2 * raid10.c : Multiple Devices driver for Linux
4 * Copyright (C) 2000-2004 Neil Brown
6 * RAID-10 support for md.
8 * Base on code in raid1.c. See raid1.c for further copyright information.
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)
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.
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>
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
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 )
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.
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.
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
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]
75 * Number of guaranteed r10bios in case of extreme VM load:
77 #define NR_RAID10_BIOS 256
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
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
89 #define IO_MADE_GOOD ((struct bio *)2)
91 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
93 /* When there are this many requests queued to be written by
94 * the raid10 thread, we become 'congested' to provide back-pressure
97 static int max_queued_requests = 1024;
99 static void allow_barrier(struct r10conf *conf);
100 static void lower_barrier(struct r10conf *conf);
101 static int _enough(struct r10conf *conf, int previous, int ignore);
102 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
104 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
105 static void end_reshape_write(struct bio *bio);
106 static void end_reshape(struct r10conf *conf);
108 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
110 struct r10conf *conf = data;
111 int size = offsetof(struct r10bio, devs[conf->copies]);
113 /* allocate a r10bio with room for raid_disks entries in the
115 return kzalloc(size, gfp_flags);
118 static void r10bio_pool_free(void *r10_bio, void *data)
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)
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)
138 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
140 struct r10conf *conf = data;
142 struct r10bio *r10_bio;
147 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
151 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
152 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
153 nalloc = conf->copies; /* resync */
155 nalloc = 2; /* recovery */
160 for (j = nalloc ; j-- ; ) {
161 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
164 r10_bio->devs[j].bio = bio;
165 if (!conf->have_replacement)
167 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
170 r10_bio->devs[j].repl_bio = bio;
173 * Allocate RESYNC_PAGES data pages and attach them
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
184 struct bio *rbio = r10_bio->devs[0].bio;
185 page = rbio->bi_io_vec[i].bv_page;
188 page = alloc_page(gfp_flags);
192 bio->bi_io_vec[i].bv_page = page;
194 rbio->bi_io_vec[i].bv_page = page;
202 safe_put_page(bio->bi_io_vec[i-1].bv_page);
204 for (i = 0; i < RESYNC_PAGES ; i++)
205 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
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);
214 r10bio_pool_free(r10_bio, conf);
218 static void r10buf_pool_free(void *__r10_bio, void *data)
221 struct r10conf *conf = data;
222 struct r10bio *r10bio = __r10_bio;
225 for (j=0; j < conf->copies; j++) {
226 struct bio *bio = r10bio->devs[j].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;
234 bio = r10bio->devs[j].repl_bio;
238 r10bio_pool_free(r10bio, conf);
241 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
245 for (i = 0; i < conf->copies; i++) {
246 struct bio **bio = & r10_bio->devs[i].bio;
247 if (!BIO_SPECIAL(*bio))
250 bio = &r10_bio->devs[i].repl_bio;
251 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
257 static void free_r10bio(struct r10bio *r10_bio)
259 struct r10conf *conf = r10_bio->mddev->private;
261 put_all_bios(conf, r10_bio);
262 mempool_free(r10_bio, conf->r10bio_pool);
265 static void put_buf(struct r10bio *r10_bio)
267 struct r10conf *conf = r10_bio->mddev->private;
269 mempool_free(r10_bio, conf->r10buf_pool);
274 static void reschedule_retry(struct r10bio *r10_bio)
277 struct mddev *mddev = r10_bio->mddev;
278 struct r10conf *conf = mddev->private;
280 spin_lock_irqsave(&conf->device_lock, flags);
281 list_add(&r10_bio->retry_list, &conf->retry_list);
283 spin_unlock_irqrestore(&conf->device_lock, flags);
285 /* wake up frozen array... */
286 wake_up(&conf->wait_barrier);
288 md_wakeup_thread(mddev->thread);
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
296 static void raid_end_bio_io(struct r10bio *r10_bio)
298 struct bio *bio = r10_bio->master_bio;
300 struct r10conf *conf = r10_bio->mddev->private;
302 if (bio->bi_phys_segments) {
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);
310 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
311 bio->bi_error = -EIO;
315 * Wake up any possible resync thread that waits for the device
320 free_r10bio(r10_bio);
324 * Update disk head position estimator based on IRQ completion info.
326 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
328 struct r10conf *conf = r10_bio->mddev->private;
330 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
331 r10_bio->devs[slot].addr + (r10_bio->sectors);
335 * Find the disk number which triggered given bio
337 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
338 struct bio *bio, int *slotp, int *replp)
343 for (slot = 0; slot < conf->copies; slot++) {
344 if (r10_bio->devs[slot].bio == bio)
346 if (r10_bio->devs[slot].repl_bio == bio) {
352 BUG_ON(slot == conf->copies);
353 update_head_pos(slot, r10_bio);
359 return r10_bio->devs[slot].devnum;
362 static void raid10_end_read_request(struct bio *bio)
364 int uptodate = !bio->bi_error;
365 struct r10bio *r10_bio = bio->bi_private;
367 struct md_rdev *rdev;
368 struct r10conf *conf = r10_bio->mddev->private;
370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
372 rdev = r10_bio->devs[slot].rdev;
374 * this branch is our 'one mirror IO has finished' event handler:
376 update_head_pos(slot, r10_bio);
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.
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.
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
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"
395 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
400 raid_end_bio_io(r10_bio);
401 rdev_dec_pending(rdev, conf->mddev);
404 * oops, read error - keep the refcount on the rdev
406 char b[BDEVNAME_SIZE];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
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);
417 static void close_write(struct r10bio *r10_bio)
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
422 !test_bit(R10BIO_Degraded, &r10_bio->state),
424 md_write_end(r10_bio->mddev);
427 static void one_write_done(struct r10bio *r10_bio)
429 if (atomic_dec_and_test(&r10_bio->remaining)) {
430 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431 reschedule_retry(r10_bio);
433 close_write(r10_bio);
434 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435 reschedule_retry(r10_bio);
437 raid_end_bio_io(r10_bio);
442 static void raid10_end_write_request(struct bio *bio)
444 struct r10bio *r10_bio = bio->bi_private;
447 struct r10conf *conf = r10_bio->mddev->private;
449 struct md_rdev *rdev = NULL;
452 discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
454 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
457 rdev = conf->mirrors[dev].replacement;
461 rdev = conf->mirrors[dev].rdev;
464 * this branch is our 'one mirror IO has finished' event handler:
466 if (bio->bi_error && !discard_error) {
468 /* Never record new bad blocks to replacement,
471 md_error(rdev->mddev, rdev);
473 set_bit(WriteErrorSeen, &rdev->flags);
474 if (!test_and_set_bit(WantReplacement, &rdev->flags))
475 set_bit(MD_RECOVERY_NEEDED,
476 &rdev->mddev->recovery);
477 set_bit(R10BIO_WriteError, &r10_bio->state);
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
494 * Do not set R10BIO_Uptodate if the current device is
495 * rebuilding or Faulty. This is because we cannot use
496 * such device for properly reading the data back (we could
497 * potentially use it, if the current write would have felt
498 * before rdev->recovery_offset, but for simplicity we don't
501 if (test_bit(In_sync, &rdev->flags) &&
502 !test_bit(Faulty, &rdev->flags))
503 set_bit(R10BIO_Uptodate, &r10_bio->state);
505 /* Maybe we can clear some bad blocks. */
506 if (is_badblock(rdev,
507 r10_bio->devs[slot].addr,
509 &first_bad, &bad_sectors) && !discard_error) {
512 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
514 r10_bio->devs[slot].bio = IO_MADE_GOOD;
516 set_bit(R10BIO_MadeGood, &r10_bio->state);
522 * Let's see if all mirrored write operations have finished
525 one_write_done(r10_bio);
527 rdev_dec_pending(rdev, conf->mddev);
531 * RAID10 layout manager
532 * As well as the chunksize and raid_disks count, there are two
533 * parameters: near_copies and far_copies.
534 * near_copies * far_copies must be <= raid_disks.
535 * Normally one of these will be 1.
536 * If both are 1, we get raid0.
537 * If near_copies == raid_disks, we get raid1.
539 * Chunks are laid out in raid0 style with near_copies copies of the
540 * first chunk, followed by near_copies copies of the next chunk and
542 * If far_copies > 1, then after 1/far_copies of the array has been assigned
543 * as described above, we start again with a device offset of near_copies.
544 * So we effectively have another copy of the whole array further down all
545 * the drives, but with blocks on different drives.
546 * With this layout, and block is never stored twice on the one device.
548 * raid10_find_phys finds the sector offset of a given virtual sector
549 * on each device that it is on.
551 * raid10_find_virt does the reverse mapping, from a device and a
552 * sector offset to a virtual address
555 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
563 int last_far_set_start, last_far_set_size;
565 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
566 last_far_set_start *= geo->far_set_size;
568 last_far_set_size = geo->far_set_size;
569 last_far_set_size += (geo->raid_disks % geo->far_set_size);
571 /* now calculate first sector/dev */
572 chunk = r10bio->sector >> geo->chunk_shift;
573 sector = r10bio->sector & geo->chunk_mask;
575 chunk *= geo->near_copies;
577 dev = sector_div(stripe, geo->raid_disks);
579 stripe *= geo->far_copies;
581 sector += stripe << geo->chunk_shift;
583 /* and calculate all the others */
584 for (n = 0; n < geo->near_copies; n++) {
588 r10bio->devs[slot].devnum = d;
589 r10bio->devs[slot].addr = s;
592 for (f = 1; f < geo->far_copies; f++) {
593 set = d / geo->far_set_size;
594 d += geo->near_copies;
596 if ((geo->raid_disks % geo->far_set_size) &&
597 (d > last_far_set_start)) {
598 d -= last_far_set_start;
599 d %= last_far_set_size;
600 d += last_far_set_start;
602 d %= geo->far_set_size;
603 d += geo->far_set_size * set;
606 r10bio->devs[slot].devnum = d;
607 r10bio->devs[slot].addr = s;
611 if (dev >= geo->raid_disks) {
613 sector += (geo->chunk_mask + 1);
618 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
620 struct geom *geo = &conf->geo;
622 if (conf->reshape_progress != MaxSector &&
623 ((r10bio->sector >= conf->reshape_progress) !=
624 conf->mddev->reshape_backwards)) {
625 set_bit(R10BIO_Previous, &r10bio->state);
628 clear_bit(R10BIO_Previous, &r10bio->state);
630 __raid10_find_phys(geo, r10bio);
633 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
635 sector_t offset, chunk, vchunk;
636 /* Never use conf->prev as this is only called during resync
637 * or recovery, so reshape isn't happening
639 struct geom *geo = &conf->geo;
640 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
641 int far_set_size = geo->far_set_size;
642 int last_far_set_start;
644 if (geo->raid_disks % geo->far_set_size) {
645 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
646 last_far_set_start *= geo->far_set_size;
648 if (dev >= last_far_set_start) {
649 far_set_size = geo->far_set_size;
650 far_set_size += (geo->raid_disks % geo->far_set_size);
651 far_set_start = last_far_set_start;
655 offset = sector & geo->chunk_mask;
656 if (geo->far_offset) {
658 chunk = sector >> geo->chunk_shift;
659 fc = sector_div(chunk, geo->far_copies);
660 dev -= fc * geo->near_copies;
661 if (dev < far_set_start)
664 while (sector >= geo->stride) {
665 sector -= geo->stride;
666 if (dev < (geo->near_copies + far_set_start))
667 dev += far_set_size - geo->near_copies;
669 dev -= geo->near_copies;
671 chunk = sector >> geo->chunk_shift;
673 vchunk = chunk * geo->raid_disks + dev;
674 sector_div(vchunk, geo->near_copies);
675 return (vchunk << geo->chunk_shift) + offset;
679 * This routine returns the disk from which the requested read should
680 * be done. There is a per-array 'next expected sequential IO' sector
681 * number - if this matches on the next IO then we use the last disk.
682 * There is also a per-disk 'last know head position' sector that is
683 * maintained from IRQ contexts, both the normal and the resync IO
684 * completion handlers update this position correctly. If there is no
685 * perfect sequential match then we pick the disk whose head is closest.
687 * If there are 2 mirrors in the same 2 devices, performance degrades
688 * because position is mirror, not device based.
690 * The rdev for the device selected will have nr_pending incremented.
694 * FIXME: possibly should rethink readbalancing and do it differently
695 * depending on near_copies / far_copies geometry.
697 static struct md_rdev *read_balance(struct r10conf *conf,
698 struct r10bio *r10_bio,
701 const sector_t this_sector = r10_bio->sector;
703 int sectors = r10_bio->sectors;
704 int best_good_sectors;
705 sector_t new_distance, best_dist;
706 struct md_rdev *best_rdev, *rdev = NULL;
709 struct geom *geo = &conf->geo;
711 raid10_find_phys(conf, r10_bio);
713 sectors = r10_bio->sectors;
716 best_dist = MaxSector;
717 best_good_sectors = 0;
720 * Check if we can balance. We can balance on the whole
721 * device if no resync is going on (recovery is ok), or below
722 * the resync window. We take the first readable disk when
723 * above the resync window.
725 if (conf->mddev->recovery_cp < MaxSector
726 && (this_sector + sectors >= conf->next_resync))
729 for (slot = 0; slot < conf->copies ; slot++) {
734 if (r10_bio->devs[slot].bio == IO_BLOCKED)
736 disk = r10_bio->devs[slot].devnum;
737 rdev = rcu_dereference(conf->mirrors[disk].replacement);
738 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
739 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
740 rdev = rcu_dereference(conf->mirrors[disk].rdev);
742 test_bit(Faulty, &rdev->flags))
744 if (!test_bit(In_sync, &rdev->flags) &&
745 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
748 dev_sector = r10_bio->devs[slot].addr;
749 if (is_badblock(rdev, dev_sector, sectors,
750 &first_bad, &bad_sectors)) {
751 if (best_dist < MaxSector)
752 /* Already have a better slot */
754 if (first_bad <= dev_sector) {
755 /* Cannot read here. If this is the
756 * 'primary' device, then we must not read
757 * beyond 'bad_sectors' from another device.
759 bad_sectors -= (dev_sector - first_bad);
760 if (!do_balance && sectors > bad_sectors)
761 sectors = bad_sectors;
762 if (best_good_sectors > sectors)
763 best_good_sectors = sectors;
765 sector_t good_sectors =
766 first_bad - dev_sector;
767 if (good_sectors > best_good_sectors) {
768 best_good_sectors = good_sectors;
773 /* Must read from here */
778 best_good_sectors = sectors;
783 /* This optimisation is debatable, and completely destroys
784 * sequential read speed for 'far copies' arrays. So only
785 * keep it for 'near' arrays, and review those later.
787 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
790 /* for far > 1 always use the lowest address */
791 if (geo->far_copies > 1)
792 new_distance = r10_bio->devs[slot].addr;
794 new_distance = abs(r10_bio->devs[slot].addr -
795 conf->mirrors[disk].head_position);
796 if (new_distance < best_dist) {
797 best_dist = new_distance;
802 if (slot >= conf->copies) {
808 atomic_inc(&rdev->nr_pending);
809 r10_bio->read_slot = slot;
813 *max_sectors = best_good_sectors;
818 static int raid10_congested(struct mddev *mddev, int bits)
820 struct r10conf *conf = mddev->private;
823 if ((bits & (1 << WB_async_congested)) &&
824 conf->pending_count >= max_queued_requests)
829 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
832 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
833 if (rdev && !test_bit(Faulty, &rdev->flags)) {
834 struct request_queue *q = bdev_get_queue(rdev->bdev);
836 ret |= bdi_congested(&q->backing_dev_info, bits);
843 static void flush_pending_writes(struct r10conf *conf)
845 /* Any writes that have been queued but are awaiting
846 * bitmap updates get flushed here.
848 spin_lock_irq(&conf->device_lock);
850 if (conf->pending_bio_list.head) {
852 bio = bio_list_get(&conf->pending_bio_list);
853 conf->pending_count = 0;
854 spin_unlock_irq(&conf->device_lock);
855 /* flush any pending bitmap writes to disk
856 * before proceeding w/ I/O */
857 bitmap_unplug(conf->mddev->bitmap);
858 wake_up(&conf->wait_barrier);
860 while (bio) { /* submit pending writes */
861 struct bio *next = bio->bi_next;
863 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
864 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
868 generic_make_request(bio);
872 spin_unlock_irq(&conf->device_lock);
876 * Sometimes we need to suspend IO while we do something else,
877 * either some resync/recovery, or reconfigure the array.
878 * To do this we raise a 'barrier'.
879 * The 'barrier' is a counter that can be raised multiple times
880 * to count how many activities are happening which preclude
882 * We can only raise the barrier if there is no pending IO.
883 * i.e. if nr_pending == 0.
884 * We choose only to raise the barrier if no-one is waiting for the
885 * barrier to go down. This means that as soon as an IO request
886 * is ready, no other operations which require a barrier will start
887 * until the IO request has had a chance.
889 * So: regular IO calls 'wait_barrier'. When that returns there
890 * is no backgroup IO happening, It must arrange to call
891 * allow_barrier when it has finished its IO.
892 * backgroup IO calls must call raise_barrier. Once that returns
893 * there is no normal IO happeing. It must arrange to call
894 * lower_barrier when the particular background IO completes.
897 static void raise_barrier(struct r10conf *conf, int force)
899 BUG_ON(force && !conf->barrier);
900 spin_lock_irq(&conf->resync_lock);
902 /* Wait until no block IO is waiting (unless 'force') */
903 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
906 /* block any new IO from starting */
909 /* Now wait for all pending IO to complete */
910 wait_event_lock_irq(conf->wait_barrier,
911 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
914 spin_unlock_irq(&conf->resync_lock);
917 static void lower_barrier(struct r10conf *conf)
920 spin_lock_irqsave(&conf->resync_lock, flags);
922 spin_unlock_irqrestore(&conf->resync_lock, flags);
923 wake_up(&conf->wait_barrier);
926 static void wait_barrier(struct r10conf *conf)
928 spin_lock_irq(&conf->resync_lock);
931 /* Wait for the barrier to drop.
932 * However if there are already pending
933 * requests (preventing the barrier from
934 * rising completely), and the
935 * pre-process bio queue isn't empty,
936 * then don't wait, as we need to empty
937 * that queue to get the nr_pending
940 wait_event_lock_irq(conf->wait_barrier,
942 (atomic_read(&conf->nr_pending) &&
944 (!bio_list_empty(¤t->bio_list[0]) ||
945 !bio_list_empty(¤t->bio_list[1]))),
948 if (!conf->nr_waiting)
949 wake_up(&conf->wait_barrier);
951 atomic_inc(&conf->nr_pending);
952 spin_unlock_irq(&conf->resync_lock);
955 static void allow_barrier(struct r10conf *conf)
957 if ((atomic_dec_and_test(&conf->nr_pending)) ||
958 (conf->array_freeze_pending))
959 wake_up(&conf->wait_barrier);
962 static void freeze_array(struct r10conf *conf, int extra)
964 /* stop syncio and normal IO and wait for everything to
966 * We increment barrier and nr_waiting, and then
967 * wait until nr_pending match nr_queued+extra
968 * This is called in the context of one normal IO request
969 * that has failed. Thus any sync request that might be pending
970 * will be blocked by nr_pending, and we need to wait for
971 * pending IO requests to complete or be queued for re-try.
972 * Thus the number queued (nr_queued) plus this request (extra)
973 * must match the number of pending IOs (nr_pending) before
976 spin_lock_irq(&conf->resync_lock);
977 conf->array_freeze_pending++;
980 wait_event_lock_irq_cmd(conf->wait_barrier,
981 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
983 flush_pending_writes(conf));
985 conf->array_freeze_pending--;
986 spin_unlock_irq(&conf->resync_lock);
989 static void unfreeze_array(struct r10conf *conf)
991 /* reverse the effect of the freeze */
992 spin_lock_irq(&conf->resync_lock);
995 wake_up(&conf->wait_barrier);
996 spin_unlock_irq(&conf->resync_lock);
999 static sector_t choose_data_offset(struct r10bio *r10_bio,
1000 struct md_rdev *rdev)
1002 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1003 test_bit(R10BIO_Previous, &r10_bio->state))
1004 return rdev->data_offset;
1006 return rdev->new_data_offset;
1009 struct raid10_plug_cb {
1010 struct blk_plug_cb cb;
1011 struct bio_list pending;
1015 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1017 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1019 struct mddev *mddev = plug->cb.data;
1020 struct r10conf *conf = mddev->private;
1023 if (from_schedule || current->bio_list) {
1024 spin_lock_irq(&conf->device_lock);
1025 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1026 conf->pending_count += plug->pending_cnt;
1027 spin_unlock_irq(&conf->device_lock);
1028 wake_up(&conf->wait_barrier);
1029 md_wakeup_thread(mddev->thread);
1034 /* we aren't scheduling, so we can do the write-out directly. */
1035 bio = bio_list_get(&plug->pending);
1036 bitmap_unplug(mddev->bitmap);
1037 wake_up(&conf->wait_barrier);
1039 while (bio) { /* submit pending writes */
1040 struct bio *next = bio->bi_next;
1041 bio->bi_next = NULL;
1042 if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1043 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1044 /* Just ignore it */
1047 generic_make_request(bio);
1053 static void __make_request(struct mddev *mddev, struct bio *bio)
1055 struct r10conf *conf = mddev->private;
1056 struct r10bio *r10_bio;
1057 struct bio *read_bio;
1059 const int op = bio_op(bio);
1060 const int rw = bio_data_dir(bio);
1061 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1062 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1063 unsigned long flags;
1064 struct md_rdev *blocked_rdev;
1065 struct blk_plug_cb *cb;
1066 struct raid10_plug_cb *plug = NULL;
1067 int sectors_handled;
1071 md_write_start(mddev, bio);
1074 * Register the new request and wait if the reconstruction
1075 * thread has put up a bar for new requests.
1076 * Continue immediately if no resync is active currently.
1080 sectors = bio_sectors(bio);
1081 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1082 bio->bi_iter.bi_sector < conf->reshape_progress &&
1083 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1084 /* IO spans the reshape position. Need to wait for
1087 allow_barrier(conf);
1088 wait_event(conf->wait_barrier,
1089 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1090 conf->reshape_progress >= bio->bi_iter.bi_sector +
1094 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1095 bio_data_dir(bio) == WRITE &&
1096 (mddev->reshape_backwards
1097 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1098 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1099 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1100 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1101 /* Need to update reshape_position in metadata */
1102 mddev->reshape_position = conf->reshape_progress;
1103 set_mask_bits(&mddev->flags, 0,
1104 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1105 md_wakeup_thread(mddev->thread);
1106 wait_event(mddev->sb_wait,
1107 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1109 conf->reshape_safe = mddev->reshape_position;
1112 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1114 r10_bio->master_bio = bio;
1115 r10_bio->sectors = sectors;
1117 r10_bio->mddev = mddev;
1118 r10_bio->sector = bio->bi_iter.bi_sector;
1121 /* We might need to issue multiple reads to different
1122 * devices if there are bad blocks around, so we keep
1123 * track of the number of reads in bio->bi_phys_segments.
1124 * If this is 0, there is only one r10_bio and no locking
1125 * will be needed when the request completes. If it is
1126 * non-zero, then it is the number of not-completed requests.
1128 bio->bi_phys_segments = 0;
1129 bio_clear_flag(bio, BIO_SEG_VALID);
1133 * read balancing logic:
1135 struct md_rdev *rdev;
1139 rdev = read_balance(conf, r10_bio, &max_sectors);
1141 raid_end_bio_io(r10_bio);
1144 slot = r10_bio->read_slot;
1146 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1147 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1150 r10_bio->devs[slot].bio = read_bio;
1151 r10_bio->devs[slot].rdev = rdev;
1153 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1154 choose_data_offset(r10_bio, rdev);
1155 read_bio->bi_bdev = rdev->bdev;
1156 read_bio->bi_end_io = raid10_end_read_request;
1157 bio_set_op_attrs(read_bio, op, do_sync);
1158 read_bio->bi_private = r10_bio;
1160 if (max_sectors < r10_bio->sectors) {
1161 /* Could not read all from this device, so we will
1162 * need another r10_bio.
1164 sectors_handled = (r10_bio->sector + max_sectors
1165 - bio->bi_iter.bi_sector);
1166 r10_bio->sectors = max_sectors;
1167 spin_lock_irq(&conf->device_lock);
1168 if (bio->bi_phys_segments == 0)
1169 bio->bi_phys_segments = 2;
1171 bio->bi_phys_segments++;
1172 spin_unlock_irq(&conf->device_lock);
1173 /* Cannot call generic_make_request directly
1174 * as that will be queued in __generic_make_request
1175 * and subsequent mempool_alloc might block
1176 * waiting for it. so hand bio over to raid10d.
1178 reschedule_retry(r10_bio);
1180 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1182 r10_bio->master_bio = bio;
1183 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1185 r10_bio->mddev = mddev;
1186 r10_bio->sector = bio->bi_iter.bi_sector +
1190 generic_make_request(read_bio);
1197 if (conf->pending_count >= max_queued_requests) {
1198 md_wakeup_thread(mddev->thread);
1199 wait_event(conf->wait_barrier,
1200 conf->pending_count < max_queued_requests);
1202 /* first select target devices under rcu_lock and
1203 * inc refcount on their rdev. Record them by setting
1205 * If there are known/acknowledged bad blocks on any device
1206 * on which we have seen a write error, we want to avoid
1207 * writing to those blocks. This potentially requires several
1208 * writes to write around the bad blocks. Each set of writes
1209 * gets its own r10_bio with a set of bios attached. The number
1210 * of r10_bios is recored in bio->bi_phys_segments just as with
1214 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1215 raid10_find_phys(conf, r10_bio);
1217 blocked_rdev = NULL;
1219 max_sectors = r10_bio->sectors;
1221 for (i = 0; i < conf->copies; i++) {
1222 int d = r10_bio->devs[i].devnum;
1223 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1224 struct md_rdev *rrdev = rcu_dereference(
1225 conf->mirrors[d].replacement);
1228 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1229 atomic_inc(&rdev->nr_pending);
1230 blocked_rdev = rdev;
1233 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1234 atomic_inc(&rrdev->nr_pending);
1235 blocked_rdev = rrdev;
1238 if (rdev && (test_bit(Faulty, &rdev->flags)))
1240 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1243 r10_bio->devs[i].bio = NULL;
1244 r10_bio->devs[i].repl_bio = NULL;
1246 if (!rdev && !rrdev) {
1247 set_bit(R10BIO_Degraded, &r10_bio->state);
1250 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1252 sector_t dev_sector = r10_bio->devs[i].addr;
1256 is_bad = is_badblock(rdev, dev_sector,
1258 &first_bad, &bad_sectors);
1260 /* Mustn't write here until the bad block
1263 atomic_inc(&rdev->nr_pending);
1264 set_bit(BlockedBadBlocks, &rdev->flags);
1265 blocked_rdev = rdev;
1268 if (is_bad && first_bad <= dev_sector) {
1269 /* Cannot write here at all */
1270 bad_sectors -= (dev_sector - first_bad);
1271 if (bad_sectors < max_sectors)
1272 /* Mustn't write more than bad_sectors
1273 * to other devices yet
1275 max_sectors = bad_sectors;
1276 /* We don't set R10BIO_Degraded as that
1277 * only applies if the disk is missing,
1278 * so it might be re-added, and we want to
1279 * know to recover this chunk.
1280 * In this case the device is here, and the
1281 * fact that this chunk is not in-sync is
1282 * recorded in the bad block log.
1287 int good_sectors = first_bad - dev_sector;
1288 if (good_sectors < max_sectors)
1289 max_sectors = good_sectors;
1293 r10_bio->devs[i].bio = bio;
1294 atomic_inc(&rdev->nr_pending);
1297 r10_bio->devs[i].repl_bio = bio;
1298 atomic_inc(&rrdev->nr_pending);
1303 if (unlikely(blocked_rdev)) {
1304 /* Have to wait for this device to get unblocked, then retry */
1308 for (j = 0; j < i; j++) {
1309 if (r10_bio->devs[j].bio) {
1310 d = r10_bio->devs[j].devnum;
1311 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1313 if (r10_bio->devs[j].repl_bio) {
1314 struct md_rdev *rdev;
1315 d = r10_bio->devs[j].devnum;
1316 rdev = conf->mirrors[d].replacement;
1318 /* Race with remove_disk */
1320 rdev = conf->mirrors[d].rdev;
1322 rdev_dec_pending(rdev, mddev);
1325 allow_barrier(conf);
1326 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1331 if (max_sectors < r10_bio->sectors) {
1332 /* We are splitting this into multiple parts, so
1333 * we need to prepare for allocating another r10_bio.
1335 r10_bio->sectors = max_sectors;
1336 spin_lock_irq(&conf->device_lock);
1337 if (bio->bi_phys_segments == 0)
1338 bio->bi_phys_segments = 2;
1340 bio->bi_phys_segments++;
1341 spin_unlock_irq(&conf->device_lock);
1343 sectors_handled = r10_bio->sector + max_sectors -
1344 bio->bi_iter.bi_sector;
1346 atomic_set(&r10_bio->remaining, 1);
1347 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1349 for (i = 0; i < conf->copies; i++) {
1351 int d = r10_bio->devs[i].devnum;
1352 if (r10_bio->devs[i].bio) {
1353 struct md_rdev *rdev = conf->mirrors[d].rdev;
1354 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1355 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1357 r10_bio->devs[i].bio = mbio;
1359 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1360 choose_data_offset(r10_bio,
1362 mbio->bi_bdev = rdev->bdev;
1363 mbio->bi_end_io = raid10_end_write_request;
1364 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1365 mbio->bi_private = r10_bio;
1367 atomic_inc(&r10_bio->remaining);
1369 cb = blk_check_plugged(raid10_unplug, mddev,
1372 plug = container_of(cb, struct raid10_plug_cb,
1376 spin_lock_irqsave(&conf->device_lock, flags);
1378 bio_list_add(&plug->pending, mbio);
1379 plug->pending_cnt++;
1381 bio_list_add(&conf->pending_bio_list, mbio);
1382 conf->pending_count++;
1384 spin_unlock_irqrestore(&conf->device_lock, flags);
1386 md_wakeup_thread(mddev->thread);
1389 if (r10_bio->devs[i].repl_bio) {
1390 struct md_rdev *rdev = conf->mirrors[d].replacement;
1392 /* Replacement just got moved to main 'rdev' */
1394 rdev = conf->mirrors[d].rdev;
1396 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1397 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1399 r10_bio->devs[i].repl_bio = mbio;
1401 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1404 mbio->bi_bdev = rdev->bdev;
1405 mbio->bi_end_io = raid10_end_write_request;
1406 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1407 mbio->bi_private = r10_bio;
1409 atomic_inc(&r10_bio->remaining);
1411 cb = blk_check_plugged(raid10_unplug, mddev,
1414 plug = container_of(cb, struct raid10_plug_cb,
1418 spin_lock_irqsave(&conf->device_lock, flags);
1420 bio_list_add(&plug->pending, mbio);
1421 plug->pending_cnt++;
1423 bio_list_add(&conf->pending_bio_list, mbio);
1424 conf->pending_count++;
1426 spin_unlock_irqrestore(&conf->device_lock, flags);
1428 md_wakeup_thread(mddev->thread);
1432 /* Don't remove the bias on 'remaining' (one_write_done) until
1433 * after checking if we need to go around again.
1436 if (sectors_handled < bio_sectors(bio)) {
1437 one_write_done(r10_bio);
1438 /* We need another r10_bio. It has already been counted
1439 * in bio->bi_phys_segments.
1441 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1443 r10_bio->master_bio = bio;
1444 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1446 r10_bio->mddev = mddev;
1447 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1451 one_write_done(r10_bio);
1454 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1456 struct r10conf *conf = mddev->private;
1457 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1458 int chunk_sects = chunk_mask + 1;
1462 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1463 md_flush_request(mddev, bio);
1470 * If this request crosses a chunk boundary, we need to split
1473 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1474 bio_sectors(bio) > chunk_sects
1475 && (conf->geo.near_copies < conf->geo.raid_disks
1476 || conf->prev.near_copies <
1477 conf->prev.raid_disks))) {
1478 split = bio_split(bio, chunk_sects -
1479 (bio->bi_iter.bi_sector &
1481 GFP_NOIO, fs_bio_set);
1482 bio_chain(split, bio);
1488 * If a bio is splitted, the first part of bio will pass
1489 * barrier but the bio is queued in current->bio_list (see
1490 * generic_make_request). If there is a raise_barrier() called
1491 * here, the second part of bio can't pass barrier. But since
1492 * the first part bio isn't dispatched to underlaying disks
1493 * yet, the barrier is never released, hence raise_barrier will
1494 * alays wait. We have a deadlock.
1495 * Note, this only happens in read path. For write path, the
1496 * first part of bio is dispatched in a schedule() call
1497 * (because of blk plug) or offloaded to raid10d.
1498 * Quitting from the function immediately can change the bio
1499 * order queued in bio_list and avoid the deadlock.
1501 __make_request(mddev, split);
1502 if (split != bio && bio_data_dir(bio) == READ) {
1503 generic_make_request(bio);
1506 } while (split != bio);
1508 /* In case raid10d snuck in to freeze_array */
1509 wake_up(&conf->wait_barrier);
1512 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1514 struct r10conf *conf = mddev->private;
1517 if (conf->geo.near_copies < conf->geo.raid_disks)
1518 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1519 if (conf->geo.near_copies > 1)
1520 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1521 if (conf->geo.far_copies > 1) {
1522 if (conf->geo.far_offset)
1523 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1525 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1526 if (conf->geo.far_set_size != conf->geo.raid_disks)
1527 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1529 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1530 conf->geo.raid_disks - mddev->degraded);
1532 for (i = 0; i < conf->geo.raid_disks; i++) {
1533 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1534 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1537 seq_printf(seq, "]");
1540 /* check if there are enough drives for
1541 * every block to appear on atleast one.
1542 * Don't consider the device numbered 'ignore'
1543 * as we might be about to remove it.
1545 static int _enough(struct r10conf *conf, int previous, int ignore)
1551 disks = conf->prev.raid_disks;
1552 ncopies = conf->prev.near_copies;
1554 disks = conf->geo.raid_disks;
1555 ncopies = conf->geo.near_copies;
1560 int n = conf->copies;
1564 struct md_rdev *rdev;
1565 if (this != ignore &&
1566 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1567 test_bit(In_sync, &rdev->flags))
1569 this = (this+1) % disks;
1573 first = (first + ncopies) % disks;
1574 } while (first != 0);
1581 static int enough(struct r10conf *conf, int ignore)
1583 /* when calling 'enough', both 'prev' and 'geo' must
1585 * This is ensured if ->reconfig_mutex or ->device_lock
1588 return _enough(conf, 0, ignore) &&
1589 _enough(conf, 1, ignore);
1592 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1594 char b[BDEVNAME_SIZE];
1595 struct r10conf *conf = mddev->private;
1596 unsigned long flags;
1599 * If it is not operational, then we have already marked it as dead
1600 * else if it is the last working disks, ignore the error, let the
1601 * next level up know.
1602 * else mark the drive as failed
1604 spin_lock_irqsave(&conf->device_lock, flags);
1605 if (test_bit(In_sync, &rdev->flags)
1606 && !enough(conf, rdev->raid_disk)) {
1608 * Don't fail the drive, just return an IO error.
1610 spin_unlock_irqrestore(&conf->device_lock, flags);
1613 if (test_and_clear_bit(In_sync, &rdev->flags))
1616 * If recovery is running, make sure it aborts.
1618 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1619 set_bit(Blocked, &rdev->flags);
1620 set_bit(Faulty, &rdev->flags);
1621 set_mask_bits(&mddev->flags, 0,
1622 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1623 spin_unlock_irqrestore(&conf->device_lock, flags);
1625 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1626 "md/raid10:%s: Operation continuing on %d devices.\n",
1627 mdname(mddev), bdevname(rdev->bdev, b),
1628 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1631 static void print_conf(struct r10conf *conf)
1634 struct md_rdev *rdev;
1636 printk(KERN_DEBUG "RAID10 conf printout:\n");
1638 printk(KERN_DEBUG "(!conf)\n");
1641 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1642 conf->geo.raid_disks);
1644 /* This is only called with ->reconfix_mutex held, so
1645 * rcu protection of rdev is not needed */
1646 for (i = 0; i < conf->geo.raid_disks; i++) {
1647 char b[BDEVNAME_SIZE];
1648 rdev = conf->mirrors[i].rdev;
1650 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1651 i, !test_bit(In_sync, &rdev->flags),
1652 !test_bit(Faulty, &rdev->flags),
1653 bdevname(rdev->bdev,b));
1657 static void close_sync(struct r10conf *conf)
1660 allow_barrier(conf);
1662 mempool_destroy(conf->r10buf_pool);
1663 conf->r10buf_pool = NULL;
1666 static int raid10_spare_active(struct mddev *mddev)
1669 struct r10conf *conf = mddev->private;
1670 struct raid10_info *tmp;
1672 unsigned long flags;
1675 * Find all non-in_sync disks within the RAID10 configuration
1676 * and mark them in_sync
1678 for (i = 0; i < conf->geo.raid_disks; i++) {
1679 tmp = conf->mirrors + i;
1680 if (tmp->replacement
1681 && tmp->replacement->recovery_offset == MaxSector
1682 && !test_bit(Faulty, &tmp->replacement->flags)
1683 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1684 /* Replacement has just become active */
1686 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1689 /* Replaced device not technically faulty,
1690 * but we need to be sure it gets removed
1691 * and never re-added.
1693 set_bit(Faulty, &tmp->rdev->flags);
1694 sysfs_notify_dirent_safe(
1695 tmp->rdev->sysfs_state);
1697 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1698 } else if (tmp->rdev
1699 && tmp->rdev->recovery_offset == MaxSector
1700 && !test_bit(Faulty, &tmp->rdev->flags)
1701 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1703 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1706 spin_lock_irqsave(&conf->device_lock, flags);
1707 mddev->degraded -= count;
1708 spin_unlock_irqrestore(&conf->device_lock, flags);
1714 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1716 struct r10conf *conf = mddev->private;
1720 int last = conf->geo.raid_disks - 1;
1722 if (mddev->recovery_cp < MaxSector)
1723 /* only hot-add to in-sync arrays, as recovery is
1724 * very different from resync
1727 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1730 if (md_integrity_add_rdev(rdev, mddev))
1733 if (rdev->raid_disk >= 0)
1734 first = last = rdev->raid_disk;
1736 if (rdev->saved_raid_disk >= first &&
1737 rdev->saved_raid_disk < conf->geo.raid_disks &&
1738 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1739 mirror = rdev->saved_raid_disk;
1742 for ( ; mirror <= last ; mirror++) {
1743 struct raid10_info *p = &conf->mirrors[mirror];
1744 if (p->recovery_disabled == mddev->recovery_disabled)
1747 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1748 p->replacement != NULL)
1750 clear_bit(In_sync, &rdev->flags);
1751 set_bit(Replacement, &rdev->flags);
1752 rdev->raid_disk = mirror;
1755 disk_stack_limits(mddev->gendisk, rdev->bdev,
1756 rdev->data_offset << 9);
1758 rcu_assign_pointer(p->replacement, rdev);
1763 disk_stack_limits(mddev->gendisk, rdev->bdev,
1764 rdev->data_offset << 9);
1766 p->head_position = 0;
1767 p->recovery_disabled = mddev->recovery_disabled - 1;
1768 rdev->raid_disk = mirror;
1770 if (rdev->saved_raid_disk != mirror)
1772 rcu_assign_pointer(p->rdev, rdev);
1775 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1776 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1782 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1784 struct r10conf *conf = mddev->private;
1786 int number = rdev->raid_disk;
1787 struct md_rdev **rdevp;
1788 struct raid10_info *p;
1791 if (unlikely(number >= mddev->raid_disks))
1793 p = conf->mirrors + number;
1794 if (rdev == p->rdev)
1796 else if (rdev == p->replacement)
1797 rdevp = &p->replacement;
1801 if (test_bit(In_sync, &rdev->flags) ||
1802 atomic_read(&rdev->nr_pending)) {
1806 /* Only remove non-faulty devices if recovery
1809 if (!test_bit(Faulty, &rdev->flags) &&
1810 mddev->recovery_disabled != p->recovery_disabled &&
1811 (!p->replacement || p->replacement == rdev) &&
1812 number < conf->geo.raid_disks &&
1818 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1820 if (atomic_read(&rdev->nr_pending)) {
1821 /* lost the race, try later */
1827 if (p->replacement) {
1828 /* We must have just cleared 'rdev' */
1829 p->rdev = p->replacement;
1830 clear_bit(Replacement, &p->replacement->flags);
1831 smp_mb(); /* Make sure other CPUs may see both as identical
1832 * but will never see neither -- if they are careful.
1834 p->replacement = NULL;
1835 clear_bit(WantReplacement, &rdev->flags);
1837 /* We might have just remove the Replacement as faulty
1838 * Clear the flag just in case
1840 clear_bit(WantReplacement, &rdev->flags);
1842 err = md_integrity_register(mddev);
1850 static void end_sync_read(struct bio *bio)
1852 struct r10bio *r10_bio = bio->bi_private;
1853 struct r10conf *conf = r10_bio->mddev->private;
1856 if (bio == r10_bio->master_bio) {
1857 /* this is a reshape read */
1858 d = r10_bio->read_slot; /* really the read dev */
1860 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1863 set_bit(R10BIO_Uptodate, &r10_bio->state);
1865 /* The write handler will notice the lack of
1866 * R10BIO_Uptodate and record any errors etc
1868 atomic_add(r10_bio->sectors,
1869 &conf->mirrors[d].rdev->corrected_errors);
1871 /* for reconstruct, we always reschedule after a read.
1872 * for resync, only after all reads
1874 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1875 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1876 atomic_dec_and_test(&r10_bio->remaining)) {
1877 /* we have read all the blocks,
1878 * do the comparison in process context in raid10d
1880 reschedule_retry(r10_bio);
1884 static void end_sync_request(struct r10bio *r10_bio)
1886 struct mddev *mddev = r10_bio->mddev;
1888 while (atomic_dec_and_test(&r10_bio->remaining)) {
1889 if (r10_bio->master_bio == NULL) {
1890 /* the primary of several recovery bios */
1891 sector_t s = r10_bio->sectors;
1892 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1893 test_bit(R10BIO_WriteError, &r10_bio->state))
1894 reschedule_retry(r10_bio);
1897 md_done_sync(mddev, s, 1);
1900 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1901 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1902 test_bit(R10BIO_WriteError, &r10_bio->state))
1903 reschedule_retry(r10_bio);
1911 static void end_sync_write(struct bio *bio)
1913 struct r10bio *r10_bio = bio->bi_private;
1914 struct mddev *mddev = r10_bio->mddev;
1915 struct r10conf *conf = mddev->private;
1921 struct md_rdev *rdev = NULL;
1923 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1925 rdev = conf->mirrors[d].replacement;
1927 rdev = conf->mirrors[d].rdev;
1929 if (bio->bi_error) {
1931 md_error(mddev, rdev);
1933 set_bit(WriteErrorSeen, &rdev->flags);
1934 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1935 set_bit(MD_RECOVERY_NEEDED,
1936 &rdev->mddev->recovery);
1937 set_bit(R10BIO_WriteError, &r10_bio->state);
1939 } else if (is_badblock(rdev,
1940 r10_bio->devs[slot].addr,
1942 &first_bad, &bad_sectors))
1943 set_bit(R10BIO_MadeGood, &r10_bio->state);
1945 rdev_dec_pending(rdev, mddev);
1947 end_sync_request(r10_bio);
1951 * Note: sync and recover and handled very differently for raid10
1952 * This code is for resync.
1953 * For resync, we read through virtual addresses and read all blocks.
1954 * If there is any error, we schedule a write. The lowest numbered
1955 * drive is authoritative.
1956 * However requests come for physical address, so we need to map.
1957 * For every physical address there are raid_disks/copies virtual addresses,
1958 * which is always are least one, but is not necessarly an integer.
1959 * This means that a physical address can span multiple chunks, so we may
1960 * have to submit multiple io requests for a single sync request.
1963 * We check if all blocks are in-sync and only write to blocks that
1966 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1968 struct r10conf *conf = mddev->private;
1970 struct bio *tbio, *fbio;
1973 atomic_set(&r10_bio->remaining, 1);
1975 /* find the first device with a block */
1976 for (i=0; i<conf->copies; i++)
1977 if (!r10_bio->devs[i].bio->bi_error)
1980 if (i == conf->copies)
1984 fbio = r10_bio->devs[i].bio;
1985 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1986 fbio->bi_iter.bi_idx = 0;
1988 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1989 /* now find blocks with errors */
1990 for (i=0 ; i < conf->copies ; i++) {
1993 tbio = r10_bio->devs[i].bio;
1995 if (tbio->bi_end_io != end_sync_read)
1999 if (!r10_bio->devs[i].bio->bi_error) {
2000 /* We know that the bi_io_vec layout is the same for
2001 * both 'first' and 'i', so we just compare them.
2002 * All vec entries are PAGE_SIZE;
2004 int sectors = r10_bio->sectors;
2005 for (j = 0; j < vcnt; j++) {
2006 int len = PAGE_SIZE;
2007 if (sectors < (len / 512))
2008 len = sectors * 512;
2009 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2010 page_address(tbio->bi_io_vec[j].bv_page),
2017 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2018 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2019 /* Don't fix anything. */
2022 /* Ok, we need to write this bio, either to correct an
2023 * inconsistency or to correct an unreadable block.
2024 * First we need to fixup bv_offset, bv_len and
2025 * bi_vecs, as the read request might have corrupted these
2029 tbio->bi_vcnt = vcnt;
2030 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
2031 tbio->bi_private = r10_bio;
2032 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2033 tbio->bi_end_io = end_sync_write;
2034 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2036 bio_copy_data(tbio, fbio);
2038 d = r10_bio->devs[i].devnum;
2039 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2040 atomic_inc(&r10_bio->remaining);
2041 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2043 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2044 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2045 generic_make_request(tbio);
2048 /* Now write out to any replacement devices
2051 for (i = 0; i < conf->copies; i++) {
2054 tbio = r10_bio->devs[i].repl_bio;
2055 if (!tbio || !tbio->bi_end_io)
2057 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2058 && r10_bio->devs[i].bio != fbio)
2059 bio_copy_data(tbio, fbio);
2060 d = r10_bio->devs[i].devnum;
2061 atomic_inc(&r10_bio->remaining);
2062 md_sync_acct(conf->mirrors[d].replacement->bdev,
2064 generic_make_request(tbio);
2068 if (atomic_dec_and_test(&r10_bio->remaining)) {
2069 md_done_sync(mddev, r10_bio->sectors, 1);
2075 * Now for the recovery code.
2076 * Recovery happens across physical sectors.
2077 * We recover all non-is_sync drives by finding the virtual address of
2078 * each, and then choose a working drive that also has that virt address.
2079 * There is a separate r10_bio for each non-in_sync drive.
2080 * Only the first two slots are in use. The first for reading,
2081 * The second for writing.
2084 static void fix_recovery_read_error(struct r10bio *r10_bio)
2086 /* We got a read error during recovery.
2087 * We repeat the read in smaller page-sized sections.
2088 * If a read succeeds, write it to the new device or record
2089 * a bad block if we cannot.
2090 * If a read fails, record a bad block on both old and
2093 struct mddev *mddev = r10_bio->mddev;
2094 struct r10conf *conf = mddev->private;
2095 struct bio *bio = r10_bio->devs[0].bio;
2097 int sectors = r10_bio->sectors;
2099 int dr = r10_bio->devs[0].devnum;
2100 int dw = r10_bio->devs[1].devnum;
2104 struct md_rdev *rdev;
2108 if (s > (PAGE_SIZE>>9))
2111 rdev = conf->mirrors[dr].rdev;
2112 addr = r10_bio->devs[0].addr + sect,
2113 ok = sync_page_io(rdev,
2116 bio->bi_io_vec[idx].bv_page,
2117 REQ_OP_READ, 0, false);
2119 rdev = conf->mirrors[dw].rdev;
2120 addr = r10_bio->devs[1].addr + sect;
2121 ok = sync_page_io(rdev,
2124 bio->bi_io_vec[idx].bv_page,
2125 REQ_OP_WRITE, 0, false);
2127 set_bit(WriteErrorSeen, &rdev->flags);
2128 if (!test_and_set_bit(WantReplacement,
2130 set_bit(MD_RECOVERY_NEEDED,
2131 &rdev->mddev->recovery);
2135 /* We don't worry if we cannot set a bad block -
2136 * it really is bad so there is no loss in not
2139 rdev_set_badblocks(rdev, addr, s, 0);
2141 if (rdev != conf->mirrors[dw].rdev) {
2142 /* need bad block on destination too */
2143 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2144 addr = r10_bio->devs[1].addr + sect;
2145 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2147 /* just abort the recovery */
2149 "md/raid10:%s: recovery aborted"
2150 " due to read error\n",
2153 conf->mirrors[dw].recovery_disabled
2154 = mddev->recovery_disabled;
2155 set_bit(MD_RECOVERY_INTR,
2168 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2170 struct r10conf *conf = mddev->private;
2172 struct bio *wbio, *wbio2;
2174 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2175 fix_recovery_read_error(r10_bio);
2176 end_sync_request(r10_bio);
2181 * share the pages with the first bio
2182 * and submit the write request
2184 d = r10_bio->devs[1].devnum;
2185 wbio = r10_bio->devs[1].bio;
2186 wbio2 = r10_bio->devs[1].repl_bio;
2187 /* Need to test wbio2->bi_end_io before we call
2188 * generic_make_request as if the former is NULL,
2189 * the latter is free to free wbio2.
2191 if (wbio2 && !wbio2->bi_end_io)
2193 if (wbio->bi_end_io) {
2194 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2195 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2196 generic_make_request(wbio);
2199 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2200 md_sync_acct(conf->mirrors[d].replacement->bdev,
2201 bio_sectors(wbio2));
2202 generic_make_request(wbio2);
2207 * Used by fix_read_error() to decay the per rdev read_errors.
2208 * We halve the read error count for every hour that has elapsed
2209 * since the last recorded read error.
2212 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2215 unsigned long hours_since_last;
2216 unsigned int read_errors = atomic_read(&rdev->read_errors);
2218 cur_time_mon = ktime_get_seconds();
2220 if (rdev->last_read_error == 0) {
2221 /* first time we've seen a read error */
2222 rdev->last_read_error = cur_time_mon;
2226 hours_since_last = (long)(cur_time_mon -
2227 rdev->last_read_error) / 3600;
2229 rdev->last_read_error = cur_time_mon;
2232 * if hours_since_last is > the number of bits in read_errors
2233 * just set read errors to 0. We do this to avoid
2234 * overflowing the shift of read_errors by hours_since_last.
2236 if (hours_since_last >= 8 * sizeof(read_errors))
2237 atomic_set(&rdev->read_errors, 0);
2239 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2242 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2243 int sectors, struct page *page, int rw)
2248 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2249 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2251 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2255 set_bit(WriteErrorSeen, &rdev->flags);
2256 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2257 set_bit(MD_RECOVERY_NEEDED,
2258 &rdev->mddev->recovery);
2260 /* need to record an error - either for the block or the device */
2261 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2262 md_error(rdev->mddev, rdev);
2267 * This is a kernel thread which:
2269 * 1. Retries failed read operations on working mirrors.
2270 * 2. Updates the raid superblock when problems encounter.
2271 * 3. Performs writes following reads for array synchronising.
2274 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2276 int sect = 0; /* Offset from r10_bio->sector */
2277 int sectors = r10_bio->sectors;
2278 struct md_rdev*rdev;
2279 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2280 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2282 /* still own a reference to this rdev, so it cannot
2283 * have been cleared recently.
2285 rdev = conf->mirrors[d].rdev;
2287 if (test_bit(Faulty, &rdev->flags))
2288 /* drive has already been failed, just ignore any
2289 more fix_read_error() attempts */
2292 check_decay_read_errors(mddev, rdev);
2293 atomic_inc(&rdev->read_errors);
2294 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2295 char b[BDEVNAME_SIZE];
2296 bdevname(rdev->bdev, b);
2299 "md/raid10:%s: %s: Raid device exceeded "
2300 "read_error threshold [cur %d:max %d]\n",
2302 atomic_read(&rdev->read_errors), max_read_errors);
2304 "md/raid10:%s: %s: Failing raid device\n",
2306 md_error(mddev, rdev);
2307 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2313 int sl = r10_bio->read_slot;
2317 if (s > (PAGE_SIZE>>9))
2325 d = r10_bio->devs[sl].devnum;
2326 rdev = rcu_dereference(conf->mirrors[d].rdev);
2328 test_bit(In_sync, &rdev->flags) &&
2329 !test_bit(Faulty, &rdev->flags) &&
2330 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2331 &first_bad, &bad_sectors) == 0) {
2332 atomic_inc(&rdev->nr_pending);
2334 success = sync_page_io(rdev,
2335 r10_bio->devs[sl].addr +
2339 REQ_OP_READ, 0, false);
2340 rdev_dec_pending(rdev, mddev);
2346 if (sl == conf->copies)
2348 } while (!success && sl != r10_bio->read_slot);
2352 /* Cannot read from anywhere, just mark the block
2353 * as bad on the first device to discourage future
2356 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2357 rdev = conf->mirrors[dn].rdev;
2359 if (!rdev_set_badblocks(
2361 r10_bio->devs[r10_bio->read_slot].addr
2364 md_error(mddev, rdev);
2365 r10_bio->devs[r10_bio->read_slot].bio
2372 /* write it back and re-read */
2374 while (sl != r10_bio->read_slot) {
2375 char b[BDEVNAME_SIZE];
2380 d = r10_bio->devs[sl].devnum;
2381 rdev = rcu_dereference(conf->mirrors[d].rdev);
2383 test_bit(Faulty, &rdev->flags) ||
2384 !test_bit(In_sync, &rdev->flags))
2387 atomic_inc(&rdev->nr_pending);
2389 if (r10_sync_page_io(rdev,
2390 r10_bio->devs[sl].addr +
2392 s, conf->tmppage, WRITE)
2394 /* Well, this device is dead */
2396 "md/raid10:%s: read correction "
2398 " (%d sectors at %llu on %s)\n",
2400 (unsigned long long)(
2402 choose_data_offset(r10_bio,
2404 bdevname(rdev->bdev, b));
2405 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2408 bdevname(rdev->bdev, b));
2410 rdev_dec_pending(rdev, mddev);
2414 while (sl != r10_bio->read_slot) {
2415 char b[BDEVNAME_SIZE];
2420 d = r10_bio->devs[sl].devnum;
2421 rdev = rcu_dereference(conf->mirrors[d].rdev);
2423 test_bit(Faulty, &rdev->flags) ||
2424 !test_bit(In_sync, &rdev->flags))
2427 atomic_inc(&rdev->nr_pending);
2429 switch (r10_sync_page_io(rdev,
2430 r10_bio->devs[sl].addr +
2435 /* Well, this device is dead */
2437 "md/raid10:%s: unable to read back "
2439 " (%d sectors at %llu on %s)\n",
2441 (unsigned long long)(
2443 choose_data_offset(r10_bio, rdev)),
2444 bdevname(rdev->bdev, b));
2445 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2448 bdevname(rdev->bdev, b));
2452 "md/raid10:%s: read error corrected"
2453 " (%d sectors at %llu on %s)\n",
2455 (unsigned long long)(
2457 choose_data_offset(r10_bio, rdev)),
2458 bdevname(rdev->bdev, b));
2459 atomic_add(s, &rdev->corrected_errors);
2462 rdev_dec_pending(rdev, mddev);
2472 static int narrow_write_error(struct r10bio *r10_bio, int i)
2474 struct bio *bio = r10_bio->master_bio;
2475 struct mddev *mddev = r10_bio->mddev;
2476 struct r10conf *conf = mddev->private;
2477 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2478 /* bio has the data to be written to slot 'i' where
2479 * we just recently had a write error.
2480 * We repeatedly clone the bio and trim down to one block,
2481 * then try the write. Where the write fails we record
2483 * It is conceivable that the bio doesn't exactly align with
2484 * blocks. We must handle this.
2486 * We currently own a reference to the rdev.
2492 int sect_to_write = r10_bio->sectors;
2495 if (rdev->badblocks.shift < 0)
2498 block_sectors = roundup(1 << rdev->badblocks.shift,
2499 bdev_logical_block_size(rdev->bdev) >> 9);
2500 sector = r10_bio->sector;
2501 sectors = ((r10_bio->sector + block_sectors)
2502 & ~(sector_t)(block_sectors - 1))
2505 while (sect_to_write) {
2508 if (sectors > sect_to_write)
2509 sectors = sect_to_write;
2510 /* Write at 'sector' for 'sectors' */
2511 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2512 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2513 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2514 wbio->bi_iter.bi_sector = wsector +
2515 choose_data_offset(r10_bio, rdev);
2516 wbio->bi_bdev = rdev->bdev;
2517 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2519 if (submit_bio_wait(wbio) < 0)
2521 ok = rdev_set_badblocks(rdev, wsector,
2526 sect_to_write -= sectors;
2528 sectors = block_sectors;
2533 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2535 int slot = r10_bio->read_slot;
2537 struct r10conf *conf = mddev->private;
2538 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2539 char b[BDEVNAME_SIZE];
2540 unsigned long do_sync;
2543 /* we got a read error. Maybe the drive is bad. Maybe just
2544 * the block and we can fix it.
2545 * We freeze all other IO, and try reading the block from
2546 * other devices. When we find one, we re-write
2547 * and check it that fixes the read error.
2548 * This is all done synchronously while the array is
2551 bio = r10_bio->devs[slot].bio;
2552 bdevname(bio->bi_bdev, b);
2554 r10_bio->devs[slot].bio = NULL;
2556 if (mddev->ro == 0) {
2557 freeze_array(conf, 1);
2558 fix_read_error(conf, mddev, r10_bio);
2559 unfreeze_array(conf);
2561 r10_bio->devs[slot].bio = IO_BLOCKED;
2563 rdev_dec_pending(rdev, mddev);
2566 rdev = read_balance(conf, r10_bio, &max_sectors);
2568 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2569 " read error for block %llu\n",
2571 (unsigned long long)r10_bio->sector);
2572 raid_end_bio_io(r10_bio);
2576 do_sync = (r10_bio->master_bio->bi_opf & REQ_SYNC);
2577 slot = r10_bio->read_slot;
2580 "md/raid10:%s: %s: redirecting "
2581 "sector %llu to another mirror\n",
2583 bdevname(rdev->bdev, b),
2584 (unsigned long long)r10_bio->sector);
2585 bio = bio_clone_mddev(r10_bio->master_bio,
2587 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2588 r10_bio->devs[slot].bio = bio;
2589 r10_bio->devs[slot].rdev = rdev;
2590 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2591 + choose_data_offset(r10_bio, rdev);
2592 bio->bi_bdev = rdev->bdev;
2593 bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2594 bio->bi_private = r10_bio;
2595 bio->bi_end_io = raid10_end_read_request;
2596 if (max_sectors < r10_bio->sectors) {
2597 /* Drat - have to split this up more */
2598 struct bio *mbio = r10_bio->master_bio;
2599 int sectors_handled =
2600 r10_bio->sector + max_sectors
2601 - mbio->bi_iter.bi_sector;
2602 r10_bio->sectors = max_sectors;
2603 spin_lock_irq(&conf->device_lock);
2604 if (mbio->bi_phys_segments == 0)
2605 mbio->bi_phys_segments = 2;
2607 mbio->bi_phys_segments++;
2608 spin_unlock_irq(&conf->device_lock);
2609 generic_make_request(bio);
2611 r10_bio = mempool_alloc(conf->r10bio_pool,
2613 r10_bio->master_bio = mbio;
2614 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2616 set_bit(R10BIO_ReadError,
2618 r10_bio->mddev = mddev;
2619 r10_bio->sector = mbio->bi_iter.bi_sector
2624 generic_make_request(bio);
2627 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2629 /* Some sort of write request has finished and it
2630 * succeeded in writing where we thought there was a
2631 * bad block. So forget the bad block.
2632 * Or possibly if failed and we need to record
2636 struct md_rdev *rdev;
2638 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2639 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2640 for (m = 0; m < conf->copies; m++) {
2641 int dev = r10_bio->devs[m].devnum;
2642 rdev = conf->mirrors[dev].rdev;
2643 if (r10_bio->devs[m].bio == NULL ||
2644 r10_bio->devs[m].bio->bi_end_io == NULL)
2646 if (!r10_bio->devs[m].bio->bi_error) {
2647 rdev_clear_badblocks(
2649 r10_bio->devs[m].addr,
2650 r10_bio->sectors, 0);
2652 if (!rdev_set_badblocks(
2654 r10_bio->devs[m].addr,
2655 r10_bio->sectors, 0))
2656 md_error(conf->mddev, rdev);
2658 rdev = conf->mirrors[dev].replacement;
2659 if (r10_bio->devs[m].repl_bio == NULL ||
2660 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2663 if (!r10_bio->devs[m].repl_bio->bi_error) {
2664 rdev_clear_badblocks(
2666 r10_bio->devs[m].addr,
2667 r10_bio->sectors, 0);
2669 if (!rdev_set_badblocks(
2671 r10_bio->devs[m].addr,
2672 r10_bio->sectors, 0))
2673 md_error(conf->mddev, rdev);
2679 for (m = 0; m < conf->copies; m++) {
2680 int dev = r10_bio->devs[m].devnum;
2681 struct bio *bio = r10_bio->devs[m].bio;
2682 rdev = conf->mirrors[dev].rdev;
2683 if (bio == IO_MADE_GOOD) {
2684 rdev_clear_badblocks(
2686 r10_bio->devs[m].addr,
2687 r10_bio->sectors, 0);
2688 rdev_dec_pending(rdev, conf->mddev);
2689 } else if (bio != NULL && bio->bi_error) {
2691 if (!narrow_write_error(r10_bio, m)) {
2692 md_error(conf->mddev, rdev);
2693 set_bit(R10BIO_Degraded,
2696 rdev_dec_pending(rdev, conf->mddev);
2698 bio = r10_bio->devs[m].repl_bio;
2699 rdev = conf->mirrors[dev].replacement;
2700 if (rdev && bio == IO_MADE_GOOD) {
2701 rdev_clear_badblocks(
2703 r10_bio->devs[m].addr,
2704 r10_bio->sectors, 0);
2705 rdev_dec_pending(rdev, conf->mddev);
2709 spin_lock_irq(&conf->device_lock);
2710 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2712 spin_unlock_irq(&conf->device_lock);
2714 * In case freeze_array() is waiting for condition
2715 * nr_pending == nr_queued + extra to be true.
2717 wake_up(&conf->wait_barrier);
2718 md_wakeup_thread(conf->mddev->thread);
2720 if (test_bit(R10BIO_WriteError,
2722 close_write(r10_bio);
2723 raid_end_bio_io(r10_bio);
2728 static void raid10d(struct md_thread *thread)
2730 struct mddev *mddev = thread->mddev;
2731 struct r10bio *r10_bio;
2732 unsigned long flags;
2733 struct r10conf *conf = mddev->private;
2734 struct list_head *head = &conf->retry_list;
2735 struct blk_plug plug;
2737 md_check_recovery(mddev);
2739 if (!list_empty_careful(&conf->bio_end_io_list) &&
2740 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2742 spin_lock_irqsave(&conf->device_lock, flags);
2743 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2744 while (!list_empty(&conf->bio_end_io_list)) {
2745 list_move(conf->bio_end_io_list.prev, &tmp);
2749 spin_unlock_irqrestore(&conf->device_lock, flags);
2750 while (!list_empty(&tmp)) {
2751 r10_bio = list_first_entry(&tmp, struct r10bio,
2753 list_del(&r10_bio->retry_list);
2754 if (mddev->degraded)
2755 set_bit(R10BIO_Degraded, &r10_bio->state);
2757 if (test_bit(R10BIO_WriteError,
2759 close_write(r10_bio);
2760 raid_end_bio_io(r10_bio);
2764 blk_start_plug(&plug);
2767 flush_pending_writes(conf);
2769 spin_lock_irqsave(&conf->device_lock, flags);
2770 if (list_empty(head)) {
2771 spin_unlock_irqrestore(&conf->device_lock, flags);
2774 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2775 list_del(head->prev);
2777 spin_unlock_irqrestore(&conf->device_lock, flags);
2779 mddev = r10_bio->mddev;
2780 conf = mddev->private;
2781 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2782 test_bit(R10BIO_WriteError, &r10_bio->state))
2783 handle_write_completed(conf, r10_bio);
2784 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2785 reshape_request_write(mddev, r10_bio);
2786 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2787 sync_request_write(mddev, r10_bio);
2788 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2789 recovery_request_write(mddev, r10_bio);
2790 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2791 handle_read_error(mddev, r10_bio);
2793 /* just a partial read to be scheduled from a
2796 int slot = r10_bio->read_slot;
2797 generic_make_request(r10_bio->devs[slot].bio);
2801 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2802 md_check_recovery(mddev);
2804 blk_finish_plug(&plug);
2807 static int init_resync(struct r10conf *conf)
2812 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2813 BUG_ON(conf->r10buf_pool);
2814 conf->have_replacement = 0;
2815 for (i = 0; i < conf->geo.raid_disks; i++)
2816 if (conf->mirrors[i].replacement)
2817 conf->have_replacement = 1;
2818 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2819 if (!conf->r10buf_pool)
2821 conf->next_resync = 0;
2826 * perform a "sync" on one "block"
2828 * We need to make sure that no normal I/O request - particularly write
2829 * requests - conflict with active sync requests.
2831 * This is achieved by tracking pending requests and a 'barrier' concept
2832 * that can be installed to exclude normal IO requests.
2834 * Resync and recovery are handled very differently.
2835 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2837 * For resync, we iterate over virtual addresses, read all copies,
2838 * and update if there are differences. If only one copy is live,
2840 * For recovery, we iterate over physical addresses, read a good
2841 * value for each non-in_sync drive, and over-write.
2843 * So, for recovery we may have several outstanding complex requests for a
2844 * given address, one for each out-of-sync device. We model this by allocating
2845 * a number of r10_bio structures, one for each out-of-sync device.
2846 * As we setup these structures, we collect all bio's together into a list
2847 * which we then process collectively to add pages, and then process again
2848 * to pass to generic_make_request.
2850 * The r10_bio structures are linked using a borrowed master_bio pointer.
2851 * This link is counted in ->remaining. When the r10_bio that points to NULL
2852 * has its remaining count decremented to 0, the whole complex operation
2857 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2860 struct r10conf *conf = mddev->private;
2861 struct r10bio *r10_bio;
2862 struct bio *biolist = NULL, *bio;
2863 sector_t max_sector, nr_sectors;
2866 sector_t sync_blocks;
2867 sector_t sectors_skipped = 0;
2868 int chunks_skipped = 0;
2869 sector_t chunk_mask = conf->geo.chunk_mask;
2871 if (!conf->r10buf_pool)
2872 if (init_resync(conf))
2876 * Allow skipping a full rebuild for incremental assembly
2877 * of a clean array, like RAID1 does.
2879 if (mddev->bitmap == NULL &&
2880 mddev->recovery_cp == MaxSector &&
2881 mddev->reshape_position == MaxSector &&
2882 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2883 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2884 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2885 conf->fullsync == 0) {
2887 return mddev->dev_sectors - sector_nr;
2891 max_sector = mddev->dev_sectors;
2892 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2893 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2894 max_sector = mddev->resync_max_sectors;
2895 if (sector_nr >= max_sector) {
2896 /* If we aborted, we need to abort the
2897 * sync on the 'current' bitmap chucks (there can
2898 * be several when recovering multiple devices).
2899 * as we may have started syncing it but not finished.
2900 * We can find the current address in
2901 * mddev->curr_resync, but for recovery,
2902 * we need to convert that to several
2903 * virtual addresses.
2905 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2911 if (mddev->curr_resync < max_sector) { /* aborted */
2912 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2913 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2915 else for (i = 0; i < conf->geo.raid_disks; i++) {
2917 raid10_find_virt(conf, mddev->curr_resync, i);
2918 bitmap_end_sync(mddev->bitmap, sect,
2922 /* completed sync */
2923 if ((!mddev->bitmap || conf->fullsync)
2924 && conf->have_replacement
2925 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2926 /* Completed a full sync so the replacements
2927 * are now fully recovered.
2930 for (i = 0; i < conf->geo.raid_disks; i++) {
2931 struct md_rdev *rdev =
2932 rcu_dereference(conf->mirrors[i].replacement);
2934 rdev->recovery_offset = MaxSector;
2940 bitmap_close_sync(mddev->bitmap);
2943 return sectors_skipped;
2946 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2947 return reshape_request(mddev, sector_nr, skipped);
2949 if (chunks_skipped >= conf->geo.raid_disks) {
2950 /* if there has been nothing to do on any drive,
2951 * then there is nothing to do at all..
2954 return (max_sector - sector_nr) + sectors_skipped;
2957 if (max_sector > mddev->resync_max)
2958 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2960 /* make sure whole request will fit in a chunk - if chunks
2963 if (conf->geo.near_copies < conf->geo.raid_disks &&
2964 max_sector > (sector_nr | chunk_mask))
2965 max_sector = (sector_nr | chunk_mask) + 1;
2968 * If there is non-resync activity waiting for a turn, then let it
2969 * though before starting on this new sync request.
2971 if (conf->nr_waiting)
2972 schedule_timeout_uninterruptible(1);
2974 /* Again, very different code for resync and recovery.
2975 * Both must result in an r10bio with a list of bios that
2976 * have bi_end_io, bi_sector, bi_bdev set,
2977 * and bi_private set to the r10bio.
2978 * For recovery, we may actually create several r10bios
2979 * with 2 bios in each, that correspond to the bios in the main one.
2980 * In this case, the subordinate r10bios link back through a
2981 * borrowed master_bio pointer, and the counter in the master
2982 * includes a ref from each subordinate.
2984 /* First, we decide what to do and set ->bi_end_io
2985 * To end_sync_read if we want to read, and
2986 * end_sync_write if we will want to write.
2989 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2990 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2991 /* recovery... the complicated one */
2995 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3001 struct raid10_info *mirror = &conf->mirrors[i];
3002 struct md_rdev *mrdev, *mreplace;
3005 mrdev = rcu_dereference(mirror->rdev);
3006 mreplace = rcu_dereference(mirror->replacement);
3008 if ((mrdev == NULL ||
3009 test_bit(Faulty, &mrdev->flags) ||
3010 test_bit(In_sync, &mrdev->flags)) &&
3011 (mreplace == NULL ||
3012 test_bit(Faulty, &mreplace->flags))) {
3018 /* want to reconstruct this device */
3020 sect = raid10_find_virt(conf, sector_nr, i);
3021 if (sect >= mddev->resync_max_sectors) {
3022 /* last stripe is not complete - don't
3023 * try to recover this sector.
3028 if (mreplace && test_bit(Faulty, &mreplace->flags))
3030 /* Unless we are doing a full sync, or a replacement
3031 * we only need to recover the block if it is set in
3034 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3036 if (sync_blocks < max_sync)
3037 max_sync = sync_blocks;
3041 /* yep, skip the sync_blocks here, but don't assume
3042 * that there will never be anything to do here
3044 chunks_skipped = -1;
3048 atomic_inc(&mrdev->nr_pending);
3050 atomic_inc(&mreplace->nr_pending);
3053 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3055 raise_barrier(conf, rb2 != NULL);
3056 atomic_set(&r10_bio->remaining, 0);
3058 r10_bio->master_bio = (struct bio*)rb2;
3060 atomic_inc(&rb2->remaining);
3061 r10_bio->mddev = mddev;
3062 set_bit(R10BIO_IsRecover, &r10_bio->state);
3063 r10_bio->sector = sect;
3065 raid10_find_phys(conf, r10_bio);
3067 /* Need to check if the array will still be
3071 for (j = 0; j < conf->geo.raid_disks; j++) {
3072 struct md_rdev *rdev = rcu_dereference(
3073 conf->mirrors[j].rdev);
3074 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3080 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3081 &sync_blocks, still_degraded);
3084 for (j=0; j<conf->copies;j++) {
3086 int d = r10_bio->devs[j].devnum;
3087 sector_t from_addr, to_addr;
3088 struct md_rdev *rdev =
3089 rcu_dereference(conf->mirrors[d].rdev);
3090 sector_t sector, first_bad;
3093 !test_bit(In_sync, &rdev->flags))
3095 /* This is where we read from */
3097 sector = r10_bio->devs[j].addr;
3099 if (is_badblock(rdev, sector, max_sync,
3100 &first_bad, &bad_sectors)) {
3101 if (first_bad > sector)
3102 max_sync = first_bad - sector;
3104 bad_sectors -= (sector
3106 if (max_sync > bad_sectors)
3107 max_sync = bad_sectors;
3111 bio = r10_bio->devs[0].bio;
3113 bio->bi_next = biolist;
3115 bio->bi_private = r10_bio;
3116 bio->bi_end_io = end_sync_read;
3117 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3118 from_addr = r10_bio->devs[j].addr;
3119 bio->bi_iter.bi_sector = from_addr +
3121 bio->bi_bdev = rdev->bdev;
3122 atomic_inc(&rdev->nr_pending);
3123 /* and we write to 'i' (if not in_sync) */
3125 for (k=0; k<conf->copies; k++)
3126 if (r10_bio->devs[k].devnum == i)
3128 BUG_ON(k == conf->copies);
3129 to_addr = r10_bio->devs[k].addr;
3130 r10_bio->devs[0].devnum = d;
3131 r10_bio->devs[0].addr = from_addr;
3132 r10_bio->devs[1].devnum = i;
3133 r10_bio->devs[1].addr = to_addr;
3135 if (!test_bit(In_sync, &mrdev->flags)) {
3136 bio = r10_bio->devs[1].bio;
3138 bio->bi_next = biolist;
3140 bio->bi_private = r10_bio;
3141 bio->bi_end_io = end_sync_write;
3142 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3143 bio->bi_iter.bi_sector = to_addr
3144 + mrdev->data_offset;
3145 bio->bi_bdev = mrdev->bdev;
3146 atomic_inc(&r10_bio->remaining);
3148 r10_bio->devs[1].bio->bi_end_io = NULL;
3150 /* and maybe write to replacement */
3151 bio = r10_bio->devs[1].repl_bio;
3153 bio->bi_end_io = NULL;
3154 /* Note: if mreplace != NULL, then bio
3155 * cannot be NULL as r10buf_pool_alloc will
3156 * have allocated it.
3157 * So the second test here is pointless.
3158 * But it keeps semantic-checkers happy, and
3159 * this comment keeps human reviewers
3162 if (mreplace == NULL || bio == NULL ||
3163 test_bit(Faulty, &mreplace->flags))
3166 bio->bi_next = biolist;
3168 bio->bi_private = r10_bio;
3169 bio->bi_end_io = end_sync_write;
3170 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3171 bio->bi_iter.bi_sector = to_addr +
3172 mreplace->data_offset;
3173 bio->bi_bdev = mreplace->bdev;
3174 atomic_inc(&r10_bio->remaining);
3178 if (j == conf->copies) {
3179 /* Cannot recover, so abort the recovery or
3180 * record a bad block */
3182 /* problem is that there are bad blocks
3183 * on other device(s)
3186 for (k = 0; k < conf->copies; k++)
3187 if (r10_bio->devs[k].devnum == i)
3189 if (!test_bit(In_sync,
3191 && !rdev_set_badblocks(
3193 r10_bio->devs[k].addr,
3197 !rdev_set_badblocks(
3199 r10_bio->devs[k].addr,
3204 if (!test_and_set_bit(MD_RECOVERY_INTR,
3206 printk(KERN_INFO "md/raid10:%s: insufficient "
3207 "working devices for recovery.\n",
3209 mirror->recovery_disabled
3210 = mddev->recovery_disabled;
3214 atomic_dec(&rb2->remaining);
3216 rdev_dec_pending(mrdev, mddev);
3218 rdev_dec_pending(mreplace, mddev);
3221 rdev_dec_pending(mrdev, mddev);
3223 rdev_dec_pending(mreplace, mddev);
3225 if (biolist == NULL) {
3227 struct r10bio *rb2 = r10_bio;
3228 r10_bio = (struct r10bio*) rb2->master_bio;
3229 rb2->master_bio = NULL;
3235 /* resync. Schedule a read for every block at this virt offset */
3238 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3240 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3241 &sync_blocks, mddev->degraded) &&
3242 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3243 &mddev->recovery)) {
3244 /* We can skip this block */
3246 return sync_blocks + sectors_skipped;
3248 if (sync_blocks < max_sync)
3249 max_sync = sync_blocks;
3250 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3253 r10_bio->mddev = mddev;
3254 atomic_set(&r10_bio->remaining, 0);
3255 raise_barrier(conf, 0);
3256 conf->next_resync = sector_nr;
3258 r10_bio->master_bio = NULL;
3259 r10_bio->sector = sector_nr;
3260 set_bit(R10BIO_IsSync, &r10_bio->state);
3261 raid10_find_phys(conf, r10_bio);
3262 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3264 for (i = 0; i < conf->copies; i++) {
3265 int d = r10_bio->devs[i].devnum;
3266 sector_t first_bad, sector;
3268 struct md_rdev *rdev;
3270 if (r10_bio->devs[i].repl_bio)
3271 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3273 bio = r10_bio->devs[i].bio;
3275 bio->bi_error = -EIO;
3277 rdev = rcu_dereference(conf->mirrors[d].rdev);
3278 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3282 sector = r10_bio->devs[i].addr;
3283 if (is_badblock(rdev, sector, max_sync,
3284 &first_bad, &bad_sectors)) {
3285 if (first_bad > sector)
3286 max_sync = first_bad - sector;
3288 bad_sectors -= (sector - first_bad);
3289 if (max_sync > bad_sectors)
3290 max_sync = bad_sectors;
3295 atomic_inc(&rdev->nr_pending);
3296 atomic_inc(&r10_bio->remaining);
3297 bio->bi_next = biolist;
3299 bio->bi_private = r10_bio;
3300 bio->bi_end_io = end_sync_read;
3301 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3302 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3303 bio->bi_bdev = rdev->bdev;
3306 rdev = rcu_dereference(conf->mirrors[d].replacement);
3307 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3311 atomic_inc(&rdev->nr_pending);
3314 /* Need to set up for writing to the replacement */
3315 bio = r10_bio->devs[i].repl_bio;
3317 bio->bi_error = -EIO;
3319 sector = r10_bio->devs[i].addr;
3320 bio->bi_next = biolist;
3322 bio->bi_private = r10_bio;
3323 bio->bi_end_io = end_sync_write;
3324 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3325 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3326 bio->bi_bdev = rdev->bdev;
3331 for (i=0; i<conf->copies; i++) {
3332 int d = r10_bio->devs[i].devnum;
3333 if (r10_bio->devs[i].bio->bi_end_io)
3334 rdev_dec_pending(conf->mirrors[d].rdev,
3336 if (r10_bio->devs[i].repl_bio &&
3337 r10_bio->devs[i].repl_bio->bi_end_io)
3339 conf->mirrors[d].replacement,
3349 if (sector_nr + max_sync < max_sector)
3350 max_sector = sector_nr + max_sync;
3353 int len = PAGE_SIZE;
3354 if (sector_nr + (len>>9) > max_sector)
3355 len = (max_sector - sector_nr) << 9;
3358 for (bio= biolist ; bio ; bio=bio->bi_next) {
3360 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3361 if (bio_add_page(bio, page, len, 0))
3365 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3366 for (bio2 = biolist;
3367 bio2 && bio2 != bio;
3368 bio2 = bio2->bi_next) {
3369 /* remove last page from this bio */
3371 bio2->bi_iter.bi_size -= len;
3372 bio_clear_flag(bio2, BIO_SEG_VALID);
3376 nr_sectors += len>>9;
3377 sector_nr += len>>9;
3378 } while (biolist->bi_vcnt < RESYNC_PAGES);
3380 r10_bio->sectors = nr_sectors;
3384 biolist = biolist->bi_next;
3386 bio->bi_next = NULL;
3387 r10_bio = bio->bi_private;
3388 r10_bio->sectors = nr_sectors;
3390 if (bio->bi_end_io == end_sync_read) {
3391 md_sync_acct(bio->bi_bdev, nr_sectors);
3393 generic_make_request(bio);
3397 if (sectors_skipped)
3398 /* pretend they weren't skipped, it makes
3399 * no important difference in this case
3401 md_done_sync(mddev, sectors_skipped, 1);
3403 return sectors_skipped + nr_sectors;
3405 /* There is nowhere to write, so all non-sync
3406 * drives must be failed or in resync, all drives
3407 * have a bad block, so try the next chunk...
3409 if (sector_nr + max_sync < max_sector)
3410 max_sector = sector_nr + max_sync;
3412 sectors_skipped += (max_sector - sector_nr);
3414 sector_nr = max_sector;
3419 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3422 struct r10conf *conf = mddev->private;
3425 raid_disks = min(conf->geo.raid_disks,
3426 conf->prev.raid_disks);
3428 sectors = conf->dev_sectors;
3430 size = sectors >> conf->geo.chunk_shift;
3431 sector_div(size, conf->geo.far_copies);
3432 size = size * raid_disks;
3433 sector_div(size, conf->geo.near_copies);
3435 return size << conf->geo.chunk_shift;
3438 static void calc_sectors(struct r10conf *conf, sector_t size)
3440 /* Calculate the number of sectors-per-device that will
3441 * actually be used, and set conf->dev_sectors and
3445 size = size >> conf->geo.chunk_shift;
3446 sector_div(size, conf->geo.far_copies);
3447 size = size * conf->geo.raid_disks;
3448 sector_div(size, conf->geo.near_copies);
3449 /* 'size' is now the number of chunks in the array */
3450 /* calculate "used chunks per device" */
3451 size = size * conf->copies;
3453 /* We need to round up when dividing by raid_disks to
3454 * get the stride size.
3456 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3458 conf->dev_sectors = size << conf->geo.chunk_shift;
3460 if (conf->geo.far_offset)
3461 conf->geo.stride = 1 << conf->geo.chunk_shift;
3463 sector_div(size, conf->geo.far_copies);
3464 conf->geo.stride = size << conf->geo.chunk_shift;
3468 enum geo_type {geo_new, geo_old, geo_start};
3469 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3472 int layout, chunk, disks;
3475 layout = mddev->layout;
3476 chunk = mddev->chunk_sectors;
3477 disks = mddev->raid_disks - mddev->delta_disks;
3480 layout = mddev->new_layout;
3481 chunk = mddev->new_chunk_sectors;
3482 disks = mddev->raid_disks;
3484 default: /* avoid 'may be unused' warnings */
3485 case geo_start: /* new when starting reshape - raid_disks not
3487 layout = mddev->new_layout;
3488 chunk = mddev->new_chunk_sectors;
3489 disks = mddev->raid_disks + mddev->delta_disks;
3494 if (chunk < (PAGE_SIZE >> 9) ||
3495 !is_power_of_2(chunk))
3498 fc = (layout >> 8) & 255;
3499 fo = layout & (1<<16);
3500 geo->raid_disks = disks;
3501 geo->near_copies = nc;
3502 geo->far_copies = fc;
3503 geo->far_offset = fo;
3504 switch (layout >> 17) {
3505 case 0: /* original layout. simple but not always optimal */
3506 geo->far_set_size = disks;
3508 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3509 * actually using this, but leave code here just in case.*/
3510 geo->far_set_size = disks/fc;
3511 WARN(geo->far_set_size < fc,
3512 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3514 case 2: /* "improved" layout fixed to match documentation */
3515 geo->far_set_size = fc * nc;
3517 default: /* Not a valid layout */
3520 geo->chunk_mask = chunk - 1;
3521 geo->chunk_shift = ffz(~chunk);
3525 static struct r10conf *setup_conf(struct mddev *mddev)
3527 struct r10conf *conf = NULL;
3532 copies = setup_geo(&geo, mddev, geo_new);
3535 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3536 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3537 mdname(mddev), PAGE_SIZE);
3541 if (copies < 2 || copies > mddev->raid_disks) {
3542 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3543 mdname(mddev), mddev->new_layout);
3548 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3552 /* FIXME calc properly */
3553 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3554 max(0,-mddev->delta_disks)),
3559 conf->tmppage = alloc_page(GFP_KERNEL);
3564 conf->copies = copies;
3565 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3566 r10bio_pool_free, conf);
3567 if (!conf->r10bio_pool)
3570 calc_sectors(conf, mddev->dev_sectors);
3571 if (mddev->reshape_position == MaxSector) {
3572 conf->prev = conf->geo;
3573 conf->reshape_progress = MaxSector;
3575 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3579 conf->reshape_progress = mddev->reshape_position;
3580 if (conf->prev.far_offset)
3581 conf->prev.stride = 1 << conf->prev.chunk_shift;
3583 /* far_copies must be 1 */
3584 conf->prev.stride = conf->dev_sectors;
3586 conf->reshape_safe = conf->reshape_progress;
3587 spin_lock_init(&conf->device_lock);
3588 INIT_LIST_HEAD(&conf->retry_list);
3589 INIT_LIST_HEAD(&conf->bio_end_io_list);
3591 spin_lock_init(&conf->resync_lock);
3592 init_waitqueue_head(&conf->wait_barrier);
3593 atomic_set(&conf->nr_pending, 0);
3595 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3599 conf->mddev = mddev;
3604 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3607 mempool_destroy(conf->r10bio_pool);
3608 kfree(conf->mirrors);
3609 safe_put_page(conf->tmppage);
3612 return ERR_PTR(err);
3615 static int raid10_run(struct mddev *mddev)
3617 struct r10conf *conf;
3618 int i, disk_idx, chunk_size;
3619 struct raid10_info *disk;
3620 struct md_rdev *rdev;
3622 sector_t min_offset_diff = 0;
3624 bool discard_supported = false;
3626 if (mddev->private == NULL) {
3627 conf = setup_conf(mddev);
3629 return PTR_ERR(conf);
3630 mddev->private = conf;
3632 conf = mddev->private;
3636 mddev->thread = conf->thread;
3637 conf->thread = NULL;
3639 chunk_size = mddev->chunk_sectors << 9;
3641 blk_queue_max_discard_sectors(mddev->queue,
3642 mddev->chunk_sectors);
3643 blk_queue_max_write_same_sectors(mddev->queue, 0);
3644 blk_queue_io_min(mddev->queue, chunk_size);
3645 if (conf->geo.raid_disks % conf->geo.near_copies)
3646 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3648 blk_queue_io_opt(mddev->queue, chunk_size *
3649 (conf->geo.raid_disks / conf->geo.near_copies));
3652 rdev_for_each(rdev, mddev) {
3654 struct request_queue *q;
3656 disk_idx = rdev->raid_disk;
3659 if (disk_idx >= conf->geo.raid_disks &&
3660 disk_idx >= conf->prev.raid_disks)
3662 disk = conf->mirrors + disk_idx;
3664 if (test_bit(Replacement, &rdev->flags)) {
3665 if (disk->replacement)
3667 disk->replacement = rdev;
3673 q = bdev_get_queue(rdev->bdev);
3674 diff = (rdev->new_data_offset - rdev->data_offset);
3675 if (!mddev->reshape_backwards)
3679 if (first || diff < min_offset_diff)
3680 min_offset_diff = diff;
3683 disk_stack_limits(mddev->gendisk, rdev->bdev,
3684 rdev->data_offset << 9);
3686 disk->head_position = 0;
3688 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3689 discard_supported = true;
3694 if (discard_supported)
3695 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3698 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3701 /* need to check that every block has at least one working mirror */
3702 if (!enough(conf, -1)) {
3703 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3708 if (conf->reshape_progress != MaxSector) {
3709 /* must ensure that shape change is supported */
3710 if (conf->geo.far_copies != 1 &&
3711 conf->geo.far_offset == 0)
3713 if (conf->prev.far_copies != 1 &&
3714 conf->prev.far_offset == 0)
3718 mddev->degraded = 0;
3720 i < conf->geo.raid_disks
3721 || i < conf->prev.raid_disks;
3724 disk = conf->mirrors + i;
3726 if (!disk->rdev && disk->replacement) {
3727 /* The replacement is all we have - use it */
3728 disk->rdev = disk->replacement;
3729 disk->replacement = NULL;
3730 clear_bit(Replacement, &disk->rdev->flags);
3734 !test_bit(In_sync, &disk->rdev->flags)) {
3735 disk->head_position = 0;
3738 disk->rdev->saved_raid_disk < 0)
3742 if (disk->replacement &&
3743 !test_bit(In_sync, &disk->replacement->flags) &&
3744 disk->replacement->saved_raid_disk < 0) {
3748 disk->recovery_disabled = mddev->recovery_disabled - 1;
3751 if (mddev->recovery_cp != MaxSector)
3752 printk(KERN_NOTICE "md/raid10:%s: not clean"
3753 " -- starting background reconstruction\n",
3756 "md/raid10:%s: active with %d out of %d devices\n",
3757 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3758 conf->geo.raid_disks);
3760 * Ok, everything is just fine now
3762 mddev->dev_sectors = conf->dev_sectors;
3763 size = raid10_size(mddev, 0, 0);
3764 md_set_array_sectors(mddev, size);
3765 mddev->resync_max_sectors = size;
3768 int stripe = conf->geo.raid_disks *
3769 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3771 /* Calculate max read-ahead size.
3772 * We need to readahead at least twice a whole stripe....
3775 stripe /= conf->geo.near_copies;
3776 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3777 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3780 if (md_integrity_register(mddev))
3783 if (conf->reshape_progress != MaxSector) {
3784 unsigned long before_length, after_length;
3786 before_length = ((1 << conf->prev.chunk_shift) *
3787 conf->prev.far_copies);
3788 after_length = ((1 << conf->geo.chunk_shift) *
3789 conf->geo.far_copies);
3791 if (max(before_length, after_length) > min_offset_diff) {
3792 /* This cannot work */
3793 printk("md/raid10: offset difference not enough to continue reshape\n");
3796 conf->offset_diff = min_offset_diff;
3798 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3799 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3800 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3801 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3802 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3804 if (!mddev->sync_thread)
3811 md_unregister_thread(&mddev->thread);
3812 mempool_destroy(conf->r10bio_pool);
3813 safe_put_page(conf->tmppage);
3814 kfree(conf->mirrors);
3816 mddev->private = NULL;
3821 static void raid10_free(struct mddev *mddev, void *priv)
3823 struct r10conf *conf = priv;
3825 mempool_destroy(conf->r10bio_pool);
3826 safe_put_page(conf->tmppage);
3827 kfree(conf->mirrors);
3828 kfree(conf->mirrors_old);
3829 kfree(conf->mirrors_new);
3833 static void raid10_quiesce(struct mddev *mddev, int state)
3835 struct r10conf *conf = mddev->private;
3839 raise_barrier(conf, 0);
3842 lower_barrier(conf);
3847 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3849 /* Resize of 'far' arrays is not supported.
3850 * For 'near' and 'offset' arrays we can set the
3851 * number of sectors used to be an appropriate multiple
3852 * of the chunk size.
3853 * For 'offset', this is far_copies*chunksize.
3854 * For 'near' the multiplier is the LCM of
3855 * near_copies and raid_disks.
3856 * So if far_copies > 1 && !far_offset, fail.
3857 * Else find LCM(raid_disks, near_copy)*far_copies and
3858 * multiply by chunk_size. Then round to this number.
3859 * This is mostly done by raid10_size()
3861 struct r10conf *conf = mddev->private;
3862 sector_t oldsize, size;
3864 if (mddev->reshape_position != MaxSector)
3867 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3870 oldsize = raid10_size(mddev, 0, 0);
3871 size = raid10_size(mddev, sectors, 0);
3872 if (mddev->external_size &&
3873 mddev->array_sectors > size)
3875 if (mddev->bitmap) {
3876 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3880 md_set_array_sectors(mddev, size);
3882 set_capacity(mddev->gendisk, mddev->array_sectors);
3883 revalidate_disk(mddev->gendisk);
3885 if (sectors > mddev->dev_sectors &&
3886 mddev->recovery_cp > oldsize) {
3887 mddev->recovery_cp = oldsize;
3888 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3890 calc_sectors(conf, sectors);
3891 mddev->dev_sectors = conf->dev_sectors;
3892 mddev->resync_max_sectors = size;
3896 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3898 struct md_rdev *rdev;
3899 struct r10conf *conf;
3901 if (mddev->degraded > 0) {
3902 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3904 return ERR_PTR(-EINVAL);
3906 sector_div(size, devs);
3908 /* Set new parameters */
3909 mddev->new_level = 10;
3910 /* new layout: far_copies = 1, near_copies = 2 */
3911 mddev->new_layout = (1<<8) + 2;
3912 mddev->new_chunk_sectors = mddev->chunk_sectors;
3913 mddev->delta_disks = mddev->raid_disks;
3914 mddev->raid_disks *= 2;
3915 /* make sure it will be not marked as dirty */
3916 mddev->recovery_cp = MaxSector;
3917 mddev->dev_sectors = size;
3919 conf = setup_conf(mddev);
3920 if (!IS_ERR(conf)) {
3921 rdev_for_each(rdev, mddev)
3922 if (rdev->raid_disk >= 0) {
3923 rdev->new_raid_disk = rdev->raid_disk * 2;
3924 rdev->sectors = size;
3932 static void *raid10_takeover(struct mddev *mddev)
3934 struct r0conf *raid0_conf;
3936 /* raid10 can take over:
3937 * raid0 - providing it has only two drives
3939 if (mddev->level == 0) {
3940 /* for raid0 takeover only one zone is supported */
3941 raid0_conf = mddev->private;
3942 if (raid0_conf->nr_strip_zones > 1) {
3943 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3944 " with more than one zone.\n",
3946 return ERR_PTR(-EINVAL);
3948 return raid10_takeover_raid0(mddev,
3949 raid0_conf->strip_zone->zone_end,
3950 raid0_conf->strip_zone->nb_dev);
3952 return ERR_PTR(-EINVAL);
3955 static int raid10_check_reshape(struct mddev *mddev)
3957 /* Called when there is a request to change
3958 * - layout (to ->new_layout)
3959 * - chunk size (to ->new_chunk_sectors)
3960 * - raid_disks (by delta_disks)
3961 * or when trying to restart a reshape that was ongoing.
3963 * We need to validate the request and possibly allocate
3964 * space if that might be an issue later.
3966 * Currently we reject any reshape of a 'far' mode array,
3967 * allow chunk size to change if new is generally acceptable,
3968 * allow raid_disks to increase, and allow
3969 * a switch between 'near' mode and 'offset' mode.
3971 struct r10conf *conf = mddev->private;
3974 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3977 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3978 /* mustn't change number of copies */
3980 if (geo.far_copies > 1 && !geo.far_offset)
3981 /* Cannot switch to 'far' mode */
3984 if (mddev->array_sectors & geo.chunk_mask)
3985 /* not factor of array size */
3988 if (!enough(conf, -1))
3991 kfree(conf->mirrors_new);
3992 conf->mirrors_new = NULL;
3993 if (mddev->delta_disks > 0) {
3994 /* allocate new 'mirrors' list */
3995 conf->mirrors_new = kzalloc(
3996 sizeof(struct raid10_info)
3997 *(mddev->raid_disks +
3998 mddev->delta_disks),
4000 if (!conf->mirrors_new)
4007 * Need to check if array has failed when deciding whether to:
4009 * - remove non-faulty devices
4012 * This determination is simple when no reshape is happening.
4013 * However if there is a reshape, we need to carefully check
4014 * both the before and after sections.
4015 * This is because some failed devices may only affect one
4016 * of the two sections, and some non-in_sync devices may
4017 * be insync in the section most affected by failed devices.
4019 static int calc_degraded(struct r10conf *conf)
4021 int degraded, degraded2;
4026 /* 'prev' section first */
4027 for (i = 0; i < conf->prev.raid_disks; i++) {
4028 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4029 if (!rdev || test_bit(Faulty, &rdev->flags))
4031 else if (!test_bit(In_sync, &rdev->flags))
4032 /* When we can reduce the number of devices in
4033 * an array, this might not contribute to
4034 * 'degraded'. It does now.
4039 if (conf->geo.raid_disks == conf->prev.raid_disks)
4043 for (i = 0; i < conf->geo.raid_disks; i++) {
4044 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4045 if (!rdev || test_bit(Faulty, &rdev->flags))
4047 else if (!test_bit(In_sync, &rdev->flags)) {
4048 /* If reshape is increasing the number of devices,
4049 * this section has already been recovered, so
4050 * it doesn't contribute to degraded.
4053 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4058 if (degraded2 > degraded)
4063 static int raid10_start_reshape(struct mddev *mddev)
4065 /* A 'reshape' has been requested. This commits
4066 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4067 * This also checks if there are enough spares and adds them
4069 * We currently require enough spares to make the final
4070 * array non-degraded. We also require that the difference
4071 * between old and new data_offset - on each device - is
4072 * enough that we never risk over-writing.
4075 unsigned long before_length, after_length;
4076 sector_t min_offset_diff = 0;
4079 struct r10conf *conf = mddev->private;
4080 struct md_rdev *rdev;
4084 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4087 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4090 before_length = ((1 << conf->prev.chunk_shift) *
4091 conf->prev.far_copies);
4092 after_length = ((1 << conf->geo.chunk_shift) *
4093 conf->geo.far_copies);
4095 rdev_for_each(rdev, mddev) {
4096 if (!test_bit(In_sync, &rdev->flags)
4097 && !test_bit(Faulty, &rdev->flags))
4099 if (rdev->raid_disk >= 0) {
4100 long long diff = (rdev->new_data_offset
4101 - rdev->data_offset);
4102 if (!mddev->reshape_backwards)
4106 if (first || diff < min_offset_diff)
4107 min_offset_diff = diff;
4112 if (max(before_length, after_length) > min_offset_diff)
4115 if (spares < mddev->delta_disks)
4118 conf->offset_diff = min_offset_diff;
4119 spin_lock_irq(&conf->device_lock);
4120 if (conf->mirrors_new) {
4121 memcpy(conf->mirrors_new, conf->mirrors,
4122 sizeof(struct raid10_info)*conf->prev.raid_disks);
4124 kfree(conf->mirrors_old);
4125 conf->mirrors_old = conf->mirrors;
4126 conf->mirrors = conf->mirrors_new;
4127 conf->mirrors_new = NULL;
4129 setup_geo(&conf->geo, mddev, geo_start);
4131 if (mddev->reshape_backwards) {
4132 sector_t size = raid10_size(mddev, 0, 0);
4133 if (size < mddev->array_sectors) {
4134 spin_unlock_irq(&conf->device_lock);
4135 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4139 mddev->resync_max_sectors = size;
4140 conf->reshape_progress = size;
4142 conf->reshape_progress = 0;
4143 conf->reshape_safe = conf->reshape_progress;
4144 spin_unlock_irq(&conf->device_lock);
4146 if (mddev->delta_disks && mddev->bitmap) {
4147 ret = bitmap_resize(mddev->bitmap,
4148 raid10_size(mddev, 0,
4149 conf->geo.raid_disks),
4154 if (mddev->delta_disks > 0) {
4155 rdev_for_each(rdev, mddev)
4156 if (rdev->raid_disk < 0 &&
4157 !test_bit(Faulty, &rdev->flags)) {
4158 if (raid10_add_disk(mddev, rdev) == 0) {
4159 if (rdev->raid_disk >=
4160 conf->prev.raid_disks)
4161 set_bit(In_sync, &rdev->flags);
4163 rdev->recovery_offset = 0;
4165 if (sysfs_link_rdev(mddev, rdev))
4166 /* Failure here is OK */;
4168 } else if (rdev->raid_disk >= conf->prev.raid_disks
4169 && !test_bit(Faulty, &rdev->flags)) {
4170 /* This is a spare that was manually added */
4171 set_bit(In_sync, &rdev->flags);
4174 /* When a reshape changes the number of devices,
4175 * ->degraded is measured against the larger of the
4176 * pre and post numbers.
4178 spin_lock_irq(&conf->device_lock);
4179 mddev->degraded = calc_degraded(conf);
4180 spin_unlock_irq(&conf->device_lock);
4181 mddev->raid_disks = conf->geo.raid_disks;
4182 mddev->reshape_position = conf->reshape_progress;
4183 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4185 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4186 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4187 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4188 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4189 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4191 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4193 if (!mddev->sync_thread) {
4197 conf->reshape_checkpoint = jiffies;
4198 md_wakeup_thread(mddev->sync_thread);
4199 md_new_event(mddev);
4203 mddev->recovery = 0;
4204 spin_lock_irq(&conf->device_lock);
4205 conf->geo = conf->prev;
4206 mddev->raid_disks = conf->geo.raid_disks;
4207 rdev_for_each(rdev, mddev)
4208 rdev->new_data_offset = rdev->data_offset;
4210 conf->reshape_progress = MaxSector;
4211 conf->reshape_safe = MaxSector;
4212 mddev->reshape_position = MaxSector;
4213 spin_unlock_irq(&conf->device_lock);
4217 /* Calculate the last device-address that could contain
4218 * any block from the chunk that includes the array-address 's'
4219 * and report the next address.
4220 * i.e. the address returned will be chunk-aligned and after
4221 * any data that is in the chunk containing 's'.
4223 static sector_t last_dev_address(sector_t s, struct geom *geo)
4225 s = (s | geo->chunk_mask) + 1;
4226 s >>= geo->chunk_shift;
4227 s *= geo->near_copies;
4228 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4229 s *= geo->far_copies;
4230 s <<= geo->chunk_shift;
4234 /* Calculate the first device-address that could contain
4235 * any block from the chunk that includes the array-address 's'.
4236 * This too will be the start of a chunk
4238 static sector_t first_dev_address(sector_t s, struct geom *geo)
4240 s >>= geo->chunk_shift;
4241 s *= geo->near_copies;
4242 sector_div(s, geo->raid_disks);
4243 s *= geo->far_copies;
4244 s <<= geo->chunk_shift;
4248 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4251 /* We simply copy at most one chunk (smallest of old and new)
4252 * at a time, possibly less if that exceeds RESYNC_PAGES,
4253 * or we hit a bad block or something.
4254 * This might mean we pause for normal IO in the middle of
4255 * a chunk, but that is not a problem as mddev->reshape_position
4256 * can record any location.
4258 * If we will want to write to a location that isn't
4259 * yet recorded as 'safe' (i.e. in metadata on disk) then
4260 * we need to flush all reshape requests and update the metadata.
4262 * When reshaping forwards (e.g. to more devices), we interpret
4263 * 'safe' as the earliest block which might not have been copied
4264 * down yet. We divide this by previous stripe size and multiply
4265 * by previous stripe length to get lowest device offset that we
4266 * cannot write to yet.
4267 * We interpret 'sector_nr' as an address that we want to write to.
4268 * From this we use last_device_address() to find where we might
4269 * write to, and first_device_address on the 'safe' position.
4270 * If this 'next' write position is after the 'safe' position,
4271 * we must update the metadata to increase the 'safe' position.
4273 * When reshaping backwards, we round in the opposite direction
4274 * and perform the reverse test: next write position must not be
4275 * less than current safe position.
4277 * In all this the minimum difference in data offsets
4278 * (conf->offset_diff - always positive) allows a bit of slack,
4279 * so next can be after 'safe', but not by more than offset_diff
4281 * We need to prepare all the bios here before we start any IO
4282 * to ensure the size we choose is acceptable to all devices.
4283 * The means one for each copy for write-out and an extra one for
4285 * We store the read-in bio in ->master_bio and the others in
4286 * ->devs[x].bio and ->devs[x].repl_bio.
4288 struct r10conf *conf = mddev->private;
4289 struct r10bio *r10_bio;
4290 sector_t next, safe, last;
4294 struct md_rdev *rdev;
4297 struct bio *bio, *read_bio;
4298 int sectors_done = 0;
4300 if (sector_nr == 0) {
4301 /* If restarting in the middle, skip the initial sectors */
4302 if (mddev->reshape_backwards &&
4303 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4304 sector_nr = (raid10_size(mddev, 0, 0)
4305 - conf->reshape_progress);
4306 } else if (!mddev->reshape_backwards &&
4307 conf->reshape_progress > 0)
4308 sector_nr = conf->reshape_progress;
4310 mddev->curr_resync_completed = sector_nr;
4311 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4317 /* We don't use sector_nr to track where we are up to
4318 * as that doesn't work well for ->reshape_backwards.
4319 * So just use ->reshape_progress.
4321 if (mddev->reshape_backwards) {
4322 /* 'next' is the earliest device address that we might
4323 * write to for this chunk in the new layout
4325 next = first_dev_address(conf->reshape_progress - 1,
4328 /* 'safe' is the last device address that we might read from
4329 * in the old layout after a restart
4331 safe = last_dev_address(conf->reshape_safe - 1,
4334 if (next + conf->offset_diff < safe)
4337 last = conf->reshape_progress - 1;
4338 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4339 & conf->prev.chunk_mask);
4340 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4341 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4343 /* 'next' is after the last device address that we
4344 * might write to for this chunk in the new layout
4346 next = last_dev_address(conf->reshape_progress, &conf->geo);
4348 /* 'safe' is the earliest device address that we might
4349 * read from in the old layout after a restart
4351 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4353 /* Need to update metadata if 'next' might be beyond 'safe'
4354 * as that would possibly corrupt data
4356 if (next > safe + conf->offset_diff)
4359 sector_nr = conf->reshape_progress;
4360 last = sector_nr | (conf->geo.chunk_mask
4361 & conf->prev.chunk_mask);
4363 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4364 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4368 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4369 /* Need to update reshape_position in metadata */
4371 mddev->reshape_position = conf->reshape_progress;
4372 if (mddev->reshape_backwards)
4373 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4374 - conf->reshape_progress;
4376 mddev->curr_resync_completed = conf->reshape_progress;
4377 conf->reshape_checkpoint = jiffies;
4378 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4379 md_wakeup_thread(mddev->thread);
4380 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4381 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4382 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4383 allow_barrier(conf);
4384 return sectors_done;
4386 conf->reshape_safe = mddev->reshape_position;
4387 allow_barrier(conf);
4390 raise_barrier(conf, 0);
4392 /* Now schedule reads for blocks from sector_nr to last */
4393 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4395 raise_barrier(conf, 1);
4396 atomic_set(&r10_bio->remaining, 0);
4397 r10_bio->mddev = mddev;
4398 r10_bio->sector = sector_nr;
4399 set_bit(R10BIO_IsReshape, &r10_bio->state);
4400 r10_bio->sectors = last - sector_nr + 1;
4401 rdev = read_balance(conf, r10_bio, &max_sectors);
4402 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4405 /* Cannot read from here, so need to record bad blocks
4406 * on all the target devices.
4409 mempool_free(r10_bio, conf->r10buf_pool);
4410 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4411 return sectors_done;
4414 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4416 read_bio->bi_bdev = rdev->bdev;
4417 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4418 + rdev->data_offset);
4419 read_bio->bi_private = r10_bio;
4420 read_bio->bi_end_io = end_sync_read;
4421 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4422 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4423 read_bio->bi_error = 0;
4424 read_bio->bi_vcnt = 0;
4425 read_bio->bi_iter.bi_size = 0;
4426 r10_bio->master_bio = read_bio;
4427 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4429 /* Now find the locations in the new layout */
4430 __raid10_find_phys(&conf->geo, r10_bio);
4433 read_bio->bi_next = NULL;
4436 for (s = 0; s < conf->copies*2; s++) {
4438 int d = r10_bio->devs[s/2].devnum;
4439 struct md_rdev *rdev2;
4441 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4442 b = r10_bio->devs[s/2].repl_bio;
4444 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4445 b = r10_bio->devs[s/2].bio;
4447 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4451 b->bi_bdev = rdev2->bdev;
4452 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4453 rdev2->new_data_offset;
4454 b->bi_private = r10_bio;
4455 b->bi_end_io = end_reshape_write;
4456 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4461 /* Now add as many pages as possible to all of these bios. */
4464 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4465 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4466 int len = (max_sectors - s) << 9;
4467 if (len > PAGE_SIZE)
4469 for (bio = blist; bio ; bio = bio->bi_next) {
4471 if (bio_add_page(bio, page, len, 0))
4474 /* Didn't fit, must stop */
4476 bio2 && bio2 != bio;
4477 bio2 = bio2->bi_next) {
4478 /* Remove last page from this bio */
4480 bio2->bi_iter.bi_size -= len;
4481 bio_clear_flag(bio2, BIO_SEG_VALID);
4485 sector_nr += len >> 9;
4486 nr_sectors += len >> 9;
4490 r10_bio->sectors = nr_sectors;
4492 /* Now submit the read */
4493 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4494 atomic_inc(&r10_bio->remaining);
4495 read_bio->bi_next = NULL;
4496 generic_make_request(read_bio);
4497 sectors_done += nr_sectors;
4498 if (sector_nr <= last)
4501 lower_barrier(conf);
4503 /* Now that we have done the whole section we can
4504 * update reshape_progress
4506 if (mddev->reshape_backwards)
4507 conf->reshape_progress -= sectors_done;
4509 conf->reshape_progress += sectors_done;
4511 return sectors_done;
4514 static void end_reshape_request(struct r10bio *r10_bio);
4515 static int handle_reshape_read_error(struct mddev *mddev,
4516 struct r10bio *r10_bio);
4517 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4519 /* Reshape read completed. Hopefully we have a block
4521 * If we got a read error then we do sync 1-page reads from
4522 * elsewhere until we find the data - or give up.
4524 struct r10conf *conf = mddev->private;
4527 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4528 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4529 /* Reshape has been aborted */
4530 md_done_sync(mddev, r10_bio->sectors, 0);
4534 /* We definitely have the data in the pages, schedule the
4537 atomic_set(&r10_bio->remaining, 1);
4538 for (s = 0; s < conf->copies*2; s++) {
4540 int d = r10_bio->devs[s/2].devnum;
4541 struct md_rdev *rdev;
4544 rdev = rcu_dereference(conf->mirrors[d].replacement);
4545 b = r10_bio->devs[s/2].repl_bio;
4547 rdev = rcu_dereference(conf->mirrors[d].rdev);
4548 b = r10_bio->devs[s/2].bio;
4550 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4554 atomic_inc(&rdev->nr_pending);
4556 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4557 atomic_inc(&r10_bio->remaining);
4559 generic_make_request(b);
4561 end_reshape_request(r10_bio);
4564 static void end_reshape(struct r10conf *conf)
4566 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4569 spin_lock_irq(&conf->device_lock);
4570 conf->prev = conf->geo;
4571 md_finish_reshape(conf->mddev);
4573 conf->reshape_progress = MaxSector;
4574 conf->reshape_safe = MaxSector;
4575 spin_unlock_irq(&conf->device_lock);
4577 /* read-ahead size must cover two whole stripes, which is
4578 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4580 if (conf->mddev->queue) {
4581 int stripe = conf->geo.raid_disks *
4582 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4583 stripe /= conf->geo.near_copies;
4584 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4585 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4590 static int handle_reshape_read_error(struct mddev *mddev,
4591 struct r10bio *r10_bio)
4593 /* Use sync reads to get the blocks from somewhere else */
4594 int sectors = r10_bio->sectors;
4595 struct r10conf *conf = mddev->private;
4597 struct r10bio r10_bio;
4598 struct r10dev devs[conf->copies];
4600 struct r10bio *r10b = &on_stack.r10_bio;
4603 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4605 r10b->sector = r10_bio->sector;
4606 __raid10_find_phys(&conf->prev, r10b);
4611 int first_slot = slot;
4613 if (s > (PAGE_SIZE >> 9))
4618 int d = r10b->devs[slot].devnum;
4619 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4622 test_bit(Faulty, &rdev->flags) ||
4623 !test_bit(In_sync, &rdev->flags))
4626 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4627 atomic_inc(&rdev->nr_pending);
4629 success = sync_page_io(rdev,
4633 REQ_OP_READ, 0, false);
4634 rdev_dec_pending(rdev, mddev);
4640 if (slot >= conf->copies)
4642 if (slot == first_slot)
4647 /* couldn't read this block, must give up */
4648 set_bit(MD_RECOVERY_INTR,
4658 static void end_reshape_write(struct bio *bio)
4660 struct r10bio *r10_bio = bio->bi_private;
4661 struct mddev *mddev = r10_bio->mddev;
4662 struct r10conf *conf = mddev->private;
4666 struct md_rdev *rdev = NULL;
4668 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4670 rdev = conf->mirrors[d].replacement;
4673 rdev = conf->mirrors[d].rdev;
4676 if (bio->bi_error) {
4677 /* FIXME should record badblock */
4678 md_error(mddev, rdev);
4681 rdev_dec_pending(rdev, mddev);
4682 end_reshape_request(r10_bio);
4685 static void end_reshape_request(struct r10bio *r10_bio)
4687 if (!atomic_dec_and_test(&r10_bio->remaining))
4689 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4690 bio_put(r10_bio->master_bio);
4694 static void raid10_finish_reshape(struct mddev *mddev)
4696 struct r10conf *conf = mddev->private;
4698 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4701 if (mddev->delta_disks > 0) {
4702 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4703 mddev->recovery_cp = mddev->resync_max_sectors;
4704 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4706 mddev->resync_max_sectors = mddev->array_sectors;
4710 for (d = conf->geo.raid_disks ;
4711 d < conf->geo.raid_disks - mddev->delta_disks;
4713 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4715 clear_bit(In_sync, &rdev->flags);
4716 rdev = rcu_dereference(conf->mirrors[d].replacement);
4718 clear_bit(In_sync, &rdev->flags);
4722 mddev->layout = mddev->new_layout;
4723 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4724 mddev->reshape_position = MaxSector;
4725 mddev->delta_disks = 0;
4726 mddev->reshape_backwards = 0;
4729 static struct md_personality raid10_personality =
4733 .owner = THIS_MODULE,
4734 .make_request = raid10_make_request,
4736 .free = raid10_free,
4737 .status = raid10_status,
4738 .error_handler = raid10_error,
4739 .hot_add_disk = raid10_add_disk,
4740 .hot_remove_disk= raid10_remove_disk,
4741 .spare_active = raid10_spare_active,
4742 .sync_request = raid10_sync_request,
4743 .quiesce = raid10_quiesce,
4744 .size = raid10_size,
4745 .resize = raid10_resize,
4746 .takeover = raid10_takeover,
4747 .check_reshape = raid10_check_reshape,
4748 .start_reshape = raid10_start_reshape,
4749 .finish_reshape = raid10_finish_reshape,
4750 .congested = raid10_congested,
4753 static int __init raid_init(void)
4755 return register_md_personality(&raid10_personality);
4758 static void raid_exit(void)
4760 unregister_md_personality(&raid10_personality);
4763 module_init(raid_init);
4764 module_exit(raid_exit);
4765 MODULE_LICENSE("GPL");
4766 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4767 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4768 MODULE_ALIAS("md-raid10");
4769 MODULE_ALIAS("md-level-10");
4771 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);