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>
25 #include <linux/ratelimit.h>
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
48 * near_copies and far_copies must be at least one, and their product is at most
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
57 * Number of guaranteed r10bios in case of extreme VM load:
59 #define NR_RAID10_BIOS 256
61 static void allow_barrier(conf_t *conf);
62 static void lower_barrier(conf_t *conf);
64 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
67 int size = offsetof(struct r10bio_s, devs[conf->copies]);
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size, gfp_flags);
73 static void r10bio_pool_free(void *r10_bio, void *data)
78 /* Maximum size of each resync request */
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81 /* amount of memory to reserve for resync requests */
82 #define RESYNC_WINDOW (1024*1024)
83 /* maximum number of concurrent requests, memory permitting */
84 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
93 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
102 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
106 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107 nalloc = conf->copies; /* resync */
109 nalloc = 2; /* recovery */
114 for (j = nalloc ; j-- ; ) {
115 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
118 r10_bio->devs[j].bio = bio;
121 * Allocate RESYNC_PAGES data pages and attach them
124 for (j = 0 ; j < nalloc; j++) {
125 bio = r10_bio->devs[j].bio;
126 for (i = 0; i < RESYNC_PAGES; i++) {
127 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128 &conf->mddev->recovery)) {
129 /* we can share bv_page's during recovery */
130 struct bio *rbio = r10_bio->devs[0].bio;
131 page = rbio->bi_io_vec[i].bv_page;
134 page = alloc_page(gfp_flags);
138 bio->bi_io_vec[i].bv_page = page;
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
158 static void r10buf_pool_free(void *__r10_bio, void *data)
162 r10bio_t *r10bio = __r10_bio;
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
175 r10bio_pool_free(r10bio, conf);
178 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (*bio && *bio != IO_BLOCKED)
190 static void free_r10bio(r10bio_t *r10_bio)
192 conf_t *conf = r10_bio->mddev->private;
195 * Wake up any possible resync thread that waits for the device
200 put_all_bios(conf, r10_bio);
201 mempool_free(r10_bio, conf->r10bio_pool);
204 static void put_buf(r10bio_t *r10_bio)
206 conf_t *conf = r10_bio->mddev->private;
208 mempool_free(r10_bio, conf->r10buf_pool);
213 static void reschedule_retry(r10bio_t *r10_bio)
216 mddev_t *mddev = r10_bio->mddev;
217 conf_t *conf = mddev->private;
219 spin_lock_irqsave(&conf->device_lock, flags);
220 list_add(&r10_bio->retry_list, &conf->retry_list);
222 spin_unlock_irqrestore(&conf->device_lock, flags);
224 /* wake up frozen array... */
225 wake_up(&conf->wait_barrier);
227 md_wakeup_thread(mddev->thread);
231 * raid_end_bio_io() is called when we have finished servicing a mirrored
232 * operation and are ready to return a success/failure code to the buffer
235 static void raid_end_bio_io(r10bio_t *r10_bio)
237 struct bio *bio = r10_bio->master_bio;
240 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
241 free_r10bio(r10_bio);
245 * Update disk head position estimator based on IRQ completion info.
247 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
249 conf_t *conf = r10_bio->mddev->private;
251 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
252 r10_bio->devs[slot].addr + (r10_bio->sectors);
256 * Find the disk number which triggered given bio
258 static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio, struct bio *bio)
262 for (slot = 0; slot < conf->copies; slot++)
263 if (r10_bio->devs[slot].bio == bio)
266 BUG_ON(slot == conf->copies);
267 update_head_pos(slot, r10_bio);
269 return r10_bio->devs[slot].devnum;
272 static void raid10_end_read_request(struct bio *bio, int error)
274 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
275 r10bio_t *r10_bio = bio->bi_private;
277 conf_t *conf = r10_bio->mddev->private;
280 slot = r10_bio->read_slot;
281 dev = r10_bio->devs[slot].devnum;
283 * this branch is our 'one mirror IO has finished' event handler:
285 update_head_pos(slot, r10_bio);
289 * Set R10BIO_Uptodate in our master bio, so that
290 * we will return a good error code to the higher
291 * levels even if IO on some other mirrored buffer fails.
293 * The 'master' represents the composite IO operation to
294 * user-side. So if something waits for IO, then it will
295 * wait for the 'master' bio.
297 set_bit(R10BIO_Uptodate, &r10_bio->state);
298 raid_end_bio_io(r10_bio);
299 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
302 * oops, read error - keep the refcount on the rdev
304 char b[BDEVNAME_SIZE];
305 printk_ratelimited(KERN_ERR
306 "md/raid10:%s: %s: rescheduling sector %llu\n",
308 bdevname(conf->mirrors[dev].rdev->bdev, b),
309 (unsigned long long)r10_bio->sector);
310 reschedule_retry(r10_bio);
314 static void raid10_end_write_request(struct bio *bio, int error)
316 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
317 r10bio_t *r10_bio = bio->bi_private;
319 conf_t *conf = r10_bio->mddev->private;
321 dev = find_bio_disk(conf, r10_bio, bio);
324 * this branch is our 'one mirror IO has finished' event handler:
327 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
328 /* an I/O failed, we can't clear the bitmap */
329 set_bit(R10BIO_Degraded, &r10_bio->state);
332 * Set R10BIO_Uptodate in our master bio, so that
333 * we will return a good error code for to the higher
334 * levels even if IO on some other mirrored buffer fails.
336 * The 'master' represents the composite IO operation to
337 * user-side. So if something waits for IO, then it will
338 * wait for the 'master' bio.
340 set_bit(R10BIO_Uptodate, &r10_bio->state);
344 * Let's see if all mirrored write operations have finished
347 if (atomic_dec_and_test(&r10_bio->remaining)) {
348 /* clear the bitmap if all writes complete successfully */
349 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
351 !test_bit(R10BIO_Degraded, &r10_bio->state),
353 md_write_end(r10_bio->mddev);
354 raid_end_bio_io(r10_bio);
357 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
362 * RAID10 layout manager
363 * As well as the chunksize and raid_disks count, there are two
364 * parameters: near_copies and far_copies.
365 * near_copies * far_copies must be <= raid_disks.
366 * Normally one of these will be 1.
367 * If both are 1, we get raid0.
368 * If near_copies == raid_disks, we get raid1.
370 * Chunks are laid out in raid0 style with near_copies copies of the
371 * first chunk, followed by near_copies copies of the next chunk and
373 * If far_copies > 1, then after 1/far_copies of the array has been assigned
374 * as described above, we start again with a device offset of near_copies.
375 * So we effectively have another copy of the whole array further down all
376 * the drives, but with blocks on different drives.
377 * With this layout, and block is never stored twice on the one device.
379 * raid10_find_phys finds the sector offset of a given virtual sector
380 * on each device that it is on.
382 * raid10_find_virt does the reverse mapping, from a device and a
383 * sector offset to a virtual address
386 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
396 /* now calculate first sector/dev */
397 chunk = r10bio->sector >> conf->chunk_shift;
398 sector = r10bio->sector & conf->chunk_mask;
400 chunk *= conf->near_copies;
402 dev = sector_div(stripe, conf->raid_disks);
403 if (conf->far_offset)
404 stripe *= conf->far_copies;
406 sector += stripe << conf->chunk_shift;
408 /* and calculate all the others */
409 for (n=0; n < conf->near_copies; n++) {
412 r10bio->devs[slot].addr = sector;
413 r10bio->devs[slot].devnum = d;
416 for (f = 1; f < conf->far_copies; f++) {
417 d += conf->near_copies;
418 if (d >= conf->raid_disks)
419 d -= conf->raid_disks;
421 r10bio->devs[slot].devnum = d;
422 r10bio->devs[slot].addr = s;
426 if (dev >= conf->raid_disks) {
428 sector += (conf->chunk_mask + 1);
431 BUG_ON(slot != conf->copies);
434 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
436 sector_t offset, chunk, vchunk;
438 offset = sector & conf->chunk_mask;
439 if (conf->far_offset) {
441 chunk = sector >> conf->chunk_shift;
442 fc = sector_div(chunk, conf->far_copies);
443 dev -= fc * conf->near_copies;
445 dev += conf->raid_disks;
447 while (sector >= conf->stride) {
448 sector -= conf->stride;
449 if (dev < conf->near_copies)
450 dev += conf->raid_disks - conf->near_copies;
452 dev -= conf->near_copies;
454 chunk = sector >> conf->chunk_shift;
456 vchunk = chunk * conf->raid_disks + dev;
457 sector_div(vchunk, conf->near_copies);
458 return (vchunk << conf->chunk_shift) + offset;
462 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
464 * @bvm: properties of new bio
465 * @biovec: the request that could be merged to it.
467 * Return amount of bytes we can accept at this offset
468 * If near_copies == raid_disk, there are no striping issues,
469 * but in that case, the function isn't called at all.
471 static int raid10_mergeable_bvec(struct request_queue *q,
472 struct bvec_merge_data *bvm,
473 struct bio_vec *biovec)
475 mddev_t *mddev = q->queuedata;
476 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
478 unsigned int chunk_sectors = mddev->chunk_sectors;
479 unsigned int bio_sectors = bvm->bi_size >> 9;
481 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
482 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
483 if (max <= biovec->bv_len && bio_sectors == 0)
484 return biovec->bv_len;
490 * This routine returns the disk from which the requested read should
491 * be done. There is a per-array 'next expected sequential IO' sector
492 * number - if this matches on the next IO then we use the last disk.
493 * There is also a per-disk 'last know head position' sector that is
494 * maintained from IRQ contexts, both the normal and the resync IO
495 * completion handlers update this position correctly. If there is no
496 * perfect sequential match then we pick the disk whose head is closest.
498 * If there are 2 mirrors in the same 2 devices, performance degrades
499 * because position is mirror, not device based.
501 * The rdev for the device selected will have nr_pending incremented.
505 * FIXME: possibly should rethink readbalancing and do it differently
506 * depending on near_copies / far_copies geometry.
508 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
510 const sector_t this_sector = r10_bio->sector;
512 const int sectors = r10_bio->sectors;
513 sector_t new_distance, best_dist;
518 raid10_find_phys(conf, r10_bio);
522 best_dist = MaxSector;
525 * Check if we can balance. We can balance on the whole
526 * device if no resync is going on (recovery is ok), or below
527 * the resync window. We take the first readable disk when
528 * above the resync window.
530 if (conf->mddev->recovery_cp < MaxSector
531 && (this_sector + sectors >= conf->next_resync))
534 for (slot = 0; slot < conf->copies ; slot++) {
535 if (r10_bio->devs[slot].bio == IO_BLOCKED)
537 disk = r10_bio->devs[slot].devnum;
538 rdev = rcu_dereference(conf->mirrors[disk].rdev);
541 if (!test_bit(In_sync, &rdev->flags))
547 /* This optimisation is debatable, and completely destroys
548 * sequential read speed for 'far copies' arrays. So only
549 * keep it for 'near' arrays, and review those later.
551 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
554 /* for far > 1 always use the lowest address */
555 if (conf->far_copies > 1)
556 new_distance = r10_bio->devs[slot].addr;
558 new_distance = abs(r10_bio->devs[slot].addr -
559 conf->mirrors[disk].head_position);
560 if (new_distance < best_dist) {
561 best_dist = new_distance;
565 if (slot == conf->copies)
569 disk = r10_bio->devs[slot].devnum;
570 rdev = rcu_dereference(conf->mirrors[disk].rdev);
573 atomic_inc(&rdev->nr_pending);
574 if (test_bit(Faulty, &rdev->flags)) {
575 /* Cannot risk returning a device that failed
576 * before we inc'ed nr_pending
578 rdev_dec_pending(rdev, conf->mddev);
581 r10_bio->read_slot = slot;
589 static int raid10_congested(void *data, int bits)
591 mddev_t *mddev = data;
592 conf_t *conf = mddev->private;
595 if (mddev_congested(mddev, bits))
598 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
599 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
600 if (rdev && !test_bit(Faulty, &rdev->flags)) {
601 struct request_queue *q = bdev_get_queue(rdev->bdev);
603 ret |= bdi_congested(&q->backing_dev_info, bits);
610 static void flush_pending_writes(conf_t *conf)
612 /* Any writes that have been queued but are awaiting
613 * bitmap updates get flushed here.
615 spin_lock_irq(&conf->device_lock);
617 if (conf->pending_bio_list.head) {
619 bio = bio_list_get(&conf->pending_bio_list);
620 spin_unlock_irq(&conf->device_lock);
621 /* flush any pending bitmap writes to disk
622 * before proceeding w/ I/O */
623 bitmap_unplug(conf->mddev->bitmap);
625 while (bio) { /* submit pending writes */
626 struct bio *next = bio->bi_next;
628 generic_make_request(bio);
632 spin_unlock_irq(&conf->device_lock);
636 * Sometimes we need to suspend IO while we do something else,
637 * either some resync/recovery, or reconfigure the array.
638 * To do this we raise a 'barrier'.
639 * The 'barrier' is a counter that can be raised multiple times
640 * to count how many activities are happening which preclude
642 * We can only raise the barrier if there is no pending IO.
643 * i.e. if nr_pending == 0.
644 * We choose only to raise the barrier if no-one is waiting for the
645 * barrier to go down. This means that as soon as an IO request
646 * is ready, no other operations which require a barrier will start
647 * until the IO request has had a chance.
649 * So: regular IO calls 'wait_barrier'. When that returns there
650 * is no backgroup IO happening, It must arrange to call
651 * allow_barrier when it has finished its IO.
652 * backgroup IO calls must call raise_barrier. Once that returns
653 * there is no normal IO happeing. It must arrange to call
654 * lower_barrier when the particular background IO completes.
657 static void raise_barrier(conf_t *conf, int force)
659 BUG_ON(force && !conf->barrier);
660 spin_lock_irq(&conf->resync_lock);
662 /* Wait until no block IO is waiting (unless 'force') */
663 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
664 conf->resync_lock, );
666 /* block any new IO from starting */
669 /* Now wait for all pending IO to complete */
670 wait_event_lock_irq(conf->wait_barrier,
671 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
672 conf->resync_lock, );
674 spin_unlock_irq(&conf->resync_lock);
677 static void lower_barrier(conf_t *conf)
680 spin_lock_irqsave(&conf->resync_lock, flags);
682 spin_unlock_irqrestore(&conf->resync_lock, flags);
683 wake_up(&conf->wait_barrier);
686 static void wait_barrier(conf_t *conf)
688 spin_lock_irq(&conf->resync_lock);
691 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
697 spin_unlock_irq(&conf->resync_lock);
700 static void allow_barrier(conf_t *conf)
703 spin_lock_irqsave(&conf->resync_lock, flags);
705 spin_unlock_irqrestore(&conf->resync_lock, flags);
706 wake_up(&conf->wait_barrier);
709 static void freeze_array(conf_t *conf)
711 /* stop syncio and normal IO and wait for everything to
713 * We increment barrier and nr_waiting, and then
714 * wait until nr_pending match nr_queued+1
715 * This is called in the context of one normal IO request
716 * that has failed. Thus any sync request that might be pending
717 * will be blocked by nr_pending, and we need to wait for
718 * pending IO requests to complete or be queued for re-try.
719 * Thus the number queued (nr_queued) plus this request (1)
720 * must match the number of pending IOs (nr_pending) before
723 spin_lock_irq(&conf->resync_lock);
726 wait_event_lock_irq(conf->wait_barrier,
727 conf->nr_pending == conf->nr_queued+1,
729 flush_pending_writes(conf));
731 spin_unlock_irq(&conf->resync_lock);
734 static void unfreeze_array(conf_t *conf)
736 /* reverse the effect of the freeze */
737 spin_lock_irq(&conf->resync_lock);
740 wake_up(&conf->wait_barrier);
741 spin_unlock_irq(&conf->resync_lock);
744 static int make_request(mddev_t *mddev, struct bio * bio)
746 conf_t *conf = mddev->private;
747 mirror_info_t *mirror;
749 struct bio *read_bio;
751 int chunk_sects = conf->chunk_mask + 1;
752 const int rw = bio_data_dir(bio);
753 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
754 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
756 mdk_rdev_t *blocked_rdev;
759 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
760 md_flush_request(mddev, bio);
764 /* If this request crosses a chunk boundary, we need to
765 * split it. This will only happen for 1 PAGE (or less) requests.
767 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
769 conf->near_copies < conf->raid_disks)) {
771 /* Sanity check -- queue functions should prevent this happening */
772 if (bio->bi_vcnt != 1 ||
775 /* This is a one page bio that upper layers
776 * refuse to split for us, so we need to split it.
779 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
781 /* Each of these 'make_request' calls will call 'wait_barrier'.
782 * If the first succeeds but the second blocks due to the resync
783 * thread raising the barrier, we will deadlock because the
784 * IO to the underlying device will be queued in generic_make_request
785 * and will never complete, so will never reduce nr_pending.
786 * So increment nr_waiting here so no new raise_barriers will
787 * succeed, and so the second wait_barrier cannot block.
789 spin_lock_irq(&conf->resync_lock);
791 spin_unlock_irq(&conf->resync_lock);
793 if (make_request(mddev, &bp->bio1))
794 generic_make_request(&bp->bio1);
795 if (make_request(mddev, &bp->bio2))
796 generic_make_request(&bp->bio2);
798 spin_lock_irq(&conf->resync_lock);
800 wake_up(&conf->wait_barrier);
801 spin_unlock_irq(&conf->resync_lock);
803 bio_pair_release(bp);
806 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
807 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
808 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
814 md_write_start(mddev, bio);
817 * Register the new request and wait if the reconstruction
818 * thread has put up a bar for new requests.
819 * Continue immediately if no resync is active currently.
823 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
825 r10_bio->master_bio = bio;
826 r10_bio->sectors = bio->bi_size >> 9;
828 r10_bio->mddev = mddev;
829 r10_bio->sector = bio->bi_sector;
834 * read balancing logic:
836 int disk = read_balance(conf, r10_bio);
837 int slot = r10_bio->read_slot;
839 raid_end_bio_io(r10_bio);
842 mirror = conf->mirrors + disk;
844 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
846 r10_bio->devs[slot].bio = read_bio;
848 read_bio->bi_sector = r10_bio->devs[slot].addr +
849 mirror->rdev->data_offset;
850 read_bio->bi_bdev = mirror->rdev->bdev;
851 read_bio->bi_end_io = raid10_end_read_request;
852 read_bio->bi_rw = READ | do_sync;
853 read_bio->bi_private = r10_bio;
855 generic_make_request(read_bio);
862 /* first select target devices under rcu_lock and
863 * inc refcount on their rdev. Record them by setting
866 plugged = mddev_check_plugged(mddev);
868 raid10_find_phys(conf, r10_bio);
872 for (i = 0; i < conf->copies; i++) {
873 int d = r10_bio->devs[i].devnum;
874 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
875 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
876 atomic_inc(&rdev->nr_pending);
880 if (rdev && !test_bit(Faulty, &rdev->flags)) {
881 atomic_inc(&rdev->nr_pending);
882 r10_bio->devs[i].bio = bio;
884 r10_bio->devs[i].bio = NULL;
885 set_bit(R10BIO_Degraded, &r10_bio->state);
890 if (unlikely(blocked_rdev)) {
891 /* Have to wait for this device to get unblocked, then retry */
895 for (j = 0; j < i; j++)
896 if (r10_bio->devs[j].bio) {
897 d = r10_bio->devs[j].devnum;
898 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
901 md_wait_for_blocked_rdev(blocked_rdev, mddev);
906 atomic_set(&r10_bio->remaining, 1);
907 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
909 for (i = 0; i < conf->copies; i++) {
911 int d = r10_bio->devs[i].devnum;
912 if (!r10_bio->devs[i].bio)
915 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
916 r10_bio->devs[i].bio = mbio;
918 mbio->bi_sector = r10_bio->devs[i].addr+
919 conf->mirrors[d].rdev->data_offset;
920 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
921 mbio->bi_end_io = raid10_end_write_request;
922 mbio->bi_rw = WRITE | do_sync | do_fua;
923 mbio->bi_private = r10_bio;
925 atomic_inc(&r10_bio->remaining);
926 spin_lock_irqsave(&conf->device_lock, flags);
927 bio_list_add(&conf->pending_bio_list, mbio);
928 spin_unlock_irqrestore(&conf->device_lock, flags);
931 if (atomic_dec_and_test(&r10_bio->remaining)) {
932 /* This matches the end of raid10_end_write_request() */
933 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
935 !test_bit(R10BIO_Degraded, &r10_bio->state),
938 raid_end_bio_io(r10_bio);
941 /* In case raid10d snuck in to freeze_array */
942 wake_up(&conf->wait_barrier);
944 if (do_sync || !mddev->bitmap || !plugged)
945 md_wakeup_thread(mddev->thread);
949 static void status(struct seq_file *seq, mddev_t *mddev)
951 conf_t *conf = mddev->private;
954 if (conf->near_copies < conf->raid_disks)
955 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
956 if (conf->near_copies > 1)
957 seq_printf(seq, " %d near-copies", conf->near_copies);
958 if (conf->far_copies > 1) {
959 if (conf->far_offset)
960 seq_printf(seq, " %d offset-copies", conf->far_copies);
962 seq_printf(seq, " %d far-copies", conf->far_copies);
964 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
965 conf->raid_disks - mddev->degraded);
966 for (i = 0; i < conf->raid_disks; i++)
967 seq_printf(seq, "%s",
968 conf->mirrors[i].rdev &&
969 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
970 seq_printf(seq, "]");
973 /* check if there are enough drives for
974 * every block to appear on atleast one.
975 * Don't consider the device numbered 'ignore'
976 * as we might be about to remove it.
978 static int enough(conf_t *conf, int ignore)
983 int n = conf->copies;
986 if (conf->mirrors[first].rdev &&
989 first = (first+1) % conf->raid_disks;
993 } while (first != 0);
997 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
999 char b[BDEVNAME_SIZE];
1000 conf_t *conf = mddev->private;
1003 * If it is not operational, then we have already marked it as dead
1004 * else if it is the last working disks, ignore the error, let the
1005 * next level up know.
1006 * else mark the drive as failed
1008 if (test_bit(In_sync, &rdev->flags)
1009 && !enough(conf, rdev->raid_disk))
1011 * Don't fail the drive, just return an IO error.
1014 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1015 unsigned long flags;
1016 spin_lock_irqsave(&conf->device_lock, flags);
1018 spin_unlock_irqrestore(&conf->device_lock, flags);
1020 * if recovery is running, make sure it aborts.
1022 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1024 set_bit(Faulty, &rdev->flags);
1025 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1027 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1028 "md/raid10:%s: Operation continuing on %d devices.\n",
1029 mdname(mddev), bdevname(rdev->bdev, b),
1030 mdname(mddev), conf->raid_disks - mddev->degraded);
1033 static void print_conf(conf_t *conf)
1038 printk(KERN_DEBUG "RAID10 conf printout:\n");
1040 printk(KERN_DEBUG "(!conf)\n");
1043 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1046 for (i = 0; i < conf->raid_disks; i++) {
1047 char b[BDEVNAME_SIZE];
1048 tmp = conf->mirrors + i;
1050 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1051 i, !test_bit(In_sync, &tmp->rdev->flags),
1052 !test_bit(Faulty, &tmp->rdev->flags),
1053 bdevname(tmp->rdev->bdev,b));
1057 static void close_sync(conf_t *conf)
1060 allow_barrier(conf);
1062 mempool_destroy(conf->r10buf_pool);
1063 conf->r10buf_pool = NULL;
1066 static int raid10_spare_active(mddev_t *mddev)
1069 conf_t *conf = mddev->private;
1072 unsigned long flags;
1075 * Find all non-in_sync disks within the RAID10 configuration
1076 * and mark them in_sync
1078 for (i = 0; i < conf->raid_disks; i++) {
1079 tmp = conf->mirrors + i;
1081 && !test_bit(Faulty, &tmp->rdev->flags)
1082 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1084 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1087 spin_lock_irqsave(&conf->device_lock, flags);
1088 mddev->degraded -= count;
1089 spin_unlock_irqrestore(&conf->device_lock, flags);
1096 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1098 conf_t *conf = mddev->private;
1102 int last = conf->raid_disks - 1;
1104 if (rdev->badblocks.count)
1107 if (mddev->recovery_cp < MaxSector)
1108 /* only hot-add to in-sync arrays, as recovery is
1109 * very different from resync
1112 if (!enough(conf, -1))
1115 if (rdev->raid_disk >= 0)
1116 first = last = rdev->raid_disk;
1118 if (rdev->saved_raid_disk >= first &&
1119 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1120 mirror = rdev->saved_raid_disk;
1123 for ( ; mirror <= last ; mirror++) {
1124 mirror_info_t *p = &conf->mirrors[mirror];
1125 if (p->recovery_disabled == mddev->recovery_disabled)
1130 disk_stack_limits(mddev->gendisk, rdev->bdev,
1131 rdev->data_offset << 9);
1132 /* as we don't honour merge_bvec_fn, we must
1133 * never risk violating it, so limit
1134 * ->max_segments to one lying with a single
1135 * page, as a one page request is never in
1138 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1139 blk_queue_max_segments(mddev->queue, 1);
1140 blk_queue_segment_boundary(mddev->queue,
1141 PAGE_CACHE_SIZE - 1);
1144 p->head_position = 0;
1145 rdev->raid_disk = mirror;
1147 if (rdev->saved_raid_disk != mirror)
1149 rcu_assign_pointer(p->rdev, rdev);
1153 md_integrity_add_rdev(rdev, mddev);
1158 static int raid10_remove_disk(mddev_t *mddev, int number)
1160 conf_t *conf = mddev->private;
1163 mirror_info_t *p = conf->mirrors+ number;
1168 if (test_bit(In_sync, &rdev->flags) ||
1169 atomic_read(&rdev->nr_pending)) {
1173 /* Only remove faulty devices in recovery
1176 if (!test_bit(Faulty, &rdev->flags) &&
1177 mddev->recovery_disabled != p->recovery_disabled &&
1184 if (atomic_read(&rdev->nr_pending)) {
1185 /* lost the race, try later */
1190 err = md_integrity_register(mddev);
1199 static void end_sync_read(struct bio *bio, int error)
1201 r10bio_t *r10_bio = bio->bi_private;
1202 conf_t *conf = r10_bio->mddev->private;
1205 d = find_bio_disk(conf, r10_bio, bio);
1207 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1208 set_bit(R10BIO_Uptodate, &r10_bio->state);
1210 atomic_add(r10_bio->sectors,
1211 &conf->mirrors[d].rdev->corrected_errors);
1212 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1213 md_error(r10_bio->mddev,
1214 conf->mirrors[d].rdev);
1217 /* for reconstruct, we always reschedule after a read.
1218 * for resync, only after all reads
1220 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1221 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1222 atomic_dec_and_test(&r10_bio->remaining)) {
1223 /* we have read all the blocks,
1224 * do the comparison in process context in raid10d
1226 reschedule_retry(r10_bio);
1230 static void end_sync_write(struct bio *bio, int error)
1232 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1233 r10bio_t *r10_bio = bio->bi_private;
1234 mddev_t *mddev = r10_bio->mddev;
1235 conf_t *conf = mddev->private;
1238 d = find_bio_disk(conf, r10_bio, bio);
1241 md_error(mddev, conf->mirrors[d].rdev);
1243 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1244 while (atomic_dec_and_test(&r10_bio->remaining)) {
1245 if (r10_bio->master_bio == NULL) {
1246 /* the primary of several recovery bios */
1247 sector_t s = r10_bio->sectors;
1249 md_done_sync(mddev, s, 1);
1252 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1260 * Note: sync and recover and handled very differently for raid10
1261 * This code is for resync.
1262 * For resync, we read through virtual addresses and read all blocks.
1263 * If there is any error, we schedule a write. The lowest numbered
1264 * drive is authoritative.
1265 * However requests come for physical address, so we need to map.
1266 * For every physical address there are raid_disks/copies virtual addresses,
1267 * which is always are least one, but is not necessarly an integer.
1268 * This means that a physical address can span multiple chunks, so we may
1269 * have to submit multiple io requests for a single sync request.
1272 * We check if all blocks are in-sync and only write to blocks that
1275 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1277 conf_t *conf = mddev->private;
1279 struct bio *tbio, *fbio;
1281 atomic_set(&r10_bio->remaining, 1);
1283 /* find the first device with a block */
1284 for (i=0; i<conf->copies; i++)
1285 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1288 if (i == conf->copies)
1292 fbio = r10_bio->devs[i].bio;
1294 /* now find blocks with errors */
1295 for (i=0 ; i < conf->copies ; i++) {
1297 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1299 tbio = r10_bio->devs[i].bio;
1301 if (tbio->bi_end_io != end_sync_read)
1305 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1306 /* We know that the bi_io_vec layout is the same for
1307 * both 'first' and 'i', so we just compare them.
1308 * All vec entries are PAGE_SIZE;
1310 for (j = 0; j < vcnt; j++)
1311 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1312 page_address(tbio->bi_io_vec[j].bv_page),
1317 mddev->resync_mismatches += r10_bio->sectors;
1319 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1320 /* Don't fix anything. */
1322 /* Ok, we need to write this bio
1323 * First we need to fixup bv_offset, bv_len and
1324 * bi_vecs, as the read request might have corrupted these
1326 tbio->bi_vcnt = vcnt;
1327 tbio->bi_size = r10_bio->sectors << 9;
1329 tbio->bi_phys_segments = 0;
1330 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1331 tbio->bi_flags |= 1 << BIO_UPTODATE;
1332 tbio->bi_next = NULL;
1333 tbio->bi_rw = WRITE;
1334 tbio->bi_private = r10_bio;
1335 tbio->bi_sector = r10_bio->devs[i].addr;
1337 for (j=0; j < vcnt ; j++) {
1338 tbio->bi_io_vec[j].bv_offset = 0;
1339 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1341 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1342 page_address(fbio->bi_io_vec[j].bv_page),
1345 tbio->bi_end_io = end_sync_write;
1347 d = r10_bio->devs[i].devnum;
1348 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1349 atomic_inc(&r10_bio->remaining);
1350 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1352 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1353 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1354 generic_make_request(tbio);
1358 if (atomic_dec_and_test(&r10_bio->remaining)) {
1359 md_done_sync(mddev, r10_bio->sectors, 1);
1365 * Now for the recovery code.
1366 * Recovery happens across physical sectors.
1367 * We recover all non-is_sync drives by finding the virtual address of
1368 * each, and then choose a working drive that also has that virt address.
1369 * There is a separate r10_bio for each non-in_sync drive.
1370 * Only the first two slots are in use. The first for reading,
1371 * The second for writing.
1375 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1377 conf_t *conf = mddev->private;
1382 * share the pages with the first bio
1383 * and submit the write request
1385 wbio = r10_bio->devs[1].bio;
1386 d = r10_bio->devs[1].devnum;
1388 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1389 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1390 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1391 generic_make_request(wbio);
1394 "md/raid10:%s: recovery aborted due to read error\n",
1396 conf->mirrors[d].recovery_disabled = mddev->recovery_disabled;
1397 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1404 * Used by fix_read_error() to decay the per rdev read_errors.
1405 * We halve the read error count for every hour that has elapsed
1406 * since the last recorded read error.
1409 static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1411 struct timespec cur_time_mon;
1412 unsigned long hours_since_last;
1413 unsigned int read_errors = atomic_read(&rdev->read_errors);
1415 ktime_get_ts(&cur_time_mon);
1417 if (rdev->last_read_error.tv_sec == 0 &&
1418 rdev->last_read_error.tv_nsec == 0) {
1419 /* first time we've seen a read error */
1420 rdev->last_read_error = cur_time_mon;
1424 hours_since_last = (cur_time_mon.tv_sec -
1425 rdev->last_read_error.tv_sec) / 3600;
1427 rdev->last_read_error = cur_time_mon;
1430 * if hours_since_last is > the number of bits in read_errors
1431 * just set read errors to 0. We do this to avoid
1432 * overflowing the shift of read_errors by hours_since_last.
1434 if (hours_since_last >= 8 * sizeof(read_errors))
1435 atomic_set(&rdev->read_errors, 0);
1437 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1441 * This is a kernel thread which:
1443 * 1. Retries failed read operations on working mirrors.
1444 * 2. Updates the raid superblock when problems encounter.
1445 * 3. Performs writes following reads for array synchronising.
1448 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1450 int sect = 0; /* Offset from r10_bio->sector */
1451 int sectors = r10_bio->sectors;
1453 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1454 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1456 /* still own a reference to this rdev, so it cannot
1457 * have been cleared recently.
1459 rdev = conf->mirrors[d].rdev;
1461 if (test_bit(Faulty, &rdev->flags))
1462 /* drive has already been failed, just ignore any
1463 more fix_read_error() attempts */
1466 check_decay_read_errors(mddev, rdev);
1467 atomic_inc(&rdev->read_errors);
1468 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1469 char b[BDEVNAME_SIZE];
1470 bdevname(rdev->bdev, b);
1473 "md/raid10:%s: %s: Raid device exceeded "
1474 "read_error threshold [cur %d:max %d]\n",
1476 atomic_read(&rdev->read_errors), max_read_errors);
1478 "md/raid10:%s: %s: Failing raid device\n",
1480 md_error(mddev, conf->mirrors[d].rdev);
1486 int sl = r10_bio->read_slot;
1490 if (s > (PAGE_SIZE>>9))
1495 d = r10_bio->devs[sl].devnum;
1496 rdev = rcu_dereference(conf->mirrors[d].rdev);
1498 test_bit(In_sync, &rdev->flags)) {
1499 atomic_inc(&rdev->nr_pending);
1501 success = sync_page_io(rdev,
1502 r10_bio->devs[sl].addr +
1505 conf->tmppage, READ, false);
1506 rdev_dec_pending(rdev, mddev);
1512 if (sl == conf->copies)
1514 } while (!success && sl != r10_bio->read_slot);
1518 /* Cannot read from anywhere -- bye bye array */
1519 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1520 md_error(mddev, conf->mirrors[dn].rdev);
1525 /* write it back and re-read */
1527 while (sl != r10_bio->read_slot) {
1528 char b[BDEVNAME_SIZE];
1533 d = r10_bio->devs[sl].devnum;
1534 rdev = rcu_dereference(conf->mirrors[d].rdev);
1536 test_bit(In_sync, &rdev->flags)) {
1537 atomic_inc(&rdev->nr_pending);
1539 if (sync_page_io(rdev,
1540 r10_bio->devs[sl].addr +
1542 s<<9, conf->tmppage, WRITE, false)
1544 /* Well, this device is dead */
1546 "md/raid10:%s: read correction "
1548 " (%d sectors at %llu on %s)\n",
1550 (unsigned long long)(
1551 sect + rdev->data_offset),
1552 bdevname(rdev->bdev, b));
1553 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1556 bdevname(rdev->bdev, b));
1557 md_error(mddev, rdev);
1559 rdev_dec_pending(rdev, mddev);
1564 while (sl != r10_bio->read_slot) {
1569 d = r10_bio->devs[sl].devnum;
1570 rdev = rcu_dereference(conf->mirrors[d].rdev);
1572 test_bit(In_sync, &rdev->flags)) {
1573 char b[BDEVNAME_SIZE];
1574 atomic_inc(&rdev->nr_pending);
1576 if (sync_page_io(rdev,
1577 r10_bio->devs[sl].addr +
1579 s<<9, conf->tmppage,
1580 READ, false) == 0) {
1581 /* Well, this device is dead */
1583 "md/raid10:%s: unable to read back "
1585 " (%d sectors at %llu on %s)\n",
1587 (unsigned long long)(
1588 sect + rdev->data_offset),
1589 bdevname(rdev->bdev, b));
1590 printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1592 bdevname(rdev->bdev, b));
1594 md_error(mddev, rdev);
1597 "md/raid10:%s: read error corrected"
1598 " (%d sectors at %llu on %s)\n",
1600 (unsigned long long)(
1601 sect + rdev->data_offset),
1602 bdevname(rdev->bdev, b));
1603 atomic_add(s, &rdev->corrected_errors);
1606 rdev_dec_pending(rdev, mddev);
1617 static void raid10d(mddev_t *mddev)
1621 unsigned long flags;
1622 conf_t *conf = mddev->private;
1623 struct list_head *head = &conf->retry_list;
1625 struct blk_plug plug;
1627 md_check_recovery(mddev);
1629 blk_start_plug(&plug);
1631 char b[BDEVNAME_SIZE];
1633 flush_pending_writes(conf);
1635 spin_lock_irqsave(&conf->device_lock, flags);
1636 if (list_empty(head)) {
1637 spin_unlock_irqrestore(&conf->device_lock, flags);
1640 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1641 list_del(head->prev);
1643 spin_unlock_irqrestore(&conf->device_lock, flags);
1645 mddev = r10_bio->mddev;
1646 conf = mddev->private;
1647 if (test_bit(R10BIO_IsSync, &r10_bio->state))
1648 sync_request_write(mddev, r10_bio);
1649 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
1650 recovery_request_write(mddev, r10_bio);
1652 int slot = r10_bio->read_slot;
1653 int mirror = r10_bio->devs[slot].devnum;
1654 /* we got a read error. Maybe the drive is bad. Maybe just
1655 * the block and we can fix it.
1656 * We freeze all other IO, and try reading the block from
1657 * other devices. When we find one, we re-write
1658 * and check it that fixes the read error.
1659 * This is all done synchronously while the array is
1662 if (mddev->ro == 0) {
1664 fix_read_error(conf, mddev, r10_bio);
1665 unfreeze_array(conf);
1667 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
1669 bio = r10_bio->devs[slot].bio;
1670 r10_bio->devs[slot].bio =
1671 mddev->ro ? IO_BLOCKED : NULL;
1672 mirror = read_balance(conf, r10_bio);
1674 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
1675 " read error for block %llu\n",
1677 bdevname(bio->bi_bdev,b),
1678 (unsigned long long)r10_bio->sector);
1679 raid_end_bio_io(r10_bio);
1682 const unsigned long do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
1684 slot = r10_bio->read_slot;
1685 rdev = conf->mirrors[mirror].rdev;
1688 "md/raid10:%s: %s: redirecting"
1689 "sector %llu to another mirror\n",
1691 bdevname(rdev->bdev, b),
1692 (unsigned long long)r10_bio->sector);
1693 bio = bio_clone_mddev(r10_bio->master_bio,
1695 r10_bio->devs[slot].bio = bio;
1696 bio->bi_sector = r10_bio->devs[slot].addr
1697 + rdev->data_offset;
1698 bio->bi_bdev = rdev->bdev;
1699 bio->bi_rw = READ | do_sync;
1700 bio->bi_private = r10_bio;
1701 bio->bi_end_io = raid10_end_read_request;
1702 generic_make_request(bio);
1707 blk_finish_plug(&plug);
1711 static int init_resync(conf_t *conf)
1715 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1716 BUG_ON(conf->r10buf_pool);
1717 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1718 if (!conf->r10buf_pool)
1720 conf->next_resync = 0;
1725 * perform a "sync" on one "block"
1727 * We need to make sure that no normal I/O request - particularly write
1728 * requests - conflict with active sync requests.
1730 * This is achieved by tracking pending requests and a 'barrier' concept
1731 * that can be installed to exclude normal IO requests.
1733 * Resync and recovery are handled very differently.
1734 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1736 * For resync, we iterate over virtual addresses, read all copies,
1737 * and update if there are differences. If only one copy is live,
1739 * For recovery, we iterate over physical addresses, read a good
1740 * value for each non-in_sync drive, and over-write.
1742 * So, for recovery we may have several outstanding complex requests for a
1743 * given address, one for each out-of-sync device. We model this by allocating
1744 * a number of r10_bio structures, one for each out-of-sync device.
1745 * As we setup these structures, we collect all bio's together into a list
1746 * which we then process collectively to add pages, and then process again
1747 * to pass to generic_make_request.
1749 * The r10_bio structures are linked using a borrowed master_bio pointer.
1750 * This link is counted in ->remaining. When the r10_bio that points to NULL
1751 * has its remaining count decremented to 0, the whole complex operation
1756 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
1757 int *skipped, int go_faster)
1759 conf_t *conf = mddev->private;
1761 struct bio *biolist = NULL, *bio;
1762 sector_t max_sector, nr_sectors;
1765 sector_t sync_blocks;
1767 sector_t sectors_skipped = 0;
1768 int chunks_skipped = 0;
1770 if (!conf->r10buf_pool)
1771 if (init_resync(conf))
1775 max_sector = mddev->dev_sectors;
1776 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1777 max_sector = mddev->resync_max_sectors;
1778 if (sector_nr >= max_sector) {
1779 /* If we aborted, we need to abort the
1780 * sync on the 'current' bitmap chucks (there can
1781 * be several when recovering multiple devices).
1782 * as we may have started syncing it but not finished.
1783 * We can find the current address in
1784 * mddev->curr_resync, but for recovery,
1785 * we need to convert that to several
1786 * virtual addresses.
1788 if (mddev->curr_resync < max_sector) { /* aborted */
1789 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1790 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1792 else for (i=0; i<conf->raid_disks; i++) {
1794 raid10_find_virt(conf, mddev->curr_resync, i);
1795 bitmap_end_sync(mddev->bitmap, sect,
1798 } else /* completed sync */
1801 bitmap_close_sync(mddev->bitmap);
1804 return sectors_skipped;
1806 if (chunks_skipped >= conf->raid_disks) {
1807 /* if there has been nothing to do on any drive,
1808 * then there is nothing to do at all..
1811 return (max_sector - sector_nr) + sectors_skipped;
1814 if (max_sector > mddev->resync_max)
1815 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1817 /* make sure whole request will fit in a chunk - if chunks
1820 if (conf->near_copies < conf->raid_disks &&
1821 max_sector > (sector_nr | conf->chunk_mask))
1822 max_sector = (sector_nr | conf->chunk_mask) + 1;
1824 * If there is non-resync activity waiting for us then
1825 * put in a delay to throttle resync.
1827 if (!go_faster && conf->nr_waiting)
1828 msleep_interruptible(1000);
1830 /* Again, very different code for resync and recovery.
1831 * Both must result in an r10bio with a list of bios that
1832 * have bi_end_io, bi_sector, bi_bdev set,
1833 * and bi_private set to the r10bio.
1834 * For recovery, we may actually create several r10bios
1835 * with 2 bios in each, that correspond to the bios in the main one.
1836 * In this case, the subordinate r10bios link back through a
1837 * borrowed master_bio pointer, and the counter in the master
1838 * includes a ref from each subordinate.
1840 /* First, we decide what to do and set ->bi_end_io
1841 * To end_sync_read if we want to read, and
1842 * end_sync_write if we will want to write.
1845 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1846 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1847 /* recovery... the complicated one */
1851 for (i=0 ; i<conf->raid_disks; i++) {
1857 if (conf->mirrors[i].rdev == NULL ||
1858 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
1862 /* want to reconstruct this device */
1864 sect = raid10_find_virt(conf, sector_nr, i);
1865 /* Unless we are doing a full sync, we only need
1866 * to recover the block if it is set in the bitmap
1868 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1870 if (sync_blocks < max_sync)
1871 max_sync = sync_blocks;
1874 /* yep, skip the sync_blocks here, but don't assume
1875 * that there will never be anything to do here
1877 chunks_skipped = -1;
1881 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1882 raise_barrier(conf, rb2 != NULL);
1883 atomic_set(&r10_bio->remaining, 0);
1885 r10_bio->master_bio = (struct bio*)rb2;
1887 atomic_inc(&rb2->remaining);
1888 r10_bio->mddev = mddev;
1889 set_bit(R10BIO_IsRecover, &r10_bio->state);
1890 r10_bio->sector = sect;
1892 raid10_find_phys(conf, r10_bio);
1894 /* Need to check if the array will still be
1897 for (j=0; j<conf->raid_disks; j++)
1898 if (conf->mirrors[j].rdev == NULL ||
1899 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1904 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1905 &sync_blocks, still_degraded);
1907 for (j=0; j<conf->copies;j++) {
1908 int d = r10_bio->devs[j].devnum;
1909 if (!conf->mirrors[d].rdev ||
1910 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
1912 /* This is where we read from */
1913 bio = r10_bio->devs[0].bio;
1914 bio->bi_next = biolist;
1916 bio->bi_private = r10_bio;
1917 bio->bi_end_io = end_sync_read;
1919 bio->bi_sector = r10_bio->devs[j].addr +
1920 conf->mirrors[d].rdev->data_offset;
1921 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1922 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1923 atomic_inc(&r10_bio->remaining);
1924 /* and we write to 'i' */
1926 for (k=0; k<conf->copies; k++)
1927 if (r10_bio->devs[k].devnum == i)
1929 BUG_ON(k == conf->copies);
1930 bio = r10_bio->devs[1].bio;
1931 bio->bi_next = biolist;
1933 bio->bi_private = r10_bio;
1934 bio->bi_end_io = end_sync_write;
1936 bio->bi_sector = r10_bio->devs[k].addr +
1937 conf->mirrors[i].rdev->data_offset;
1938 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1940 r10_bio->devs[0].devnum = d;
1941 r10_bio->devs[1].devnum = i;
1945 if (j == conf->copies) {
1946 /* Cannot recover, so abort the recovery */
1949 atomic_dec(&rb2->remaining);
1951 if (!test_and_set_bit(MD_RECOVERY_INTR,
1953 printk(KERN_INFO "md/raid10:%s: insufficient "
1954 "working devices for recovery.\n",
1959 if (biolist == NULL) {
1961 r10bio_t *rb2 = r10_bio;
1962 r10_bio = (r10bio_t*) rb2->master_bio;
1963 rb2->master_bio = NULL;
1969 /* resync. Schedule a read for every block at this virt offset */
1972 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1974 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1975 &sync_blocks, mddev->degraded) &&
1976 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
1977 &mddev->recovery)) {
1978 /* We can skip this block */
1980 return sync_blocks + sectors_skipped;
1982 if (sync_blocks < max_sync)
1983 max_sync = sync_blocks;
1984 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1986 r10_bio->mddev = mddev;
1987 atomic_set(&r10_bio->remaining, 0);
1988 raise_barrier(conf, 0);
1989 conf->next_resync = sector_nr;
1991 r10_bio->master_bio = NULL;
1992 r10_bio->sector = sector_nr;
1993 set_bit(R10BIO_IsSync, &r10_bio->state);
1994 raid10_find_phys(conf, r10_bio);
1995 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1997 for (i=0; i<conf->copies; i++) {
1998 int d = r10_bio->devs[i].devnum;
1999 bio = r10_bio->devs[i].bio;
2000 bio->bi_end_io = NULL;
2001 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2002 if (conf->mirrors[d].rdev == NULL ||
2003 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2005 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2006 atomic_inc(&r10_bio->remaining);
2007 bio->bi_next = biolist;
2009 bio->bi_private = r10_bio;
2010 bio->bi_end_io = end_sync_read;
2012 bio->bi_sector = r10_bio->devs[i].addr +
2013 conf->mirrors[d].rdev->data_offset;
2014 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2019 for (i=0; i<conf->copies; i++) {
2020 int d = r10_bio->devs[i].devnum;
2021 if (r10_bio->devs[i].bio->bi_end_io)
2022 rdev_dec_pending(conf->mirrors[d].rdev,
2031 for (bio = biolist; bio ; bio=bio->bi_next) {
2033 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2035 bio->bi_flags |= 1 << BIO_UPTODATE;
2038 bio->bi_phys_segments = 0;
2043 if (sector_nr + max_sync < max_sector)
2044 max_sector = sector_nr + max_sync;
2047 int len = PAGE_SIZE;
2048 if (sector_nr + (len>>9) > max_sector)
2049 len = (max_sector - sector_nr) << 9;
2052 for (bio= biolist ; bio ; bio=bio->bi_next) {
2054 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2055 if (bio_add_page(bio, page, len, 0))
2059 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2060 for (bio2 = biolist;
2061 bio2 && bio2 != bio;
2062 bio2 = bio2->bi_next) {
2063 /* remove last page from this bio */
2065 bio2->bi_size -= len;
2066 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2070 nr_sectors += len>>9;
2071 sector_nr += len>>9;
2072 } while (biolist->bi_vcnt < RESYNC_PAGES);
2074 r10_bio->sectors = nr_sectors;
2078 biolist = biolist->bi_next;
2080 bio->bi_next = NULL;
2081 r10_bio = bio->bi_private;
2082 r10_bio->sectors = nr_sectors;
2084 if (bio->bi_end_io == end_sync_read) {
2085 md_sync_acct(bio->bi_bdev, nr_sectors);
2086 generic_make_request(bio);
2090 if (sectors_skipped)
2091 /* pretend they weren't skipped, it makes
2092 * no important difference in this case
2094 md_done_sync(mddev, sectors_skipped, 1);
2096 return sectors_skipped + nr_sectors;
2098 /* There is nowhere to write, so all non-sync
2099 * drives must be failed, so try the next chunk...
2101 if (sector_nr + max_sync < max_sector)
2102 max_sector = sector_nr + max_sync;
2104 sectors_skipped += (max_sector - sector_nr);
2106 sector_nr = max_sector;
2111 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2114 conf_t *conf = mddev->private;
2117 raid_disks = conf->raid_disks;
2119 sectors = conf->dev_sectors;
2121 size = sectors >> conf->chunk_shift;
2122 sector_div(size, conf->far_copies);
2123 size = size * raid_disks;
2124 sector_div(size, conf->near_copies);
2126 return size << conf->chunk_shift;
2130 static conf_t *setup_conf(mddev_t *mddev)
2132 conf_t *conf = NULL;
2134 sector_t stride, size;
2137 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2138 !is_power_of_2(mddev->new_chunk_sectors)) {
2139 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2140 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2141 mdname(mddev), PAGE_SIZE);
2145 nc = mddev->new_layout & 255;
2146 fc = (mddev->new_layout >> 8) & 255;
2147 fo = mddev->new_layout & (1<<16);
2149 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2150 (mddev->new_layout >> 17)) {
2151 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2152 mdname(mddev), mddev->new_layout);
2157 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2161 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2166 conf->tmppage = alloc_page(GFP_KERNEL);
2171 conf->raid_disks = mddev->raid_disks;
2172 conf->near_copies = nc;
2173 conf->far_copies = fc;
2174 conf->copies = nc*fc;
2175 conf->far_offset = fo;
2176 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2177 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2179 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2180 r10bio_pool_free, conf);
2181 if (!conf->r10bio_pool)
2184 size = mddev->dev_sectors >> conf->chunk_shift;
2185 sector_div(size, fc);
2186 size = size * conf->raid_disks;
2187 sector_div(size, nc);
2188 /* 'size' is now the number of chunks in the array */
2189 /* calculate "used chunks per device" in 'stride' */
2190 stride = size * conf->copies;
2192 /* We need to round up when dividing by raid_disks to
2193 * get the stride size.
2195 stride += conf->raid_disks - 1;
2196 sector_div(stride, conf->raid_disks);
2198 conf->dev_sectors = stride << conf->chunk_shift;
2203 sector_div(stride, fc);
2204 conf->stride = stride << conf->chunk_shift;
2207 spin_lock_init(&conf->device_lock);
2208 INIT_LIST_HEAD(&conf->retry_list);
2210 spin_lock_init(&conf->resync_lock);
2211 init_waitqueue_head(&conf->wait_barrier);
2213 conf->thread = md_register_thread(raid10d, mddev, NULL);
2217 conf->mddev = mddev;
2221 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2224 if (conf->r10bio_pool)
2225 mempool_destroy(conf->r10bio_pool);
2226 kfree(conf->mirrors);
2227 safe_put_page(conf->tmppage);
2230 return ERR_PTR(err);
2233 static int run(mddev_t *mddev)
2236 int i, disk_idx, chunk_size;
2237 mirror_info_t *disk;
2242 * copy the already verified devices into our private RAID10
2243 * bookkeeping area. [whatever we allocate in run(),
2244 * should be freed in stop()]
2247 if (mddev->private == NULL) {
2248 conf = setup_conf(mddev);
2250 return PTR_ERR(conf);
2251 mddev->private = conf;
2253 conf = mddev->private;
2257 mddev->thread = conf->thread;
2258 conf->thread = NULL;
2260 chunk_size = mddev->chunk_sectors << 9;
2261 blk_queue_io_min(mddev->queue, chunk_size);
2262 if (conf->raid_disks % conf->near_copies)
2263 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2265 blk_queue_io_opt(mddev->queue, chunk_size *
2266 (conf->raid_disks / conf->near_copies));
2268 list_for_each_entry(rdev, &mddev->disks, same_set) {
2270 if (rdev->badblocks.count) {
2271 printk(KERN_ERR "md/raid10: cannot handle bad blocks yet\n");
2274 disk_idx = rdev->raid_disk;
2275 if (disk_idx >= conf->raid_disks
2278 disk = conf->mirrors + disk_idx;
2281 disk_stack_limits(mddev->gendisk, rdev->bdev,
2282 rdev->data_offset << 9);
2283 /* as we don't honour merge_bvec_fn, we must never risk
2284 * violating it, so limit max_segments to 1 lying
2285 * within a single page.
2287 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2288 blk_queue_max_segments(mddev->queue, 1);
2289 blk_queue_segment_boundary(mddev->queue,
2290 PAGE_CACHE_SIZE - 1);
2293 disk->head_position = 0;
2295 /* need to check that every block has at least one working mirror */
2296 if (!enough(conf, -1)) {
2297 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2302 mddev->degraded = 0;
2303 for (i = 0; i < conf->raid_disks; i++) {
2305 disk = conf->mirrors + i;
2308 !test_bit(In_sync, &disk->rdev->flags)) {
2309 disk->head_position = 0;
2316 if (mddev->recovery_cp != MaxSector)
2317 printk(KERN_NOTICE "md/raid10:%s: not clean"
2318 " -- starting background reconstruction\n",
2321 "md/raid10:%s: active with %d out of %d devices\n",
2322 mdname(mddev), conf->raid_disks - mddev->degraded,
2325 * Ok, everything is just fine now
2327 mddev->dev_sectors = conf->dev_sectors;
2328 size = raid10_size(mddev, 0, 0);
2329 md_set_array_sectors(mddev, size);
2330 mddev->resync_max_sectors = size;
2332 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2333 mddev->queue->backing_dev_info.congested_data = mddev;
2335 /* Calculate max read-ahead size.
2336 * We need to readahead at least twice a whole stripe....
2340 int stripe = conf->raid_disks *
2341 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2342 stripe /= conf->near_copies;
2343 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2344 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2347 if (conf->near_copies < conf->raid_disks)
2348 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2350 if (md_integrity_register(mddev))
2356 md_unregister_thread(mddev->thread);
2357 if (conf->r10bio_pool)
2358 mempool_destroy(conf->r10bio_pool);
2359 safe_put_page(conf->tmppage);
2360 kfree(conf->mirrors);
2362 mddev->private = NULL;
2367 static int stop(mddev_t *mddev)
2369 conf_t *conf = mddev->private;
2371 raise_barrier(conf, 0);
2372 lower_barrier(conf);
2374 md_unregister_thread(mddev->thread);
2375 mddev->thread = NULL;
2376 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2377 if (conf->r10bio_pool)
2378 mempool_destroy(conf->r10bio_pool);
2379 kfree(conf->mirrors);
2381 mddev->private = NULL;
2385 static void raid10_quiesce(mddev_t *mddev, int state)
2387 conf_t *conf = mddev->private;
2391 raise_barrier(conf, 0);
2394 lower_barrier(conf);
2399 static void *raid10_takeover_raid0(mddev_t *mddev)
2404 if (mddev->degraded > 0) {
2405 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2407 return ERR_PTR(-EINVAL);
2410 /* Set new parameters */
2411 mddev->new_level = 10;
2412 /* new layout: far_copies = 1, near_copies = 2 */
2413 mddev->new_layout = (1<<8) + 2;
2414 mddev->new_chunk_sectors = mddev->chunk_sectors;
2415 mddev->delta_disks = mddev->raid_disks;
2416 mddev->raid_disks *= 2;
2417 /* make sure it will be not marked as dirty */
2418 mddev->recovery_cp = MaxSector;
2420 conf = setup_conf(mddev);
2421 if (!IS_ERR(conf)) {
2422 list_for_each_entry(rdev, &mddev->disks, same_set)
2423 if (rdev->raid_disk >= 0)
2424 rdev->new_raid_disk = rdev->raid_disk * 2;
2431 static void *raid10_takeover(mddev_t *mddev)
2433 struct raid0_private_data *raid0_priv;
2435 /* raid10 can take over:
2436 * raid0 - providing it has only two drives
2438 if (mddev->level == 0) {
2439 /* for raid0 takeover only one zone is supported */
2440 raid0_priv = mddev->private;
2441 if (raid0_priv->nr_strip_zones > 1) {
2442 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2443 " with more than one zone.\n",
2445 return ERR_PTR(-EINVAL);
2447 return raid10_takeover_raid0(mddev);
2449 return ERR_PTR(-EINVAL);
2452 static struct mdk_personality raid10_personality =
2456 .owner = THIS_MODULE,
2457 .make_request = make_request,
2461 .error_handler = error,
2462 .hot_add_disk = raid10_add_disk,
2463 .hot_remove_disk= raid10_remove_disk,
2464 .spare_active = raid10_spare_active,
2465 .sync_request = sync_request,
2466 .quiesce = raid10_quiesce,
2467 .size = raid10_size,
2468 .takeover = raid10_takeover,
2471 static int __init raid_init(void)
2473 return register_md_personality(&raid10_personality);
2476 static void raid_exit(void)
2478 unregister_md_personality(&raid10_personality);
2481 module_init(raid_init);
2482 module_exit(raid_exit);
2483 MODULE_LICENSE("GPL");
2484 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
2485 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2486 MODULE_ALIAS("md-raid10");
2487 MODULE_ALIAS("md-level-10");