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;
451 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
454 rdev = conf->mirrors[dev].replacement;
458 rdev = conf->mirrors[dev].rdev;
461 * this branch is our 'one mirror IO has finished' event handler:
465 /* Never record new bad blocks to replacement,
468 md_error(rdev->mddev, rdev);
470 set_bit(WriteErrorSeen, &rdev->flags);
471 if (!test_and_set_bit(WantReplacement, &rdev->flags))
472 set_bit(MD_RECOVERY_NEEDED,
473 &rdev->mddev->recovery);
474 set_bit(R10BIO_WriteError, &r10_bio->state);
479 * Set R10BIO_Uptodate in our master bio, so that
480 * we will return a good error code for to the higher
481 * levels even if IO on some other mirrored buffer fails.
483 * The 'master' represents the composite IO operation to
484 * user-side. So if something waits for IO, then it will
485 * wait for the 'master' bio.
491 * Do not set R10BIO_Uptodate if the current device is
492 * rebuilding or Faulty. This is because we cannot use
493 * such device for properly reading the data back (we could
494 * potentially use it, if the current write would have felt
495 * before rdev->recovery_offset, but for simplicity we don't
498 if (test_bit(In_sync, &rdev->flags) &&
499 !test_bit(Faulty, &rdev->flags))
500 set_bit(R10BIO_Uptodate, &r10_bio->state);
502 /* Maybe we can clear some bad blocks. */
503 if (is_badblock(rdev,
504 r10_bio->devs[slot].addr,
506 &first_bad, &bad_sectors)) {
509 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
511 r10_bio->devs[slot].bio = IO_MADE_GOOD;
513 set_bit(R10BIO_MadeGood, &r10_bio->state);
519 * Let's see if all mirrored write operations have finished
522 one_write_done(r10_bio);
524 rdev_dec_pending(rdev, conf->mddev);
528 * RAID10 layout manager
529 * As well as the chunksize and raid_disks count, there are two
530 * parameters: near_copies and far_copies.
531 * near_copies * far_copies must be <= raid_disks.
532 * Normally one of these will be 1.
533 * If both are 1, we get raid0.
534 * If near_copies == raid_disks, we get raid1.
536 * Chunks are laid out in raid0 style with near_copies copies of the
537 * first chunk, followed by near_copies copies of the next chunk and
539 * If far_copies > 1, then after 1/far_copies of the array has been assigned
540 * as described above, we start again with a device offset of near_copies.
541 * So we effectively have another copy of the whole array further down all
542 * the drives, but with blocks on different drives.
543 * With this layout, and block is never stored twice on the one device.
545 * raid10_find_phys finds the sector offset of a given virtual sector
546 * on each device that it is on.
548 * raid10_find_virt does the reverse mapping, from a device and a
549 * sector offset to a virtual address
552 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
560 int last_far_set_start, last_far_set_size;
562 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
563 last_far_set_start *= geo->far_set_size;
565 last_far_set_size = geo->far_set_size;
566 last_far_set_size += (geo->raid_disks % geo->far_set_size);
568 /* now calculate first sector/dev */
569 chunk = r10bio->sector >> geo->chunk_shift;
570 sector = r10bio->sector & geo->chunk_mask;
572 chunk *= geo->near_copies;
574 dev = sector_div(stripe, geo->raid_disks);
576 stripe *= geo->far_copies;
578 sector += stripe << geo->chunk_shift;
580 /* and calculate all the others */
581 for (n = 0; n < geo->near_copies; n++) {
585 r10bio->devs[slot].devnum = d;
586 r10bio->devs[slot].addr = s;
589 for (f = 1; f < geo->far_copies; f++) {
590 set = d / geo->far_set_size;
591 d += geo->near_copies;
593 if ((geo->raid_disks % geo->far_set_size) &&
594 (d > last_far_set_start)) {
595 d -= last_far_set_start;
596 d %= last_far_set_size;
597 d += last_far_set_start;
599 d %= geo->far_set_size;
600 d += geo->far_set_size * set;
603 r10bio->devs[slot].devnum = d;
604 r10bio->devs[slot].addr = s;
608 if (dev >= geo->raid_disks) {
610 sector += (geo->chunk_mask + 1);
615 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
617 struct geom *geo = &conf->geo;
619 if (conf->reshape_progress != MaxSector &&
620 ((r10bio->sector >= conf->reshape_progress) !=
621 conf->mddev->reshape_backwards)) {
622 set_bit(R10BIO_Previous, &r10bio->state);
625 clear_bit(R10BIO_Previous, &r10bio->state);
627 __raid10_find_phys(geo, r10bio);
630 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
632 sector_t offset, chunk, vchunk;
633 /* Never use conf->prev as this is only called during resync
634 * or recovery, so reshape isn't happening
636 struct geom *geo = &conf->geo;
637 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
638 int far_set_size = geo->far_set_size;
639 int last_far_set_start;
641 if (geo->raid_disks % geo->far_set_size) {
642 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
643 last_far_set_start *= geo->far_set_size;
645 if (dev >= last_far_set_start) {
646 far_set_size = geo->far_set_size;
647 far_set_size += (geo->raid_disks % geo->far_set_size);
648 far_set_start = last_far_set_start;
652 offset = sector & geo->chunk_mask;
653 if (geo->far_offset) {
655 chunk = sector >> geo->chunk_shift;
656 fc = sector_div(chunk, geo->far_copies);
657 dev -= fc * geo->near_copies;
658 if (dev < far_set_start)
661 while (sector >= geo->stride) {
662 sector -= geo->stride;
663 if (dev < (geo->near_copies + far_set_start))
664 dev += far_set_size - geo->near_copies;
666 dev -= geo->near_copies;
668 chunk = sector >> geo->chunk_shift;
670 vchunk = chunk * geo->raid_disks + dev;
671 sector_div(vchunk, geo->near_copies);
672 return (vchunk << geo->chunk_shift) + offset;
676 * This routine returns the disk from which the requested read should
677 * be done. There is a per-array 'next expected sequential IO' sector
678 * number - if this matches on the next IO then we use the last disk.
679 * There is also a per-disk 'last know head position' sector that is
680 * maintained from IRQ contexts, both the normal and the resync IO
681 * completion handlers update this position correctly. If there is no
682 * perfect sequential match then we pick the disk whose head is closest.
684 * If there are 2 mirrors in the same 2 devices, performance degrades
685 * because position is mirror, not device based.
687 * The rdev for the device selected will have nr_pending incremented.
691 * FIXME: possibly should rethink readbalancing and do it differently
692 * depending on near_copies / far_copies geometry.
694 static struct md_rdev *read_balance(struct r10conf *conf,
695 struct r10bio *r10_bio,
698 const sector_t this_sector = r10_bio->sector;
700 int sectors = r10_bio->sectors;
701 int best_good_sectors;
702 sector_t new_distance, best_dist;
703 struct md_rdev *best_rdev, *rdev = NULL;
706 struct geom *geo = &conf->geo;
708 raid10_find_phys(conf, r10_bio);
710 sectors = r10_bio->sectors;
713 best_dist = MaxSector;
714 best_good_sectors = 0;
717 * Check if we can balance. We can balance on the whole
718 * device if no resync is going on (recovery is ok), or below
719 * the resync window. We take the first readable disk when
720 * above the resync window.
722 if (conf->mddev->recovery_cp < MaxSector
723 && (this_sector + sectors >= conf->next_resync))
726 for (slot = 0; slot < conf->copies ; slot++) {
731 if (r10_bio->devs[slot].bio == IO_BLOCKED)
733 disk = r10_bio->devs[slot].devnum;
734 rdev = rcu_dereference(conf->mirrors[disk].replacement);
735 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
736 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
737 rdev = rcu_dereference(conf->mirrors[disk].rdev);
739 test_bit(Faulty, &rdev->flags))
741 if (!test_bit(In_sync, &rdev->flags) &&
742 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
745 dev_sector = r10_bio->devs[slot].addr;
746 if (is_badblock(rdev, dev_sector, sectors,
747 &first_bad, &bad_sectors)) {
748 if (best_dist < MaxSector)
749 /* Already have a better slot */
751 if (first_bad <= dev_sector) {
752 /* Cannot read here. If this is the
753 * 'primary' device, then we must not read
754 * beyond 'bad_sectors' from another device.
756 bad_sectors -= (dev_sector - first_bad);
757 if (!do_balance && sectors > bad_sectors)
758 sectors = bad_sectors;
759 if (best_good_sectors > sectors)
760 best_good_sectors = sectors;
762 sector_t good_sectors =
763 first_bad - dev_sector;
764 if (good_sectors > best_good_sectors) {
765 best_good_sectors = good_sectors;
770 /* Must read from here */
775 best_good_sectors = sectors;
780 /* This optimisation is debatable, and completely destroys
781 * sequential read speed for 'far copies' arrays. So only
782 * keep it for 'near' arrays, and review those later.
784 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
787 /* for far > 1 always use the lowest address */
788 if (geo->far_copies > 1)
789 new_distance = r10_bio->devs[slot].addr;
791 new_distance = abs(r10_bio->devs[slot].addr -
792 conf->mirrors[disk].head_position);
793 if (new_distance < best_dist) {
794 best_dist = new_distance;
799 if (slot >= conf->copies) {
805 atomic_inc(&rdev->nr_pending);
806 r10_bio->read_slot = slot;
810 *max_sectors = best_good_sectors;
815 static int raid10_congested(struct mddev *mddev, int bits)
817 struct r10conf *conf = mddev->private;
820 if ((bits & (1 << WB_async_congested)) &&
821 conf->pending_count >= max_queued_requests)
826 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
829 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
830 if (rdev && !test_bit(Faulty, &rdev->flags)) {
831 struct request_queue *q = bdev_get_queue(rdev->bdev);
833 ret |= bdi_congested(&q->backing_dev_info, bits);
840 static void flush_pending_writes(struct r10conf *conf)
842 /* Any writes that have been queued but are awaiting
843 * bitmap updates get flushed here.
845 spin_lock_irq(&conf->device_lock);
847 if (conf->pending_bio_list.head) {
849 bio = bio_list_get(&conf->pending_bio_list);
850 conf->pending_count = 0;
851 spin_unlock_irq(&conf->device_lock);
852 /* flush any pending bitmap writes to disk
853 * before proceeding w/ I/O */
854 bitmap_unplug(conf->mddev->bitmap);
855 wake_up(&conf->wait_barrier);
857 while (bio) { /* submit pending writes */
858 struct bio *next = bio->bi_next;
860 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
861 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
865 generic_make_request(bio);
869 spin_unlock_irq(&conf->device_lock);
873 * Sometimes we need to suspend IO while we do something else,
874 * either some resync/recovery, or reconfigure the array.
875 * To do this we raise a 'barrier'.
876 * The 'barrier' is a counter that can be raised multiple times
877 * to count how many activities are happening which preclude
879 * We can only raise the barrier if there is no pending IO.
880 * i.e. if nr_pending == 0.
881 * We choose only to raise the barrier if no-one is waiting for the
882 * barrier to go down. This means that as soon as an IO request
883 * is ready, no other operations which require a barrier will start
884 * until the IO request has had a chance.
886 * So: regular IO calls 'wait_barrier'. When that returns there
887 * is no backgroup IO happening, It must arrange to call
888 * allow_barrier when it has finished its IO.
889 * backgroup IO calls must call raise_barrier. Once that returns
890 * there is no normal IO happeing. It must arrange to call
891 * lower_barrier when the particular background IO completes.
894 static void raise_barrier(struct r10conf *conf, int force)
896 BUG_ON(force && !conf->barrier);
897 spin_lock_irq(&conf->resync_lock);
899 /* Wait until no block IO is waiting (unless 'force') */
900 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
903 /* block any new IO from starting */
906 /* Now wait for all pending IO to complete */
907 wait_event_lock_irq(conf->wait_barrier,
908 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
911 spin_unlock_irq(&conf->resync_lock);
914 static void lower_barrier(struct r10conf *conf)
917 spin_lock_irqsave(&conf->resync_lock, flags);
919 spin_unlock_irqrestore(&conf->resync_lock, flags);
920 wake_up(&conf->wait_barrier);
923 static void wait_barrier(struct r10conf *conf)
925 spin_lock_irq(&conf->resync_lock);
928 /* Wait for the barrier to drop.
929 * However if there are already pending
930 * requests (preventing the barrier from
931 * rising completely), and the
932 * pre-process bio queue isn't empty,
933 * then don't wait, as we need to empty
934 * that queue to get the nr_pending
937 wait_event_lock_irq(conf->wait_barrier,
941 !bio_list_empty(current->bio_list)),
946 spin_unlock_irq(&conf->resync_lock);
949 static void allow_barrier(struct r10conf *conf)
952 spin_lock_irqsave(&conf->resync_lock, flags);
954 spin_unlock_irqrestore(&conf->resync_lock, flags);
955 wake_up(&conf->wait_barrier);
958 static void freeze_array(struct r10conf *conf, int extra)
960 /* stop syncio and normal IO and wait for everything to
962 * We increment barrier and nr_waiting, and then
963 * wait until nr_pending match nr_queued+extra
964 * This is called in the context of one normal IO request
965 * that has failed. Thus any sync request that might be pending
966 * will be blocked by nr_pending, and we need to wait for
967 * pending IO requests to complete or be queued for re-try.
968 * Thus the number queued (nr_queued) plus this request (extra)
969 * must match the number of pending IOs (nr_pending) before
972 spin_lock_irq(&conf->resync_lock);
975 wait_event_lock_irq_cmd(conf->wait_barrier,
976 conf->nr_pending == conf->nr_queued+extra,
978 flush_pending_writes(conf));
980 spin_unlock_irq(&conf->resync_lock);
983 static void unfreeze_array(struct r10conf *conf)
985 /* reverse the effect of the freeze */
986 spin_lock_irq(&conf->resync_lock);
989 wake_up(&conf->wait_barrier);
990 spin_unlock_irq(&conf->resync_lock);
993 static sector_t choose_data_offset(struct r10bio *r10_bio,
994 struct md_rdev *rdev)
996 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
997 test_bit(R10BIO_Previous, &r10_bio->state))
998 return rdev->data_offset;
1000 return rdev->new_data_offset;
1003 struct raid10_plug_cb {
1004 struct blk_plug_cb cb;
1005 struct bio_list pending;
1009 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1011 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1013 struct mddev *mddev = plug->cb.data;
1014 struct r10conf *conf = mddev->private;
1017 if (from_schedule || current->bio_list) {
1018 spin_lock_irq(&conf->device_lock);
1019 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1020 conf->pending_count += plug->pending_cnt;
1021 spin_unlock_irq(&conf->device_lock);
1022 wake_up(&conf->wait_barrier);
1023 md_wakeup_thread(mddev->thread);
1028 /* we aren't scheduling, so we can do the write-out directly. */
1029 bio = bio_list_get(&plug->pending);
1030 bitmap_unplug(mddev->bitmap);
1031 wake_up(&conf->wait_barrier);
1033 while (bio) { /* submit pending writes */
1034 struct bio *next = bio->bi_next;
1035 bio->bi_next = NULL;
1036 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1037 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1038 /* Just ignore it */
1041 generic_make_request(bio);
1047 static void __make_request(struct mddev *mddev, struct bio *bio)
1049 struct r10conf *conf = mddev->private;
1050 struct r10bio *r10_bio;
1051 struct bio *read_bio;
1053 const int rw = bio_data_dir(bio);
1054 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1055 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1056 const unsigned long do_discard = (bio->bi_rw
1057 & (REQ_DISCARD | REQ_SECURE));
1058 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1059 unsigned long flags;
1060 struct md_rdev *blocked_rdev;
1061 struct blk_plug_cb *cb;
1062 struct raid10_plug_cb *plug = NULL;
1063 int sectors_handled;
1068 * Register the new request and wait if the reconstruction
1069 * thread has put up a bar for new requests.
1070 * Continue immediately if no resync is active currently.
1074 sectors = bio_sectors(bio);
1075 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1076 bio->bi_iter.bi_sector < conf->reshape_progress &&
1077 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1078 /* IO spans the reshape position. Need to wait for
1081 allow_barrier(conf);
1082 wait_event(conf->wait_barrier,
1083 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1084 conf->reshape_progress >= bio->bi_iter.bi_sector +
1088 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1089 bio_data_dir(bio) == WRITE &&
1090 (mddev->reshape_backwards
1091 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1092 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1093 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1094 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1095 /* Need to update reshape_position in metadata */
1096 mddev->reshape_position = conf->reshape_progress;
1097 set_mask_bits(&mddev->flags, 0,
1098 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1099 md_wakeup_thread(mddev->thread);
1100 wait_event(mddev->sb_wait,
1101 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1103 conf->reshape_safe = mddev->reshape_position;
1106 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1108 r10_bio->master_bio = bio;
1109 r10_bio->sectors = sectors;
1111 r10_bio->mddev = mddev;
1112 r10_bio->sector = bio->bi_iter.bi_sector;
1115 /* We might need to issue multiple reads to different
1116 * devices if there are bad blocks around, so we keep
1117 * track of the number of reads in bio->bi_phys_segments.
1118 * If this is 0, there is only one r10_bio and no locking
1119 * will be needed when the request completes. If it is
1120 * non-zero, then it is the number of not-completed requests.
1122 bio->bi_phys_segments = 0;
1123 bio_clear_flag(bio, BIO_SEG_VALID);
1127 * read balancing logic:
1129 struct md_rdev *rdev;
1133 rdev = read_balance(conf, r10_bio, &max_sectors);
1135 raid_end_bio_io(r10_bio);
1138 slot = r10_bio->read_slot;
1140 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1141 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1144 r10_bio->devs[slot].bio = read_bio;
1145 r10_bio->devs[slot].rdev = rdev;
1147 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1148 choose_data_offset(r10_bio, rdev);
1149 read_bio->bi_bdev = rdev->bdev;
1150 read_bio->bi_end_io = raid10_end_read_request;
1151 read_bio->bi_rw = READ | do_sync;
1152 read_bio->bi_private = r10_bio;
1154 if (max_sectors < r10_bio->sectors) {
1155 /* Could not read all from this device, so we will
1156 * need another r10_bio.
1158 sectors_handled = (r10_bio->sector + max_sectors
1159 - bio->bi_iter.bi_sector);
1160 r10_bio->sectors = max_sectors;
1161 spin_lock_irq(&conf->device_lock);
1162 if (bio->bi_phys_segments == 0)
1163 bio->bi_phys_segments = 2;
1165 bio->bi_phys_segments++;
1166 spin_unlock_irq(&conf->device_lock);
1167 /* Cannot call generic_make_request directly
1168 * as that will be queued in __generic_make_request
1169 * and subsequent mempool_alloc might block
1170 * waiting for it. so hand bio over to raid10d.
1172 reschedule_retry(r10_bio);
1174 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1176 r10_bio->master_bio = bio;
1177 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1179 r10_bio->mddev = mddev;
1180 r10_bio->sector = bio->bi_iter.bi_sector +
1184 generic_make_request(read_bio);
1191 if (conf->pending_count >= max_queued_requests) {
1192 md_wakeup_thread(mddev->thread);
1193 wait_event(conf->wait_barrier,
1194 conf->pending_count < max_queued_requests);
1196 /* first select target devices under rcu_lock and
1197 * inc refcount on their rdev. Record them by setting
1199 * If there are known/acknowledged bad blocks on any device
1200 * on which we have seen a write error, we want to avoid
1201 * writing to those blocks. This potentially requires several
1202 * writes to write around the bad blocks. Each set of writes
1203 * gets its own r10_bio with a set of bios attached. The number
1204 * of r10_bios is recored in bio->bi_phys_segments just as with
1208 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1209 raid10_find_phys(conf, r10_bio);
1211 blocked_rdev = NULL;
1213 max_sectors = r10_bio->sectors;
1215 for (i = 0; i < conf->copies; i++) {
1216 int d = r10_bio->devs[i].devnum;
1217 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1218 struct md_rdev *rrdev = rcu_dereference(
1219 conf->mirrors[d].replacement);
1222 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1223 atomic_inc(&rdev->nr_pending);
1224 blocked_rdev = rdev;
1227 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1228 atomic_inc(&rrdev->nr_pending);
1229 blocked_rdev = rrdev;
1232 if (rdev && (test_bit(Faulty, &rdev->flags)))
1234 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1237 r10_bio->devs[i].bio = NULL;
1238 r10_bio->devs[i].repl_bio = NULL;
1240 if (!rdev && !rrdev) {
1241 set_bit(R10BIO_Degraded, &r10_bio->state);
1244 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1246 sector_t dev_sector = r10_bio->devs[i].addr;
1250 is_bad = is_badblock(rdev, dev_sector,
1252 &first_bad, &bad_sectors);
1254 /* Mustn't write here until the bad block
1257 atomic_inc(&rdev->nr_pending);
1258 set_bit(BlockedBadBlocks, &rdev->flags);
1259 blocked_rdev = rdev;
1262 if (is_bad && first_bad <= dev_sector) {
1263 /* Cannot write here at all */
1264 bad_sectors -= (dev_sector - first_bad);
1265 if (bad_sectors < max_sectors)
1266 /* Mustn't write more than bad_sectors
1267 * to other devices yet
1269 max_sectors = bad_sectors;
1270 /* We don't set R10BIO_Degraded as that
1271 * only applies if the disk is missing,
1272 * so it might be re-added, and we want to
1273 * know to recover this chunk.
1274 * In this case the device is here, and the
1275 * fact that this chunk is not in-sync is
1276 * recorded in the bad block log.
1281 int good_sectors = first_bad - dev_sector;
1282 if (good_sectors < max_sectors)
1283 max_sectors = good_sectors;
1287 r10_bio->devs[i].bio = bio;
1288 atomic_inc(&rdev->nr_pending);
1291 r10_bio->devs[i].repl_bio = bio;
1292 atomic_inc(&rrdev->nr_pending);
1297 if (unlikely(blocked_rdev)) {
1298 /* Have to wait for this device to get unblocked, then retry */
1302 for (j = 0; j < i; j++) {
1303 if (r10_bio->devs[j].bio) {
1304 d = r10_bio->devs[j].devnum;
1305 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1307 if (r10_bio->devs[j].repl_bio) {
1308 struct md_rdev *rdev;
1309 d = r10_bio->devs[j].devnum;
1310 rdev = conf->mirrors[d].replacement;
1312 /* Race with remove_disk */
1314 rdev = conf->mirrors[d].rdev;
1316 rdev_dec_pending(rdev, mddev);
1319 allow_barrier(conf);
1320 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1325 if (max_sectors < r10_bio->sectors) {
1326 /* We are splitting this into multiple parts, so
1327 * we need to prepare for allocating another r10_bio.
1329 r10_bio->sectors = max_sectors;
1330 spin_lock_irq(&conf->device_lock);
1331 if (bio->bi_phys_segments == 0)
1332 bio->bi_phys_segments = 2;
1334 bio->bi_phys_segments++;
1335 spin_unlock_irq(&conf->device_lock);
1337 sectors_handled = r10_bio->sector + max_sectors -
1338 bio->bi_iter.bi_sector;
1340 atomic_set(&r10_bio->remaining, 1);
1341 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1343 for (i = 0; i < conf->copies; i++) {
1345 int d = r10_bio->devs[i].devnum;
1346 if (r10_bio->devs[i].bio) {
1347 struct md_rdev *rdev = conf->mirrors[d].rdev;
1348 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1349 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1351 r10_bio->devs[i].bio = mbio;
1353 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1354 choose_data_offset(r10_bio,
1356 mbio->bi_bdev = rdev->bdev;
1357 mbio->bi_end_io = raid10_end_write_request;
1359 WRITE | do_sync | do_fua | do_discard | do_same;
1360 mbio->bi_private = r10_bio;
1362 atomic_inc(&r10_bio->remaining);
1364 cb = blk_check_plugged(raid10_unplug, mddev,
1367 plug = container_of(cb, struct raid10_plug_cb,
1371 spin_lock_irqsave(&conf->device_lock, flags);
1373 bio_list_add(&plug->pending, mbio);
1374 plug->pending_cnt++;
1376 bio_list_add(&conf->pending_bio_list, mbio);
1377 conf->pending_count++;
1379 spin_unlock_irqrestore(&conf->device_lock, flags);
1381 md_wakeup_thread(mddev->thread);
1384 if (r10_bio->devs[i].repl_bio) {
1385 struct md_rdev *rdev = conf->mirrors[d].replacement;
1387 /* Replacement just got moved to main 'rdev' */
1389 rdev = conf->mirrors[d].rdev;
1391 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1392 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1394 r10_bio->devs[i].repl_bio = mbio;
1396 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1399 mbio->bi_bdev = rdev->bdev;
1400 mbio->bi_end_io = raid10_end_write_request;
1402 WRITE | do_sync | do_fua | do_discard | do_same;
1403 mbio->bi_private = r10_bio;
1405 atomic_inc(&r10_bio->remaining);
1406 spin_lock_irqsave(&conf->device_lock, flags);
1407 bio_list_add(&conf->pending_bio_list, mbio);
1408 conf->pending_count++;
1409 spin_unlock_irqrestore(&conf->device_lock, flags);
1410 if (!mddev_check_plugged(mddev))
1411 md_wakeup_thread(mddev->thread);
1415 /* Don't remove the bias on 'remaining' (one_write_done) until
1416 * after checking if we need to go around again.
1419 if (sectors_handled < bio_sectors(bio)) {
1420 one_write_done(r10_bio);
1421 /* We need another r10_bio. It has already been counted
1422 * in bio->bi_phys_segments.
1424 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1426 r10_bio->master_bio = bio;
1427 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1429 r10_bio->mddev = mddev;
1430 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1434 one_write_done(r10_bio);
1437 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1439 struct r10conf *conf = mddev->private;
1440 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1441 int chunk_sects = chunk_mask + 1;
1445 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1446 md_flush_request(mddev, bio);
1450 md_write_start(mddev, bio);
1455 * If this request crosses a chunk boundary, we need to split
1458 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1459 bio_sectors(bio) > chunk_sects
1460 && (conf->geo.near_copies < conf->geo.raid_disks
1461 || conf->prev.near_copies <
1462 conf->prev.raid_disks))) {
1463 split = bio_split(bio, chunk_sects -
1464 (bio->bi_iter.bi_sector &
1466 GFP_NOIO, fs_bio_set);
1467 bio_chain(split, bio);
1472 __make_request(mddev, split);
1473 } while (split != bio);
1475 /* In case raid10d snuck in to freeze_array */
1476 wake_up(&conf->wait_barrier);
1479 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1481 struct r10conf *conf = mddev->private;
1484 if (conf->geo.near_copies < conf->geo.raid_disks)
1485 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1486 if (conf->geo.near_copies > 1)
1487 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1488 if (conf->geo.far_copies > 1) {
1489 if (conf->geo.far_offset)
1490 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1492 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1493 if (conf->geo.far_set_size != conf->geo.raid_disks)
1494 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1496 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1497 conf->geo.raid_disks - mddev->degraded);
1499 for (i = 0; i < conf->geo.raid_disks; i++) {
1500 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1501 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1504 seq_printf(seq, "]");
1507 /* check if there are enough drives for
1508 * every block to appear on atleast one.
1509 * Don't consider the device numbered 'ignore'
1510 * as we might be about to remove it.
1512 static int _enough(struct r10conf *conf, int previous, int ignore)
1518 disks = conf->prev.raid_disks;
1519 ncopies = conf->prev.near_copies;
1521 disks = conf->geo.raid_disks;
1522 ncopies = conf->geo.near_copies;
1527 int n = conf->copies;
1531 struct md_rdev *rdev;
1532 if (this != ignore &&
1533 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1534 test_bit(In_sync, &rdev->flags))
1536 this = (this+1) % disks;
1540 first = (first + ncopies) % disks;
1541 } while (first != 0);
1548 static int enough(struct r10conf *conf, int ignore)
1550 /* when calling 'enough', both 'prev' and 'geo' must
1552 * This is ensured if ->reconfig_mutex or ->device_lock
1555 return _enough(conf, 0, ignore) &&
1556 _enough(conf, 1, ignore);
1559 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1561 char b[BDEVNAME_SIZE];
1562 struct r10conf *conf = mddev->private;
1563 unsigned long flags;
1566 * If it is not operational, then we have already marked it as dead
1567 * else if it is the last working disks, ignore the error, let the
1568 * next level up know.
1569 * else mark the drive as failed
1571 spin_lock_irqsave(&conf->device_lock, flags);
1572 if (test_bit(In_sync, &rdev->flags)
1573 && !enough(conf, rdev->raid_disk)) {
1575 * Don't fail the drive, just return an IO error.
1577 spin_unlock_irqrestore(&conf->device_lock, flags);
1580 if (test_and_clear_bit(In_sync, &rdev->flags))
1583 * If recovery is running, make sure it aborts.
1585 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1586 set_bit(Blocked, &rdev->flags);
1587 set_bit(Faulty, &rdev->flags);
1588 set_mask_bits(&mddev->flags, 0,
1589 BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1590 spin_unlock_irqrestore(&conf->device_lock, flags);
1592 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1593 "md/raid10:%s: Operation continuing on %d devices.\n",
1594 mdname(mddev), bdevname(rdev->bdev, b),
1595 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1598 static void print_conf(struct r10conf *conf)
1601 struct raid10_info *tmp;
1603 printk(KERN_DEBUG "RAID10 conf printout:\n");
1605 printk(KERN_DEBUG "(!conf)\n");
1608 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1609 conf->geo.raid_disks);
1611 for (i = 0; i < conf->geo.raid_disks; i++) {
1612 char b[BDEVNAME_SIZE];
1613 tmp = conf->mirrors + i;
1615 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1616 i, !test_bit(In_sync, &tmp->rdev->flags),
1617 !test_bit(Faulty, &tmp->rdev->flags),
1618 bdevname(tmp->rdev->bdev,b));
1622 static void close_sync(struct r10conf *conf)
1625 allow_barrier(conf);
1627 mempool_destroy(conf->r10buf_pool);
1628 conf->r10buf_pool = NULL;
1631 static int raid10_spare_active(struct mddev *mddev)
1634 struct r10conf *conf = mddev->private;
1635 struct raid10_info *tmp;
1637 unsigned long flags;
1640 * Find all non-in_sync disks within the RAID10 configuration
1641 * and mark them in_sync
1643 for (i = 0; i < conf->geo.raid_disks; i++) {
1644 tmp = conf->mirrors + i;
1645 if (tmp->replacement
1646 && tmp->replacement->recovery_offset == MaxSector
1647 && !test_bit(Faulty, &tmp->replacement->flags)
1648 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1649 /* Replacement has just become active */
1651 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1654 /* Replaced device not technically faulty,
1655 * but we need to be sure it gets removed
1656 * and never re-added.
1658 set_bit(Faulty, &tmp->rdev->flags);
1659 sysfs_notify_dirent_safe(
1660 tmp->rdev->sysfs_state);
1662 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1663 } else if (tmp->rdev
1664 && tmp->rdev->recovery_offset == MaxSector
1665 && !test_bit(Faulty, &tmp->rdev->flags)
1666 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1668 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1671 spin_lock_irqsave(&conf->device_lock, flags);
1672 mddev->degraded -= count;
1673 spin_unlock_irqrestore(&conf->device_lock, flags);
1679 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1681 struct r10conf *conf = mddev->private;
1685 int last = conf->geo.raid_disks - 1;
1687 if (mddev->recovery_cp < MaxSector)
1688 /* only hot-add to in-sync arrays, as recovery is
1689 * very different from resync
1692 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1695 if (md_integrity_add_rdev(rdev, mddev))
1698 if (rdev->raid_disk >= 0)
1699 first = last = rdev->raid_disk;
1701 if (rdev->saved_raid_disk >= first &&
1702 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1703 mirror = rdev->saved_raid_disk;
1706 for ( ; mirror <= last ; mirror++) {
1707 struct raid10_info *p = &conf->mirrors[mirror];
1708 if (p->recovery_disabled == mddev->recovery_disabled)
1711 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1712 p->replacement != NULL)
1714 clear_bit(In_sync, &rdev->flags);
1715 set_bit(Replacement, &rdev->flags);
1716 rdev->raid_disk = mirror;
1719 disk_stack_limits(mddev->gendisk, rdev->bdev,
1720 rdev->data_offset << 9);
1722 rcu_assign_pointer(p->replacement, rdev);
1727 disk_stack_limits(mddev->gendisk, rdev->bdev,
1728 rdev->data_offset << 9);
1730 p->head_position = 0;
1731 p->recovery_disabled = mddev->recovery_disabled - 1;
1732 rdev->raid_disk = mirror;
1734 if (rdev->saved_raid_disk != mirror)
1736 rcu_assign_pointer(p->rdev, rdev);
1739 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1740 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1746 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1748 struct r10conf *conf = mddev->private;
1750 int number = rdev->raid_disk;
1751 struct md_rdev **rdevp;
1752 struct raid10_info *p = conf->mirrors + number;
1755 if (rdev == p->rdev)
1757 else if (rdev == p->replacement)
1758 rdevp = &p->replacement;
1762 if (test_bit(In_sync, &rdev->flags) ||
1763 atomic_read(&rdev->nr_pending)) {
1767 /* Only remove faulty devices if recovery
1770 if (!test_bit(Faulty, &rdev->flags) &&
1771 mddev->recovery_disabled != p->recovery_disabled &&
1772 (!p->replacement || p->replacement == rdev) &&
1773 number < conf->geo.raid_disks &&
1780 if (atomic_read(&rdev->nr_pending)) {
1781 /* lost the race, try later */
1785 } else if (p->replacement) {
1786 /* We must have just cleared 'rdev' */
1787 p->rdev = p->replacement;
1788 clear_bit(Replacement, &p->replacement->flags);
1789 smp_mb(); /* Make sure other CPUs may see both as identical
1790 * but will never see neither -- if they are careful.
1792 p->replacement = NULL;
1793 clear_bit(WantReplacement, &rdev->flags);
1795 /* We might have just remove the Replacement as faulty
1796 * Clear the flag just in case
1798 clear_bit(WantReplacement, &rdev->flags);
1800 err = md_integrity_register(mddev);
1808 static void end_sync_read(struct bio *bio)
1810 struct r10bio *r10_bio = bio->bi_private;
1811 struct r10conf *conf = r10_bio->mddev->private;
1814 if (bio == r10_bio->master_bio) {
1815 /* this is a reshape read */
1816 d = r10_bio->read_slot; /* really the read dev */
1818 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1821 set_bit(R10BIO_Uptodate, &r10_bio->state);
1823 /* The write handler will notice the lack of
1824 * R10BIO_Uptodate and record any errors etc
1826 atomic_add(r10_bio->sectors,
1827 &conf->mirrors[d].rdev->corrected_errors);
1829 /* for reconstruct, we always reschedule after a read.
1830 * for resync, only after all reads
1832 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1833 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1834 atomic_dec_and_test(&r10_bio->remaining)) {
1835 /* we have read all the blocks,
1836 * do the comparison in process context in raid10d
1838 reschedule_retry(r10_bio);
1842 static void end_sync_request(struct r10bio *r10_bio)
1844 struct mddev *mddev = r10_bio->mddev;
1846 while (atomic_dec_and_test(&r10_bio->remaining)) {
1847 if (r10_bio->master_bio == NULL) {
1848 /* the primary of several recovery bios */
1849 sector_t s = r10_bio->sectors;
1850 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1851 test_bit(R10BIO_WriteError, &r10_bio->state))
1852 reschedule_retry(r10_bio);
1855 md_done_sync(mddev, s, 1);
1858 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1859 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1860 test_bit(R10BIO_WriteError, &r10_bio->state))
1861 reschedule_retry(r10_bio);
1869 static void end_sync_write(struct bio *bio)
1871 struct r10bio *r10_bio = bio->bi_private;
1872 struct mddev *mddev = r10_bio->mddev;
1873 struct r10conf *conf = mddev->private;
1879 struct md_rdev *rdev = NULL;
1881 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1883 rdev = conf->mirrors[d].replacement;
1885 rdev = conf->mirrors[d].rdev;
1887 if (bio->bi_error) {
1889 md_error(mddev, rdev);
1891 set_bit(WriteErrorSeen, &rdev->flags);
1892 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1893 set_bit(MD_RECOVERY_NEEDED,
1894 &rdev->mddev->recovery);
1895 set_bit(R10BIO_WriteError, &r10_bio->state);
1897 } else if (is_badblock(rdev,
1898 r10_bio->devs[slot].addr,
1900 &first_bad, &bad_sectors))
1901 set_bit(R10BIO_MadeGood, &r10_bio->state);
1903 rdev_dec_pending(rdev, mddev);
1905 end_sync_request(r10_bio);
1909 * Note: sync and recover and handled very differently for raid10
1910 * This code is for resync.
1911 * For resync, we read through virtual addresses and read all blocks.
1912 * If there is any error, we schedule a write. The lowest numbered
1913 * drive is authoritative.
1914 * However requests come for physical address, so we need to map.
1915 * For every physical address there are raid_disks/copies virtual addresses,
1916 * which is always are least one, but is not necessarly an integer.
1917 * This means that a physical address can span multiple chunks, so we may
1918 * have to submit multiple io requests for a single sync request.
1921 * We check if all blocks are in-sync and only write to blocks that
1924 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1926 struct r10conf *conf = mddev->private;
1928 struct bio *tbio, *fbio;
1931 atomic_set(&r10_bio->remaining, 1);
1933 /* find the first device with a block */
1934 for (i=0; i<conf->copies; i++)
1935 if (!r10_bio->devs[i].bio->bi_error)
1938 if (i == conf->copies)
1942 fbio = r10_bio->devs[i].bio;
1943 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1944 fbio->bi_iter.bi_idx = 0;
1946 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1947 /* now find blocks with errors */
1948 for (i=0 ; i < conf->copies ; i++) {
1951 tbio = r10_bio->devs[i].bio;
1953 if (tbio->bi_end_io != end_sync_read)
1957 if (!r10_bio->devs[i].bio->bi_error) {
1958 /* We know that the bi_io_vec layout is the same for
1959 * both 'first' and 'i', so we just compare them.
1960 * All vec entries are PAGE_SIZE;
1962 int sectors = r10_bio->sectors;
1963 for (j = 0; j < vcnt; j++) {
1964 int len = PAGE_SIZE;
1965 if (sectors < (len / 512))
1966 len = sectors * 512;
1967 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1968 page_address(tbio->bi_io_vec[j].bv_page),
1975 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1976 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1977 /* Don't fix anything. */
1980 /* Ok, we need to write this bio, either to correct an
1981 * inconsistency or to correct an unreadable block.
1982 * First we need to fixup bv_offset, bv_len and
1983 * bi_vecs, as the read request might have corrupted these
1987 tbio->bi_vcnt = vcnt;
1988 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
1989 tbio->bi_rw = WRITE;
1990 tbio->bi_private = r10_bio;
1991 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1992 tbio->bi_end_io = end_sync_write;
1994 bio_copy_data(tbio, fbio);
1996 d = r10_bio->devs[i].devnum;
1997 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1998 atomic_inc(&r10_bio->remaining);
1999 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2001 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2002 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2003 generic_make_request(tbio);
2006 /* Now write out to any replacement devices
2009 for (i = 0; i < conf->copies; i++) {
2012 tbio = r10_bio->devs[i].repl_bio;
2013 if (!tbio || !tbio->bi_end_io)
2015 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2016 && r10_bio->devs[i].bio != fbio)
2017 bio_copy_data(tbio, fbio);
2018 d = r10_bio->devs[i].devnum;
2019 atomic_inc(&r10_bio->remaining);
2020 md_sync_acct(conf->mirrors[d].replacement->bdev,
2022 generic_make_request(tbio);
2026 if (atomic_dec_and_test(&r10_bio->remaining)) {
2027 md_done_sync(mddev, r10_bio->sectors, 1);
2033 * Now for the recovery code.
2034 * Recovery happens across physical sectors.
2035 * We recover all non-is_sync drives by finding the virtual address of
2036 * each, and then choose a working drive that also has that virt address.
2037 * There is a separate r10_bio for each non-in_sync drive.
2038 * Only the first two slots are in use. The first for reading,
2039 * The second for writing.
2042 static void fix_recovery_read_error(struct r10bio *r10_bio)
2044 /* We got a read error during recovery.
2045 * We repeat the read in smaller page-sized sections.
2046 * If a read succeeds, write it to the new device or record
2047 * a bad block if we cannot.
2048 * If a read fails, record a bad block on both old and
2051 struct mddev *mddev = r10_bio->mddev;
2052 struct r10conf *conf = mddev->private;
2053 struct bio *bio = r10_bio->devs[0].bio;
2055 int sectors = r10_bio->sectors;
2057 int dr = r10_bio->devs[0].devnum;
2058 int dw = r10_bio->devs[1].devnum;
2062 struct md_rdev *rdev;
2066 if (s > (PAGE_SIZE>>9))
2069 rdev = conf->mirrors[dr].rdev;
2070 addr = r10_bio->devs[0].addr + sect,
2071 ok = sync_page_io(rdev,
2074 bio->bi_io_vec[idx].bv_page,
2077 rdev = conf->mirrors[dw].rdev;
2078 addr = r10_bio->devs[1].addr + sect;
2079 ok = sync_page_io(rdev,
2082 bio->bi_io_vec[idx].bv_page,
2085 set_bit(WriteErrorSeen, &rdev->flags);
2086 if (!test_and_set_bit(WantReplacement,
2088 set_bit(MD_RECOVERY_NEEDED,
2089 &rdev->mddev->recovery);
2093 /* We don't worry if we cannot set a bad block -
2094 * it really is bad so there is no loss in not
2097 rdev_set_badblocks(rdev, addr, s, 0);
2099 if (rdev != conf->mirrors[dw].rdev) {
2100 /* need bad block on destination too */
2101 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2102 addr = r10_bio->devs[1].addr + sect;
2103 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2105 /* just abort the recovery */
2107 "md/raid10:%s: recovery aborted"
2108 " due to read error\n",
2111 conf->mirrors[dw].recovery_disabled
2112 = mddev->recovery_disabled;
2113 set_bit(MD_RECOVERY_INTR,
2126 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2128 struct r10conf *conf = mddev->private;
2130 struct bio *wbio, *wbio2;
2132 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2133 fix_recovery_read_error(r10_bio);
2134 end_sync_request(r10_bio);
2139 * share the pages with the first bio
2140 * and submit the write request
2142 d = r10_bio->devs[1].devnum;
2143 wbio = r10_bio->devs[1].bio;
2144 wbio2 = r10_bio->devs[1].repl_bio;
2145 /* Need to test wbio2->bi_end_io before we call
2146 * generic_make_request as if the former is NULL,
2147 * the latter is free to free wbio2.
2149 if (wbio2 && !wbio2->bi_end_io)
2151 if (wbio->bi_end_io) {
2152 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2153 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2154 generic_make_request(wbio);
2157 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2158 md_sync_acct(conf->mirrors[d].replacement->bdev,
2159 bio_sectors(wbio2));
2160 generic_make_request(wbio2);
2165 * Used by fix_read_error() to decay the per rdev read_errors.
2166 * We halve the read error count for every hour that has elapsed
2167 * since the last recorded read error.
2170 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2172 struct timespec cur_time_mon;
2173 unsigned long hours_since_last;
2174 unsigned int read_errors = atomic_read(&rdev->read_errors);
2176 ktime_get_ts(&cur_time_mon);
2178 if (rdev->last_read_error.tv_sec == 0 &&
2179 rdev->last_read_error.tv_nsec == 0) {
2180 /* first time we've seen a read error */
2181 rdev->last_read_error = cur_time_mon;
2185 hours_since_last = (cur_time_mon.tv_sec -
2186 rdev->last_read_error.tv_sec) / 3600;
2188 rdev->last_read_error = cur_time_mon;
2191 * if hours_since_last is > the number of bits in read_errors
2192 * just set read errors to 0. We do this to avoid
2193 * overflowing the shift of read_errors by hours_since_last.
2195 if (hours_since_last >= 8 * sizeof(read_errors))
2196 atomic_set(&rdev->read_errors, 0);
2198 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2201 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2202 int sectors, struct page *page, int rw)
2207 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2208 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2210 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2214 set_bit(WriteErrorSeen, &rdev->flags);
2215 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2216 set_bit(MD_RECOVERY_NEEDED,
2217 &rdev->mddev->recovery);
2219 /* need to record an error - either for the block or the device */
2220 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2221 md_error(rdev->mddev, rdev);
2226 * This is a kernel thread which:
2228 * 1. Retries failed read operations on working mirrors.
2229 * 2. Updates the raid superblock when problems encounter.
2230 * 3. Performs writes following reads for array synchronising.
2233 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2235 int sect = 0; /* Offset from r10_bio->sector */
2236 int sectors = r10_bio->sectors;
2237 struct md_rdev*rdev;
2238 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2239 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2241 /* still own a reference to this rdev, so it cannot
2242 * have been cleared recently.
2244 rdev = conf->mirrors[d].rdev;
2246 if (test_bit(Faulty, &rdev->flags))
2247 /* drive has already been failed, just ignore any
2248 more fix_read_error() attempts */
2251 check_decay_read_errors(mddev, rdev);
2252 atomic_inc(&rdev->read_errors);
2253 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2254 char b[BDEVNAME_SIZE];
2255 bdevname(rdev->bdev, b);
2258 "md/raid10:%s: %s: Raid device exceeded "
2259 "read_error threshold [cur %d:max %d]\n",
2261 atomic_read(&rdev->read_errors), max_read_errors);
2263 "md/raid10:%s: %s: Failing raid device\n",
2265 md_error(mddev, conf->mirrors[d].rdev);
2266 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2272 int sl = r10_bio->read_slot;
2276 if (s > (PAGE_SIZE>>9))
2284 d = r10_bio->devs[sl].devnum;
2285 rdev = rcu_dereference(conf->mirrors[d].rdev);
2287 test_bit(In_sync, &rdev->flags) &&
2288 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2289 &first_bad, &bad_sectors) == 0) {
2290 atomic_inc(&rdev->nr_pending);
2292 success = sync_page_io(rdev,
2293 r10_bio->devs[sl].addr +
2296 conf->tmppage, READ, false);
2297 rdev_dec_pending(rdev, mddev);
2303 if (sl == conf->copies)
2305 } while (!success && sl != r10_bio->read_slot);
2309 /* Cannot read from anywhere, just mark the block
2310 * as bad on the first device to discourage future
2313 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2314 rdev = conf->mirrors[dn].rdev;
2316 if (!rdev_set_badblocks(
2318 r10_bio->devs[r10_bio->read_slot].addr
2321 md_error(mddev, rdev);
2322 r10_bio->devs[r10_bio->read_slot].bio
2329 /* write it back and re-read */
2331 while (sl != r10_bio->read_slot) {
2332 char b[BDEVNAME_SIZE];
2337 d = r10_bio->devs[sl].devnum;
2338 rdev = rcu_dereference(conf->mirrors[d].rdev);
2340 !test_bit(In_sync, &rdev->flags))
2343 atomic_inc(&rdev->nr_pending);
2345 if (r10_sync_page_io(rdev,
2346 r10_bio->devs[sl].addr +
2348 s, conf->tmppage, WRITE)
2350 /* Well, this device is dead */
2352 "md/raid10:%s: read correction "
2354 " (%d sectors at %llu on %s)\n",
2356 (unsigned long long)(
2358 choose_data_offset(r10_bio,
2360 bdevname(rdev->bdev, b));
2361 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2364 bdevname(rdev->bdev, b));
2366 rdev_dec_pending(rdev, mddev);
2370 while (sl != r10_bio->read_slot) {
2371 char b[BDEVNAME_SIZE];
2376 d = r10_bio->devs[sl].devnum;
2377 rdev = rcu_dereference(conf->mirrors[d].rdev);
2379 !test_bit(In_sync, &rdev->flags))
2382 atomic_inc(&rdev->nr_pending);
2384 switch (r10_sync_page_io(rdev,
2385 r10_bio->devs[sl].addr +
2390 /* Well, this device is dead */
2392 "md/raid10:%s: unable to read back "
2394 " (%d sectors at %llu on %s)\n",
2396 (unsigned long long)(
2398 choose_data_offset(r10_bio, rdev)),
2399 bdevname(rdev->bdev, b));
2400 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2403 bdevname(rdev->bdev, b));
2407 "md/raid10:%s: read error corrected"
2408 " (%d sectors at %llu on %s)\n",
2410 (unsigned long long)(
2412 choose_data_offset(r10_bio, rdev)),
2413 bdevname(rdev->bdev, b));
2414 atomic_add(s, &rdev->corrected_errors);
2417 rdev_dec_pending(rdev, mddev);
2427 static int narrow_write_error(struct r10bio *r10_bio, int i)
2429 struct bio *bio = r10_bio->master_bio;
2430 struct mddev *mddev = r10_bio->mddev;
2431 struct r10conf *conf = mddev->private;
2432 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2433 /* bio has the data to be written to slot 'i' where
2434 * we just recently had a write error.
2435 * We repeatedly clone the bio and trim down to one block,
2436 * then try the write. Where the write fails we record
2438 * It is conceivable that the bio doesn't exactly align with
2439 * blocks. We must handle this.
2441 * We currently own a reference to the rdev.
2447 int sect_to_write = r10_bio->sectors;
2450 if (rdev->badblocks.shift < 0)
2453 block_sectors = roundup(1 << rdev->badblocks.shift,
2454 bdev_logical_block_size(rdev->bdev) >> 9);
2455 sector = r10_bio->sector;
2456 sectors = ((r10_bio->sector + block_sectors)
2457 & ~(sector_t)(block_sectors - 1))
2460 while (sect_to_write) {
2462 if (sectors > sect_to_write)
2463 sectors = sect_to_write;
2464 /* Write at 'sector' for 'sectors' */
2465 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2466 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2467 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2468 choose_data_offset(r10_bio, rdev) +
2469 (sector - r10_bio->sector));
2470 wbio->bi_bdev = rdev->bdev;
2471 if (submit_bio_wait(WRITE, wbio) < 0)
2473 ok = rdev_set_badblocks(rdev, sector,
2478 sect_to_write -= sectors;
2480 sectors = block_sectors;
2485 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2487 int slot = r10_bio->read_slot;
2489 struct r10conf *conf = mddev->private;
2490 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2491 char b[BDEVNAME_SIZE];
2492 unsigned long do_sync;
2495 /* we got a read error. Maybe the drive is bad. Maybe just
2496 * the block and we can fix it.
2497 * We freeze all other IO, and try reading the block from
2498 * other devices. When we find one, we re-write
2499 * and check it that fixes the read error.
2500 * This is all done synchronously while the array is
2503 bio = r10_bio->devs[slot].bio;
2504 bdevname(bio->bi_bdev, b);
2506 r10_bio->devs[slot].bio = NULL;
2508 if (mddev->ro == 0) {
2509 freeze_array(conf, 1);
2510 fix_read_error(conf, mddev, r10_bio);
2511 unfreeze_array(conf);
2513 r10_bio->devs[slot].bio = IO_BLOCKED;
2515 rdev_dec_pending(rdev, mddev);
2518 rdev = read_balance(conf, r10_bio, &max_sectors);
2520 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2521 " read error for block %llu\n",
2523 (unsigned long long)r10_bio->sector);
2524 raid_end_bio_io(r10_bio);
2528 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2529 slot = r10_bio->read_slot;
2532 "md/raid10:%s: %s: redirecting "
2533 "sector %llu to another mirror\n",
2535 bdevname(rdev->bdev, b),
2536 (unsigned long long)r10_bio->sector);
2537 bio = bio_clone_mddev(r10_bio->master_bio,
2539 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2540 r10_bio->devs[slot].bio = bio;
2541 r10_bio->devs[slot].rdev = rdev;
2542 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2543 + choose_data_offset(r10_bio, rdev);
2544 bio->bi_bdev = rdev->bdev;
2545 bio->bi_rw = READ | do_sync;
2546 bio->bi_private = r10_bio;
2547 bio->bi_end_io = raid10_end_read_request;
2548 if (max_sectors < r10_bio->sectors) {
2549 /* Drat - have to split this up more */
2550 struct bio *mbio = r10_bio->master_bio;
2551 int sectors_handled =
2552 r10_bio->sector + max_sectors
2553 - mbio->bi_iter.bi_sector;
2554 r10_bio->sectors = max_sectors;
2555 spin_lock_irq(&conf->device_lock);
2556 if (mbio->bi_phys_segments == 0)
2557 mbio->bi_phys_segments = 2;
2559 mbio->bi_phys_segments++;
2560 spin_unlock_irq(&conf->device_lock);
2561 generic_make_request(bio);
2563 r10_bio = mempool_alloc(conf->r10bio_pool,
2565 r10_bio->master_bio = mbio;
2566 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2568 set_bit(R10BIO_ReadError,
2570 r10_bio->mddev = mddev;
2571 r10_bio->sector = mbio->bi_iter.bi_sector
2576 generic_make_request(bio);
2579 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2581 /* Some sort of write request has finished and it
2582 * succeeded in writing where we thought there was a
2583 * bad block. So forget the bad block.
2584 * Or possibly if failed and we need to record
2588 struct md_rdev *rdev;
2590 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2591 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2592 for (m = 0; m < conf->copies; m++) {
2593 int dev = r10_bio->devs[m].devnum;
2594 rdev = conf->mirrors[dev].rdev;
2595 if (r10_bio->devs[m].bio == NULL)
2597 if (!r10_bio->devs[m].bio->bi_error) {
2598 rdev_clear_badblocks(
2600 r10_bio->devs[m].addr,
2601 r10_bio->sectors, 0);
2603 if (!rdev_set_badblocks(
2605 r10_bio->devs[m].addr,
2606 r10_bio->sectors, 0))
2607 md_error(conf->mddev, rdev);
2609 rdev = conf->mirrors[dev].replacement;
2610 if (r10_bio->devs[m].repl_bio == NULL)
2613 if (!r10_bio->devs[m].repl_bio->bi_error) {
2614 rdev_clear_badblocks(
2616 r10_bio->devs[m].addr,
2617 r10_bio->sectors, 0);
2619 if (!rdev_set_badblocks(
2621 r10_bio->devs[m].addr,
2622 r10_bio->sectors, 0))
2623 md_error(conf->mddev, rdev);
2629 for (m = 0; m < conf->copies; m++) {
2630 int dev = r10_bio->devs[m].devnum;
2631 struct bio *bio = r10_bio->devs[m].bio;
2632 rdev = conf->mirrors[dev].rdev;
2633 if (bio == IO_MADE_GOOD) {
2634 rdev_clear_badblocks(
2636 r10_bio->devs[m].addr,
2637 r10_bio->sectors, 0);
2638 rdev_dec_pending(rdev, conf->mddev);
2639 } else if (bio != NULL && bio->bi_error) {
2641 if (!narrow_write_error(r10_bio, m)) {
2642 md_error(conf->mddev, rdev);
2643 set_bit(R10BIO_Degraded,
2646 rdev_dec_pending(rdev, conf->mddev);
2648 bio = r10_bio->devs[m].repl_bio;
2649 rdev = conf->mirrors[dev].replacement;
2650 if (rdev && bio == IO_MADE_GOOD) {
2651 rdev_clear_badblocks(
2653 r10_bio->devs[m].addr,
2654 r10_bio->sectors, 0);
2655 rdev_dec_pending(rdev, conf->mddev);
2659 spin_lock_irq(&conf->device_lock);
2660 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2662 spin_unlock_irq(&conf->device_lock);
2663 md_wakeup_thread(conf->mddev->thread);
2665 if (test_bit(R10BIO_WriteError,
2667 close_write(r10_bio);
2668 raid_end_bio_io(r10_bio);
2673 static void raid10d(struct md_thread *thread)
2675 struct mddev *mddev = thread->mddev;
2676 struct r10bio *r10_bio;
2677 unsigned long flags;
2678 struct r10conf *conf = mddev->private;
2679 struct list_head *head = &conf->retry_list;
2680 struct blk_plug plug;
2682 md_check_recovery(mddev);
2684 if (!list_empty_careful(&conf->bio_end_io_list) &&
2685 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2687 spin_lock_irqsave(&conf->device_lock, flags);
2688 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2689 while (!list_empty(&conf->bio_end_io_list)) {
2690 list_move(conf->bio_end_io_list.prev, &tmp);
2694 spin_unlock_irqrestore(&conf->device_lock, flags);
2695 while (!list_empty(&tmp)) {
2696 r10_bio = list_first_entry(&tmp, struct r10bio,
2698 list_del(&r10_bio->retry_list);
2699 if (mddev->degraded)
2700 set_bit(R10BIO_Degraded, &r10_bio->state);
2702 if (test_bit(R10BIO_WriteError,
2704 close_write(r10_bio);
2705 raid_end_bio_io(r10_bio);
2709 blk_start_plug(&plug);
2712 flush_pending_writes(conf);
2714 spin_lock_irqsave(&conf->device_lock, flags);
2715 if (list_empty(head)) {
2716 spin_unlock_irqrestore(&conf->device_lock, flags);
2719 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2720 list_del(head->prev);
2722 spin_unlock_irqrestore(&conf->device_lock, flags);
2724 mddev = r10_bio->mddev;
2725 conf = mddev->private;
2726 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2727 test_bit(R10BIO_WriteError, &r10_bio->state))
2728 handle_write_completed(conf, r10_bio);
2729 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2730 reshape_request_write(mddev, r10_bio);
2731 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2732 sync_request_write(mddev, r10_bio);
2733 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2734 recovery_request_write(mddev, r10_bio);
2735 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2736 handle_read_error(mddev, r10_bio);
2738 /* just a partial read to be scheduled from a
2741 int slot = r10_bio->read_slot;
2742 generic_make_request(r10_bio->devs[slot].bio);
2746 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2747 md_check_recovery(mddev);
2749 blk_finish_plug(&plug);
2752 static int init_resync(struct r10conf *conf)
2757 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2758 BUG_ON(conf->r10buf_pool);
2759 conf->have_replacement = 0;
2760 for (i = 0; i < conf->geo.raid_disks; i++)
2761 if (conf->mirrors[i].replacement)
2762 conf->have_replacement = 1;
2763 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2764 if (!conf->r10buf_pool)
2766 conf->next_resync = 0;
2771 * perform a "sync" on one "block"
2773 * We need to make sure that no normal I/O request - particularly write
2774 * requests - conflict with active sync requests.
2776 * This is achieved by tracking pending requests and a 'barrier' concept
2777 * that can be installed to exclude normal IO requests.
2779 * Resync and recovery are handled very differently.
2780 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2782 * For resync, we iterate over virtual addresses, read all copies,
2783 * and update if there are differences. If only one copy is live,
2785 * For recovery, we iterate over physical addresses, read a good
2786 * value for each non-in_sync drive, and over-write.
2788 * So, for recovery we may have several outstanding complex requests for a
2789 * given address, one for each out-of-sync device. We model this by allocating
2790 * a number of r10_bio structures, one for each out-of-sync device.
2791 * As we setup these structures, we collect all bio's together into a list
2792 * which we then process collectively to add pages, and then process again
2793 * to pass to generic_make_request.
2795 * The r10_bio structures are linked using a borrowed master_bio pointer.
2796 * This link is counted in ->remaining. When the r10_bio that points to NULL
2797 * has its remaining count decremented to 0, the whole complex operation
2802 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2805 struct r10conf *conf = mddev->private;
2806 struct r10bio *r10_bio;
2807 struct bio *biolist = NULL, *bio;
2808 sector_t max_sector, nr_sectors;
2811 sector_t sync_blocks;
2812 sector_t sectors_skipped = 0;
2813 int chunks_skipped = 0;
2814 sector_t chunk_mask = conf->geo.chunk_mask;
2816 if (!conf->r10buf_pool)
2817 if (init_resync(conf))
2821 * Allow skipping a full rebuild for incremental assembly
2822 * of a clean array, like RAID1 does.
2824 if (mddev->bitmap == NULL &&
2825 mddev->recovery_cp == MaxSector &&
2826 mddev->reshape_position == MaxSector &&
2827 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2828 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2829 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2830 conf->fullsync == 0) {
2832 return mddev->dev_sectors - sector_nr;
2836 max_sector = mddev->dev_sectors;
2837 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2838 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2839 max_sector = mddev->resync_max_sectors;
2840 if (sector_nr >= max_sector) {
2841 /* If we aborted, we need to abort the
2842 * sync on the 'current' bitmap chucks (there can
2843 * be several when recovering multiple devices).
2844 * as we may have started syncing it but not finished.
2845 * We can find the current address in
2846 * mddev->curr_resync, but for recovery,
2847 * we need to convert that to several
2848 * virtual addresses.
2850 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2856 if (mddev->curr_resync < max_sector) { /* aborted */
2857 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2858 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2860 else for (i = 0; i < conf->geo.raid_disks; i++) {
2862 raid10_find_virt(conf, mddev->curr_resync, i);
2863 bitmap_end_sync(mddev->bitmap, sect,
2867 /* completed sync */
2868 if ((!mddev->bitmap || conf->fullsync)
2869 && conf->have_replacement
2870 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2871 /* Completed a full sync so the replacements
2872 * are now fully recovered.
2875 for (i = 0; i < conf->geo.raid_disks; i++) {
2876 struct md_rdev *rdev =
2877 rcu_dereference(conf->mirrors[i].replacement);
2879 rdev->recovery_offset = MaxSector;
2885 bitmap_close_sync(mddev->bitmap);
2888 return sectors_skipped;
2891 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2892 return reshape_request(mddev, sector_nr, skipped);
2894 if (chunks_skipped >= conf->geo.raid_disks) {
2895 /* if there has been nothing to do on any drive,
2896 * then there is nothing to do at all..
2899 return (max_sector - sector_nr) + sectors_skipped;
2902 if (max_sector > mddev->resync_max)
2903 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2905 /* make sure whole request will fit in a chunk - if chunks
2908 if (conf->geo.near_copies < conf->geo.raid_disks &&
2909 max_sector > (sector_nr | chunk_mask))
2910 max_sector = (sector_nr | chunk_mask) + 1;
2913 * If there is non-resync activity waiting for a turn, then let it
2914 * though before starting on this new sync request.
2916 if (conf->nr_waiting)
2917 schedule_timeout_uninterruptible(1);
2919 /* Again, very different code for resync and recovery.
2920 * Both must result in an r10bio with a list of bios that
2921 * have bi_end_io, bi_sector, bi_bdev set,
2922 * and bi_private set to the r10bio.
2923 * For recovery, we may actually create several r10bios
2924 * with 2 bios in each, that correspond to the bios in the main one.
2925 * In this case, the subordinate r10bios link back through a
2926 * borrowed master_bio pointer, and the counter in the master
2927 * includes a ref from each subordinate.
2929 /* First, we decide what to do and set ->bi_end_io
2930 * To end_sync_read if we want to read, and
2931 * end_sync_write if we will want to write.
2934 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2935 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2936 /* recovery... the complicated one */
2940 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2946 struct raid10_info *mirror = &conf->mirrors[i];
2947 struct md_rdev *mrdev, *mreplace;
2950 mrdev = rcu_dereference(mirror->rdev);
2951 mreplace = rcu_dereference(mirror->replacement);
2953 if ((mrdev == NULL ||
2954 test_bit(In_sync, &mrdev->flags)) &&
2955 (mreplace == NULL ||
2956 test_bit(Faulty, &mreplace->flags))) {
2962 /* want to reconstruct this device */
2964 sect = raid10_find_virt(conf, sector_nr, i);
2965 if (sect >= mddev->resync_max_sectors) {
2966 /* last stripe is not complete - don't
2967 * try to recover this sector.
2972 /* Unless we are doing a full sync, or a replacement
2973 * we only need to recover the block if it is set in
2976 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2978 if (sync_blocks < max_sync)
2979 max_sync = sync_blocks;
2983 /* yep, skip the sync_blocks here, but don't assume
2984 * that there will never be anything to do here
2986 chunks_skipped = -1;
2990 atomic_inc(&mrdev->nr_pending);
2992 atomic_inc(&mreplace->nr_pending);
2995 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2997 raise_barrier(conf, rb2 != NULL);
2998 atomic_set(&r10_bio->remaining, 0);
3000 r10_bio->master_bio = (struct bio*)rb2;
3002 atomic_inc(&rb2->remaining);
3003 r10_bio->mddev = mddev;
3004 set_bit(R10BIO_IsRecover, &r10_bio->state);
3005 r10_bio->sector = sect;
3007 raid10_find_phys(conf, r10_bio);
3009 /* Need to check if the array will still be
3013 for (j = 0; j < conf->geo.raid_disks; j++) {
3014 struct md_rdev *rdev = rcu_dereference(
3015 conf->mirrors[j].rdev);
3016 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3022 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3023 &sync_blocks, still_degraded);
3026 for (j=0; j<conf->copies;j++) {
3028 int d = r10_bio->devs[j].devnum;
3029 sector_t from_addr, to_addr;
3030 struct md_rdev *rdev =
3031 rcu_dereference(conf->mirrors[d].rdev);
3032 sector_t sector, first_bad;
3035 !test_bit(In_sync, &rdev->flags))
3037 /* This is where we read from */
3039 sector = r10_bio->devs[j].addr;
3041 if (is_badblock(rdev, sector, max_sync,
3042 &first_bad, &bad_sectors)) {
3043 if (first_bad > sector)
3044 max_sync = first_bad - sector;
3046 bad_sectors -= (sector
3048 if (max_sync > bad_sectors)
3049 max_sync = bad_sectors;
3053 bio = r10_bio->devs[0].bio;
3055 bio->bi_next = biolist;
3057 bio->bi_private = r10_bio;
3058 bio->bi_end_io = end_sync_read;
3060 from_addr = r10_bio->devs[j].addr;
3061 bio->bi_iter.bi_sector = from_addr +
3063 bio->bi_bdev = rdev->bdev;
3064 atomic_inc(&rdev->nr_pending);
3065 /* and we write to 'i' (if not in_sync) */
3067 for (k=0; k<conf->copies; k++)
3068 if (r10_bio->devs[k].devnum == i)
3070 BUG_ON(k == conf->copies);
3071 to_addr = r10_bio->devs[k].addr;
3072 r10_bio->devs[0].devnum = d;
3073 r10_bio->devs[0].addr = from_addr;
3074 r10_bio->devs[1].devnum = i;
3075 r10_bio->devs[1].addr = to_addr;
3077 if (!test_bit(In_sync, &mrdev->flags)) {
3078 bio = r10_bio->devs[1].bio;
3080 bio->bi_next = biolist;
3082 bio->bi_private = r10_bio;
3083 bio->bi_end_io = end_sync_write;
3085 bio->bi_iter.bi_sector = to_addr
3086 + mrdev->data_offset;
3087 bio->bi_bdev = mrdev->bdev;
3088 atomic_inc(&r10_bio->remaining);
3090 r10_bio->devs[1].bio->bi_end_io = NULL;
3092 /* and maybe write to replacement */
3093 bio = r10_bio->devs[1].repl_bio;
3095 bio->bi_end_io = NULL;
3096 /* Note: if mreplace != NULL, then bio
3097 * cannot be NULL as r10buf_pool_alloc will
3098 * have allocated it.
3099 * So the second test here is pointless.
3100 * But it keeps semantic-checkers happy, and
3101 * this comment keeps human reviewers
3104 if (mreplace == NULL || bio == NULL ||
3105 test_bit(Faulty, &mreplace->flags))
3108 bio->bi_next = biolist;
3110 bio->bi_private = r10_bio;
3111 bio->bi_end_io = end_sync_write;
3113 bio->bi_iter.bi_sector = to_addr +
3114 mreplace->data_offset;
3115 bio->bi_bdev = mreplace->bdev;
3116 atomic_inc(&r10_bio->remaining);
3120 if (j == conf->copies) {
3121 /* Cannot recover, so abort the recovery or
3122 * record a bad block */
3124 /* problem is that there are bad blocks
3125 * on other device(s)
3128 for (k = 0; k < conf->copies; k++)
3129 if (r10_bio->devs[k].devnum == i)
3131 if (!test_bit(In_sync,
3133 && !rdev_set_badblocks(
3135 r10_bio->devs[k].addr,
3139 !rdev_set_badblocks(
3141 r10_bio->devs[k].addr,
3146 if (!test_and_set_bit(MD_RECOVERY_INTR,
3148 printk(KERN_INFO "md/raid10:%s: insufficient "
3149 "working devices for recovery.\n",
3151 mirror->recovery_disabled
3152 = mddev->recovery_disabled;
3156 atomic_dec(&rb2->remaining);
3158 rdev_dec_pending(mrdev, mddev);
3160 rdev_dec_pending(mreplace, mddev);
3163 rdev_dec_pending(mrdev, mddev);
3165 rdev_dec_pending(mreplace, mddev);
3167 if (biolist == NULL) {
3169 struct r10bio *rb2 = r10_bio;
3170 r10_bio = (struct r10bio*) rb2->master_bio;
3171 rb2->master_bio = NULL;
3177 /* resync. Schedule a read for every block at this virt offset */
3180 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3182 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3183 &sync_blocks, mddev->degraded) &&
3184 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3185 &mddev->recovery)) {
3186 /* We can skip this block */
3188 return sync_blocks + sectors_skipped;
3190 if (sync_blocks < max_sync)
3191 max_sync = sync_blocks;
3192 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3195 r10_bio->mddev = mddev;
3196 atomic_set(&r10_bio->remaining, 0);
3197 raise_barrier(conf, 0);
3198 conf->next_resync = sector_nr;
3200 r10_bio->master_bio = NULL;
3201 r10_bio->sector = sector_nr;
3202 set_bit(R10BIO_IsSync, &r10_bio->state);
3203 raid10_find_phys(conf, r10_bio);
3204 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3206 for (i = 0; i < conf->copies; i++) {
3207 int d = r10_bio->devs[i].devnum;
3208 sector_t first_bad, sector;
3210 struct md_rdev *rdev;
3212 if (r10_bio->devs[i].repl_bio)
3213 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3215 bio = r10_bio->devs[i].bio;
3217 bio->bi_error = -EIO;
3219 rdev = rcu_dereference(conf->mirrors[d].rdev);
3220 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3224 sector = r10_bio->devs[i].addr;
3225 if (is_badblock(rdev, sector, max_sync,
3226 &first_bad, &bad_sectors)) {
3227 if (first_bad > sector)
3228 max_sync = first_bad - sector;
3230 bad_sectors -= (sector - first_bad);
3231 if (max_sync > bad_sectors)
3232 max_sync = bad_sectors;
3237 atomic_inc(&rdev->nr_pending);
3238 atomic_inc(&r10_bio->remaining);
3239 bio->bi_next = biolist;
3241 bio->bi_private = r10_bio;
3242 bio->bi_end_io = end_sync_read;
3244 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3245 bio->bi_bdev = rdev->bdev;
3248 rdev = rcu_dereference(conf->mirrors[d].replacement);
3249 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3253 atomic_inc(&rdev->nr_pending);
3256 /* Need to set up for writing to the replacement */
3257 bio = r10_bio->devs[i].repl_bio;
3259 bio->bi_error = -EIO;
3261 sector = r10_bio->devs[i].addr;
3262 bio->bi_next = biolist;
3264 bio->bi_private = r10_bio;
3265 bio->bi_end_io = end_sync_write;
3267 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3268 bio->bi_bdev = rdev->bdev;
3273 for (i=0; i<conf->copies; i++) {
3274 int d = r10_bio->devs[i].devnum;
3275 if (r10_bio->devs[i].bio->bi_end_io)
3276 rdev_dec_pending(conf->mirrors[d].rdev,
3278 if (r10_bio->devs[i].repl_bio &&
3279 r10_bio->devs[i].repl_bio->bi_end_io)
3281 conf->mirrors[d].replacement,
3291 if (sector_nr + max_sync < max_sector)
3292 max_sector = sector_nr + max_sync;
3295 int len = PAGE_SIZE;
3296 if (sector_nr + (len>>9) > max_sector)
3297 len = (max_sector - sector_nr) << 9;
3300 for (bio= biolist ; bio ; bio=bio->bi_next) {
3302 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3303 if (bio_add_page(bio, page, len, 0))
3307 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3308 for (bio2 = biolist;
3309 bio2 && bio2 != bio;
3310 bio2 = bio2->bi_next) {
3311 /* remove last page from this bio */
3313 bio2->bi_iter.bi_size -= len;
3314 bio_clear_flag(bio2, BIO_SEG_VALID);
3318 nr_sectors += len>>9;
3319 sector_nr += len>>9;
3320 } while (biolist->bi_vcnt < RESYNC_PAGES);
3322 r10_bio->sectors = nr_sectors;
3326 biolist = biolist->bi_next;
3328 bio->bi_next = NULL;
3329 r10_bio = bio->bi_private;
3330 r10_bio->sectors = nr_sectors;
3332 if (bio->bi_end_io == end_sync_read) {
3333 md_sync_acct(bio->bi_bdev, nr_sectors);
3335 generic_make_request(bio);
3339 if (sectors_skipped)
3340 /* pretend they weren't skipped, it makes
3341 * no important difference in this case
3343 md_done_sync(mddev, sectors_skipped, 1);
3345 return sectors_skipped + nr_sectors;
3347 /* There is nowhere to write, so all non-sync
3348 * drives must be failed or in resync, all drives
3349 * have a bad block, so try the next chunk...
3351 if (sector_nr + max_sync < max_sector)
3352 max_sector = sector_nr + max_sync;
3354 sectors_skipped += (max_sector - sector_nr);
3356 sector_nr = max_sector;
3361 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3364 struct r10conf *conf = mddev->private;
3367 raid_disks = min(conf->geo.raid_disks,
3368 conf->prev.raid_disks);
3370 sectors = conf->dev_sectors;
3372 size = sectors >> conf->geo.chunk_shift;
3373 sector_div(size, conf->geo.far_copies);
3374 size = size * raid_disks;
3375 sector_div(size, conf->geo.near_copies);
3377 return size << conf->geo.chunk_shift;
3380 static void calc_sectors(struct r10conf *conf, sector_t size)
3382 /* Calculate the number of sectors-per-device that will
3383 * actually be used, and set conf->dev_sectors and
3387 size = size >> conf->geo.chunk_shift;
3388 sector_div(size, conf->geo.far_copies);
3389 size = size * conf->geo.raid_disks;
3390 sector_div(size, conf->geo.near_copies);
3391 /* 'size' is now the number of chunks in the array */
3392 /* calculate "used chunks per device" */
3393 size = size * conf->copies;
3395 /* We need to round up when dividing by raid_disks to
3396 * get the stride size.
3398 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3400 conf->dev_sectors = size << conf->geo.chunk_shift;
3402 if (conf->geo.far_offset)
3403 conf->geo.stride = 1 << conf->geo.chunk_shift;
3405 sector_div(size, conf->geo.far_copies);
3406 conf->geo.stride = size << conf->geo.chunk_shift;
3410 enum geo_type {geo_new, geo_old, geo_start};
3411 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3414 int layout, chunk, disks;
3417 layout = mddev->layout;
3418 chunk = mddev->chunk_sectors;
3419 disks = mddev->raid_disks - mddev->delta_disks;
3422 layout = mddev->new_layout;
3423 chunk = mddev->new_chunk_sectors;
3424 disks = mddev->raid_disks;
3426 default: /* avoid 'may be unused' warnings */
3427 case geo_start: /* new when starting reshape - raid_disks not
3429 layout = mddev->new_layout;
3430 chunk = mddev->new_chunk_sectors;
3431 disks = mddev->raid_disks + mddev->delta_disks;
3436 if (chunk < (PAGE_SIZE >> 9) ||
3437 !is_power_of_2(chunk))
3440 fc = (layout >> 8) & 255;
3441 fo = layout & (1<<16);
3442 geo->raid_disks = disks;
3443 geo->near_copies = nc;
3444 geo->far_copies = fc;
3445 geo->far_offset = fo;
3446 switch (layout >> 17) {
3447 case 0: /* original layout. simple but not always optimal */
3448 geo->far_set_size = disks;
3450 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3451 * actually using this, but leave code here just in case.*/
3452 geo->far_set_size = disks/fc;
3453 WARN(geo->far_set_size < fc,
3454 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3456 case 2: /* "improved" layout fixed to match documentation */
3457 geo->far_set_size = fc * nc;
3459 default: /* Not a valid layout */
3462 geo->chunk_mask = chunk - 1;
3463 geo->chunk_shift = ffz(~chunk);
3467 static struct r10conf *setup_conf(struct mddev *mddev)
3469 struct r10conf *conf = NULL;
3474 copies = setup_geo(&geo, mddev, geo_new);
3477 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3478 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3479 mdname(mddev), PAGE_SIZE);
3483 if (copies < 2 || copies > mddev->raid_disks) {
3484 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3485 mdname(mddev), mddev->new_layout);
3490 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3494 /* FIXME calc properly */
3495 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3496 max(0,-mddev->delta_disks)),
3501 conf->tmppage = alloc_page(GFP_KERNEL);
3506 conf->copies = copies;
3507 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3508 r10bio_pool_free, conf);
3509 if (!conf->r10bio_pool)
3512 calc_sectors(conf, mddev->dev_sectors);
3513 if (mddev->reshape_position == MaxSector) {
3514 conf->prev = conf->geo;
3515 conf->reshape_progress = MaxSector;
3517 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3521 conf->reshape_progress = mddev->reshape_position;
3522 if (conf->prev.far_offset)
3523 conf->prev.stride = 1 << conf->prev.chunk_shift;
3525 /* far_copies must be 1 */
3526 conf->prev.stride = conf->dev_sectors;
3528 conf->reshape_safe = conf->reshape_progress;
3529 spin_lock_init(&conf->device_lock);
3530 INIT_LIST_HEAD(&conf->retry_list);
3531 INIT_LIST_HEAD(&conf->bio_end_io_list);
3533 spin_lock_init(&conf->resync_lock);
3534 init_waitqueue_head(&conf->wait_barrier);
3536 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3540 conf->mddev = mddev;
3545 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3548 mempool_destroy(conf->r10bio_pool);
3549 kfree(conf->mirrors);
3550 safe_put_page(conf->tmppage);
3553 return ERR_PTR(err);
3556 static int raid10_run(struct mddev *mddev)
3558 struct r10conf *conf;
3559 int i, disk_idx, chunk_size;
3560 struct raid10_info *disk;
3561 struct md_rdev *rdev;
3563 sector_t min_offset_diff = 0;
3565 bool discard_supported = false;
3567 if (mddev->private == NULL) {
3568 conf = setup_conf(mddev);
3570 return PTR_ERR(conf);
3571 mddev->private = conf;
3573 conf = mddev->private;
3577 mddev->thread = conf->thread;
3578 conf->thread = NULL;
3580 chunk_size = mddev->chunk_sectors << 9;
3582 blk_queue_max_discard_sectors(mddev->queue,
3583 mddev->chunk_sectors);
3584 blk_queue_max_write_same_sectors(mddev->queue, 0);
3585 blk_queue_io_min(mddev->queue, chunk_size);
3586 if (conf->geo.raid_disks % conf->geo.near_copies)
3587 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3589 blk_queue_io_opt(mddev->queue, chunk_size *
3590 (conf->geo.raid_disks / conf->geo.near_copies));
3593 rdev_for_each(rdev, mddev) {
3595 struct request_queue *q;
3597 disk_idx = rdev->raid_disk;
3600 if (disk_idx >= conf->geo.raid_disks &&
3601 disk_idx >= conf->prev.raid_disks)
3603 disk = conf->mirrors + disk_idx;
3605 if (test_bit(Replacement, &rdev->flags)) {
3606 if (disk->replacement)
3608 disk->replacement = rdev;
3614 q = bdev_get_queue(rdev->bdev);
3615 diff = (rdev->new_data_offset - rdev->data_offset);
3616 if (!mddev->reshape_backwards)
3620 if (first || diff < min_offset_diff)
3621 min_offset_diff = diff;
3624 disk_stack_limits(mddev->gendisk, rdev->bdev,
3625 rdev->data_offset << 9);
3627 disk->head_position = 0;
3629 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3630 discard_supported = true;
3634 if (discard_supported)
3635 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3638 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3641 /* need to check that every block has at least one working mirror */
3642 if (!enough(conf, -1)) {
3643 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3648 if (conf->reshape_progress != MaxSector) {
3649 /* must ensure that shape change is supported */
3650 if (conf->geo.far_copies != 1 &&
3651 conf->geo.far_offset == 0)
3653 if (conf->prev.far_copies != 1 &&
3654 conf->prev.far_offset == 0)
3658 mddev->degraded = 0;
3660 i < conf->geo.raid_disks
3661 || i < conf->prev.raid_disks;
3664 disk = conf->mirrors + i;
3666 if (!disk->rdev && disk->replacement) {
3667 /* The replacement is all we have - use it */
3668 disk->rdev = disk->replacement;
3669 disk->replacement = NULL;
3670 clear_bit(Replacement, &disk->rdev->flags);
3674 !test_bit(In_sync, &disk->rdev->flags)) {
3675 disk->head_position = 0;
3678 disk->rdev->saved_raid_disk < 0)
3681 disk->recovery_disabled = mddev->recovery_disabled - 1;
3684 if (mddev->recovery_cp != MaxSector)
3685 printk(KERN_NOTICE "md/raid10:%s: not clean"
3686 " -- starting background reconstruction\n",
3689 "md/raid10:%s: active with %d out of %d devices\n",
3690 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3691 conf->geo.raid_disks);
3693 * Ok, everything is just fine now
3695 mddev->dev_sectors = conf->dev_sectors;
3696 size = raid10_size(mddev, 0, 0);
3697 md_set_array_sectors(mddev, size);
3698 mddev->resync_max_sectors = size;
3701 int stripe = conf->geo.raid_disks *
3702 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3704 /* Calculate max read-ahead size.
3705 * We need to readahead at least twice a whole stripe....
3708 stripe /= conf->geo.near_copies;
3709 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3710 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3713 if (md_integrity_register(mddev))
3716 if (conf->reshape_progress != MaxSector) {
3717 unsigned long before_length, after_length;
3719 before_length = ((1 << conf->prev.chunk_shift) *
3720 conf->prev.far_copies);
3721 after_length = ((1 << conf->geo.chunk_shift) *
3722 conf->geo.far_copies);
3724 if (max(before_length, after_length) > min_offset_diff) {
3725 /* This cannot work */
3726 printk("md/raid10: offset difference not enough to continue reshape\n");
3729 conf->offset_diff = min_offset_diff;
3731 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3732 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3733 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3734 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3735 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3742 md_unregister_thread(&mddev->thread);
3743 mempool_destroy(conf->r10bio_pool);
3744 safe_put_page(conf->tmppage);
3745 kfree(conf->mirrors);
3747 mddev->private = NULL;
3752 static void raid10_free(struct mddev *mddev, void *priv)
3754 struct r10conf *conf = priv;
3756 mempool_destroy(conf->r10bio_pool);
3757 safe_put_page(conf->tmppage);
3758 kfree(conf->mirrors);
3759 kfree(conf->mirrors_old);
3760 kfree(conf->mirrors_new);
3764 static void raid10_quiesce(struct mddev *mddev, int state)
3766 struct r10conf *conf = mddev->private;
3770 raise_barrier(conf, 0);
3773 lower_barrier(conf);
3778 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3780 /* Resize of 'far' arrays is not supported.
3781 * For 'near' and 'offset' arrays we can set the
3782 * number of sectors used to be an appropriate multiple
3783 * of the chunk size.
3784 * For 'offset', this is far_copies*chunksize.
3785 * For 'near' the multiplier is the LCM of
3786 * near_copies and raid_disks.
3787 * So if far_copies > 1 && !far_offset, fail.
3788 * Else find LCM(raid_disks, near_copy)*far_copies and
3789 * multiply by chunk_size. Then round to this number.
3790 * This is mostly done by raid10_size()
3792 struct r10conf *conf = mddev->private;
3793 sector_t oldsize, size;
3795 if (mddev->reshape_position != MaxSector)
3798 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3801 oldsize = raid10_size(mddev, 0, 0);
3802 size = raid10_size(mddev, sectors, 0);
3803 if (mddev->external_size &&
3804 mddev->array_sectors > size)
3806 if (mddev->bitmap) {
3807 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3811 md_set_array_sectors(mddev, size);
3813 set_capacity(mddev->gendisk, mddev->array_sectors);
3814 revalidate_disk(mddev->gendisk);
3816 if (sectors > mddev->dev_sectors &&
3817 mddev->recovery_cp > oldsize) {
3818 mddev->recovery_cp = oldsize;
3819 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3821 calc_sectors(conf, sectors);
3822 mddev->dev_sectors = conf->dev_sectors;
3823 mddev->resync_max_sectors = size;
3827 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3829 struct md_rdev *rdev;
3830 struct r10conf *conf;
3832 if (mddev->degraded > 0) {
3833 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3835 return ERR_PTR(-EINVAL);
3837 sector_div(size, devs);
3839 /* Set new parameters */
3840 mddev->new_level = 10;
3841 /* new layout: far_copies = 1, near_copies = 2 */
3842 mddev->new_layout = (1<<8) + 2;
3843 mddev->new_chunk_sectors = mddev->chunk_sectors;
3844 mddev->delta_disks = mddev->raid_disks;
3845 mddev->raid_disks *= 2;
3846 /* make sure it will be not marked as dirty */
3847 mddev->recovery_cp = MaxSector;
3848 mddev->dev_sectors = size;
3850 conf = setup_conf(mddev);
3851 if (!IS_ERR(conf)) {
3852 rdev_for_each(rdev, mddev)
3853 if (rdev->raid_disk >= 0) {
3854 rdev->new_raid_disk = rdev->raid_disk * 2;
3855 rdev->sectors = size;
3863 static void *raid10_takeover(struct mddev *mddev)
3865 struct r0conf *raid0_conf;
3867 /* raid10 can take over:
3868 * raid0 - providing it has only two drives
3870 if (mddev->level == 0) {
3871 /* for raid0 takeover only one zone is supported */
3872 raid0_conf = mddev->private;
3873 if (raid0_conf->nr_strip_zones > 1) {
3874 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3875 " with more than one zone.\n",
3877 return ERR_PTR(-EINVAL);
3879 return raid10_takeover_raid0(mddev,
3880 raid0_conf->strip_zone->zone_end,
3881 raid0_conf->strip_zone->nb_dev);
3883 return ERR_PTR(-EINVAL);
3886 static int raid10_check_reshape(struct mddev *mddev)
3888 /* Called when there is a request to change
3889 * - layout (to ->new_layout)
3890 * - chunk size (to ->new_chunk_sectors)
3891 * - raid_disks (by delta_disks)
3892 * or when trying to restart a reshape that was ongoing.
3894 * We need to validate the request and possibly allocate
3895 * space if that might be an issue later.
3897 * Currently we reject any reshape of a 'far' mode array,
3898 * allow chunk size to change if new is generally acceptable,
3899 * allow raid_disks to increase, and allow
3900 * a switch between 'near' mode and 'offset' mode.
3902 struct r10conf *conf = mddev->private;
3905 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3908 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3909 /* mustn't change number of copies */
3911 if (geo.far_copies > 1 && !geo.far_offset)
3912 /* Cannot switch to 'far' mode */
3915 if (mddev->array_sectors & geo.chunk_mask)
3916 /* not factor of array size */
3919 if (!enough(conf, -1))
3922 kfree(conf->mirrors_new);
3923 conf->mirrors_new = NULL;
3924 if (mddev->delta_disks > 0) {
3925 /* allocate new 'mirrors' list */
3926 conf->mirrors_new = kzalloc(
3927 sizeof(struct raid10_info)
3928 *(mddev->raid_disks +
3929 mddev->delta_disks),
3931 if (!conf->mirrors_new)
3938 * Need to check if array has failed when deciding whether to:
3940 * - remove non-faulty devices
3943 * This determination is simple when no reshape is happening.
3944 * However if there is a reshape, we need to carefully check
3945 * both the before and after sections.
3946 * This is because some failed devices may only affect one
3947 * of the two sections, and some non-in_sync devices may
3948 * be insync in the section most affected by failed devices.
3950 static int calc_degraded(struct r10conf *conf)
3952 int degraded, degraded2;
3957 /* 'prev' section first */
3958 for (i = 0; i < conf->prev.raid_disks; i++) {
3959 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3960 if (!rdev || test_bit(Faulty, &rdev->flags))
3962 else if (!test_bit(In_sync, &rdev->flags))
3963 /* When we can reduce the number of devices in
3964 * an array, this might not contribute to
3965 * 'degraded'. It does now.
3970 if (conf->geo.raid_disks == conf->prev.raid_disks)
3974 for (i = 0; i < conf->geo.raid_disks; i++) {
3975 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3976 if (!rdev || test_bit(Faulty, &rdev->flags))
3978 else if (!test_bit(In_sync, &rdev->flags)) {
3979 /* If reshape is increasing the number of devices,
3980 * this section has already been recovered, so
3981 * it doesn't contribute to degraded.
3984 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3989 if (degraded2 > degraded)
3994 static int raid10_start_reshape(struct mddev *mddev)
3996 /* A 'reshape' has been requested. This commits
3997 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3998 * This also checks if there are enough spares and adds them
4000 * We currently require enough spares to make the final
4001 * array non-degraded. We also require that the difference
4002 * between old and new data_offset - on each device - is
4003 * enough that we never risk over-writing.
4006 unsigned long before_length, after_length;
4007 sector_t min_offset_diff = 0;
4010 struct r10conf *conf = mddev->private;
4011 struct md_rdev *rdev;
4015 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4018 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4021 before_length = ((1 << conf->prev.chunk_shift) *
4022 conf->prev.far_copies);
4023 after_length = ((1 << conf->geo.chunk_shift) *
4024 conf->geo.far_copies);
4026 rdev_for_each(rdev, mddev) {
4027 if (!test_bit(In_sync, &rdev->flags)
4028 && !test_bit(Faulty, &rdev->flags))
4030 if (rdev->raid_disk >= 0) {
4031 long long diff = (rdev->new_data_offset
4032 - rdev->data_offset);
4033 if (!mddev->reshape_backwards)
4037 if (first || diff < min_offset_diff)
4038 min_offset_diff = diff;
4042 if (max(before_length, after_length) > min_offset_diff)
4045 if (spares < mddev->delta_disks)
4048 conf->offset_diff = min_offset_diff;
4049 spin_lock_irq(&conf->device_lock);
4050 if (conf->mirrors_new) {
4051 memcpy(conf->mirrors_new, conf->mirrors,
4052 sizeof(struct raid10_info)*conf->prev.raid_disks);
4054 kfree(conf->mirrors_old);
4055 conf->mirrors_old = conf->mirrors;
4056 conf->mirrors = conf->mirrors_new;
4057 conf->mirrors_new = NULL;
4059 setup_geo(&conf->geo, mddev, geo_start);
4061 if (mddev->reshape_backwards) {
4062 sector_t size = raid10_size(mddev, 0, 0);
4063 if (size < mddev->array_sectors) {
4064 spin_unlock_irq(&conf->device_lock);
4065 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4069 mddev->resync_max_sectors = size;
4070 conf->reshape_progress = size;
4072 conf->reshape_progress = 0;
4073 conf->reshape_safe = conf->reshape_progress;
4074 spin_unlock_irq(&conf->device_lock);
4076 if (mddev->delta_disks && mddev->bitmap) {
4077 ret = bitmap_resize(mddev->bitmap,
4078 raid10_size(mddev, 0,
4079 conf->geo.raid_disks),
4084 if (mddev->delta_disks > 0) {
4085 rdev_for_each(rdev, mddev)
4086 if (rdev->raid_disk < 0 &&
4087 !test_bit(Faulty, &rdev->flags)) {
4088 if (raid10_add_disk(mddev, rdev) == 0) {
4089 if (rdev->raid_disk >=
4090 conf->prev.raid_disks)
4091 set_bit(In_sync, &rdev->flags);
4093 rdev->recovery_offset = 0;
4095 if (sysfs_link_rdev(mddev, rdev))
4096 /* Failure here is OK */;
4098 } else if (rdev->raid_disk >= conf->prev.raid_disks
4099 && !test_bit(Faulty, &rdev->flags)) {
4100 /* This is a spare that was manually added */
4101 set_bit(In_sync, &rdev->flags);
4104 /* When a reshape changes the number of devices,
4105 * ->degraded is measured against the larger of the
4106 * pre and post numbers.
4108 spin_lock_irq(&conf->device_lock);
4109 mddev->degraded = calc_degraded(conf);
4110 spin_unlock_irq(&conf->device_lock);
4111 mddev->raid_disks = conf->geo.raid_disks;
4112 mddev->reshape_position = conf->reshape_progress;
4113 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4115 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4116 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4117 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4118 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4119 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4121 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4123 if (!mddev->sync_thread) {
4127 conf->reshape_checkpoint = jiffies;
4128 md_wakeup_thread(mddev->sync_thread);
4129 md_new_event(mddev);
4133 mddev->recovery = 0;
4134 spin_lock_irq(&conf->device_lock);
4135 conf->geo = conf->prev;
4136 mddev->raid_disks = conf->geo.raid_disks;
4137 rdev_for_each(rdev, mddev)
4138 rdev->new_data_offset = rdev->data_offset;
4140 conf->reshape_progress = MaxSector;
4141 conf->reshape_safe = MaxSector;
4142 mddev->reshape_position = MaxSector;
4143 spin_unlock_irq(&conf->device_lock);
4147 /* Calculate the last device-address that could contain
4148 * any block from the chunk that includes the array-address 's'
4149 * and report the next address.
4150 * i.e. the address returned will be chunk-aligned and after
4151 * any data that is in the chunk containing 's'.
4153 static sector_t last_dev_address(sector_t s, struct geom *geo)
4155 s = (s | geo->chunk_mask) + 1;
4156 s >>= geo->chunk_shift;
4157 s *= geo->near_copies;
4158 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4159 s *= geo->far_copies;
4160 s <<= geo->chunk_shift;
4164 /* Calculate the first device-address that could contain
4165 * any block from the chunk that includes the array-address 's'.
4166 * This too will be the start of a chunk
4168 static sector_t first_dev_address(sector_t s, struct geom *geo)
4170 s >>= geo->chunk_shift;
4171 s *= geo->near_copies;
4172 sector_div(s, geo->raid_disks);
4173 s *= geo->far_copies;
4174 s <<= geo->chunk_shift;
4178 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4181 /* We simply copy at most one chunk (smallest of old and new)
4182 * at a time, possibly less if that exceeds RESYNC_PAGES,
4183 * or we hit a bad block or something.
4184 * This might mean we pause for normal IO in the middle of
4185 * a chunk, but that is not a problem as mddev->reshape_position
4186 * can record any location.
4188 * If we will want to write to a location that isn't
4189 * yet recorded as 'safe' (i.e. in metadata on disk) then
4190 * we need to flush all reshape requests and update the metadata.
4192 * When reshaping forwards (e.g. to more devices), we interpret
4193 * 'safe' as the earliest block which might not have been copied
4194 * down yet. We divide this by previous stripe size and multiply
4195 * by previous stripe length to get lowest device offset that we
4196 * cannot write to yet.
4197 * We interpret 'sector_nr' as an address that we want to write to.
4198 * From this we use last_device_address() to find where we might
4199 * write to, and first_device_address on the 'safe' position.
4200 * If this 'next' write position is after the 'safe' position,
4201 * we must update the metadata to increase the 'safe' position.
4203 * When reshaping backwards, we round in the opposite direction
4204 * and perform the reverse test: next write position must not be
4205 * less than current safe position.
4207 * In all this the minimum difference in data offsets
4208 * (conf->offset_diff - always positive) allows a bit of slack,
4209 * so next can be after 'safe', but not by more than offset_diff
4211 * We need to prepare all the bios here before we start any IO
4212 * to ensure the size we choose is acceptable to all devices.
4213 * The means one for each copy for write-out and an extra one for
4215 * We store the read-in bio in ->master_bio and the others in
4216 * ->devs[x].bio and ->devs[x].repl_bio.
4218 struct r10conf *conf = mddev->private;
4219 struct r10bio *r10_bio;
4220 sector_t next, safe, last;
4224 struct md_rdev *rdev;
4227 struct bio *bio, *read_bio;
4228 int sectors_done = 0;
4230 if (sector_nr == 0) {
4231 /* If restarting in the middle, skip the initial sectors */
4232 if (mddev->reshape_backwards &&
4233 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4234 sector_nr = (raid10_size(mddev, 0, 0)
4235 - conf->reshape_progress);
4236 } else if (!mddev->reshape_backwards &&
4237 conf->reshape_progress > 0)
4238 sector_nr = conf->reshape_progress;
4240 mddev->curr_resync_completed = sector_nr;
4241 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4247 /* We don't use sector_nr to track where we are up to
4248 * as that doesn't work well for ->reshape_backwards.
4249 * So just use ->reshape_progress.
4251 if (mddev->reshape_backwards) {
4252 /* 'next' is the earliest device address that we might
4253 * write to for this chunk in the new layout
4255 next = first_dev_address(conf->reshape_progress - 1,
4258 /* 'safe' is the last device address that we might read from
4259 * in the old layout after a restart
4261 safe = last_dev_address(conf->reshape_safe - 1,
4264 if (next + conf->offset_diff < safe)
4267 last = conf->reshape_progress - 1;
4268 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4269 & conf->prev.chunk_mask);
4270 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4271 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4273 /* 'next' is after the last device address that we
4274 * might write to for this chunk in the new layout
4276 next = last_dev_address(conf->reshape_progress, &conf->geo);
4278 /* 'safe' is the earliest device address that we might
4279 * read from in the old layout after a restart
4281 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4283 /* Need to update metadata if 'next' might be beyond 'safe'
4284 * as that would possibly corrupt data
4286 if (next > safe + conf->offset_diff)
4289 sector_nr = conf->reshape_progress;
4290 last = sector_nr | (conf->geo.chunk_mask
4291 & conf->prev.chunk_mask);
4293 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4294 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4298 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4299 /* Need to update reshape_position in metadata */
4301 mddev->reshape_position = conf->reshape_progress;
4302 if (mddev->reshape_backwards)
4303 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4304 - conf->reshape_progress;
4306 mddev->curr_resync_completed = conf->reshape_progress;
4307 conf->reshape_checkpoint = jiffies;
4308 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4309 md_wakeup_thread(mddev->thread);
4310 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4311 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4312 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4313 allow_barrier(conf);
4314 return sectors_done;
4316 conf->reshape_safe = mddev->reshape_position;
4317 allow_barrier(conf);
4321 /* Now schedule reads for blocks from sector_nr to last */
4322 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4324 raise_barrier(conf, sectors_done != 0);
4325 atomic_set(&r10_bio->remaining, 0);
4326 r10_bio->mddev = mddev;
4327 r10_bio->sector = sector_nr;
4328 set_bit(R10BIO_IsReshape, &r10_bio->state);
4329 r10_bio->sectors = last - sector_nr + 1;
4330 rdev = read_balance(conf, r10_bio, &max_sectors);
4331 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4334 /* Cannot read from here, so need to record bad blocks
4335 * on all the target devices.
4338 mempool_free(r10_bio, conf->r10buf_pool);
4339 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4340 return sectors_done;
4343 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4345 read_bio->bi_bdev = rdev->bdev;
4346 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4347 + rdev->data_offset);
4348 read_bio->bi_private = r10_bio;
4349 read_bio->bi_end_io = end_sync_read;
4350 read_bio->bi_rw = READ;
4351 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4352 read_bio->bi_error = 0;
4353 read_bio->bi_vcnt = 0;
4354 read_bio->bi_iter.bi_size = 0;
4355 r10_bio->master_bio = read_bio;
4356 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4358 /* Now find the locations in the new layout */
4359 __raid10_find_phys(&conf->geo, r10_bio);
4362 read_bio->bi_next = NULL;
4364 for (s = 0; s < conf->copies*2; s++) {
4366 int d = r10_bio->devs[s/2].devnum;
4367 struct md_rdev *rdev2;
4369 rdev2 = conf->mirrors[d].replacement;
4370 b = r10_bio->devs[s/2].repl_bio;
4372 rdev2 = conf->mirrors[d].rdev;
4373 b = r10_bio->devs[s/2].bio;
4375 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4379 b->bi_bdev = rdev2->bdev;
4380 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4381 rdev2->new_data_offset;
4382 b->bi_private = r10_bio;
4383 b->bi_end_io = end_reshape_write;
4389 /* Now add as many pages as possible to all of these bios. */
4392 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4393 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4394 int len = (max_sectors - s) << 9;
4395 if (len > PAGE_SIZE)
4397 for (bio = blist; bio ; bio = bio->bi_next) {
4399 if (bio_add_page(bio, page, len, 0))
4402 /* Didn't fit, must stop */
4404 bio2 && bio2 != bio;
4405 bio2 = bio2->bi_next) {
4406 /* Remove last page from this bio */
4408 bio2->bi_iter.bi_size -= len;
4409 bio_clear_flag(bio2, BIO_SEG_VALID);
4413 sector_nr += len >> 9;
4414 nr_sectors += len >> 9;
4417 r10_bio->sectors = nr_sectors;
4419 /* Now submit the read */
4420 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4421 atomic_inc(&r10_bio->remaining);
4422 read_bio->bi_next = NULL;
4423 generic_make_request(read_bio);
4424 sector_nr += nr_sectors;
4425 sectors_done += nr_sectors;
4426 if (sector_nr <= last)
4429 /* Now that we have done the whole section we can
4430 * update reshape_progress
4432 if (mddev->reshape_backwards)
4433 conf->reshape_progress -= sectors_done;
4435 conf->reshape_progress += sectors_done;
4437 return sectors_done;
4440 static void end_reshape_request(struct r10bio *r10_bio);
4441 static int handle_reshape_read_error(struct mddev *mddev,
4442 struct r10bio *r10_bio);
4443 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4445 /* Reshape read completed. Hopefully we have a block
4447 * If we got a read error then we do sync 1-page reads from
4448 * elsewhere until we find the data - or give up.
4450 struct r10conf *conf = mddev->private;
4453 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4454 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4455 /* Reshape has been aborted */
4456 md_done_sync(mddev, r10_bio->sectors, 0);
4460 /* We definitely have the data in the pages, schedule the
4463 atomic_set(&r10_bio->remaining, 1);
4464 for (s = 0; s < conf->copies*2; s++) {
4466 int d = r10_bio->devs[s/2].devnum;
4467 struct md_rdev *rdev;
4469 rdev = conf->mirrors[d].replacement;
4470 b = r10_bio->devs[s/2].repl_bio;
4472 rdev = conf->mirrors[d].rdev;
4473 b = r10_bio->devs[s/2].bio;
4475 if (!rdev || test_bit(Faulty, &rdev->flags))
4477 atomic_inc(&rdev->nr_pending);
4478 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4479 atomic_inc(&r10_bio->remaining);
4481 generic_make_request(b);
4483 end_reshape_request(r10_bio);
4486 static void end_reshape(struct r10conf *conf)
4488 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4491 spin_lock_irq(&conf->device_lock);
4492 conf->prev = conf->geo;
4493 md_finish_reshape(conf->mddev);
4495 conf->reshape_progress = MaxSector;
4496 conf->reshape_safe = MaxSector;
4497 spin_unlock_irq(&conf->device_lock);
4499 /* read-ahead size must cover two whole stripes, which is
4500 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4502 if (conf->mddev->queue) {
4503 int stripe = conf->geo.raid_disks *
4504 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4505 stripe /= conf->geo.near_copies;
4506 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4507 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4512 static int handle_reshape_read_error(struct mddev *mddev,
4513 struct r10bio *r10_bio)
4515 /* Use sync reads to get the blocks from somewhere else */
4516 int sectors = r10_bio->sectors;
4517 struct r10conf *conf = mddev->private;
4519 struct r10bio r10_bio;
4520 struct r10dev devs[conf->copies];
4522 struct r10bio *r10b = &on_stack.r10_bio;
4525 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4527 r10b->sector = r10_bio->sector;
4528 __raid10_find_phys(&conf->prev, r10b);
4533 int first_slot = slot;
4535 if (s > (PAGE_SIZE >> 9))
4539 int d = r10b->devs[slot].devnum;
4540 struct md_rdev *rdev = conf->mirrors[d].rdev;
4543 test_bit(Faulty, &rdev->flags) ||
4544 !test_bit(In_sync, &rdev->flags))
4547 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4548 success = sync_page_io(rdev,
4557 if (slot >= conf->copies)
4559 if (slot == first_slot)
4563 /* couldn't read this block, must give up */
4564 set_bit(MD_RECOVERY_INTR,
4574 static void end_reshape_write(struct bio *bio)
4576 struct r10bio *r10_bio = bio->bi_private;
4577 struct mddev *mddev = r10_bio->mddev;
4578 struct r10conf *conf = mddev->private;
4582 struct md_rdev *rdev = NULL;
4584 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4586 rdev = conf->mirrors[d].replacement;
4589 rdev = conf->mirrors[d].rdev;
4592 if (bio->bi_error) {
4593 /* FIXME should record badblock */
4594 md_error(mddev, rdev);
4597 rdev_dec_pending(rdev, mddev);
4598 end_reshape_request(r10_bio);
4601 static void end_reshape_request(struct r10bio *r10_bio)
4603 if (!atomic_dec_and_test(&r10_bio->remaining))
4605 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4606 bio_put(r10_bio->master_bio);
4610 static void raid10_finish_reshape(struct mddev *mddev)
4612 struct r10conf *conf = mddev->private;
4614 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4617 if (mddev->delta_disks > 0) {
4618 sector_t size = raid10_size(mddev, 0, 0);
4619 md_set_array_sectors(mddev, size);
4620 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4621 mddev->recovery_cp = mddev->resync_max_sectors;
4622 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4624 mddev->resync_max_sectors = size;
4626 set_capacity(mddev->gendisk, mddev->array_sectors);
4627 revalidate_disk(mddev->gendisk);
4631 for (d = conf->geo.raid_disks ;
4632 d < conf->geo.raid_disks - mddev->delta_disks;
4634 struct md_rdev *rdev = conf->mirrors[d].rdev;
4636 clear_bit(In_sync, &rdev->flags);
4637 rdev = conf->mirrors[d].replacement;
4639 clear_bit(In_sync, &rdev->flags);
4642 mddev->layout = mddev->new_layout;
4643 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4644 mddev->reshape_position = MaxSector;
4645 mddev->delta_disks = 0;
4646 mddev->reshape_backwards = 0;
4649 static struct md_personality raid10_personality =
4653 .owner = THIS_MODULE,
4654 .make_request = raid10_make_request,
4656 .free = raid10_free,
4657 .status = raid10_status,
4658 .error_handler = raid10_error,
4659 .hot_add_disk = raid10_add_disk,
4660 .hot_remove_disk= raid10_remove_disk,
4661 .spare_active = raid10_spare_active,
4662 .sync_request = raid10_sync_request,
4663 .quiesce = raid10_quiesce,
4664 .size = raid10_size,
4665 .resize = raid10_resize,
4666 .takeover = raid10_takeover,
4667 .check_reshape = raid10_check_reshape,
4668 .start_reshape = raid10_start_reshape,
4669 .finish_reshape = raid10_finish_reshape,
4670 .congested = raid10_congested,
4673 static int __init raid_init(void)
4675 return register_md_personality(&raid10_personality);
4678 static void raid_exit(void)
4680 unregister_md_personality(&raid10_personality);
4683 module_init(raid_init);
4684 module_exit(raid_exit);
4685 MODULE_LICENSE("GPL");
4686 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4687 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4688 MODULE_ALIAS("md-raid10");
4689 MODULE_ALIAS("md-level-10");
4691 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);