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/seq_file.h>
31 * RAID10 provides a combination of RAID0 and RAID1 functionality.
32 * The layout of data is defined by
35 * near_copies (stored in low byte of layout)
36 * far_copies (stored in second byte of layout)
37 * far_offset (stored in bit 16 of layout )
39 * The data to be stored is divided into chunks using chunksize.
40 * Each device is divided into far_copies sections.
41 * In each section, chunks are laid out in a style similar to raid0, but
42 * near_copies copies of each chunk is stored (each on a different drive).
43 * The starting device for each section is offset near_copies from the starting
44 * device of the previous section.
45 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47 * near_copies and far_copies must be at least one, and their product is at most
50 * If far_offset is true, then the far_copies are handled a bit differently.
51 * The copies are still in different stripes, but instead of be very far apart
52 * on disk, there are adjacent stripes.
56 * Number of guaranteed r10bios in case of extreme VM load:
58 #define NR_RAID10_BIOS 256
60 static void allow_barrier(conf_t *conf);
61 static void lower_barrier(conf_t *conf);
63 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
66 int size = offsetof(struct r10bio_s, devs[conf->copies]);
68 /* allocate a r10bio with room for raid_disks entries in the bios array */
69 return kzalloc(size, gfp_flags);
72 static void r10bio_pool_free(void *r10_bio, void *data)
77 /* Maximum size of each resync request */
78 #define RESYNC_BLOCK_SIZE (64*1024)
79 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
80 /* amount of memory to reserve for resync requests */
81 #define RESYNC_WINDOW (1024*1024)
82 /* maximum number of concurrent requests, memory permitting */
83 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
86 * When performing a resync, we need to read and compare, so
87 * we need as many pages are there are copies.
88 * When performing a recovery, we need 2 bios, one for read,
89 * one for write (we recover only one drive per r10buf)
92 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
101 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
105 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
106 nalloc = conf->copies; /* resync */
108 nalloc = 2; /* recovery */
113 for (j = nalloc ; j-- ; ) {
114 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
117 r10_bio->devs[j].bio = bio;
120 * Allocate RESYNC_PAGES data pages and attach them
123 for (j = 0 ; j < nalloc; j++) {
124 bio = r10_bio->devs[j].bio;
125 for (i = 0; i < RESYNC_PAGES; i++) {
126 page = alloc_page(gfp_flags);
130 bio->bi_io_vec[i].bv_page = page;
138 safe_put_page(bio->bi_io_vec[i-1].bv_page);
140 for (i = 0; i < RESYNC_PAGES ; i++)
141 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
144 while ( ++j < nalloc )
145 bio_put(r10_bio->devs[j].bio);
146 r10bio_pool_free(r10_bio, conf);
150 static void r10buf_pool_free(void *__r10_bio, void *data)
154 r10bio_t *r10bio = __r10_bio;
157 for (j=0; j < conf->copies; j++) {
158 struct bio *bio = r10bio->devs[j].bio;
160 for (i = 0; i < RESYNC_PAGES; i++) {
161 safe_put_page(bio->bi_io_vec[i].bv_page);
162 bio->bi_io_vec[i].bv_page = NULL;
167 r10bio_pool_free(r10bio, conf);
170 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
174 for (i = 0; i < conf->copies; i++) {
175 struct bio **bio = & r10_bio->devs[i].bio;
176 if (*bio && *bio != IO_BLOCKED)
182 static void free_r10bio(r10bio_t *r10_bio)
184 conf_t *conf = r10_bio->mddev->private;
187 * Wake up any possible resync thread that waits for the device
192 put_all_bios(conf, r10_bio);
193 mempool_free(r10_bio, conf->r10bio_pool);
196 static void put_buf(r10bio_t *r10_bio)
198 conf_t *conf = r10_bio->mddev->private;
200 mempool_free(r10_bio, conf->r10buf_pool);
205 static void reschedule_retry(r10bio_t *r10_bio)
208 mddev_t *mddev = r10_bio->mddev;
209 conf_t *conf = mddev->private;
211 spin_lock_irqsave(&conf->device_lock, flags);
212 list_add(&r10_bio->retry_list, &conf->retry_list);
214 spin_unlock_irqrestore(&conf->device_lock, flags);
216 /* wake up frozen array... */
217 wake_up(&conf->wait_barrier);
219 md_wakeup_thread(mddev->thread);
223 * raid_end_bio_io() is called when we have finished servicing a mirrored
224 * operation and are ready to return a success/failure code to the buffer
227 static void raid_end_bio_io(r10bio_t *r10_bio)
229 struct bio *bio = r10_bio->master_bio;
232 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
233 free_r10bio(r10_bio);
237 * Update disk head position estimator based on IRQ completion info.
239 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
241 conf_t *conf = r10_bio->mddev->private;
243 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
244 r10_bio->devs[slot].addr + (r10_bio->sectors);
247 static void raid10_end_read_request(struct bio *bio, int error)
249 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
250 r10bio_t *r10_bio = bio->bi_private;
252 conf_t *conf = r10_bio->mddev->private;
255 slot = r10_bio->read_slot;
256 dev = r10_bio->devs[slot].devnum;
258 * this branch is our 'one mirror IO has finished' event handler:
260 update_head_pos(slot, r10_bio);
264 * Set R10BIO_Uptodate in our master bio, so that
265 * we will return a good error code to the higher
266 * levels even if IO on some other mirrored buffer fails.
268 * The 'master' represents the composite IO operation to
269 * user-side. So if something waits for IO, then it will
270 * wait for the 'master' bio.
272 set_bit(R10BIO_Uptodate, &r10_bio->state);
273 raid_end_bio_io(r10_bio);
278 char b[BDEVNAME_SIZE];
279 if (printk_ratelimit())
280 printk(KERN_ERR "md/raid10:%s: %s: rescheduling sector %llu\n",
282 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
283 reschedule_retry(r10_bio);
286 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
289 static void raid10_end_write_request(struct bio *bio, int error)
291 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
292 r10bio_t *r10_bio = bio->bi_private;
294 conf_t *conf = r10_bio->mddev->private;
296 for (slot = 0; slot < conf->copies; slot++)
297 if (r10_bio->devs[slot].bio == bio)
299 dev = r10_bio->devs[slot].devnum;
302 * this branch is our 'one mirror IO has finished' event handler:
305 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
306 /* an I/O failed, we can't clear the bitmap */
307 set_bit(R10BIO_Degraded, &r10_bio->state);
310 * Set R10BIO_Uptodate in our master bio, so that
311 * we will return a good error code for to the higher
312 * levels even if IO on some other mirrored buffer fails.
314 * The 'master' represents the composite IO operation to
315 * user-side. So if something waits for IO, then it will
316 * wait for the 'master' bio.
318 set_bit(R10BIO_Uptodate, &r10_bio->state);
320 update_head_pos(slot, r10_bio);
324 * Let's see if all mirrored write operations have finished
327 if (atomic_dec_and_test(&r10_bio->remaining)) {
328 /* clear the bitmap if all writes complete successfully */
329 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
331 !test_bit(R10BIO_Degraded, &r10_bio->state),
333 md_write_end(r10_bio->mddev);
334 raid_end_bio_io(r10_bio);
337 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
342 * RAID10 layout manager
343 * As well as the chunksize and raid_disks count, there are two
344 * parameters: near_copies and far_copies.
345 * near_copies * far_copies must be <= raid_disks.
346 * Normally one of these will be 1.
347 * If both are 1, we get raid0.
348 * If near_copies == raid_disks, we get raid1.
350 * Chunks are laid out in raid0 style with near_copies copies of the
351 * first chunk, followed by near_copies copies of the next chunk and
353 * If far_copies > 1, then after 1/far_copies of the array has been assigned
354 * as described above, we start again with a device offset of near_copies.
355 * So we effectively have another copy of the whole array further down all
356 * the drives, but with blocks on different drives.
357 * With this layout, and block is never stored twice on the one device.
359 * raid10_find_phys finds the sector offset of a given virtual sector
360 * on each device that it is on.
362 * raid10_find_virt does the reverse mapping, from a device and a
363 * sector offset to a virtual address
366 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
376 /* now calculate first sector/dev */
377 chunk = r10bio->sector >> conf->chunk_shift;
378 sector = r10bio->sector & conf->chunk_mask;
380 chunk *= conf->near_copies;
382 dev = sector_div(stripe, conf->raid_disks);
383 if (conf->far_offset)
384 stripe *= conf->far_copies;
386 sector += stripe << conf->chunk_shift;
388 /* and calculate all the others */
389 for (n=0; n < conf->near_copies; n++) {
392 r10bio->devs[slot].addr = sector;
393 r10bio->devs[slot].devnum = d;
396 for (f = 1; f < conf->far_copies; f++) {
397 d += conf->near_copies;
398 if (d >= conf->raid_disks)
399 d -= conf->raid_disks;
401 r10bio->devs[slot].devnum = d;
402 r10bio->devs[slot].addr = s;
406 if (dev >= conf->raid_disks) {
408 sector += (conf->chunk_mask + 1);
411 BUG_ON(slot != conf->copies);
414 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
416 sector_t offset, chunk, vchunk;
418 offset = sector & conf->chunk_mask;
419 if (conf->far_offset) {
421 chunk = sector >> conf->chunk_shift;
422 fc = sector_div(chunk, conf->far_copies);
423 dev -= fc * conf->near_copies;
425 dev += conf->raid_disks;
427 while (sector >= conf->stride) {
428 sector -= conf->stride;
429 if (dev < conf->near_copies)
430 dev += conf->raid_disks - conf->near_copies;
432 dev -= conf->near_copies;
434 chunk = sector >> conf->chunk_shift;
436 vchunk = chunk * conf->raid_disks + dev;
437 sector_div(vchunk, conf->near_copies);
438 return (vchunk << conf->chunk_shift) + offset;
442 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
444 * @bvm: properties of new bio
445 * @biovec: the request that could be merged to it.
447 * Return amount of bytes we can accept at this offset
448 * If near_copies == raid_disk, there are no striping issues,
449 * but in that case, the function isn't called at all.
451 static int raid10_mergeable_bvec(struct request_queue *q,
452 struct bvec_merge_data *bvm,
453 struct bio_vec *biovec)
455 mddev_t *mddev = q->queuedata;
456 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
458 unsigned int chunk_sectors = mddev->chunk_sectors;
459 unsigned int bio_sectors = bvm->bi_size >> 9;
461 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
462 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
463 if (max <= biovec->bv_len && bio_sectors == 0)
464 return biovec->bv_len;
470 * This routine returns the disk from which the requested read should
471 * be done. There is a per-array 'next expected sequential IO' sector
472 * number - if this matches on the next IO then we use the last disk.
473 * There is also a per-disk 'last know head position' sector that is
474 * maintained from IRQ contexts, both the normal and the resync IO
475 * completion handlers update this position correctly. If there is no
476 * perfect sequential match then we pick the disk whose head is closest.
478 * If there are 2 mirrors in the same 2 devices, performance degrades
479 * because position is mirror, not device based.
481 * The rdev for the device selected will have nr_pending incremented.
485 * FIXME: possibly should rethink readbalancing and do it differently
486 * depending on near_copies / far_copies geometry.
488 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
490 const sector_t this_sector = r10_bio->sector;
492 const int sectors = r10_bio->sectors;
493 sector_t new_distance, best_dist;
498 raid10_find_phys(conf, r10_bio);
502 best_dist = MaxSector;
505 * Check if we can balance. We can balance on the whole
506 * device if no resync is going on (recovery is ok), or below
507 * the resync window. We take the first readable disk when
508 * above the resync window.
510 if (conf->mddev->recovery_cp < MaxSector
511 && (this_sector + sectors >= conf->next_resync))
514 for (slot = 0; slot < conf->copies ; slot++) {
515 if (r10_bio->devs[slot].bio == IO_BLOCKED)
517 disk = r10_bio->devs[slot].devnum;
518 rdev = rcu_dereference(conf->mirrors[disk].rdev);
521 if (!test_bit(In_sync, &rdev->flags))
527 /* This optimisation is debatable, and completely destroys
528 * sequential read speed for 'far copies' arrays. So only
529 * keep it for 'near' arrays, and review those later.
531 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
534 /* for far > 1 always use the lowest address */
535 if (conf->far_copies > 1)
536 new_distance = r10_bio->devs[slot].addr;
538 new_distance = abs(r10_bio->devs[slot].addr -
539 conf->mirrors[disk].head_position);
540 if (new_distance < best_dist) {
541 best_dist = new_distance;
545 if (slot == conf->copies)
549 disk = r10_bio->devs[slot].devnum;
550 rdev = rcu_dereference(conf->mirrors[disk].rdev);
553 atomic_inc(&rdev->nr_pending);
554 if (test_bit(Faulty, &rdev->flags)) {
555 /* Cannot risk returning a device that failed
556 * before we inc'ed nr_pending
558 rdev_dec_pending(rdev, conf->mddev);
561 r10_bio->read_slot = slot;
569 static int raid10_congested(void *data, int bits)
571 mddev_t *mddev = data;
572 conf_t *conf = mddev->private;
575 if (mddev_congested(mddev, bits))
578 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
579 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
580 if (rdev && !test_bit(Faulty, &rdev->flags)) {
581 struct request_queue *q = bdev_get_queue(rdev->bdev);
583 ret |= bdi_congested(&q->backing_dev_info, bits);
590 static void flush_pending_writes(conf_t *conf)
592 /* Any writes that have been queued but are awaiting
593 * bitmap updates get flushed here.
595 spin_lock_irq(&conf->device_lock);
597 if (conf->pending_bio_list.head) {
599 bio = bio_list_get(&conf->pending_bio_list);
600 spin_unlock_irq(&conf->device_lock);
601 /* flush any pending bitmap writes to disk
602 * before proceeding w/ I/O */
603 bitmap_unplug(conf->mddev->bitmap);
605 while (bio) { /* submit pending writes */
606 struct bio *next = bio->bi_next;
608 generic_make_request(bio);
612 spin_unlock_irq(&conf->device_lock);
616 * Sometimes we need to suspend IO while we do something else,
617 * either some resync/recovery, or reconfigure the array.
618 * To do this we raise a 'barrier'.
619 * The 'barrier' is a counter that can be raised multiple times
620 * to count how many activities are happening which preclude
622 * We can only raise the barrier if there is no pending IO.
623 * i.e. if nr_pending == 0.
624 * We choose only to raise the barrier if no-one is waiting for the
625 * barrier to go down. This means that as soon as an IO request
626 * is ready, no other operations which require a barrier will start
627 * until the IO request has had a chance.
629 * So: regular IO calls 'wait_barrier'. When that returns there
630 * is no backgroup IO happening, It must arrange to call
631 * allow_barrier when it has finished its IO.
632 * backgroup IO calls must call raise_barrier. Once that returns
633 * there is no normal IO happeing. It must arrange to call
634 * lower_barrier when the particular background IO completes.
637 static void raise_barrier(conf_t *conf, int force)
639 BUG_ON(force && !conf->barrier);
640 spin_lock_irq(&conf->resync_lock);
642 /* Wait until no block IO is waiting (unless 'force') */
643 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
644 conf->resync_lock, );
646 /* block any new IO from starting */
649 /* Now wait for all pending IO to complete */
650 wait_event_lock_irq(conf->wait_barrier,
651 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
652 conf->resync_lock, );
654 spin_unlock_irq(&conf->resync_lock);
657 static void lower_barrier(conf_t *conf)
660 spin_lock_irqsave(&conf->resync_lock, flags);
662 spin_unlock_irqrestore(&conf->resync_lock, flags);
663 wake_up(&conf->wait_barrier);
666 static void wait_barrier(conf_t *conf)
668 spin_lock_irq(&conf->resync_lock);
671 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
677 spin_unlock_irq(&conf->resync_lock);
680 static void allow_barrier(conf_t *conf)
683 spin_lock_irqsave(&conf->resync_lock, flags);
685 spin_unlock_irqrestore(&conf->resync_lock, flags);
686 wake_up(&conf->wait_barrier);
689 static void freeze_array(conf_t *conf)
691 /* stop syncio and normal IO and wait for everything to
693 * We increment barrier and nr_waiting, and then
694 * wait until nr_pending match nr_queued+1
695 * This is called in the context of one normal IO request
696 * that has failed. Thus any sync request that might be pending
697 * will be blocked by nr_pending, and we need to wait for
698 * pending IO requests to complete or be queued for re-try.
699 * Thus the number queued (nr_queued) plus this request (1)
700 * must match the number of pending IOs (nr_pending) before
703 spin_lock_irq(&conf->resync_lock);
706 wait_event_lock_irq(conf->wait_barrier,
707 conf->nr_pending == conf->nr_queued+1,
709 flush_pending_writes(conf));
711 spin_unlock_irq(&conf->resync_lock);
714 static void unfreeze_array(conf_t *conf)
716 /* reverse the effect of the freeze */
717 spin_lock_irq(&conf->resync_lock);
720 wake_up(&conf->wait_barrier);
721 spin_unlock_irq(&conf->resync_lock);
724 static int make_request(mddev_t *mddev, struct bio * bio)
726 conf_t *conf = mddev->private;
727 mirror_info_t *mirror;
729 struct bio *read_bio;
731 int chunk_sects = conf->chunk_mask + 1;
732 const int rw = bio_data_dir(bio);
733 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
734 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
736 mdk_rdev_t *blocked_rdev;
739 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
740 md_flush_request(mddev, bio);
744 /* If this request crosses a chunk boundary, we need to
745 * split it. This will only happen for 1 PAGE (or less) requests.
747 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
749 conf->near_copies < conf->raid_disks)) {
751 /* Sanity check -- queue functions should prevent this happening */
752 if (bio->bi_vcnt != 1 ||
755 /* This is a one page bio that upper layers
756 * refuse to split for us, so we need to split it.
759 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
761 /* Each of these 'make_request' calls will call 'wait_barrier'.
762 * If the first succeeds but the second blocks due to the resync
763 * thread raising the barrier, we will deadlock because the
764 * IO to the underlying device will be queued in generic_make_request
765 * and will never complete, so will never reduce nr_pending.
766 * So increment nr_waiting here so no new raise_barriers will
767 * succeed, and so the second wait_barrier cannot block.
769 spin_lock_irq(&conf->resync_lock);
771 spin_unlock_irq(&conf->resync_lock);
773 if (make_request(mddev, &bp->bio1))
774 generic_make_request(&bp->bio1);
775 if (make_request(mddev, &bp->bio2))
776 generic_make_request(&bp->bio2);
778 spin_lock_irq(&conf->resync_lock);
780 wake_up(&conf->wait_barrier);
781 spin_unlock_irq(&conf->resync_lock);
783 bio_pair_release(bp);
786 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
787 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
788 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
794 md_write_start(mddev, bio);
797 * Register the new request and wait if the reconstruction
798 * thread has put up a bar for new requests.
799 * Continue immediately if no resync is active currently.
803 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
805 r10_bio->master_bio = bio;
806 r10_bio->sectors = bio->bi_size >> 9;
808 r10_bio->mddev = mddev;
809 r10_bio->sector = bio->bi_sector;
814 * read balancing logic:
816 int disk = read_balance(conf, r10_bio);
817 int slot = r10_bio->read_slot;
819 raid_end_bio_io(r10_bio);
822 mirror = conf->mirrors + disk;
824 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
826 r10_bio->devs[slot].bio = read_bio;
828 read_bio->bi_sector = r10_bio->devs[slot].addr +
829 mirror->rdev->data_offset;
830 read_bio->bi_bdev = mirror->rdev->bdev;
831 read_bio->bi_end_io = raid10_end_read_request;
832 read_bio->bi_rw = READ | do_sync;
833 read_bio->bi_private = r10_bio;
835 generic_make_request(read_bio);
842 /* first select target devices under rcu_lock and
843 * inc refcount on their rdev. Record them by setting
846 plugged = mddev_check_plugged(mddev);
848 raid10_find_phys(conf, r10_bio);
852 for (i = 0; i < conf->copies; i++) {
853 int d = r10_bio->devs[i].devnum;
854 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
855 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
856 atomic_inc(&rdev->nr_pending);
860 if (rdev && !test_bit(Faulty, &rdev->flags)) {
861 atomic_inc(&rdev->nr_pending);
862 r10_bio->devs[i].bio = bio;
864 r10_bio->devs[i].bio = NULL;
865 set_bit(R10BIO_Degraded, &r10_bio->state);
870 if (unlikely(blocked_rdev)) {
871 /* Have to wait for this device to get unblocked, then retry */
875 for (j = 0; j < i; j++)
876 if (r10_bio->devs[j].bio) {
877 d = r10_bio->devs[j].devnum;
878 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
881 md_wait_for_blocked_rdev(blocked_rdev, mddev);
886 atomic_set(&r10_bio->remaining, 1);
887 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
889 for (i = 0; i < conf->copies; i++) {
891 int d = r10_bio->devs[i].devnum;
892 if (!r10_bio->devs[i].bio)
895 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
896 r10_bio->devs[i].bio = mbio;
898 mbio->bi_sector = r10_bio->devs[i].addr+
899 conf->mirrors[d].rdev->data_offset;
900 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
901 mbio->bi_end_io = raid10_end_write_request;
902 mbio->bi_rw = WRITE | do_sync | do_fua;
903 mbio->bi_private = r10_bio;
905 atomic_inc(&r10_bio->remaining);
906 spin_lock_irqsave(&conf->device_lock, flags);
907 bio_list_add(&conf->pending_bio_list, mbio);
908 spin_unlock_irqrestore(&conf->device_lock, flags);
911 if (atomic_dec_and_test(&r10_bio->remaining)) {
912 /* This matches the end of raid10_end_write_request() */
913 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
915 !test_bit(R10BIO_Degraded, &r10_bio->state),
918 raid_end_bio_io(r10_bio);
921 /* In case raid10d snuck in to freeze_array */
922 wake_up(&conf->wait_barrier);
924 if (do_sync || !mddev->bitmap || !plugged)
925 md_wakeup_thread(mddev->thread);
929 static void status(struct seq_file *seq, mddev_t *mddev)
931 conf_t *conf = mddev->private;
934 if (conf->near_copies < conf->raid_disks)
935 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
936 if (conf->near_copies > 1)
937 seq_printf(seq, " %d near-copies", conf->near_copies);
938 if (conf->far_copies > 1) {
939 if (conf->far_offset)
940 seq_printf(seq, " %d offset-copies", conf->far_copies);
942 seq_printf(seq, " %d far-copies", conf->far_copies);
944 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
945 conf->raid_disks - mddev->degraded);
946 for (i = 0; i < conf->raid_disks; i++)
947 seq_printf(seq, "%s",
948 conf->mirrors[i].rdev &&
949 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
950 seq_printf(seq, "]");
953 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
955 char b[BDEVNAME_SIZE];
956 conf_t *conf = mddev->private;
959 * If it is not operational, then we have already marked it as dead
960 * else if it is the last working disks, ignore the error, let the
961 * next level up know.
962 * else mark the drive as failed
964 if (test_bit(In_sync, &rdev->flags)
965 && conf->raid_disks-mddev->degraded == 1)
967 * Don't fail the drive, just return an IO error.
968 * The test should really be more sophisticated than
969 * "working_disks == 1", but it isn't critical, and
970 * can wait until we do more sophisticated "is the drive
971 * really dead" tests...
974 if (test_and_clear_bit(In_sync, &rdev->flags)) {
976 spin_lock_irqsave(&conf->device_lock, flags);
978 spin_unlock_irqrestore(&conf->device_lock, flags);
980 * if recovery is running, make sure it aborts.
982 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
984 set_bit(Faulty, &rdev->flags);
985 set_bit(MD_CHANGE_DEVS, &mddev->flags);
987 "md/raid10:%s: Disk failure on %s, disabling device.\n"
988 "md/raid10:%s: Operation continuing on %d devices.\n",
989 mdname(mddev), bdevname(rdev->bdev, b),
990 mdname(mddev), conf->raid_disks - mddev->degraded);
993 static void print_conf(conf_t *conf)
998 printk(KERN_DEBUG "RAID10 conf printout:\n");
1000 printk(KERN_DEBUG "(!conf)\n");
1003 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1006 for (i = 0; i < conf->raid_disks; i++) {
1007 char b[BDEVNAME_SIZE];
1008 tmp = conf->mirrors + i;
1010 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1011 i, !test_bit(In_sync, &tmp->rdev->flags),
1012 !test_bit(Faulty, &tmp->rdev->flags),
1013 bdevname(tmp->rdev->bdev,b));
1017 static void close_sync(conf_t *conf)
1020 allow_barrier(conf);
1022 mempool_destroy(conf->r10buf_pool);
1023 conf->r10buf_pool = NULL;
1026 /* check if there are enough drives for
1027 * every block to appear on atleast one
1029 static int enough(conf_t *conf)
1034 int n = conf->copies;
1037 if (conf->mirrors[first].rdev)
1039 first = (first+1) % conf->raid_disks;
1043 } while (first != 0);
1047 static int raid10_spare_active(mddev_t *mddev)
1050 conf_t *conf = mddev->private;
1053 unsigned long flags;
1056 * Find all non-in_sync disks within the RAID10 configuration
1057 * and mark them in_sync
1059 for (i = 0; i < conf->raid_disks; i++) {
1060 tmp = conf->mirrors + i;
1062 && !test_bit(Faulty, &tmp->rdev->flags)
1063 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1065 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1068 spin_lock_irqsave(&conf->device_lock, flags);
1069 mddev->degraded -= count;
1070 spin_unlock_irqrestore(&conf->device_lock, flags);
1077 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1079 conf_t *conf = mddev->private;
1084 int last = conf->raid_disks - 1;
1086 if (mddev->recovery_cp < MaxSector)
1087 /* only hot-add to in-sync arrays, as recovery is
1088 * very different from resync
1094 if (rdev->raid_disk >= 0)
1095 first = last = rdev->raid_disk;
1097 if (rdev->saved_raid_disk >= 0 &&
1098 rdev->saved_raid_disk >= first &&
1099 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1100 mirror = rdev->saved_raid_disk;
1103 for ( ; mirror <= last ; mirror++)
1104 if ( !(p=conf->mirrors+mirror)->rdev) {
1106 disk_stack_limits(mddev->gendisk, rdev->bdev,
1107 rdev->data_offset << 9);
1108 /* as we don't honour merge_bvec_fn, we must
1109 * never risk violating it, so limit
1110 * ->max_segments to one lying with a single
1111 * page, as a one page request is never in
1114 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1115 blk_queue_max_segments(mddev->queue, 1);
1116 blk_queue_segment_boundary(mddev->queue,
1117 PAGE_CACHE_SIZE - 1);
1120 p->head_position = 0;
1121 rdev->raid_disk = mirror;
1123 if (rdev->saved_raid_disk != mirror)
1125 rcu_assign_pointer(p->rdev, rdev);
1129 md_integrity_add_rdev(rdev, mddev);
1134 static int raid10_remove_disk(mddev_t *mddev, int number)
1136 conf_t *conf = mddev->private;
1139 mirror_info_t *p = conf->mirrors+ number;
1144 if (test_bit(In_sync, &rdev->flags) ||
1145 atomic_read(&rdev->nr_pending)) {
1149 /* Only remove faulty devices in recovery
1152 if (!test_bit(Faulty, &rdev->flags) &&
1159 if (atomic_read(&rdev->nr_pending)) {
1160 /* lost the race, try later */
1165 err = md_integrity_register(mddev);
1174 static void end_sync_read(struct bio *bio, int error)
1176 r10bio_t *r10_bio = bio->bi_private;
1177 conf_t *conf = r10_bio->mddev->private;
1180 for (i=0; i<conf->copies; i++)
1181 if (r10_bio->devs[i].bio == bio)
1183 BUG_ON(i == conf->copies);
1184 update_head_pos(i, r10_bio);
1185 d = r10_bio->devs[i].devnum;
1187 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1188 set_bit(R10BIO_Uptodate, &r10_bio->state);
1190 atomic_add(r10_bio->sectors,
1191 &conf->mirrors[d].rdev->corrected_errors);
1192 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1193 md_error(r10_bio->mddev,
1194 conf->mirrors[d].rdev);
1197 /* for reconstruct, we always reschedule after a read.
1198 * for resync, only after all reads
1200 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1201 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1202 atomic_dec_and_test(&r10_bio->remaining)) {
1203 /* we have read all the blocks,
1204 * do the comparison in process context in raid10d
1206 reschedule_retry(r10_bio);
1210 static void end_sync_write(struct bio *bio, int error)
1212 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1213 r10bio_t *r10_bio = bio->bi_private;
1214 mddev_t *mddev = r10_bio->mddev;
1215 conf_t *conf = mddev->private;
1218 for (i = 0; i < conf->copies; i++)
1219 if (r10_bio->devs[i].bio == bio)
1221 d = r10_bio->devs[i].devnum;
1224 md_error(mddev, conf->mirrors[d].rdev);
1226 update_head_pos(i, r10_bio);
1228 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1229 while (atomic_dec_and_test(&r10_bio->remaining)) {
1230 if (r10_bio->master_bio == NULL) {
1231 /* the primary of several recovery bios */
1232 sector_t s = r10_bio->sectors;
1234 md_done_sync(mddev, s, 1);
1237 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1245 * Note: sync and recover and handled very differently for raid10
1246 * This code is for resync.
1247 * For resync, we read through virtual addresses and read all blocks.
1248 * If there is any error, we schedule a write. The lowest numbered
1249 * drive is authoritative.
1250 * However requests come for physical address, so we need to map.
1251 * For every physical address there are raid_disks/copies virtual addresses,
1252 * which is always are least one, but is not necessarly an integer.
1253 * This means that a physical address can span multiple chunks, so we may
1254 * have to submit multiple io requests for a single sync request.
1257 * We check if all blocks are in-sync and only write to blocks that
1260 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1262 conf_t *conf = mddev->private;
1264 struct bio *tbio, *fbio;
1266 atomic_set(&r10_bio->remaining, 1);
1268 /* find the first device with a block */
1269 for (i=0; i<conf->copies; i++)
1270 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1273 if (i == conf->copies)
1277 fbio = r10_bio->devs[i].bio;
1279 /* now find blocks with errors */
1280 for (i=0 ; i < conf->copies ; i++) {
1282 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1284 tbio = r10_bio->devs[i].bio;
1286 if (tbio->bi_end_io != end_sync_read)
1290 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1291 /* We know that the bi_io_vec layout is the same for
1292 * both 'first' and 'i', so we just compare them.
1293 * All vec entries are PAGE_SIZE;
1295 for (j = 0; j < vcnt; j++)
1296 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1297 page_address(tbio->bi_io_vec[j].bv_page),
1302 mddev->resync_mismatches += r10_bio->sectors;
1304 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1305 /* Don't fix anything. */
1307 /* Ok, we need to write this bio
1308 * First we need to fixup bv_offset, bv_len and
1309 * bi_vecs, as the read request might have corrupted these
1311 tbio->bi_vcnt = vcnt;
1312 tbio->bi_size = r10_bio->sectors << 9;
1314 tbio->bi_phys_segments = 0;
1315 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1316 tbio->bi_flags |= 1 << BIO_UPTODATE;
1317 tbio->bi_next = NULL;
1318 tbio->bi_rw = WRITE;
1319 tbio->bi_private = r10_bio;
1320 tbio->bi_sector = r10_bio->devs[i].addr;
1322 for (j=0; j < vcnt ; j++) {
1323 tbio->bi_io_vec[j].bv_offset = 0;
1324 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1326 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1327 page_address(fbio->bi_io_vec[j].bv_page),
1330 tbio->bi_end_io = end_sync_write;
1332 d = r10_bio->devs[i].devnum;
1333 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1334 atomic_inc(&r10_bio->remaining);
1335 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1337 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1338 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1339 generic_make_request(tbio);
1343 if (atomic_dec_and_test(&r10_bio->remaining)) {
1344 md_done_sync(mddev, r10_bio->sectors, 1);
1350 * Now for the recovery code.
1351 * Recovery happens across physical sectors.
1352 * We recover all non-is_sync drives by finding the virtual address of
1353 * each, and then choose a working drive that also has that virt address.
1354 * There is a separate r10_bio for each non-in_sync drive.
1355 * Only the first two slots are in use. The first for reading,
1356 * The second for writing.
1360 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1362 conf_t *conf = mddev->private;
1364 struct bio *bio, *wbio;
1367 /* move the pages across to the second bio
1368 * and submit the write request
1370 bio = r10_bio->devs[0].bio;
1371 wbio = r10_bio->devs[1].bio;
1372 for (i=0; i < wbio->bi_vcnt; i++) {
1373 struct page *p = bio->bi_io_vec[i].bv_page;
1374 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1375 wbio->bi_io_vec[i].bv_page = p;
1377 d = r10_bio->devs[1].devnum;
1379 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1380 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1381 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1382 generic_make_request(wbio);
1384 bio_endio(wbio, -EIO);
1389 * Used by fix_read_error() to decay the per rdev read_errors.
1390 * We halve the read error count for every hour that has elapsed
1391 * since the last recorded read error.
1394 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1396 struct timespec cur_time_mon;
1397 unsigned long hours_since_last;
1398 unsigned int read_errors = atomic_read(&rdev->read_errors);
1400 ktime_get_ts(&cur_time_mon);
1402 if (rdev->last_read_error.tv_sec == 0 &&
1403 rdev->last_read_error.tv_nsec == 0) {
1404 /* first time we've seen a read error */
1405 rdev->last_read_error = cur_time_mon;
1409 hours_since_last = (cur_time_mon.tv_sec -
1410 rdev->last_read_error.tv_sec) / 3600;
1412 rdev->last_read_error = cur_time_mon;
1415 * if hours_since_last is > the number of bits in read_errors
1416 * just set read errors to 0. We do this to avoid
1417 * overflowing the shift of read_errors by hours_since_last.
1419 if (hours_since_last >= 8 * sizeof(read_errors))
1420 atomic_set(&rdev->read_errors, 0);
1422 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1426 * This is a kernel thread which:
1428 * 1. Retries failed read operations on working mirrors.
1429 * 2. Updates the raid superblock when problems encounter.
1430 * 3. Performs writes following reads for array synchronising.
1433 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1435 int sect = 0; /* Offset from r10_bio->sector */
1436 int sectors = r10_bio->sectors;
1438 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1439 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1442 rdev = rcu_dereference(conf->mirrors[d].rdev);
1443 if (rdev) { /* If rdev is not NULL */
1444 char b[BDEVNAME_SIZE];
1445 int cur_read_error_count = 0;
1447 bdevname(rdev->bdev, b);
1449 if (test_bit(Faulty, &rdev->flags)) {
1451 /* drive has already been failed, just ignore any
1452 more fix_read_error() attempts */
1456 check_decay_read_errors(mddev, rdev);
1457 atomic_inc(&rdev->read_errors);
1458 cur_read_error_count = atomic_read(&rdev->read_errors);
1459 if (cur_read_error_count > max_read_errors) {
1462 "md/raid10:%s: %s: Raid device exceeded "
1463 "read_error threshold "
1464 "[cur %d:max %d]\n",
1466 b, cur_read_error_count, max_read_errors);
1468 "md/raid10:%s: %s: Failing raid "
1469 "device\n", mdname(mddev), b);
1470 md_error(mddev, conf->mirrors[d].rdev);
1478 int sl = r10_bio->read_slot;
1482 if (s > (PAGE_SIZE>>9))
1487 d = r10_bio->devs[sl].devnum;
1488 rdev = rcu_dereference(conf->mirrors[d].rdev);
1490 test_bit(In_sync, &rdev->flags)) {
1491 atomic_inc(&rdev->nr_pending);
1493 success = sync_page_io(rdev,
1494 r10_bio->devs[sl].addr +
1497 conf->tmppage, READ, false);
1498 rdev_dec_pending(rdev, mddev);
1504 if (sl == conf->copies)
1506 } while (!success && sl != r10_bio->read_slot);
1510 /* Cannot read from anywhere -- bye bye array */
1511 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1512 md_error(mddev, conf->mirrors[dn].rdev);
1517 /* write it back and re-read */
1519 while (sl != r10_bio->read_slot) {
1520 char b[BDEVNAME_SIZE];
1525 d = r10_bio->devs[sl].devnum;
1526 rdev = rcu_dereference(conf->mirrors[d].rdev);
1528 test_bit(In_sync, &rdev->flags)) {
1529 atomic_inc(&rdev->nr_pending);
1531 atomic_add(s, &rdev->corrected_errors);
1532 if (sync_page_io(rdev,
1533 r10_bio->devs[sl].addr +
1535 s<<9, conf->tmppage, WRITE, false)
1537 /* Well, this device is dead */
1539 "md/raid10:%s: read correction "
1541 " (%d sectors at %llu on %s)\n",
1543 (unsigned long long)(sect+
1545 bdevname(rdev->bdev, b));
1546 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1549 bdevname(rdev->bdev, b));
1550 md_error(mddev, rdev);
1552 rdev_dec_pending(rdev, mddev);
1557 while (sl != r10_bio->read_slot) {
1562 d = r10_bio->devs[sl].devnum;
1563 rdev = rcu_dereference(conf->mirrors[d].rdev);
1565 test_bit(In_sync, &rdev->flags)) {
1566 char b[BDEVNAME_SIZE];
1567 atomic_inc(&rdev->nr_pending);
1569 if (sync_page_io(rdev,
1570 r10_bio->devs[sl].addr +
1572 s<<9, conf->tmppage,
1573 READ, false) == 0) {
1574 /* Well, this device is dead */
1576 "md/raid10:%s: unable to read back "
1578 " (%d sectors at %llu on %s)\n",
1580 (unsigned long long)(sect+
1582 bdevname(rdev->bdev, b));
1583 printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1585 bdevname(rdev->bdev, b));
1587 md_error(mddev, rdev);
1590 "md/raid10:%s: read error corrected"
1591 " (%d sectors at %llu on %s)\n",
1593 (unsigned long long)(sect+
1595 bdevname(rdev->bdev, b));
1598 rdev_dec_pending(rdev, mddev);
1609 static void raid10d(mddev_t *mddev)
1613 unsigned long flags;
1614 conf_t *conf = mddev->private;
1615 struct list_head *head = &conf->retry_list;
1617 struct blk_plug plug;
1619 md_check_recovery(mddev);
1621 blk_start_plug(&plug);
1623 char b[BDEVNAME_SIZE];
1625 flush_pending_writes(conf);
1627 spin_lock_irqsave(&conf->device_lock, flags);
1628 if (list_empty(head)) {
1629 spin_unlock_irqrestore(&conf->device_lock, flags);
1632 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1633 list_del(head->prev);
1635 spin_unlock_irqrestore(&conf->device_lock, flags);
1637 mddev = r10_bio->mddev;
1638 conf = mddev->private;
1639 if (test_bit(R10BIO_IsSync, &r10_bio->state))
1640 sync_request_write(mddev, r10_bio);
1641 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
1642 recovery_request_write(mddev, r10_bio);
1645 /* we got a read error. Maybe the drive is bad. Maybe just
1646 * the block and we can fix it.
1647 * We freeze all other IO, and try reading the block from
1648 * other devices. When we find one, we re-write
1649 * and check it that fixes the read error.
1650 * This is all done synchronously while the array is
1653 if (mddev->ro == 0) {
1655 fix_read_error(conf, mddev, r10_bio);
1656 unfreeze_array(conf);
1659 bio = r10_bio->devs[r10_bio->read_slot].bio;
1660 r10_bio->devs[r10_bio->read_slot].bio =
1661 mddev->ro ? IO_BLOCKED : NULL;
1662 mirror = read_balance(conf, r10_bio);
1664 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1665 " read error for block %llu\n",
1667 bdevname(bio->bi_bdev,b),
1668 (unsigned long long)r10_bio->sector);
1669 raid_end_bio_io(r10_bio);
1672 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
1674 rdev = conf->mirrors[mirror].rdev;
1675 if (printk_ratelimit())
1676 printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to"
1677 " another mirror\n",
1679 bdevname(rdev->bdev,b),
1680 (unsigned long long)r10_bio->sector);
1681 bio = bio_clone_mddev(r10_bio->master_bio,
1683 r10_bio->devs[r10_bio->read_slot].bio = bio;
1684 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1685 + rdev->data_offset;
1686 bio->bi_bdev = rdev->bdev;
1687 bio->bi_rw = READ | do_sync;
1688 bio->bi_private = r10_bio;
1689 bio->bi_end_io = raid10_end_read_request;
1690 generic_make_request(bio);
1695 blk_finish_plug(&plug);
1699 static int init_resync(conf_t *conf)
1703 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1704 BUG_ON(conf->r10buf_pool);
1705 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1706 if (!conf->r10buf_pool)
1708 conf->next_resync = 0;
1713 * perform a "sync" on one "block"
1715 * We need to make sure that no normal I/O request - particularly write
1716 * requests - conflict with active sync requests.
1718 * This is achieved by tracking pending requests and a 'barrier' concept
1719 * that can be installed to exclude normal IO requests.
1721 * Resync and recovery are handled very differently.
1722 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1724 * For resync, we iterate over virtual addresses, read all copies,
1725 * and update if there are differences. If only one copy is live,
1727 * For recovery, we iterate over physical addresses, read a good
1728 * value for each non-in_sync drive, and over-write.
1730 * So, for recovery we may have several outstanding complex requests for a
1731 * given address, one for each out-of-sync device. We model this by allocating
1732 * a number of r10_bio structures, one for each out-of-sync device.
1733 * As we setup these structures, we collect all bio's together into a list
1734 * which we then process collectively to add pages, and then process again
1735 * to pass to generic_make_request.
1737 * The r10_bio structures are linked using a borrowed master_bio pointer.
1738 * This link is counted in ->remaining. When the r10_bio that points to NULL
1739 * has its remaining count decremented to 0, the whole complex operation
1744 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1746 conf_t *conf = mddev->private;
1748 struct bio *biolist = NULL, *bio;
1749 sector_t max_sector, nr_sectors;
1753 sector_t sync_blocks;
1755 sector_t sectors_skipped = 0;
1756 int chunks_skipped = 0;
1758 if (!conf->r10buf_pool)
1759 if (init_resync(conf))
1763 max_sector = mddev->dev_sectors;
1764 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1765 max_sector = mddev->resync_max_sectors;
1766 if (sector_nr >= max_sector) {
1767 /* If we aborted, we need to abort the
1768 * sync on the 'current' bitmap chucks (there can
1769 * be several when recovering multiple devices).
1770 * as we may have started syncing it but not finished.
1771 * We can find the current address in
1772 * mddev->curr_resync, but for recovery,
1773 * we need to convert that to several
1774 * virtual addresses.
1776 if (mddev->curr_resync < max_sector) { /* aborted */
1777 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1778 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1780 else for (i=0; i<conf->raid_disks; i++) {
1782 raid10_find_virt(conf, mddev->curr_resync, i);
1783 bitmap_end_sync(mddev->bitmap, sect,
1786 } else /* completed sync */
1789 bitmap_close_sync(mddev->bitmap);
1792 return sectors_skipped;
1794 if (chunks_skipped >= conf->raid_disks) {
1795 /* if there has been nothing to do on any drive,
1796 * then there is nothing to do at all..
1799 return (max_sector - sector_nr) + sectors_skipped;
1802 if (max_sector > mddev->resync_max)
1803 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1805 /* make sure whole request will fit in a chunk - if chunks
1808 if (conf->near_copies < conf->raid_disks &&
1809 max_sector > (sector_nr | conf->chunk_mask))
1810 max_sector = (sector_nr | conf->chunk_mask) + 1;
1812 * If there is non-resync activity waiting for us then
1813 * put in a delay to throttle resync.
1815 if (!go_faster && conf->nr_waiting)
1816 msleep_interruptible(1000);
1818 /* Again, very different code for resync and recovery.
1819 * Both must result in an r10bio with a list of bios that
1820 * have bi_end_io, bi_sector, bi_bdev set,
1821 * and bi_private set to the r10bio.
1822 * For recovery, we may actually create several r10bios
1823 * with 2 bios in each, that correspond to the bios in the main one.
1824 * In this case, the subordinate r10bios link back through a
1825 * borrowed master_bio pointer, and the counter in the master
1826 * includes a ref from each subordinate.
1828 /* First, we decide what to do and set ->bi_end_io
1829 * To end_sync_read if we want to read, and
1830 * end_sync_write if we will want to write.
1833 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1834 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1835 /* recovery... the complicated one */
1839 for (i=0 ; i<conf->raid_disks; i++)
1840 if (conf->mirrors[i].rdev &&
1841 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1842 int still_degraded = 0;
1843 /* want to reconstruct this device */
1844 r10bio_t *rb2 = r10_bio;
1845 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1847 /* Unless we are doing a full sync, we only need
1848 * to recover the block if it is set in the bitmap
1850 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1852 if (sync_blocks < max_sync)
1853 max_sync = sync_blocks;
1856 /* yep, skip the sync_blocks here, but don't assume
1857 * that there will never be anything to do here
1859 chunks_skipped = -1;
1863 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1864 raise_barrier(conf, rb2 != NULL);
1865 atomic_set(&r10_bio->remaining, 0);
1867 r10_bio->master_bio = (struct bio*)rb2;
1869 atomic_inc(&rb2->remaining);
1870 r10_bio->mddev = mddev;
1871 set_bit(R10BIO_IsRecover, &r10_bio->state);
1872 r10_bio->sector = sect;
1874 raid10_find_phys(conf, r10_bio);
1876 /* Need to check if the array will still be
1879 for (j=0; j<conf->raid_disks; j++)
1880 if (conf->mirrors[j].rdev == NULL ||
1881 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1886 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1887 &sync_blocks, still_degraded);
1889 for (j=0; j<conf->copies;j++) {
1890 int d = r10_bio->devs[j].devnum;
1891 if (conf->mirrors[d].rdev &&
1892 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1893 /* This is where we read from */
1894 bio = r10_bio->devs[0].bio;
1895 bio->bi_next = biolist;
1897 bio->bi_private = r10_bio;
1898 bio->bi_end_io = end_sync_read;
1900 bio->bi_sector = r10_bio->devs[j].addr +
1901 conf->mirrors[d].rdev->data_offset;
1902 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1903 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1904 atomic_inc(&r10_bio->remaining);
1905 /* and we write to 'i' */
1907 for (k=0; k<conf->copies; k++)
1908 if (r10_bio->devs[k].devnum == i)
1910 BUG_ON(k == conf->copies);
1911 bio = r10_bio->devs[1].bio;
1912 bio->bi_next = biolist;
1914 bio->bi_private = r10_bio;
1915 bio->bi_end_io = end_sync_write;
1917 bio->bi_sector = r10_bio->devs[k].addr +
1918 conf->mirrors[i].rdev->data_offset;
1919 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1921 r10_bio->devs[0].devnum = d;
1922 r10_bio->devs[1].devnum = i;
1927 if (j == conf->copies) {
1928 /* Cannot recover, so abort the recovery */
1931 atomic_dec(&rb2->remaining);
1933 if (!test_and_set_bit(MD_RECOVERY_INTR,
1935 printk(KERN_INFO "md/raid10:%s: insufficient "
1936 "working devices for recovery.\n",
1941 if (biolist == NULL) {
1943 r10bio_t *rb2 = r10_bio;
1944 r10_bio = (r10bio_t*) rb2->master_bio;
1945 rb2->master_bio = NULL;
1951 /* resync. Schedule a read for every block at this virt offset */
1954 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1956 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1957 &sync_blocks, mddev->degraded) &&
1958 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1959 /* We can skip this block */
1961 return sync_blocks + sectors_skipped;
1963 if (sync_blocks < max_sync)
1964 max_sync = sync_blocks;
1965 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1967 r10_bio->mddev = mddev;
1968 atomic_set(&r10_bio->remaining, 0);
1969 raise_barrier(conf, 0);
1970 conf->next_resync = sector_nr;
1972 r10_bio->master_bio = NULL;
1973 r10_bio->sector = sector_nr;
1974 set_bit(R10BIO_IsSync, &r10_bio->state);
1975 raid10_find_phys(conf, r10_bio);
1976 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1978 for (i=0; i<conf->copies; i++) {
1979 int d = r10_bio->devs[i].devnum;
1980 bio = r10_bio->devs[i].bio;
1981 bio->bi_end_io = NULL;
1982 clear_bit(BIO_UPTODATE, &bio->bi_flags);
1983 if (conf->mirrors[d].rdev == NULL ||
1984 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1986 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1987 atomic_inc(&r10_bio->remaining);
1988 bio->bi_next = biolist;
1990 bio->bi_private = r10_bio;
1991 bio->bi_end_io = end_sync_read;
1993 bio->bi_sector = r10_bio->devs[i].addr +
1994 conf->mirrors[d].rdev->data_offset;
1995 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2000 for (i=0; i<conf->copies; i++) {
2001 int d = r10_bio->devs[i].devnum;
2002 if (r10_bio->devs[i].bio->bi_end_io)
2003 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
2011 for (bio = biolist; bio ; bio=bio->bi_next) {
2013 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2015 bio->bi_flags |= 1 << BIO_UPTODATE;
2018 bio->bi_phys_segments = 0;
2023 if (sector_nr + max_sync < max_sector)
2024 max_sector = sector_nr + max_sync;
2027 int len = PAGE_SIZE;
2029 if (sector_nr + (len>>9) > max_sector)
2030 len = (max_sector - sector_nr) << 9;
2033 for (bio= biolist ; bio ; bio=bio->bi_next) {
2034 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2035 if (bio_add_page(bio, page, len, 0) == 0) {
2038 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2039 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
2040 /* remove last page from this bio */
2042 bio2->bi_size -= len;
2043 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2049 nr_sectors += len>>9;
2050 sector_nr += len>>9;
2051 } while (biolist->bi_vcnt < RESYNC_PAGES);
2053 r10_bio->sectors = nr_sectors;
2057 biolist = biolist->bi_next;
2059 bio->bi_next = NULL;
2060 r10_bio = bio->bi_private;
2061 r10_bio->sectors = nr_sectors;
2063 if (bio->bi_end_io == end_sync_read) {
2064 md_sync_acct(bio->bi_bdev, nr_sectors);
2065 generic_make_request(bio);
2069 if (sectors_skipped)
2070 /* pretend they weren't skipped, it makes
2071 * no important difference in this case
2073 md_done_sync(mddev, sectors_skipped, 1);
2075 return sectors_skipped + nr_sectors;
2077 /* There is nowhere to write, so all non-sync
2078 * drives must be failed, so try the next chunk...
2080 if (sector_nr + max_sync < max_sector)
2081 max_sector = sector_nr + max_sync;
2083 sectors_skipped += (max_sector - sector_nr);
2085 sector_nr = max_sector;
2090 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2093 conf_t *conf = mddev->private;
2096 raid_disks = conf->raid_disks;
2098 sectors = conf->dev_sectors;
2100 size = sectors >> conf->chunk_shift;
2101 sector_div(size, conf->far_copies);
2102 size = size * raid_disks;
2103 sector_div(size, conf->near_copies);
2105 return size << conf->chunk_shift;
2109 static conf_t *setup_conf(mddev_t *mddev)
2111 conf_t *conf = NULL;
2113 sector_t stride, size;
2116 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2117 !is_power_of_2(mddev->new_chunk_sectors)) {
2118 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2119 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2120 mdname(mddev), PAGE_SIZE);
2124 nc = mddev->new_layout & 255;
2125 fc = (mddev->new_layout >> 8) & 255;
2126 fo = mddev->new_layout & (1<<16);
2128 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2129 (mddev->new_layout >> 17)) {
2130 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2131 mdname(mddev), mddev->new_layout);
2136 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2140 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2145 conf->tmppage = alloc_page(GFP_KERNEL);
2150 conf->raid_disks = mddev->raid_disks;
2151 conf->near_copies = nc;
2152 conf->far_copies = fc;
2153 conf->copies = nc*fc;
2154 conf->far_offset = fo;
2155 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2156 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2158 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2159 r10bio_pool_free, conf);
2160 if (!conf->r10bio_pool)
2163 size = mddev->dev_sectors >> conf->chunk_shift;
2164 sector_div(size, fc);
2165 size = size * conf->raid_disks;
2166 sector_div(size, nc);
2167 /* 'size' is now the number of chunks in the array */
2168 /* calculate "used chunks per device" in 'stride' */
2169 stride = size * conf->copies;
2171 /* We need to round up when dividing by raid_disks to
2172 * get the stride size.
2174 stride += conf->raid_disks - 1;
2175 sector_div(stride, conf->raid_disks);
2177 conf->dev_sectors = stride << conf->chunk_shift;
2182 sector_div(stride, fc);
2183 conf->stride = stride << conf->chunk_shift;
2186 spin_lock_init(&conf->device_lock);
2187 INIT_LIST_HEAD(&conf->retry_list);
2189 spin_lock_init(&conf->resync_lock);
2190 init_waitqueue_head(&conf->wait_barrier);
2192 conf->thread = md_register_thread(raid10d, mddev, NULL);
2196 conf->mddev = mddev;
2200 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2203 if (conf->r10bio_pool)
2204 mempool_destroy(conf->r10bio_pool);
2205 kfree(conf->mirrors);
2206 safe_put_page(conf->tmppage);
2209 return ERR_PTR(err);
2212 static int run(mddev_t *mddev)
2215 int i, disk_idx, chunk_size;
2216 mirror_info_t *disk;
2221 * copy the already verified devices into our private RAID10
2222 * bookkeeping area. [whatever we allocate in run(),
2223 * should be freed in stop()]
2226 if (mddev->private == NULL) {
2227 conf = setup_conf(mddev);
2229 return PTR_ERR(conf);
2230 mddev->private = conf;
2232 conf = mddev->private;
2236 mddev->thread = conf->thread;
2237 conf->thread = NULL;
2239 chunk_size = mddev->chunk_sectors << 9;
2240 blk_queue_io_min(mddev->queue, chunk_size);
2241 if (conf->raid_disks % conf->near_copies)
2242 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2244 blk_queue_io_opt(mddev->queue, chunk_size *
2245 (conf->raid_disks / conf->near_copies));
2247 list_for_each_entry(rdev, &mddev->disks, same_set) {
2248 disk_idx = rdev->raid_disk;
2249 if (disk_idx >= conf->raid_disks
2252 disk = conf->mirrors + disk_idx;
2255 disk_stack_limits(mddev->gendisk, rdev->bdev,
2256 rdev->data_offset << 9);
2257 /* as we don't honour merge_bvec_fn, we must never risk
2258 * violating it, so limit max_segments to 1 lying
2259 * within a single page.
2261 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2262 blk_queue_max_segments(mddev->queue, 1);
2263 blk_queue_segment_boundary(mddev->queue,
2264 PAGE_CACHE_SIZE - 1);
2267 disk->head_position = 0;
2269 /* need to check that every block has at least one working mirror */
2270 if (!enough(conf)) {
2271 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2276 mddev->degraded = 0;
2277 for (i = 0; i < conf->raid_disks; i++) {
2279 disk = conf->mirrors + i;
2282 !test_bit(In_sync, &disk->rdev->flags)) {
2283 disk->head_position = 0;
2290 if (mddev->recovery_cp != MaxSector)
2291 printk(KERN_NOTICE "md/raid10:%s: not clean"
2292 " -- starting background reconstruction\n",
2295 "md/raid10:%s: active with %d out of %d devices\n",
2296 mdname(mddev), conf->raid_disks - mddev->degraded,
2299 * Ok, everything is just fine now
2301 mddev->dev_sectors = conf->dev_sectors;
2302 size = raid10_size(mddev, 0, 0);
2303 md_set_array_sectors(mddev, size);
2304 mddev->resync_max_sectors = size;
2306 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2307 mddev->queue->backing_dev_info.congested_data = mddev;
2309 /* Calculate max read-ahead size.
2310 * We need to readahead at least twice a whole stripe....
2314 int stripe = conf->raid_disks *
2315 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2316 stripe /= conf->near_copies;
2317 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2318 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2321 if (conf->near_copies < conf->raid_disks)
2322 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2324 if (md_integrity_register(mddev))
2330 md_unregister_thread(mddev->thread);
2331 if (conf->r10bio_pool)
2332 mempool_destroy(conf->r10bio_pool);
2333 safe_put_page(conf->tmppage);
2334 kfree(conf->mirrors);
2336 mddev->private = NULL;
2341 static int stop(mddev_t *mddev)
2343 conf_t *conf = mddev->private;
2345 raise_barrier(conf, 0);
2346 lower_barrier(conf);
2348 md_unregister_thread(mddev->thread);
2349 mddev->thread = NULL;
2350 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2351 if (conf->r10bio_pool)
2352 mempool_destroy(conf->r10bio_pool);
2353 kfree(conf->mirrors);
2355 mddev->private = NULL;
2359 static void raid10_quiesce(mddev_t *mddev, int state)
2361 conf_t *conf = mddev->private;
2365 raise_barrier(conf, 0);
2368 lower_barrier(conf);
2373 static void *raid10_takeover_raid0(mddev_t *mddev)
2378 if (mddev->degraded > 0) {
2379 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2381 return ERR_PTR(-EINVAL);
2384 /* Set new parameters */
2385 mddev->new_level = 10;
2386 /* new layout: far_copies = 1, near_copies = 2 */
2387 mddev->new_layout = (1<<8) + 2;
2388 mddev->new_chunk_sectors = mddev->chunk_sectors;
2389 mddev->delta_disks = mddev->raid_disks;
2390 mddev->raid_disks *= 2;
2391 /* make sure it will be not marked as dirty */
2392 mddev->recovery_cp = MaxSector;
2394 conf = setup_conf(mddev);
2395 if (!IS_ERR(conf)) {
2396 list_for_each_entry(rdev, &mddev->disks, same_set)
2397 if (rdev->raid_disk >= 0)
2398 rdev->new_raid_disk = rdev->raid_disk * 2;
2405 static void *raid10_takeover(mddev_t *mddev)
2407 struct raid0_private_data *raid0_priv;
2409 /* raid10 can take over:
2410 * raid0 - providing it has only two drives
2412 if (mddev->level == 0) {
2413 /* for raid0 takeover only one zone is supported */
2414 raid0_priv = mddev->private;
2415 if (raid0_priv->nr_strip_zones > 1) {
2416 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2417 " with more than one zone.\n",
2419 return ERR_PTR(-EINVAL);
2421 return raid10_takeover_raid0(mddev);
2423 return ERR_PTR(-EINVAL);
2426 static struct mdk_personality raid10_personality =
2430 .owner = THIS_MODULE,
2431 .make_request = make_request,
2435 .error_handler = error,
2436 .hot_add_disk = raid10_add_disk,
2437 .hot_remove_disk= raid10_remove_disk,
2438 .spare_active = raid10_spare_active,
2439 .sync_request = sync_request,
2440 .quiesce = raid10_quiesce,
2441 .size = raid10_size,
2442 .takeover = raid10_takeover,
2445 static int __init raid_init(void)
2447 return register_md_personality(&raid10_personality);
2450 static void raid_exit(void)
2452 unregister_md_personality(&raid10_personality);
2455 module_init(raid_init);
2456 module_exit(raid_exit);
2457 MODULE_LICENSE("GPL");
2458 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2459 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2460 MODULE_ALIAS("md-raid10");
2461 MODULE_ALIAS("md-level-10");