0;
4065
4066 if (sysfs_link_rdev(mddev, rdev))
4067 /* Failure here is OK */;
4068 }
4069 } else if (rdev->raid_disk >= conf->prev.raid_disks
4070 && !test_bit(Faulty, &rdev->flags)) {
4071 /* This is a spare that was manually added */
4072 set_bit(In_sync, &rdev->flags);
4073 }
4074 }
4075 /* When a reshape changes the number of devices,
4076 * ->degraded is measured against the larger of the
4077 * pre and post numbers.
4078 */
4079 spin_lock_irq(&conf->device_lock);
4080 mddev->degraded = calc_degraded(conf);
4081 spin_unlock_irq(&conf->device_lock);
4082 mddev->raid_disks = conf->geo.raid_disks;
4083 mddev->reshape_position = conf->reshape_progress;
4084 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4085
4086 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4087 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4088 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4089 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4090 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4091
4092 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4093 "reshape");
4094 if (!mddev->sync_thread) {
4095 ret = -EAGAIN;
4096 goto abort;
4097 }
4098 conf->reshape_checkpoint = jiffies;
4099 md_wakeup_thread(mddev->sync_thread);
4100 md_new_event(mddev);
4101 return 0;
4102
4103abort:
4104 mddev->recovery = 0;
4105 spin_lock_irq(&conf->device_lock);
4106 conf->geo = conf->prev;
4107 mddev->raid_disks = conf->geo.raid_disks;
4108 rdev_for_each(rdev, mddev)
4109 rdev->new_data_offset = rdev->data_offset;
4110 smp_wmb();
4111 conf->reshape_progress = MaxSector;
4112 conf->reshape_safe = MaxSector;
4113 mddev->reshape_position = MaxSector;
4114 spin_unlock_irq(&conf->device_lock);
4115 return ret;
4116}
4117
4118/* Calculate the last device-address that could contain
4119 * any block from the chunk that includes the array-address 's'
4120 * and report the next address.
4121 * i.e. the address returned will be chunk-aligned and after
4122 * any data that is in the chunk containing 's'.
4123 */
4124static sector_t last_dev_address(sector_t s, struct geom *geo)
4125{
4126 s = (s | geo->chunk_mask) + 1;
4127 s >>= geo->chunk_shift;
4128 s *= geo->near_copies;
4129 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4130 s *= geo->far_copies;
4131 s <<= geo->chunk_shift;
4132 return s;
4133}
4134
4135/* Calculate the first device-address that could contain
4136 * any block from the chunk that includes the array-address 's'.
4137 * This too will be the start of a chunk
4138 */
4139static sector_t first_dev_address(sector_t s, struct geom *geo)
4140{
4141 s >>= geo->chunk_shift;
4142 s *= geo->near_copies;
4143 sector_div(s, geo->raid_disks);
4144 s *= geo->far_copies;
4145 s <<= geo->chunk_shift;
4146 return s;
4147}
4148
4149static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4150 int *skipped)
4151{
4152 /* We simply copy at most one chunk (smallest of old and new)
4153 * at a time, possibly less if that exceeds RESYNC_PAGES,
4154 * or we hit a bad block or something.
4155 * This might mean we pause for normal IO in the middle of
4156 * a chunk, but that is not a problem as mddev->reshape_position
4157 * can record any location.
4158 *
4159 * If we will want to write to a location that isn't
4160 * yet recorded as 'safe' (i.e. in metadata on disk) then
4161 * we need to flush all reshape requests and update the metadata.
4162 *
4163 * When reshaping forwards (e.g. to more devices), we interpret
4164 * 'safe' as the earliest block which might not have been copied
4165 * down yet. We divide this by previous stripe size and multiply
4166 * by previous stripe length to get lowest device offset that we
4167 * cannot write to yet.
4168 * We interpret 'sector_nr' as an address that we want to write to.
4169 * From this we use last_device_address() to find where we might
4170 * write to, and first_device_address on the 'safe' position.
4171 * If this 'next' write position is after the 'safe' position,
4172 * we must update the metadata to increase the 'safe' position.
4173 *
4174 * When reshaping backwards, we round in the opposite direction
4175 * and perform the reverse test: next write position must not be
4176 * less than current safe position.
4177 *
4178 * In all this the minimum difference in data offsets
4179 * (conf->offset_diff - always positive) allows a bit of slack,
4180 * so next can be after 'safe', but not by more than offset_diff
4181 *
4182 * We need to prepare all the bios here before we start any IO
4183 * to ensure the size we choose is acceptable to all devices.
4184 * The means one for each copy for write-out and an extra one for
4185 * read-in.
4186 * We store the read-in bio in ->master_bio and the others in
4187 * ->devs[x].bio and ->devs[x].repl_bio.
4188 */
4189 struct r10conf *conf = mddev->private;
4190 struct r10bio *r10_bio;
4191 sector_t next, safe, last;
4192 int max_sectors;
4193 int nr_sectors;
4194 int s;
4195 struct md_rdev *rdev;
4196 int need_flush = 0;
4197 struct bio *blist;
4198 struct bio *bio, *read_bio;
4199 int sectors_done = 0;
4200
4201 if (sector_nr == 0) {
4202 /* If restarting in the middle, skip the initial sectors */
4203 if (mddev->reshape_backwards &&
4204 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4205 sector_nr = (raid10_size(mddev, 0, 0)
4206 - conf->reshape_progress);
4207 } else if (!mddev->reshape_backwards &&
4208 conf->reshape_progress > 0)
4209 sector_nr = conf->reshape_progress;
4210 if (sector_nr) {
4211 mddev->curr_resync_completed = sector_nr;
4212 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4213 *skipped = 1;
4214 return sector_nr;
4215 }
4216 }
4217
4218 /* We don't use sector_nr to track where we are up to
4219 * as that doesn't work well for ->reshape_backwards.
4220 * So just use ->reshape_progress.
4221 */
4222 if (mddev->reshape_backwards) {
4223 /* 'next' is the earliest device address that we might
4224 * write to for this chunk in the new layout
4225 */
4226 next = first_dev_address(conf->reshape_progress - 1,
4227 &conf->geo);
4228
4229 /* 'safe' is the last device address that we might read from
4230 * in the old layout after a restart
4231 */
4232 safe = last_dev_address(conf->reshape_safe - 1,
4233 &conf->prev);
4234
4235 if (next + conf->offset_diff < safe)
4236 need_flush = 1;
4237
4238 last = conf->reshape_progress - 1;
4239 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4240 & conf->prev.chunk_mask);
4241 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4242 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4243 } else {
4244 /* 'next' is after the last device address that we
4245 * might write to for this chunk in the new layout
4246 */
4247 next = last_dev_address(conf->reshape_progress, &conf->geo);
4248
4249 /* 'safe' is the earliest device address that we might
4250 * read from in the old layout after a restart
4251 */
4252 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4253
4254 /* Need to update metadata if 'next' might be beyond 'safe'
4255 * as that would possibly corrupt data
4256 */
4257 if (next > safe + conf->offset_diff)
4258 need_flush = 1;
4259
4260 sector_nr = conf->reshape_progress;
4261 last = sector_nr | (conf->geo.chunk_mask
4262 & conf->prev.chunk_mask);
4263
4264 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4265 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4266 }
4267
4268 if (need_flush ||
4269 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4270 /* Need to update reshape_position in metadata */
4271 wait_barrier(conf);
4272 mddev->reshape_position = conf->reshape_progress;
4273 if (mddev->reshape_backwards)
4274 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4275 - conf->reshape_progress;
4276 else
4277 mddev->curr_resync_completed = conf->reshape_progress;
4278 conf->reshape_checkpoint = jiffies;
4279 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4280 md_wakeup_thread(mddev->thread);
4281 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4282 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4283 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4284 allow_barrier(conf);
4285 return sectors_done;
4286 }
4287 conf->reshape_safe = mddev->reshape_position;
4288 allow_barrier(conf);
4289 }
4290
4291read_more:
4292 /* Now schedule reads for blocks from sector_nr to last */
4293 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4294 r10_bio->state = 0;
4295 raise_barrier(conf, sectors_done != 0);
4296 atomic_set(&r10_bio->remaining, 0);
4297 r10_bio->mddev = mddev;
4298 r10_bio->sector = sector_nr;
4299 set_bit(R10BIO_IsReshape, &r10_bio->state);
4300 r10_bio->sectors = last - sector_nr + 1;
4301 rdev = read_balance(conf, r10_bio, &max_sectors);
4302 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4303
4304 if (!rdev) {
4305 /* Cannot read from here, so need to record bad blocks
4306 * on all the target devices.
4307 */
4308 // FIXME
4309 mempool_free(r10_bio, conf->r10buf_pool);
4310 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4311 return sectors_done;
4312 }
4313
4314 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4315
4316 read_bio->bi_bdev = rdev->bdev;
4317 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4318 + rdev->data_offset);
4319 read_bio->bi_private = r10_bio;
4320 read_bio->bi_end_io = end_sync_read;
4321 read_bio->bi_rw = READ;
4322 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4323 read_bio->bi_error = 0;
4324 read_bio->bi_vcnt = 0;
4325 read_bio->bi_iter.bi_size = 0;
4326 r10_bio->master_bio = read_bio;
4327 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4328
4329 /* Now find the locations in the new layout */
4330 __raid10_find_phys(&conf->geo, r10_bio);
4331
4332 blist = read_bio;
4333 read_bio->bi_next = NULL;
4334
4335 for (s = 0; s < conf->copies*2; s++) {
4336 struct bio *b;
4337 int d = r10_bio->devs[s/2].devnum;
4338 struct md_rdev *rdev2;
4339 if (s&1) {
4340 rdev2 = conf->mirrors[d].replacement;
4341 b = r10_bio->devs[s/2].repl_bio;
4342 } else {
4343 rdev2 = conf->mirrors[d].rdev;
4344 b = r10_bio->devs[s/2].bio;
4345 }
4346 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4347 continue;
4348
4349 bio_reset(b);
4350 b->bi_bdev = rdev2->bdev;
4351 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4352 rdev2->new_data_offset;
4353 b->bi_private = r10_bio;
4354 b->bi_end_io = end_reshape_write;
4355 b->bi_rw = WRITE;
4356 b->bi_next = blist;
4357 blist = b;
4358 }
4359
4360 /* Now add as many pages as possible to all of these bios. */
4361
4362 nr_sectors = 0;
4363 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4364 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4365 int len = (max_sectors - s) << 9;
4366 if (len > PAGE_SIZE)
4367 len = PAGE_SIZE;
4368 for (bio = blist; bio ; bio = bio->bi_next) {
4369 struct bio *bio2;
4370 if (bio_add_page(bio, page, len, 0))
4371 continue;
4372
4373 /* Didn't fit, must stop */
4374 for (bio2 = blist;
4375 bio2 && bio2 != bio;
4376 bio2 = bio2->bi_next) {
4377 /* Remove last page from this bio */
4378 bio2->bi_vcnt--;
4379 bio2->bi_iter.bi_size -= len;
4380 bio_clear_flag(bio2, BIO_SEG_VALID);
4381 }
4382 goto bio_full;
4383 }
4384 sector_nr += len >> 9;
4385 nr_sectors += len >> 9;
4386 }
4387bio_full:
4388 r10_bio->sectors = nr_sectors;
4389
4390 /* Now submit the read */
4391 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4392 atomic_inc(&r10_bio->remaining);
4393 read_bio->bi_next = NULL;
4394 generic_make_request(read_bio);
4395 sector_nr += nr_sectors;
4396 sectors_done += nr_sectors;
4397 if (sector_nr <= last)
4398 goto read_more;
4399
4400 /* Now that we have done the whole section we can
4401 * update reshape_progress
4402 */
4403 if (mddev->reshape_backwards)
4404 conf->reshape_progress -= sectors_done;
4405 else
4406 conf->reshape_progress += sectors_done;
4407
4408 return sectors_done;
4409}
4410
4411static void end_reshape_request(struct r10bio *r10_bio);
4412static int handle_reshape_read_error(struct mddev *mddev,
4413 struct r10bio *r10_bio);
4414static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4415{
4416 /* Reshape read completed. Hopefully we have a block
4417 * to write out.
4418 * If we got a read error then we do sync 1-page reads from
4419 * elsewhere until we find the data - or give up.
4420 */
4421 struct r10conf *conf = mddev->private;
4422 int s;
4423
4424 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4425 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4426 /* Reshape has been aborted */
4427 md_done_sync(mddev, r10_bio->sectors, 0);
4428 return;
4429 }
4430
4431 /* We definitely have the data in the pages, schedule the
4432 * writes.
4433 */
4434 atomic_set(&r10_bio->remaining, 1);
4435 for (s = 0; s < conf->copies*2; s++) {
4436 struct bio *b;
4437 int d = r10_bio->devs[s/2].devnum;
4438 struct md_rdev *rdev;
4439 if (s&1) {
4440 rdev = conf->mirrors[d].replacement;
4441 b = r10_bio->devs[s/2].repl_bio;
4442 } else {
4443 rdev = conf->mirrors[d].rdev;
4444 b = r10_bio->devs[s/2].bio;
4445 }
4446 if (!rdev || test_bit(Faulty, &rdev->flags))
4447 continue;
4448 atomic_inc(&rdev->nr_pending);
4449 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4450 atomic_inc(&r10_bio->remaining);
4451 b->bi_next = NULL;
4452 generic_make_request(b);
4453 }
4454 end_reshape_request(r10_bio);
4455}
4456
4457static void end_reshape(struct r10conf *conf)
4458{
4459 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4460 return;
4461
4462 spin_lock_irq(&conf->device_lock);
4463 conf->prev = conf->geo;
4464 md_finish_reshape(conf->mddev);
4465 smp_wmb();
4466 conf->reshape_progress = MaxSector;
4467 conf->reshape_safe = MaxSector;
4468 spin_unlock_irq(&conf->device_lock);
4469
4470 /* read-ahead size must cover two whole stripes, which is
4471 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4472 */
4473 if (conf->mddev->queue) {
4474 int stripe = conf->geo.raid_disks *
4475 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4476 stripe /= conf->geo.near_copies;
4477 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4478 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4479 }
4480 conf->fullsync = 0;
4481}
4482
4483static int handle_reshape_read_error(struct mddev *mddev,
4484 struct r10bio *r10_bio)
4485{
4486 /* Use sync reads to get the blocks from somewhere else */
4487 int sectors = r10_bio->sectors;
4488 struct r10conf *conf = mddev->private;
4489 struct {
4490 struct r10bio r10_bio;
4491 struct r10dev devs[conf->copies];
4492 } on_stack;
4493 struct r10bio *r10b = &on_stack.r10_bio;
4494 int slot = 0;
4495 int idx = 0;
4496 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4497
4498 r10b->sector = r10_bio->sector;
4499 __raid10_find_phys(&conf->prev, r10b);
4500
4501 while (sectors) {
4502 int s = sectors;
4503 int success = 0;
4504 int first_slot = slot;
4505
4506 if (s > (PAGE_SIZE >> 9))
4507 s = PAGE_SIZE >> 9;
4508
4509 while (!success) {
4510 int d = r10b->devs[slot].devnum;
4511 struct md_rdev *rdev = conf->mirrors[d].rdev;
4512 sector_t addr;
4513 if (rdev == NULL ||
4514 test_bit(Faulty, &rdev->flags) ||
4515 !test_bit(In_sync, &rdev->flags))
4516 goto failed;
4517
4518 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4519 success = sync_page_io(rdev,
4520 addr,
4521 s << 9,
4522 bvec[idx].bv_page,
4523 READ, false);
4524 if (success)
4525 break;
4526 failed:
4527 slot++;
4528 if (slot >= conf->copies)
4529 slot = 0;
4530 if (slot == first_slot)
4531 break;
4532 }
4533 if (!success) {
4534 /* couldn't read this block, must give up */
4535 set_bit(MD_RECOVERY_INTR,
4536 &mddev->recovery);
4537 return -EIO;
4538 }
4539 sectors -= s;
4540 idx++;
4541 }
4542 return 0;
4543}
4544
4545static void end_reshape_write(struct bio *bio)
4546{
4547 struct r10bio *r10_bio = bio->bi_private;
4548 struct mddev *mddev = r10_bio->mddev;
4549 struct r10conf *conf = mddev->private;
4550 int d;
4551 int slot;
4552 int repl;
4553 struct md_rdev *rdev = NULL;
4554
4555 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4556 if (repl)
4557 rdev = conf->mirrors[d].replacement;
4558 if (!rdev) {
4559 smp_mb();
4560 rdev = conf->mirrors[d].rdev;
4561 }
4562
4563 if (bio->bi_error) {
4564 /* FIXME should record badblock */
4565 md_error(mddev, rdev);
4566 }
4567
4568 rdev_dec_pending(rdev, mddev);
4569 end_reshape_request(r10_bio);
4570}
4571
4572static void end_reshape_request(struct r10bio *r10_bio)
4573{
4574 if (!atomic_dec_and_test(&r10_bio->remaining))
4575 return;
4576 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4577 bio_put(r10_bio->master_bio);
4578 put_buf(r10_bio);
4579}
4580
4581static void raid10_finish_reshape(struct mddev *mddev)
4582{
4583 struct r10conf *conf = mddev->private;
4584
4585 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4586 return;
4587
4588 if (mddev->delta_disks > 0) {
4589 sector_t size = raid10_size(mddev, 0, 0);
4590 md_set_array_sectors(mddev, size);
4591 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4592 mddev->recovery_cp = mddev->resync_max_sectors;
4593 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4594 }
4595 mddev->resync_max_sectors = size;
4596 set_capacity(mddev->gendisk, mddev->array_sectors);
4597 revalidate_disk(mddev->gendisk);
4598 } else {
4599 int d;
4600 for (d = conf->geo.raid_disks ;
4601 d < conf->geo.raid_disks - mddev->delta_disks;
4602 d++) {
4603 struct md_rdev *rdev = conf->mirrors[d].rdev;
4604 if (rdev)
4605 clear_bit(In_sync, &rdev->flags);
4606 rdev = conf->mirrors[d].replacement;
4607 if (rdev)
4608 clear_bit(In_sync, &rdev->flags);
4609 }
4610 }
4611 mddev->layout = mddev->new_layout;
4612 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4613 mddev->reshape_position = MaxSector;
4614 mddev->delta_disks = 0;
4615 mddev->reshape_backwards = 0;
4616}
4617
4618static struct md_personality raid10_personality =
4619{
4620 .name = "raid10",
4621 .level = 10,
4622 .owner = THIS_MODULE,
4623 .make_request = raid10_make_request,
4624 .run = raid10_run,
4625 .free = raid10_free,
4626 .status = raid10_status,
4627 .error_handler = raid10_error,
4628 .hot_add_disk = raid10_add_disk,
4629 .hot_remove_disk= raid10_remove_disk,
4630 .spare_active = raid10_spare_active,
4631 .sync_request = raid10_sync_request,
4632 .quiesce = raid10_quiesce,
4633 .size = raid10_size,
4634 .resize = raid10_resize,
4635 .takeover = raid10_takeover,
4636 .check_reshape = raid10_check_reshape,
4637 .start_reshape = raid10_start_reshape,
4638 .finish_reshape = raid10_finish_reshape,
4639 .congested = raid10_congested,
4640};
4641
4642static int __init raid_init(void)
4643{
4644 return register_md_personality(&raid10_personality);
4645}
4646
4647static void raid_exit(void)
4648{
4649 unregister_md_personality(&raid10_personality);
4650}
4651
4652module_init(raid_init);
4653module_exit(raid_exit);
4654MODULE_LICENSE("GPL");
4655MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4656MODULE_ALIAS("md-personality-9"); /* RAID10 */
4657MODULE_ALIAS("md-raid10");
4658MODULE_ALIAS("md-level-10");
4659
4660module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);