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>
33 * RAID10 provides a combination of RAID0 and RAID1 functionality.
34 * The layout of data is defined by
37 * near_copies (stored in low byte of layout)
38 * far_copies (stored in second byte of layout)
39 * far_offset (stored in bit 16 of layout )
41 * The data to be stored is divided into chunks using chunksize.
42 * Each device is divided into far_copies sections.
43 * In each section, chunks are laid out in a style similar to raid0, but
44 * near_copies copies of each chunk is stored (each on a different drive).
45 * The starting device for each section is offset near_copies from the starting
46 * device of the previous section.
47 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
49 * near_copies and far_copies must be at least one, and their product is at most
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of be very far apart
54 * on disk, there are adjacent stripes.
58 * Number of guaranteed r10bios in case of extreme VM load:
60 #define NR_RAID10_BIOS 256
62 /* When there are this many requests queue to be written by
63 * the raid10 thread, we become 'congested' to provide back-pressure
66 static int max_queued_requests = 1024;
68 static void allow_barrier(struct r10conf *conf);
69 static void lower_barrier(struct r10conf *conf);
71 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
73 struct r10conf *conf = data;
74 int size = offsetof(struct r10bio, devs[conf->copies]);
76 /* allocate a r10bio with room for raid_disks entries in the
78 return kzalloc(size, gfp_flags);
81 static void r10bio_pool_free(void *r10_bio, void *data)
86 /* Maximum size of each resync request */
87 #define RESYNC_BLOCK_SIZE (64*1024)
88 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
89 /* amount of memory to reserve for resync requests */
90 #define RESYNC_WINDOW (1024*1024)
91 /* maximum number of concurrent requests, memory permitting */
92 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
95 * When performing a resync, we need to read and compare, so
96 * we need as many pages are there are copies.
97 * When performing a recovery, we need 2 bios, one for read,
98 * one for write (we recover only one drive per r10buf)
101 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
103 struct r10conf *conf = data;
105 struct r10bio *r10_bio;
110 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
114 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
115 nalloc = conf->copies; /* resync */
117 nalloc = 2; /* recovery */
122 for (j = nalloc ; j-- ; ) {
123 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
126 r10_bio->devs[j].bio = bio;
127 if (!conf->have_replacement)
129 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
132 r10_bio->devs[j].repl_bio = bio;
135 * Allocate RESYNC_PAGES data pages and attach them
138 for (j = 0 ; j < nalloc; j++) {
139 struct bio *rbio = r10_bio->devs[j].repl_bio;
140 bio = r10_bio->devs[j].bio;
141 for (i = 0; i < RESYNC_PAGES; i++) {
142 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
143 &conf->mddev->recovery)) {
144 /* we can share bv_page's during recovery */
145 struct bio *rbio = r10_bio->devs[0].bio;
146 page = rbio->bi_io_vec[i].bv_page;
149 page = alloc_page(gfp_flags);
153 bio->bi_io_vec[i].bv_page = page;
155 rbio->bi_io_vec[i].bv_page = page;
163 safe_put_page(bio->bi_io_vec[i-1].bv_page);
165 for (i = 0; i < RESYNC_PAGES ; i++)
166 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
169 while (++j < nalloc) {
170 bio_put(r10_bio->devs[j].bio);
171 if (r10_bio->devs[j].repl_bio)
172 bio_put(r10_bio->devs[j].repl_bio);
174 r10bio_pool_free(r10_bio, conf);
178 static void r10buf_pool_free(void *__r10_bio, void *data)
181 struct r10conf *conf = data;
182 struct r10bio *r10bio = __r10_bio;
185 for (j=0; j < conf->copies; j++) {
186 struct bio *bio = r10bio->devs[j].bio;
188 for (i = 0; i < RESYNC_PAGES; i++) {
189 safe_put_page(bio->bi_io_vec[i].bv_page);
190 bio->bi_io_vec[i].bv_page = NULL;
194 bio = r10bio->devs[j].repl_bio;
198 r10bio_pool_free(r10bio, conf);
201 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
205 for (i = 0; i < conf->copies; i++) {
206 struct bio **bio = & r10_bio->devs[i].bio;
207 if (!BIO_SPECIAL(*bio))
210 bio = &r10_bio->devs[i].repl_bio;
211 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
217 static void free_r10bio(struct r10bio *r10_bio)
219 struct r10conf *conf = r10_bio->mddev->private;
221 put_all_bios(conf, r10_bio);
222 mempool_free(r10_bio, conf->r10bio_pool);
225 static void put_buf(struct r10bio *r10_bio)
227 struct r10conf *conf = r10_bio->mddev->private;
229 mempool_free(r10_bio, conf->r10buf_pool);
234 static void reschedule_retry(struct r10bio *r10_bio)
237 struct mddev *mddev = r10_bio->mddev;
238 struct r10conf *conf = mddev->private;
240 spin_lock_irqsave(&conf->device_lock, flags);
241 list_add(&r10_bio->retry_list, &conf->retry_list);
243 spin_unlock_irqrestore(&conf->device_lock, flags);
245 /* wake up frozen array... */
246 wake_up(&conf->wait_barrier);
248 md_wakeup_thread(mddev->thread);
252 * raid_end_bio_io() is called when we have finished servicing a mirrored
253 * operation and are ready to return a success/failure code to the buffer
256 static void raid_end_bio_io(struct r10bio *r10_bio)
258 struct bio *bio = r10_bio->master_bio;
260 struct r10conf *conf = r10_bio->mddev->private;
262 if (bio->bi_phys_segments) {
264 spin_lock_irqsave(&conf->device_lock, flags);
265 bio->bi_phys_segments--;
266 done = (bio->bi_phys_segments == 0);
267 spin_unlock_irqrestore(&conf->device_lock, flags);
270 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
271 clear_bit(BIO_UPTODATE, &bio->bi_flags);
275 * Wake up any possible resync thread that waits for the device
280 free_r10bio(r10_bio);
284 * Update disk head position estimator based on IRQ completion info.
286 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
288 struct r10conf *conf = r10_bio->mddev->private;
290 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
291 r10_bio->devs[slot].addr + (r10_bio->sectors);
295 * Find the disk number which triggered given bio
297 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
298 struct bio *bio, int *slotp, int *replp)
303 for (slot = 0; slot < conf->copies; slot++) {
304 if (r10_bio->devs[slot].bio == bio)
306 if (r10_bio->devs[slot].repl_bio == bio) {
312 BUG_ON(slot == conf->copies);
313 update_head_pos(slot, r10_bio);
319 return r10_bio->devs[slot].devnum;
322 static void raid10_end_read_request(struct bio *bio, int error)
324 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
325 struct r10bio *r10_bio = bio->bi_private;
327 struct r10conf *conf = r10_bio->mddev->private;
330 slot = r10_bio->read_slot;
331 dev = r10_bio->devs[slot].devnum;
333 * this branch is our 'one mirror IO has finished' event handler:
335 update_head_pos(slot, r10_bio);
339 * Set R10BIO_Uptodate in our master bio, so that
340 * we will return a good error code to the higher
341 * levels even if IO on some other mirrored buffer fails.
343 * The 'master' represents the composite IO operation to
344 * user-side. So if something waits for IO, then it will
345 * wait for the 'master' bio.
347 set_bit(R10BIO_Uptodate, &r10_bio->state);
348 raid_end_bio_io(r10_bio);
349 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
352 * oops, read error - keep the refcount on the rdev
354 char b[BDEVNAME_SIZE];
355 printk_ratelimited(KERN_ERR
356 "md/raid10:%s: %s: rescheduling sector %llu\n",
358 bdevname(conf->mirrors[dev].rdev->bdev, b),
359 (unsigned long long)r10_bio->sector);
360 set_bit(R10BIO_ReadError, &r10_bio->state);
361 reschedule_retry(r10_bio);
365 static void close_write(struct r10bio *r10_bio)
367 /* clear the bitmap if all writes complete successfully */
368 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
370 !test_bit(R10BIO_Degraded, &r10_bio->state),
372 md_write_end(r10_bio->mddev);
375 static void one_write_done(struct r10bio *r10_bio)
377 if (atomic_dec_and_test(&r10_bio->remaining)) {
378 if (test_bit(R10BIO_WriteError, &r10_bio->state))
379 reschedule_retry(r10_bio);
381 close_write(r10_bio);
382 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
383 reschedule_retry(r10_bio);
385 raid_end_bio_io(r10_bio);
390 static void raid10_end_write_request(struct bio *bio, int error)
392 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
393 struct r10bio *r10_bio = bio->bi_private;
396 struct r10conf *conf = r10_bio->mddev->private;
399 dev = find_bio_disk(conf, r10_bio, bio, &slot, NULL);
402 * this branch is our 'one mirror IO has finished' event handler:
405 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
406 set_bit(R10BIO_WriteError, &r10_bio->state);
410 * Set R10BIO_Uptodate in our master bio, so that
411 * we will return a good error code for to the higher
412 * levels even if IO on some other mirrored buffer fails.
414 * The 'master' represents the composite IO operation to
415 * user-side. So if something waits for IO, then it will
416 * wait for the 'master' bio.
421 set_bit(R10BIO_Uptodate, &r10_bio->state);
423 /* Maybe we can clear some bad blocks. */
424 if (is_badblock(conf->mirrors[dev].rdev,
425 r10_bio->devs[slot].addr,
427 &first_bad, &bad_sectors)) {
429 r10_bio->devs[slot].bio = IO_MADE_GOOD;
431 set_bit(R10BIO_MadeGood, &r10_bio->state);
437 * Let's see if all mirrored write operations have finished
440 one_write_done(r10_bio);
442 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
447 * RAID10 layout manager
448 * As well as the chunksize and raid_disks count, there are two
449 * parameters: near_copies and far_copies.
450 * near_copies * far_copies must be <= raid_disks.
451 * Normally one of these will be 1.
452 * If both are 1, we get raid0.
453 * If near_copies == raid_disks, we get raid1.
455 * Chunks are laid out in raid0 style with near_copies copies of the
456 * first chunk, followed by near_copies copies of the next chunk and
458 * If far_copies > 1, then after 1/far_copies of the array has been assigned
459 * as described above, we start again with a device offset of near_copies.
460 * So we effectively have another copy of the whole array further down all
461 * the drives, but with blocks on different drives.
462 * With this layout, and block is never stored twice on the one device.
464 * raid10_find_phys finds the sector offset of a given virtual sector
465 * on each device that it is on.
467 * raid10_find_virt does the reverse mapping, from a device and a
468 * sector offset to a virtual address
471 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
481 /* now calculate first sector/dev */
482 chunk = r10bio->sector >> conf->chunk_shift;
483 sector = r10bio->sector & conf->chunk_mask;
485 chunk *= conf->near_copies;
487 dev = sector_div(stripe, conf->raid_disks);
488 if (conf->far_offset)
489 stripe *= conf->far_copies;
491 sector += stripe << conf->chunk_shift;
493 /* and calculate all the others */
494 for (n=0; n < conf->near_copies; n++) {
497 r10bio->devs[slot].addr = sector;
498 r10bio->devs[slot].devnum = d;
501 for (f = 1; f < conf->far_copies; f++) {
502 d += conf->near_copies;
503 if (d >= conf->raid_disks)
504 d -= conf->raid_disks;
506 r10bio->devs[slot].devnum = d;
507 r10bio->devs[slot].addr = s;
511 if (dev >= conf->raid_disks) {
513 sector += (conf->chunk_mask + 1);
516 BUG_ON(slot != conf->copies);
519 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
521 sector_t offset, chunk, vchunk;
523 offset = sector & conf->chunk_mask;
524 if (conf->far_offset) {
526 chunk = sector >> conf->chunk_shift;
527 fc = sector_div(chunk, conf->far_copies);
528 dev -= fc * conf->near_copies;
530 dev += conf->raid_disks;
532 while (sector >= conf->stride) {
533 sector -= conf->stride;
534 if (dev < conf->near_copies)
535 dev += conf->raid_disks - conf->near_copies;
537 dev -= conf->near_copies;
539 chunk = sector >> conf->chunk_shift;
541 vchunk = chunk * conf->raid_disks + dev;
542 sector_div(vchunk, conf->near_copies);
543 return (vchunk << conf->chunk_shift) + offset;
547 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
549 * @bvm: properties of new bio
550 * @biovec: the request that could be merged to it.
552 * Return amount of bytes we can accept at this offset
553 * If near_copies == raid_disk, there are no striping issues,
554 * but in that case, the function isn't called at all.
556 static int raid10_mergeable_bvec(struct request_queue *q,
557 struct bvec_merge_data *bvm,
558 struct bio_vec *biovec)
560 struct mddev *mddev = q->queuedata;
561 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
563 unsigned int chunk_sectors = mddev->chunk_sectors;
564 unsigned int bio_sectors = bvm->bi_size >> 9;
566 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
567 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
568 if (max <= biovec->bv_len && bio_sectors == 0)
569 return biovec->bv_len;
575 * This routine returns the disk from which the requested read should
576 * be done. There is a per-array 'next expected sequential IO' sector
577 * number - if this matches on the next IO then we use the last disk.
578 * There is also a per-disk 'last know head position' sector that is
579 * maintained from IRQ contexts, both the normal and the resync IO
580 * completion handlers update this position correctly. If there is no
581 * perfect sequential match then we pick the disk whose head is closest.
583 * If there are 2 mirrors in the same 2 devices, performance degrades
584 * because position is mirror, not device based.
586 * The rdev for the device selected will have nr_pending incremented.
590 * FIXME: possibly should rethink readbalancing and do it differently
591 * depending on near_copies / far_copies geometry.
593 static int read_balance(struct r10conf *conf, struct r10bio *r10_bio, int *max_sectors)
595 const sector_t this_sector = r10_bio->sector;
597 int sectors = r10_bio->sectors;
598 int best_good_sectors;
599 sector_t new_distance, best_dist;
600 struct md_rdev *rdev;
604 raid10_find_phys(conf, r10_bio);
607 sectors = r10_bio->sectors;
609 best_dist = MaxSector;
610 best_good_sectors = 0;
613 * Check if we can balance. We can balance on the whole
614 * device if no resync is going on (recovery is ok), or below
615 * the resync window. We take the first readable disk when
616 * above the resync window.
618 if (conf->mddev->recovery_cp < MaxSector
619 && (this_sector + sectors >= conf->next_resync))
622 for (slot = 0; slot < conf->copies ; slot++) {
627 if (r10_bio->devs[slot].bio == IO_BLOCKED)
629 disk = r10_bio->devs[slot].devnum;
630 rdev = rcu_dereference(conf->mirrors[disk].rdev);
633 if (!test_bit(In_sync, &rdev->flags))
636 dev_sector = r10_bio->devs[slot].addr;
637 if (is_badblock(rdev, dev_sector, sectors,
638 &first_bad, &bad_sectors)) {
639 if (best_dist < MaxSector)
640 /* Already have a better slot */
642 if (first_bad <= dev_sector) {
643 /* Cannot read here. If this is the
644 * 'primary' device, then we must not read
645 * beyond 'bad_sectors' from another device.
647 bad_sectors -= (dev_sector - first_bad);
648 if (!do_balance && sectors > bad_sectors)
649 sectors = bad_sectors;
650 if (best_good_sectors > sectors)
651 best_good_sectors = sectors;
653 sector_t good_sectors =
654 first_bad - dev_sector;
655 if (good_sectors > best_good_sectors) {
656 best_good_sectors = good_sectors;
660 /* Must read from here */
665 best_good_sectors = sectors;
670 /* This optimisation is debatable, and completely destroys
671 * sequential read speed for 'far copies' arrays. So only
672 * keep it for 'near' arrays, and review those later.
674 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
677 /* for far > 1 always use the lowest address */
678 if (conf->far_copies > 1)
679 new_distance = r10_bio->devs[slot].addr;
681 new_distance = abs(r10_bio->devs[slot].addr -
682 conf->mirrors[disk].head_position);
683 if (new_distance < best_dist) {
684 best_dist = new_distance;
688 if (slot == conf->copies)
692 disk = r10_bio->devs[slot].devnum;
693 rdev = rcu_dereference(conf->mirrors[disk].rdev);
696 atomic_inc(&rdev->nr_pending);
697 if (test_bit(Faulty, &rdev->flags)) {
698 /* Cannot risk returning a device that failed
699 * before we inc'ed nr_pending
701 rdev_dec_pending(rdev, conf->mddev);
704 r10_bio->read_slot = slot;
708 *max_sectors = best_good_sectors;
713 static int raid10_congested(void *data, int bits)
715 struct mddev *mddev = data;
716 struct r10conf *conf = mddev->private;
719 if ((bits & (1 << BDI_async_congested)) &&
720 conf->pending_count >= max_queued_requests)
723 if (mddev_congested(mddev, bits))
726 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
727 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
728 if (rdev && !test_bit(Faulty, &rdev->flags)) {
729 struct request_queue *q = bdev_get_queue(rdev->bdev);
731 ret |= bdi_congested(&q->backing_dev_info, bits);
738 static void flush_pending_writes(struct r10conf *conf)
740 /* Any writes that have been queued but are awaiting
741 * bitmap updates get flushed here.
743 spin_lock_irq(&conf->device_lock);
745 if (conf->pending_bio_list.head) {
747 bio = bio_list_get(&conf->pending_bio_list);
748 conf->pending_count = 0;
749 spin_unlock_irq(&conf->device_lock);
750 /* flush any pending bitmap writes to disk
751 * before proceeding w/ I/O */
752 bitmap_unplug(conf->mddev->bitmap);
753 wake_up(&conf->wait_barrier);
755 while (bio) { /* submit pending writes */
756 struct bio *next = bio->bi_next;
758 generic_make_request(bio);
762 spin_unlock_irq(&conf->device_lock);
766 * Sometimes we need to suspend IO while we do something else,
767 * either some resync/recovery, or reconfigure the array.
768 * To do this we raise a 'barrier'.
769 * The 'barrier' is a counter that can be raised multiple times
770 * to count how many activities are happening which preclude
772 * We can only raise the barrier if there is no pending IO.
773 * i.e. if nr_pending == 0.
774 * We choose only to raise the barrier if no-one is waiting for the
775 * barrier to go down. This means that as soon as an IO request
776 * is ready, no other operations which require a barrier will start
777 * until the IO request has had a chance.
779 * So: regular IO calls 'wait_barrier'. When that returns there
780 * is no backgroup IO happening, It must arrange to call
781 * allow_barrier when it has finished its IO.
782 * backgroup IO calls must call raise_barrier. Once that returns
783 * there is no normal IO happeing. It must arrange to call
784 * lower_barrier when the particular background IO completes.
787 static void raise_barrier(struct r10conf *conf, int force)
789 BUG_ON(force && !conf->barrier);
790 spin_lock_irq(&conf->resync_lock);
792 /* Wait until no block IO is waiting (unless 'force') */
793 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
794 conf->resync_lock, );
796 /* block any new IO from starting */
799 /* Now wait for all pending IO to complete */
800 wait_event_lock_irq(conf->wait_barrier,
801 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
802 conf->resync_lock, );
804 spin_unlock_irq(&conf->resync_lock);
807 static void lower_barrier(struct r10conf *conf)
810 spin_lock_irqsave(&conf->resync_lock, flags);
812 spin_unlock_irqrestore(&conf->resync_lock, flags);
813 wake_up(&conf->wait_barrier);
816 static void wait_barrier(struct r10conf *conf)
818 spin_lock_irq(&conf->resync_lock);
821 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
827 spin_unlock_irq(&conf->resync_lock);
830 static void allow_barrier(struct r10conf *conf)
833 spin_lock_irqsave(&conf->resync_lock, flags);
835 spin_unlock_irqrestore(&conf->resync_lock, flags);
836 wake_up(&conf->wait_barrier);
839 static void freeze_array(struct r10conf *conf)
841 /* stop syncio and normal IO and wait for everything to
843 * We increment barrier and nr_waiting, and then
844 * wait until nr_pending match nr_queued+1
845 * This is called in the context of one normal IO request
846 * that has failed. Thus any sync request that might be pending
847 * will be blocked by nr_pending, and we need to wait for
848 * pending IO requests to complete or be queued for re-try.
849 * Thus the number queued (nr_queued) plus this request (1)
850 * must match the number of pending IOs (nr_pending) before
853 spin_lock_irq(&conf->resync_lock);
856 wait_event_lock_irq(conf->wait_barrier,
857 conf->nr_pending == conf->nr_queued+1,
859 flush_pending_writes(conf));
861 spin_unlock_irq(&conf->resync_lock);
864 static void unfreeze_array(struct r10conf *conf)
866 /* reverse the effect of the freeze */
867 spin_lock_irq(&conf->resync_lock);
870 wake_up(&conf->wait_barrier);
871 spin_unlock_irq(&conf->resync_lock);
874 static void make_request(struct mddev *mddev, struct bio * bio)
876 struct r10conf *conf = mddev->private;
877 struct mirror_info *mirror;
878 struct r10bio *r10_bio;
879 struct bio *read_bio;
881 int chunk_sects = conf->chunk_mask + 1;
882 const int rw = bio_data_dir(bio);
883 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
884 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
886 struct md_rdev *blocked_rdev;
891 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
892 md_flush_request(mddev, bio);
896 /* If this request crosses a chunk boundary, we need to
897 * split it. This will only happen for 1 PAGE (or less) requests.
899 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
901 conf->near_copies < conf->raid_disks)) {
903 /* Sanity check -- queue functions should prevent this happening */
904 if (bio->bi_vcnt != 1 ||
907 /* This is a one page bio that upper layers
908 * refuse to split for us, so we need to split it.
911 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
913 /* Each of these 'make_request' calls will call 'wait_barrier'.
914 * If the first succeeds but the second blocks due to the resync
915 * thread raising the barrier, we will deadlock because the
916 * IO to the underlying device will be queued in generic_make_request
917 * and will never complete, so will never reduce nr_pending.
918 * So increment nr_waiting here so no new raise_barriers will
919 * succeed, and so the second wait_barrier cannot block.
921 spin_lock_irq(&conf->resync_lock);
923 spin_unlock_irq(&conf->resync_lock);
925 make_request(mddev, &bp->bio1);
926 make_request(mddev, &bp->bio2);
928 spin_lock_irq(&conf->resync_lock);
930 wake_up(&conf->wait_barrier);
931 spin_unlock_irq(&conf->resync_lock);
933 bio_pair_release(bp);
936 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
937 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
938 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
944 md_write_start(mddev, bio);
947 * Register the new request and wait if the reconstruction
948 * thread has put up a bar for new requests.
949 * Continue immediately if no resync is active currently.
953 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
955 r10_bio->master_bio = bio;
956 r10_bio->sectors = bio->bi_size >> 9;
958 r10_bio->mddev = mddev;
959 r10_bio->sector = bio->bi_sector;
962 /* We might need to issue multiple reads to different
963 * devices if there are bad blocks around, so we keep
964 * track of the number of reads in bio->bi_phys_segments.
965 * If this is 0, there is only one r10_bio and no locking
966 * will be needed when the request completes. If it is
967 * non-zero, then it is the number of not-completed requests.
969 bio->bi_phys_segments = 0;
970 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
974 * read balancing logic:
980 disk = read_balance(conf, r10_bio, &max_sectors);
981 slot = r10_bio->read_slot;
983 raid_end_bio_io(r10_bio);
986 mirror = conf->mirrors + disk;
988 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
989 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
992 r10_bio->devs[slot].bio = read_bio;
994 read_bio->bi_sector = r10_bio->devs[slot].addr +
995 mirror->rdev->data_offset;
996 read_bio->bi_bdev = mirror->rdev->bdev;
997 read_bio->bi_end_io = raid10_end_read_request;
998 read_bio->bi_rw = READ | do_sync;
999 read_bio->bi_private = r10_bio;
1001 if (max_sectors < r10_bio->sectors) {
1002 /* Could not read all from this device, so we will
1003 * need another r10_bio.
1005 sectors_handled = (r10_bio->sectors + max_sectors
1007 r10_bio->sectors = max_sectors;
1008 spin_lock_irq(&conf->device_lock);
1009 if (bio->bi_phys_segments == 0)
1010 bio->bi_phys_segments = 2;
1012 bio->bi_phys_segments++;
1013 spin_unlock(&conf->device_lock);
1014 /* Cannot call generic_make_request directly
1015 * as that will be queued in __generic_make_request
1016 * and subsequent mempool_alloc might block
1017 * waiting for it. so hand bio over to raid10d.
1019 reschedule_retry(r10_bio);
1021 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1023 r10_bio->master_bio = bio;
1024 r10_bio->sectors = ((bio->bi_size >> 9)
1027 r10_bio->mddev = mddev;
1028 r10_bio->sector = bio->bi_sector + sectors_handled;
1031 generic_make_request(read_bio);
1038 if (conf->pending_count >= max_queued_requests) {
1039 md_wakeup_thread(mddev->thread);
1040 wait_event(conf->wait_barrier,
1041 conf->pending_count < max_queued_requests);
1043 /* first select target devices under rcu_lock and
1044 * inc refcount on their rdev. Record them by setting
1046 * If there are known/acknowledged bad blocks on any device
1047 * on which we have seen a write error, we want to avoid
1048 * writing to those blocks. This potentially requires several
1049 * writes to write around the bad blocks. Each set of writes
1050 * gets its own r10_bio with a set of bios attached. The number
1051 * of r10_bios is recored in bio->bi_phys_segments just as with
1054 plugged = mddev_check_plugged(mddev);
1056 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1057 raid10_find_phys(conf, r10_bio);
1059 blocked_rdev = NULL;
1061 max_sectors = r10_bio->sectors;
1063 for (i = 0; i < conf->copies; i++) {
1064 int d = r10_bio->devs[i].devnum;
1065 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1066 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1067 atomic_inc(&rdev->nr_pending);
1068 blocked_rdev = rdev;
1071 r10_bio->devs[i].bio = NULL;
1072 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1073 set_bit(R10BIO_Degraded, &r10_bio->state);
1076 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1078 sector_t dev_sector = r10_bio->devs[i].addr;
1082 is_bad = is_badblock(rdev, dev_sector,
1084 &first_bad, &bad_sectors);
1086 /* Mustn't write here until the bad block
1089 atomic_inc(&rdev->nr_pending);
1090 set_bit(BlockedBadBlocks, &rdev->flags);
1091 blocked_rdev = rdev;
1094 if (is_bad && first_bad <= dev_sector) {
1095 /* Cannot write here at all */
1096 bad_sectors -= (dev_sector - first_bad);
1097 if (bad_sectors < max_sectors)
1098 /* Mustn't write more than bad_sectors
1099 * to other devices yet
1101 max_sectors = bad_sectors;
1102 /* We don't set R10BIO_Degraded as that
1103 * only applies if the disk is missing,
1104 * so it might be re-added, and we want to
1105 * know to recover this chunk.
1106 * In this case the device is here, and the
1107 * fact that this chunk is not in-sync is
1108 * recorded in the bad block log.
1113 int good_sectors = first_bad - dev_sector;
1114 if (good_sectors < max_sectors)
1115 max_sectors = good_sectors;
1118 r10_bio->devs[i].bio = bio;
1119 atomic_inc(&rdev->nr_pending);
1123 if (unlikely(blocked_rdev)) {
1124 /* Have to wait for this device to get unblocked, then retry */
1128 for (j = 0; j < i; j++)
1129 if (r10_bio->devs[j].bio) {
1130 d = r10_bio->devs[j].devnum;
1131 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1133 allow_barrier(conf);
1134 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1139 if (max_sectors < r10_bio->sectors) {
1140 /* We are splitting this into multiple parts, so
1141 * we need to prepare for allocating another r10_bio.
1143 r10_bio->sectors = max_sectors;
1144 spin_lock_irq(&conf->device_lock);
1145 if (bio->bi_phys_segments == 0)
1146 bio->bi_phys_segments = 2;
1148 bio->bi_phys_segments++;
1149 spin_unlock_irq(&conf->device_lock);
1151 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1153 atomic_set(&r10_bio->remaining, 1);
1154 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1156 for (i = 0; i < conf->copies; i++) {
1158 int d = r10_bio->devs[i].devnum;
1159 if (!r10_bio->devs[i].bio)
1162 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1163 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1165 r10_bio->devs[i].bio = mbio;
1167 mbio->bi_sector = (r10_bio->devs[i].addr+
1168 conf->mirrors[d].rdev->data_offset);
1169 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1170 mbio->bi_end_io = raid10_end_write_request;
1171 mbio->bi_rw = WRITE | do_sync | do_fua;
1172 mbio->bi_private = r10_bio;
1174 atomic_inc(&r10_bio->remaining);
1175 spin_lock_irqsave(&conf->device_lock, flags);
1176 bio_list_add(&conf->pending_bio_list, mbio);
1177 conf->pending_count++;
1178 spin_unlock_irqrestore(&conf->device_lock, flags);
1181 /* Don't remove the bias on 'remaining' (one_write_done) until
1182 * after checking if we need to go around again.
1185 if (sectors_handled < (bio->bi_size >> 9)) {
1186 one_write_done(r10_bio);
1187 /* We need another r10_bio. It has already been counted
1188 * in bio->bi_phys_segments.
1190 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1192 r10_bio->master_bio = bio;
1193 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1195 r10_bio->mddev = mddev;
1196 r10_bio->sector = bio->bi_sector + sectors_handled;
1200 one_write_done(r10_bio);
1202 /* In case raid10d snuck in to freeze_array */
1203 wake_up(&conf->wait_barrier);
1205 if (do_sync || !mddev->bitmap || !plugged)
1206 md_wakeup_thread(mddev->thread);
1209 static void status(struct seq_file *seq, struct mddev *mddev)
1211 struct r10conf *conf = mddev->private;
1214 if (conf->near_copies < conf->raid_disks)
1215 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1216 if (conf->near_copies > 1)
1217 seq_printf(seq, " %d near-copies", conf->near_copies);
1218 if (conf->far_copies > 1) {
1219 if (conf->far_offset)
1220 seq_printf(seq, " %d offset-copies", conf->far_copies);
1222 seq_printf(seq, " %d far-copies", conf->far_copies);
1224 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1225 conf->raid_disks - mddev->degraded);
1226 for (i = 0; i < conf->raid_disks; i++)
1227 seq_printf(seq, "%s",
1228 conf->mirrors[i].rdev &&
1229 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1230 seq_printf(seq, "]");
1233 /* check if there are enough drives for
1234 * every block to appear on atleast one.
1235 * Don't consider the device numbered 'ignore'
1236 * as we might be about to remove it.
1238 static int enough(struct r10conf *conf, int ignore)
1243 int n = conf->copies;
1246 if (conf->mirrors[first].rdev &&
1249 first = (first+1) % conf->raid_disks;
1253 } while (first != 0);
1257 static void error(struct mddev *mddev, struct md_rdev *rdev)
1259 char b[BDEVNAME_SIZE];
1260 struct r10conf *conf = mddev->private;
1263 * If it is not operational, then we have already marked it as dead
1264 * else if it is the last working disks, ignore the error, let the
1265 * next level up know.
1266 * else mark the drive as failed
1268 if (test_bit(In_sync, &rdev->flags)
1269 && !enough(conf, rdev->raid_disk))
1271 * Don't fail the drive, just return an IO error.
1274 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1275 unsigned long flags;
1276 spin_lock_irqsave(&conf->device_lock, flags);
1278 spin_unlock_irqrestore(&conf->device_lock, flags);
1280 * if recovery is running, make sure it aborts.
1282 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1284 set_bit(Blocked, &rdev->flags);
1285 set_bit(Faulty, &rdev->flags);
1286 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1288 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1289 "md/raid10:%s: Operation continuing on %d devices.\n",
1290 mdname(mddev), bdevname(rdev->bdev, b),
1291 mdname(mddev), conf->raid_disks - mddev->degraded);
1294 static void print_conf(struct r10conf *conf)
1297 struct mirror_info *tmp;
1299 printk(KERN_DEBUG "RAID10 conf printout:\n");
1301 printk(KERN_DEBUG "(!conf)\n");
1304 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1307 for (i = 0; i < conf->raid_disks; i++) {
1308 char b[BDEVNAME_SIZE];
1309 tmp = conf->mirrors + i;
1311 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1312 i, !test_bit(In_sync, &tmp->rdev->flags),
1313 !test_bit(Faulty, &tmp->rdev->flags),
1314 bdevname(tmp->rdev->bdev,b));
1318 static void close_sync(struct r10conf *conf)
1321 allow_barrier(conf);
1323 mempool_destroy(conf->r10buf_pool);
1324 conf->r10buf_pool = NULL;
1327 static int raid10_spare_active(struct mddev *mddev)
1330 struct r10conf *conf = mddev->private;
1331 struct mirror_info *tmp;
1333 unsigned long flags;
1336 * Find all non-in_sync disks within the RAID10 configuration
1337 * and mark them in_sync
1339 for (i = 0; i < conf->raid_disks; i++) {
1340 tmp = conf->mirrors + i;
1342 && !test_bit(Faulty, &tmp->rdev->flags)
1343 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1345 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1348 spin_lock_irqsave(&conf->device_lock, flags);
1349 mddev->degraded -= count;
1350 spin_unlock_irqrestore(&conf->device_lock, flags);
1357 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1359 struct r10conf *conf = mddev->private;
1363 int last = conf->raid_disks - 1;
1365 if (mddev->recovery_cp < MaxSector)
1366 /* only hot-add to in-sync arrays, as recovery is
1367 * very different from resync
1370 if (!enough(conf, -1))
1373 if (rdev->raid_disk >= 0)
1374 first = last = rdev->raid_disk;
1376 if (rdev->saved_raid_disk >= first &&
1377 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1378 mirror = rdev->saved_raid_disk;
1381 for ( ; mirror <= last ; mirror++) {
1382 struct mirror_info *p = &conf->mirrors[mirror];
1383 if (p->recovery_disabled == mddev->recovery_disabled)
1388 disk_stack_limits(mddev->gendisk, rdev->bdev,
1389 rdev->data_offset << 9);
1390 /* as we don't honour merge_bvec_fn, we must
1391 * never risk violating it, so limit
1392 * ->max_segments to one lying with a single
1393 * page, as a one page request is never in
1396 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1397 blk_queue_max_segments(mddev->queue, 1);
1398 blk_queue_segment_boundary(mddev->queue,
1399 PAGE_CACHE_SIZE - 1);
1402 p->head_position = 0;
1403 p->recovery_disabled = mddev->recovery_disabled - 1;
1404 rdev->raid_disk = mirror;
1406 if (rdev->saved_raid_disk != mirror)
1408 rcu_assign_pointer(p->rdev, rdev);
1412 md_integrity_add_rdev(rdev, mddev);
1417 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1419 struct r10conf *conf = mddev->private;
1421 int number = rdev->raid_disk;
1422 struct mirror_info *p = conf->mirrors+ number;
1425 if (rdev == p->rdev) {
1426 if (test_bit(In_sync, &rdev->flags) ||
1427 atomic_read(&rdev->nr_pending)) {
1431 /* Only remove faulty devices in recovery
1434 if (!test_bit(Faulty, &rdev->flags) &&
1435 mddev->recovery_disabled != p->recovery_disabled &&
1442 if (atomic_read(&rdev->nr_pending)) {
1443 /* lost the race, try later */
1448 err = md_integrity_register(mddev);
1457 static void end_sync_read(struct bio *bio, int error)
1459 struct r10bio *r10_bio = bio->bi_private;
1460 struct r10conf *conf = r10_bio->mddev->private;
1463 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1465 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1466 set_bit(R10BIO_Uptodate, &r10_bio->state);
1468 /* The write handler will notice the lack of
1469 * R10BIO_Uptodate and record any errors etc
1471 atomic_add(r10_bio->sectors,
1472 &conf->mirrors[d].rdev->corrected_errors);
1474 /* for reconstruct, we always reschedule after a read.
1475 * for resync, only after all reads
1477 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1478 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1479 atomic_dec_and_test(&r10_bio->remaining)) {
1480 /* we have read all the blocks,
1481 * do the comparison in process context in raid10d
1483 reschedule_retry(r10_bio);
1487 static void end_sync_request(struct r10bio *r10_bio)
1489 struct mddev *mddev = r10_bio->mddev;
1491 while (atomic_dec_and_test(&r10_bio->remaining)) {
1492 if (r10_bio->master_bio == NULL) {
1493 /* the primary of several recovery bios */
1494 sector_t s = r10_bio->sectors;
1495 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1496 test_bit(R10BIO_WriteError, &r10_bio->state))
1497 reschedule_retry(r10_bio);
1500 md_done_sync(mddev, s, 1);
1503 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1504 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1505 test_bit(R10BIO_WriteError, &r10_bio->state))
1506 reschedule_retry(r10_bio);
1514 static void end_sync_write(struct bio *bio, int error)
1516 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1517 struct r10bio *r10_bio = bio->bi_private;
1518 struct mddev *mddev = r10_bio->mddev;
1519 struct r10conf *conf = mddev->private;
1525 d = find_bio_disk(conf, r10_bio, bio, &slot, NULL);
1528 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1529 set_bit(R10BIO_WriteError, &r10_bio->state);
1530 } else if (is_badblock(conf->mirrors[d].rdev,
1531 r10_bio->devs[slot].addr,
1533 &first_bad, &bad_sectors))
1534 set_bit(R10BIO_MadeGood, &r10_bio->state);
1536 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1538 end_sync_request(r10_bio);
1542 * Note: sync and recover and handled very differently for raid10
1543 * This code is for resync.
1544 * For resync, we read through virtual addresses and read all blocks.
1545 * If there is any error, we schedule a write. The lowest numbered
1546 * drive is authoritative.
1547 * However requests come for physical address, so we need to map.
1548 * For every physical address there are raid_disks/copies virtual addresses,
1549 * which is always are least one, but is not necessarly an integer.
1550 * This means that a physical address can span multiple chunks, so we may
1551 * have to submit multiple io requests for a single sync request.
1554 * We check if all blocks are in-sync and only write to blocks that
1557 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1559 struct r10conf *conf = mddev->private;
1561 struct bio *tbio, *fbio;
1563 atomic_set(&r10_bio->remaining, 1);
1565 /* find the first device with a block */
1566 for (i=0; i<conf->copies; i++)
1567 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1570 if (i == conf->copies)
1574 fbio = r10_bio->devs[i].bio;
1576 /* now find blocks with errors */
1577 for (i=0 ; i < conf->copies ; i++) {
1579 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1581 tbio = r10_bio->devs[i].bio;
1583 if (tbio->bi_end_io != end_sync_read)
1587 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1588 /* We know that the bi_io_vec layout is the same for
1589 * both 'first' and 'i', so we just compare them.
1590 * All vec entries are PAGE_SIZE;
1592 for (j = 0; j < vcnt; j++)
1593 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1594 page_address(tbio->bi_io_vec[j].bv_page),
1599 mddev->resync_mismatches += r10_bio->sectors;
1600 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1601 /* Don't fix anything. */
1604 /* Ok, we need to write this bio, either to correct an
1605 * inconsistency or to correct an unreadable block.
1606 * First we need to fixup bv_offset, bv_len and
1607 * bi_vecs, as the read request might have corrupted these
1609 tbio->bi_vcnt = vcnt;
1610 tbio->bi_size = r10_bio->sectors << 9;
1612 tbio->bi_phys_segments = 0;
1613 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1614 tbio->bi_flags |= 1 << BIO_UPTODATE;
1615 tbio->bi_next = NULL;
1616 tbio->bi_rw = WRITE;
1617 tbio->bi_private = r10_bio;
1618 tbio->bi_sector = r10_bio->devs[i].addr;
1620 for (j=0; j < vcnt ; j++) {
1621 tbio->bi_io_vec[j].bv_offset = 0;
1622 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1624 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1625 page_address(fbio->bi_io_vec[j].bv_page),
1628 tbio->bi_end_io = end_sync_write;
1630 d = r10_bio->devs[i].devnum;
1631 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1632 atomic_inc(&r10_bio->remaining);
1633 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1635 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1636 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1637 generic_make_request(tbio);
1641 if (atomic_dec_and_test(&r10_bio->remaining)) {
1642 md_done_sync(mddev, r10_bio->sectors, 1);
1648 * Now for the recovery code.
1649 * Recovery happens across physical sectors.
1650 * We recover all non-is_sync drives by finding the virtual address of
1651 * each, and then choose a working drive that also has that virt address.
1652 * There is a separate r10_bio for each non-in_sync drive.
1653 * Only the first two slots are in use. The first for reading,
1654 * The second for writing.
1657 static void fix_recovery_read_error(struct r10bio *r10_bio)
1659 /* We got a read error during recovery.
1660 * We repeat the read in smaller page-sized sections.
1661 * If a read succeeds, write it to the new device or record
1662 * a bad block if we cannot.
1663 * If a read fails, record a bad block on both old and
1666 struct mddev *mddev = r10_bio->mddev;
1667 struct r10conf *conf = mddev->private;
1668 struct bio *bio = r10_bio->devs[0].bio;
1670 int sectors = r10_bio->sectors;
1672 int dr = r10_bio->devs[0].devnum;
1673 int dw = r10_bio->devs[1].devnum;
1677 struct md_rdev *rdev;
1681 if (s > (PAGE_SIZE>>9))
1684 rdev = conf->mirrors[dr].rdev;
1685 addr = r10_bio->devs[0].addr + sect,
1686 ok = sync_page_io(rdev,
1689 bio->bi_io_vec[idx].bv_page,
1692 rdev = conf->mirrors[dw].rdev;
1693 addr = r10_bio->devs[1].addr + sect;
1694 ok = sync_page_io(rdev,
1697 bio->bi_io_vec[idx].bv_page,
1700 set_bit(WriteErrorSeen, &rdev->flags);
1703 /* We don't worry if we cannot set a bad block -
1704 * it really is bad so there is no loss in not
1707 rdev_set_badblocks(rdev, addr, s, 0);
1709 if (rdev != conf->mirrors[dw].rdev) {
1710 /* need bad block on destination too */
1711 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
1712 addr = r10_bio->devs[1].addr + sect;
1713 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1715 /* just abort the recovery */
1717 "md/raid10:%s: recovery aborted"
1718 " due to read error\n",
1721 conf->mirrors[dw].recovery_disabled
1722 = mddev->recovery_disabled;
1723 set_bit(MD_RECOVERY_INTR,
1736 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1738 struct r10conf *conf = mddev->private;
1742 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1743 fix_recovery_read_error(r10_bio);
1744 end_sync_request(r10_bio);
1749 * share the pages with the first bio
1750 * and submit the write request
1752 wbio = r10_bio->devs[1].bio;
1753 d = r10_bio->devs[1].devnum;
1755 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1756 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1757 generic_make_request(wbio);
1762 * Used by fix_read_error() to decay the per rdev read_errors.
1763 * We halve the read error count for every hour that has elapsed
1764 * since the last recorded read error.
1767 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
1769 struct timespec cur_time_mon;
1770 unsigned long hours_since_last;
1771 unsigned int read_errors = atomic_read(&rdev->read_errors);
1773 ktime_get_ts(&cur_time_mon);
1775 if (rdev->last_read_error.tv_sec == 0 &&
1776 rdev->last_read_error.tv_nsec == 0) {
1777 /* first time we've seen a read error */
1778 rdev->last_read_error = cur_time_mon;
1782 hours_since_last = (cur_time_mon.tv_sec -
1783 rdev->last_read_error.tv_sec) / 3600;
1785 rdev->last_read_error = cur_time_mon;
1788 * if hours_since_last is > the number of bits in read_errors
1789 * just set read errors to 0. We do this to avoid
1790 * overflowing the shift of read_errors by hours_since_last.
1792 if (hours_since_last >= 8 * sizeof(read_errors))
1793 atomic_set(&rdev->read_errors, 0);
1795 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1798 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
1799 int sectors, struct page *page, int rw)
1804 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1805 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1807 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1811 set_bit(WriteErrorSeen, &rdev->flags);
1812 /* need to record an error - either for the block or the device */
1813 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1814 md_error(rdev->mddev, rdev);
1819 * This is a kernel thread which:
1821 * 1. Retries failed read operations on working mirrors.
1822 * 2. Updates the raid superblock when problems encounter.
1823 * 3. Performs writes following reads for array synchronising.
1826 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
1828 int sect = 0; /* Offset from r10_bio->sector */
1829 int sectors = r10_bio->sectors;
1830 struct md_rdev*rdev;
1831 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1832 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1834 /* still own a reference to this rdev, so it cannot
1835 * have been cleared recently.
1837 rdev = conf->mirrors[d].rdev;
1839 if (test_bit(Faulty, &rdev->flags))
1840 /* drive has already been failed, just ignore any
1841 more fix_read_error() attempts */
1844 check_decay_read_errors(mddev, rdev);
1845 atomic_inc(&rdev->read_errors);
1846 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1847 char b[BDEVNAME_SIZE];
1848 bdevname(rdev->bdev, b);
1851 "md/raid10:%s: %s: Raid device exceeded "
1852 "read_error threshold [cur %d:max %d]\n",
1854 atomic_read(&rdev->read_errors), max_read_errors);
1856 "md/raid10:%s: %s: Failing raid device\n",
1858 md_error(mddev, conf->mirrors[d].rdev);
1864 int sl = r10_bio->read_slot;
1868 if (s > (PAGE_SIZE>>9))
1876 d = r10_bio->devs[sl].devnum;
1877 rdev = rcu_dereference(conf->mirrors[d].rdev);
1879 test_bit(In_sync, &rdev->flags) &&
1880 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1881 &first_bad, &bad_sectors) == 0) {
1882 atomic_inc(&rdev->nr_pending);
1884 success = sync_page_io(rdev,
1885 r10_bio->devs[sl].addr +
1888 conf->tmppage, READ, false);
1889 rdev_dec_pending(rdev, mddev);
1895 if (sl == conf->copies)
1897 } while (!success && sl != r10_bio->read_slot);
1901 /* Cannot read from anywhere, just mark the block
1902 * as bad on the first device to discourage future
1905 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1906 rdev = conf->mirrors[dn].rdev;
1908 if (!rdev_set_badblocks(
1910 r10_bio->devs[r10_bio->read_slot].addr
1913 md_error(mddev, rdev);
1918 /* write it back and re-read */
1920 while (sl != r10_bio->read_slot) {
1921 char b[BDEVNAME_SIZE];
1926 d = r10_bio->devs[sl].devnum;
1927 rdev = rcu_dereference(conf->mirrors[d].rdev);
1929 !test_bit(In_sync, &rdev->flags))
1932 atomic_inc(&rdev->nr_pending);
1934 if (r10_sync_page_io(rdev,
1935 r10_bio->devs[sl].addr +
1937 s<<9, conf->tmppage, WRITE)
1939 /* Well, this device is dead */
1941 "md/raid10:%s: read correction "
1943 " (%d sectors at %llu on %s)\n",
1945 (unsigned long long)(
1946 sect + rdev->data_offset),
1947 bdevname(rdev->bdev, b));
1948 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1951 bdevname(rdev->bdev, b));
1953 rdev_dec_pending(rdev, mddev);
1957 while (sl != r10_bio->read_slot) {
1958 char b[BDEVNAME_SIZE];
1963 d = r10_bio->devs[sl].devnum;
1964 rdev = rcu_dereference(conf->mirrors[d].rdev);
1966 !test_bit(In_sync, &rdev->flags))
1969 atomic_inc(&rdev->nr_pending);
1971 switch (r10_sync_page_io(rdev,
1972 r10_bio->devs[sl].addr +
1974 s<<9, conf->tmppage,
1977 /* Well, this device is dead */
1979 "md/raid10:%s: unable to read back "
1981 " (%d sectors at %llu on %s)\n",
1983 (unsigned long long)(
1984 sect + rdev->data_offset),
1985 bdevname(rdev->bdev, b));
1986 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1989 bdevname(rdev->bdev, b));
1993 "md/raid10:%s: read error corrected"
1994 " (%d sectors at %llu on %s)\n",
1996 (unsigned long long)(
1997 sect + rdev->data_offset),
1998 bdevname(rdev->bdev, b));
1999 atomic_add(s, &rdev->corrected_errors);
2002 rdev_dec_pending(rdev, mddev);
2012 static void bi_complete(struct bio *bio, int error)
2014 complete((struct completion *)bio->bi_private);
2017 static int submit_bio_wait(int rw, struct bio *bio)
2019 struct completion event;
2022 init_completion(&event);
2023 bio->bi_private = &event;
2024 bio->bi_end_io = bi_complete;
2025 submit_bio(rw, bio);
2026 wait_for_completion(&event);
2028 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2031 static int narrow_write_error(struct r10bio *r10_bio, int i)
2033 struct bio *bio = r10_bio->master_bio;
2034 struct mddev *mddev = r10_bio->mddev;
2035 struct r10conf *conf = mddev->private;
2036 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2037 /* bio has the data to be written to slot 'i' where
2038 * we just recently had a write error.
2039 * We repeatedly clone the bio and trim down to one block,
2040 * then try the write. Where the write fails we record
2042 * It is conceivable that the bio doesn't exactly align with
2043 * blocks. We must handle this.
2045 * We currently own a reference to the rdev.
2051 int sect_to_write = r10_bio->sectors;
2054 if (rdev->badblocks.shift < 0)
2057 block_sectors = 1 << rdev->badblocks.shift;
2058 sector = r10_bio->sector;
2059 sectors = ((r10_bio->sector + block_sectors)
2060 & ~(sector_t)(block_sectors - 1))
2063 while (sect_to_write) {
2065 if (sectors > sect_to_write)
2066 sectors = sect_to_write;
2067 /* Write at 'sector' for 'sectors' */
2068 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2069 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2070 wbio->bi_sector = (r10_bio->devs[i].addr+
2072 (sector - r10_bio->sector));
2073 wbio->bi_bdev = rdev->bdev;
2074 if (submit_bio_wait(WRITE, wbio) == 0)
2076 ok = rdev_set_badblocks(rdev, sector,
2081 sect_to_write -= sectors;
2083 sectors = block_sectors;
2088 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2090 int slot = r10_bio->read_slot;
2091 int mirror = r10_bio->devs[slot].devnum;
2093 struct r10conf *conf = mddev->private;
2094 struct md_rdev *rdev;
2095 char b[BDEVNAME_SIZE];
2096 unsigned long do_sync;
2099 /* we got a read error. Maybe the drive is bad. Maybe just
2100 * the block and we can fix it.
2101 * We freeze all other IO, and try reading the block from
2102 * other devices. When we find one, we re-write
2103 * and check it that fixes the read error.
2104 * This is all done synchronously while the array is
2107 if (mddev->ro == 0) {
2109 fix_read_error(conf, mddev, r10_bio);
2110 unfreeze_array(conf);
2112 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2114 bio = r10_bio->devs[slot].bio;
2115 bdevname(bio->bi_bdev, b);
2116 r10_bio->devs[slot].bio =
2117 mddev->ro ? IO_BLOCKED : NULL;
2119 mirror = read_balance(conf, r10_bio, &max_sectors);
2121 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2122 " read error for block %llu\n",
2124 (unsigned long long)r10_bio->sector);
2125 raid_end_bio_io(r10_bio);
2130 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2133 slot = r10_bio->read_slot;
2134 rdev = conf->mirrors[mirror].rdev;
2137 "md/raid10:%s: %s: redirecting"
2138 "sector %llu to another mirror\n",
2140 bdevname(rdev->bdev, b),
2141 (unsigned long long)r10_bio->sector);
2142 bio = bio_clone_mddev(r10_bio->master_bio,
2145 r10_bio->sector - bio->bi_sector,
2147 r10_bio->devs[slot].bio = bio;
2148 bio->bi_sector = r10_bio->devs[slot].addr
2149 + rdev->data_offset;
2150 bio->bi_bdev = rdev->bdev;
2151 bio->bi_rw = READ | do_sync;
2152 bio->bi_private = r10_bio;
2153 bio->bi_end_io = raid10_end_read_request;
2154 if (max_sectors < r10_bio->sectors) {
2155 /* Drat - have to split this up more */
2156 struct bio *mbio = r10_bio->master_bio;
2157 int sectors_handled =
2158 r10_bio->sector + max_sectors
2160 r10_bio->sectors = max_sectors;
2161 spin_lock_irq(&conf->device_lock);
2162 if (mbio->bi_phys_segments == 0)
2163 mbio->bi_phys_segments = 2;
2165 mbio->bi_phys_segments++;
2166 spin_unlock_irq(&conf->device_lock);
2167 generic_make_request(bio);
2170 r10_bio = mempool_alloc(conf->r10bio_pool,
2172 r10_bio->master_bio = mbio;
2173 r10_bio->sectors = (mbio->bi_size >> 9)
2176 set_bit(R10BIO_ReadError,
2178 r10_bio->mddev = mddev;
2179 r10_bio->sector = mbio->bi_sector
2184 generic_make_request(bio);
2187 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2189 /* Some sort of write request has finished and it
2190 * succeeded in writing where we thought there was a
2191 * bad block. So forget the bad block.
2192 * Or possibly if failed and we need to record
2196 struct md_rdev *rdev;
2198 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2199 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2200 for (m = 0; m < conf->copies; m++) {
2201 int dev = r10_bio->devs[m].devnum;
2202 rdev = conf->mirrors[dev].rdev;
2203 if (r10_bio->devs[m].bio == NULL)
2205 if (test_bit(BIO_UPTODATE,
2206 &r10_bio->devs[m].bio->bi_flags)) {
2207 rdev_clear_badblocks(
2209 r10_bio->devs[m].addr,
2212 if (!rdev_set_badblocks(
2214 r10_bio->devs[m].addr,
2215 r10_bio->sectors, 0))
2216 md_error(conf->mddev, rdev);
2221 for (m = 0; m < conf->copies; m++) {
2222 int dev = r10_bio->devs[m].devnum;
2223 struct bio *bio = r10_bio->devs[m].bio;
2224 rdev = conf->mirrors[dev].rdev;
2225 if (bio == IO_MADE_GOOD) {
2226 rdev_clear_badblocks(
2228 r10_bio->devs[m].addr,
2230 rdev_dec_pending(rdev, conf->mddev);
2231 } else if (bio != NULL &&
2232 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2233 if (!narrow_write_error(r10_bio, m)) {
2234 md_error(conf->mddev, rdev);
2235 set_bit(R10BIO_Degraded,
2238 rdev_dec_pending(rdev, conf->mddev);
2241 if (test_bit(R10BIO_WriteError,
2243 close_write(r10_bio);
2244 raid_end_bio_io(r10_bio);
2248 static void raid10d(struct mddev *mddev)
2250 struct r10bio *r10_bio;
2251 unsigned long flags;
2252 struct r10conf *conf = mddev->private;
2253 struct list_head *head = &conf->retry_list;
2254 struct blk_plug plug;
2256 md_check_recovery(mddev);
2258 blk_start_plug(&plug);
2261 flush_pending_writes(conf);
2263 spin_lock_irqsave(&conf->device_lock, flags);
2264 if (list_empty(head)) {
2265 spin_unlock_irqrestore(&conf->device_lock, flags);
2268 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2269 list_del(head->prev);
2271 spin_unlock_irqrestore(&conf->device_lock, flags);
2273 mddev = r10_bio->mddev;
2274 conf = mddev->private;
2275 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2276 test_bit(R10BIO_WriteError, &r10_bio->state))
2277 handle_write_completed(conf, r10_bio);
2278 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2279 sync_request_write(mddev, r10_bio);
2280 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2281 recovery_request_write(mddev, r10_bio);
2282 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2283 handle_read_error(mddev, r10_bio);
2285 /* just a partial read to be scheduled from a
2288 int slot = r10_bio->read_slot;
2289 generic_make_request(r10_bio->devs[slot].bio);
2293 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2294 md_check_recovery(mddev);
2296 blk_finish_plug(&plug);
2300 static int init_resync(struct r10conf *conf)
2305 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2306 BUG_ON(conf->r10buf_pool);
2307 conf->have_replacement = 0;
2308 for (i = 0; i < conf->raid_disks; i++)
2309 if (conf->mirrors[i].replacement)
2310 conf->have_replacement = 1;
2311 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2312 if (!conf->r10buf_pool)
2314 conf->next_resync = 0;
2319 * perform a "sync" on one "block"
2321 * We need to make sure that no normal I/O request - particularly write
2322 * requests - conflict with active sync requests.
2324 * This is achieved by tracking pending requests and a 'barrier' concept
2325 * that can be installed to exclude normal IO requests.
2327 * Resync and recovery are handled very differently.
2328 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2330 * For resync, we iterate over virtual addresses, read all copies,
2331 * and update if there are differences. If only one copy is live,
2333 * For recovery, we iterate over physical addresses, read a good
2334 * value for each non-in_sync drive, and over-write.
2336 * So, for recovery we may have several outstanding complex requests for a
2337 * given address, one for each out-of-sync device. We model this by allocating
2338 * a number of r10_bio structures, one for each out-of-sync device.
2339 * As we setup these structures, we collect all bio's together into a list
2340 * which we then process collectively to add pages, and then process again
2341 * to pass to generic_make_request.
2343 * The r10_bio structures are linked using a borrowed master_bio pointer.
2344 * This link is counted in ->remaining. When the r10_bio that points to NULL
2345 * has its remaining count decremented to 0, the whole complex operation
2350 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2351 int *skipped, int go_faster)
2353 struct r10conf *conf = mddev->private;
2354 struct r10bio *r10_bio;
2355 struct bio *biolist = NULL, *bio;
2356 sector_t max_sector, nr_sectors;
2359 sector_t sync_blocks;
2360 sector_t sectors_skipped = 0;
2361 int chunks_skipped = 0;
2363 if (!conf->r10buf_pool)
2364 if (init_resync(conf))
2368 max_sector = mddev->dev_sectors;
2369 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2370 max_sector = mddev->resync_max_sectors;
2371 if (sector_nr >= max_sector) {
2372 /* If we aborted, we need to abort the
2373 * sync on the 'current' bitmap chucks (there can
2374 * be several when recovering multiple devices).
2375 * as we may have started syncing it but not finished.
2376 * We can find the current address in
2377 * mddev->curr_resync, but for recovery,
2378 * we need to convert that to several
2379 * virtual addresses.
2381 if (mddev->curr_resync < max_sector) { /* aborted */
2382 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2383 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2385 else for (i=0; i<conf->raid_disks; i++) {
2387 raid10_find_virt(conf, mddev->curr_resync, i);
2388 bitmap_end_sync(mddev->bitmap, sect,
2391 } else /* completed sync */
2394 bitmap_close_sync(mddev->bitmap);
2397 return sectors_skipped;
2399 if (chunks_skipped >= conf->raid_disks) {
2400 /* if there has been nothing to do on any drive,
2401 * then there is nothing to do at all..
2404 return (max_sector - sector_nr) + sectors_skipped;
2407 if (max_sector > mddev->resync_max)
2408 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2410 /* make sure whole request will fit in a chunk - if chunks
2413 if (conf->near_copies < conf->raid_disks &&
2414 max_sector > (sector_nr | conf->chunk_mask))
2415 max_sector = (sector_nr | conf->chunk_mask) + 1;
2417 * If there is non-resync activity waiting for us then
2418 * put in a delay to throttle resync.
2420 if (!go_faster && conf->nr_waiting)
2421 msleep_interruptible(1000);
2423 /* Again, very different code for resync and recovery.
2424 * Both must result in an r10bio with a list of bios that
2425 * have bi_end_io, bi_sector, bi_bdev set,
2426 * and bi_private set to the r10bio.
2427 * For recovery, we may actually create several r10bios
2428 * with 2 bios in each, that correspond to the bios in the main one.
2429 * In this case, the subordinate r10bios link back through a
2430 * borrowed master_bio pointer, and the counter in the master
2431 * includes a ref from each subordinate.
2433 /* First, we decide what to do and set ->bi_end_io
2434 * To end_sync_read if we want to read, and
2435 * end_sync_write if we will want to write.
2438 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2439 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2440 /* recovery... the complicated one */
2444 for (i=0 ; i<conf->raid_disks; i++) {
2451 if (conf->mirrors[i].rdev == NULL ||
2452 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
2456 /* want to reconstruct this device */
2458 sect = raid10_find_virt(conf, sector_nr, i);
2459 /* Unless we are doing a full sync, we only need
2460 * to recover the block if it is set in the bitmap
2462 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2464 if (sync_blocks < max_sync)
2465 max_sync = sync_blocks;
2468 /* yep, skip the sync_blocks here, but don't assume
2469 * that there will never be anything to do here
2471 chunks_skipped = -1;
2475 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2476 raise_barrier(conf, rb2 != NULL);
2477 atomic_set(&r10_bio->remaining, 0);
2479 r10_bio->master_bio = (struct bio*)rb2;
2481 atomic_inc(&rb2->remaining);
2482 r10_bio->mddev = mddev;
2483 set_bit(R10BIO_IsRecover, &r10_bio->state);
2484 r10_bio->sector = sect;
2486 raid10_find_phys(conf, r10_bio);
2488 /* Need to check if the array will still be
2491 for (j=0; j<conf->raid_disks; j++)
2492 if (conf->mirrors[j].rdev == NULL ||
2493 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2498 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2499 &sync_blocks, still_degraded);
2502 for (j=0; j<conf->copies;j++) {
2504 int d = r10_bio->devs[j].devnum;
2505 sector_t from_addr, to_addr;
2506 struct md_rdev *rdev;
2507 sector_t sector, first_bad;
2509 if (!conf->mirrors[d].rdev ||
2510 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2512 /* This is where we read from */
2514 rdev = conf->mirrors[d].rdev;
2515 sector = r10_bio->devs[j].addr;
2517 if (is_badblock(rdev, sector, max_sync,
2518 &first_bad, &bad_sectors)) {
2519 if (first_bad > sector)
2520 max_sync = first_bad - sector;
2522 bad_sectors -= (sector
2524 if (max_sync > bad_sectors)
2525 max_sync = bad_sectors;
2529 bio = r10_bio->devs[0].bio;
2530 bio->bi_next = biolist;
2532 bio->bi_private = r10_bio;
2533 bio->bi_end_io = end_sync_read;
2535 from_addr = r10_bio->devs[j].addr;
2536 bio->bi_sector = from_addr +
2537 conf->mirrors[d].rdev->data_offset;
2538 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2539 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2540 atomic_inc(&r10_bio->remaining);
2541 /* and we write to 'i' */
2543 for (k=0; k<conf->copies; k++)
2544 if (r10_bio->devs[k].devnum == i)
2546 BUG_ON(k == conf->copies);
2547 bio = r10_bio->devs[1].bio;
2548 bio->bi_next = biolist;
2550 bio->bi_private = r10_bio;
2551 bio->bi_end_io = end_sync_write;
2553 to_addr = r10_bio->devs[k].addr;
2554 bio->bi_sector = to_addr +
2555 conf->mirrors[i].rdev->data_offset;
2556 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2558 r10_bio->devs[0].devnum = d;
2559 r10_bio->devs[0].addr = from_addr;
2560 r10_bio->devs[1].devnum = i;
2561 r10_bio->devs[1].addr = to_addr;
2565 if (j == conf->copies) {
2566 /* Cannot recover, so abort the recovery or
2567 * record a bad block */
2570 atomic_dec(&rb2->remaining);
2573 /* problem is that there are bad blocks
2574 * on other device(s)
2577 for (k = 0; k < conf->copies; k++)
2578 if (r10_bio->devs[k].devnum == i)
2580 if (!rdev_set_badblocks(
2581 conf->mirrors[i].rdev,
2582 r10_bio->devs[k].addr,
2587 if (!test_and_set_bit(MD_RECOVERY_INTR,
2589 printk(KERN_INFO "md/raid10:%s: insufficient "
2590 "working devices for recovery.\n",
2592 conf->mirrors[i].recovery_disabled
2593 = mddev->recovery_disabled;
2598 if (biolist == NULL) {
2600 struct r10bio *rb2 = r10_bio;
2601 r10_bio = (struct r10bio*) rb2->master_bio;
2602 rb2->master_bio = NULL;
2608 /* resync. Schedule a read for every block at this virt offset */
2611 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2613 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2614 &sync_blocks, mddev->degraded) &&
2615 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2616 &mddev->recovery)) {
2617 /* We can skip this block */
2619 return sync_blocks + sectors_skipped;
2621 if (sync_blocks < max_sync)
2622 max_sync = sync_blocks;
2623 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2625 r10_bio->mddev = mddev;
2626 atomic_set(&r10_bio->remaining, 0);
2627 raise_barrier(conf, 0);
2628 conf->next_resync = sector_nr;
2630 r10_bio->master_bio = NULL;
2631 r10_bio->sector = sector_nr;
2632 set_bit(R10BIO_IsSync, &r10_bio->state);
2633 raid10_find_phys(conf, r10_bio);
2634 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2636 for (i=0; i<conf->copies; i++) {
2637 int d = r10_bio->devs[i].devnum;
2638 sector_t first_bad, sector;
2641 bio = r10_bio->devs[i].bio;
2642 bio->bi_end_io = NULL;
2643 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2644 if (conf->mirrors[d].rdev == NULL ||
2645 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2647 sector = r10_bio->devs[i].addr;
2648 if (is_badblock(conf->mirrors[d].rdev,
2650 &first_bad, &bad_sectors)) {
2651 if (first_bad > sector)
2652 max_sync = first_bad - sector;
2654 bad_sectors -= (sector - first_bad);
2655 if (max_sync > bad_sectors)
2656 max_sync = max_sync;
2660 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2661 atomic_inc(&r10_bio->remaining);
2662 bio->bi_next = biolist;
2664 bio->bi_private = r10_bio;
2665 bio->bi_end_io = end_sync_read;
2667 bio->bi_sector = sector +
2668 conf->mirrors[d].rdev->data_offset;
2669 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2674 for (i=0; i<conf->copies; i++) {
2675 int d = r10_bio->devs[i].devnum;
2676 if (r10_bio->devs[i].bio->bi_end_io)
2677 rdev_dec_pending(conf->mirrors[d].rdev,
2686 for (bio = biolist; bio ; bio=bio->bi_next) {
2688 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2690 bio->bi_flags |= 1 << BIO_UPTODATE;
2693 bio->bi_phys_segments = 0;
2698 if (sector_nr + max_sync < max_sector)
2699 max_sector = sector_nr + max_sync;
2702 int len = PAGE_SIZE;
2703 if (sector_nr + (len>>9) > max_sector)
2704 len = (max_sector - sector_nr) << 9;
2707 for (bio= biolist ; bio ; bio=bio->bi_next) {
2709 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2710 if (bio_add_page(bio, page, len, 0))
2714 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2715 for (bio2 = biolist;
2716 bio2 && bio2 != bio;
2717 bio2 = bio2->bi_next) {
2718 /* remove last page from this bio */
2720 bio2->bi_size -= len;
2721 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2725 nr_sectors += len>>9;
2726 sector_nr += len>>9;
2727 } while (biolist->bi_vcnt < RESYNC_PAGES);
2729 r10_bio->sectors = nr_sectors;
2733 biolist = biolist->bi_next;
2735 bio->bi_next = NULL;
2736 r10_bio = bio->bi_private;
2737 r10_bio->sectors = nr_sectors;
2739 if (bio->bi_end_io == end_sync_read) {
2740 md_sync_acct(bio->bi_bdev, nr_sectors);
2741 generic_make_request(bio);
2745 if (sectors_skipped)
2746 /* pretend they weren't skipped, it makes
2747 * no important difference in this case
2749 md_done_sync(mddev, sectors_skipped, 1);
2751 return sectors_skipped + nr_sectors;
2753 /* There is nowhere to write, so all non-sync
2754 * drives must be failed or in resync, all drives
2755 * have a bad block, so try the next chunk...
2757 if (sector_nr + max_sync < max_sector)
2758 max_sector = sector_nr + max_sync;
2760 sectors_skipped += (max_sector - sector_nr);
2762 sector_nr = max_sector;
2767 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2770 struct r10conf *conf = mddev->private;
2773 raid_disks = conf->raid_disks;
2775 sectors = conf->dev_sectors;
2777 size = sectors >> conf->chunk_shift;
2778 sector_div(size, conf->far_copies);
2779 size = size * raid_disks;
2780 sector_div(size, conf->near_copies);
2782 return size << conf->chunk_shift;
2786 static struct r10conf *setup_conf(struct mddev *mddev)
2788 struct r10conf *conf = NULL;
2790 sector_t stride, size;
2793 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2794 !is_power_of_2(mddev->new_chunk_sectors)) {
2795 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2796 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2797 mdname(mddev), PAGE_SIZE);
2801 nc = mddev->new_layout & 255;
2802 fc = (mddev->new_layout >> 8) & 255;
2803 fo = mddev->new_layout & (1<<16);
2805 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2806 (mddev->new_layout >> 17)) {
2807 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2808 mdname(mddev), mddev->new_layout);
2813 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
2817 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2822 conf->tmppage = alloc_page(GFP_KERNEL);
2827 conf->raid_disks = mddev->raid_disks;
2828 conf->near_copies = nc;
2829 conf->far_copies = fc;
2830 conf->copies = nc*fc;
2831 conf->far_offset = fo;
2832 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2833 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2835 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2836 r10bio_pool_free, conf);
2837 if (!conf->r10bio_pool)
2840 size = mddev->dev_sectors >> conf->chunk_shift;
2841 sector_div(size, fc);
2842 size = size * conf->raid_disks;
2843 sector_div(size, nc);
2844 /* 'size' is now the number of chunks in the array */
2845 /* calculate "used chunks per device" in 'stride' */
2846 stride = size * conf->copies;
2848 /* We need to round up when dividing by raid_disks to
2849 * get the stride size.
2851 stride += conf->raid_disks - 1;
2852 sector_div(stride, conf->raid_disks);
2854 conf->dev_sectors = stride << conf->chunk_shift;
2859 sector_div(stride, fc);
2860 conf->stride = stride << conf->chunk_shift;
2863 spin_lock_init(&conf->device_lock);
2864 INIT_LIST_HEAD(&conf->retry_list);
2866 spin_lock_init(&conf->resync_lock);
2867 init_waitqueue_head(&conf->wait_barrier);
2869 conf->thread = md_register_thread(raid10d, mddev, NULL);
2873 conf->mddev = mddev;
2877 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2880 if (conf->r10bio_pool)
2881 mempool_destroy(conf->r10bio_pool);
2882 kfree(conf->mirrors);
2883 safe_put_page(conf->tmppage);
2886 return ERR_PTR(err);
2889 static int run(struct mddev *mddev)
2891 struct r10conf *conf;
2892 int i, disk_idx, chunk_size;
2893 struct mirror_info *disk;
2894 struct md_rdev *rdev;
2898 * copy the already verified devices into our private RAID10
2899 * bookkeeping area. [whatever we allocate in run(),
2900 * should be freed in stop()]
2903 if (mddev->private == NULL) {
2904 conf = setup_conf(mddev);
2906 return PTR_ERR(conf);
2907 mddev->private = conf;
2909 conf = mddev->private;
2913 mddev->thread = conf->thread;
2914 conf->thread = NULL;
2916 chunk_size = mddev->chunk_sectors << 9;
2917 blk_queue_io_min(mddev->queue, chunk_size);
2918 if (conf->raid_disks % conf->near_copies)
2919 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2921 blk_queue_io_opt(mddev->queue, chunk_size *
2922 (conf->raid_disks / conf->near_copies));
2924 list_for_each_entry(rdev, &mddev->disks, same_set) {
2926 disk_idx = rdev->raid_disk;
2927 if (disk_idx >= conf->raid_disks
2930 disk = conf->mirrors + disk_idx;
2933 disk_stack_limits(mddev->gendisk, rdev->bdev,
2934 rdev->data_offset << 9);
2935 /* as we don't honour merge_bvec_fn, we must never risk
2936 * violating it, so limit max_segments to 1 lying
2937 * within a single page.
2939 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2940 blk_queue_max_segments(mddev->queue, 1);
2941 blk_queue_segment_boundary(mddev->queue,
2942 PAGE_CACHE_SIZE - 1);
2945 disk->head_position = 0;
2947 /* need to check that every block has at least one working mirror */
2948 if (!enough(conf, -1)) {
2949 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2954 mddev->degraded = 0;
2955 for (i = 0; i < conf->raid_disks; i++) {
2957 disk = conf->mirrors + i;
2960 !test_bit(In_sync, &disk->rdev->flags)) {
2961 disk->head_position = 0;
2966 disk->recovery_disabled = mddev->recovery_disabled - 1;
2969 if (mddev->recovery_cp != MaxSector)
2970 printk(KERN_NOTICE "md/raid10:%s: not clean"
2971 " -- starting background reconstruction\n",
2974 "md/raid10:%s: active with %d out of %d devices\n",
2975 mdname(mddev), conf->raid_disks - mddev->degraded,
2978 * Ok, everything is just fine now
2980 mddev->dev_sectors = conf->dev_sectors;
2981 size = raid10_size(mddev, 0, 0);
2982 md_set_array_sectors(mddev, size);
2983 mddev->resync_max_sectors = size;
2985 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2986 mddev->queue->backing_dev_info.congested_data = mddev;
2988 /* Calculate max read-ahead size.
2989 * We need to readahead at least twice a whole stripe....
2993 int stripe = conf->raid_disks *
2994 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2995 stripe /= conf->near_copies;
2996 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2997 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
3000 if (conf->near_copies < conf->raid_disks)
3001 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3003 if (md_integrity_register(mddev))
3009 md_unregister_thread(&mddev->thread);
3010 if (conf->r10bio_pool)
3011 mempool_destroy(conf->r10bio_pool);
3012 safe_put_page(conf->tmppage);
3013 kfree(conf->mirrors);
3015 mddev->private = NULL;
3020 static int stop(struct mddev *mddev)
3022 struct r10conf *conf = mddev->private;
3024 raise_barrier(conf, 0);
3025 lower_barrier(conf);
3027 md_unregister_thread(&mddev->thread);
3028 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3029 if (conf->r10bio_pool)
3030 mempool_destroy(conf->r10bio_pool);
3031 kfree(conf->mirrors);
3033 mddev->private = NULL;
3037 static void raid10_quiesce(struct mddev *mddev, int state)
3039 struct r10conf *conf = mddev->private;
3043 raise_barrier(conf, 0);
3046 lower_barrier(conf);
3051 static void *raid10_takeover_raid0(struct mddev *mddev)
3053 struct md_rdev *rdev;
3054 struct r10conf *conf;
3056 if (mddev->degraded > 0) {
3057 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3059 return ERR_PTR(-EINVAL);
3062 /* Set new parameters */
3063 mddev->new_level = 10;
3064 /* new layout: far_copies = 1, near_copies = 2 */
3065 mddev->new_layout = (1<<8) + 2;
3066 mddev->new_chunk_sectors = mddev->chunk_sectors;
3067 mddev->delta_disks = mddev->raid_disks;
3068 mddev->raid_disks *= 2;
3069 /* make sure it will be not marked as dirty */
3070 mddev->recovery_cp = MaxSector;
3072 conf = setup_conf(mddev);
3073 if (!IS_ERR(conf)) {
3074 list_for_each_entry(rdev, &mddev->disks, same_set)
3075 if (rdev->raid_disk >= 0)
3076 rdev->new_raid_disk = rdev->raid_disk * 2;
3083 static void *raid10_takeover(struct mddev *mddev)
3085 struct r0conf *raid0_conf;
3087 /* raid10 can take over:
3088 * raid0 - providing it has only two drives
3090 if (mddev->level == 0) {
3091 /* for raid0 takeover only one zone is supported */
3092 raid0_conf = mddev->private;
3093 if (raid0_conf->nr_strip_zones > 1) {
3094 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3095 " with more than one zone.\n",
3097 return ERR_PTR(-EINVAL);
3099 return raid10_takeover_raid0(mddev);
3101 return ERR_PTR(-EINVAL);
3104 static struct md_personality raid10_personality =
3108 .owner = THIS_MODULE,
3109 .make_request = make_request,
3113 .error_handler = error,
3114 .hot_add_disk = raid10_add_disk,
3115 .hot_remove_disk= raid10_remove_disk,
3116 .spare_active = raid10_spare_active,
3117 .sync_request = sync_request,
3118 .quiesce = raid10_quiesce,
3119 .size = raid10_size,
3120 .takeover = raid10_takeover,
3123 static int __init raid_init(void)
3125 return register_md_personality(&raid10_personality);
3128 static void raid_exit(void)
3130 unregister_md_personality(&raid10_personality);
3133 module_init(raid_init);
3134 module_exit(raid_exit);
3135 MODULE_LICENSE("GPL");
3136 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3137 MODULE_ALIAS("md-personality-9"); /* RAID10 */
3138 MODULE_ALIAS("md-raid10");
3139 MODULE_ALIAS("md-level-10");
3141 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);